Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
Multiple Access Protocols
Performance and analysis
This book is made available with the consent of SpringerVerlag.
The book or any part of it cannot be reproduced for commercial purposes
without the explicit written permission of Springer Verlag. Individual printed
copies are permitted as long as the entire book is left intact. The electronic
version of this book may not be stored except on www.comnet.technion.ac.il.
The authors would welcome feedback and comments.
SpringerVerlag
New York Berlin Heidelberg
London Paris Tokyo Hong Kong
PREFACE
Computer communication networks have come of age. Today, there is hardly any profes
sional, particularly in engineering, that has not been the user of such a network. This pro
liferation requires the thorough understanding of the behavior of networks by those who
are responsible for their operation as well as by those whose task it is to design such net
works. This is probably the reason for the large number of books, monographs, and arti
cles treating relevant issues, problems, and solutions.
Among all computer network architectures, those based on broadcast multiaccess chan
nels stand out in their uniqueness. These networks appear naturally in environments
requiring user mobility where the use of any ﬁxed wiring is impossible and a wireless
channel is the only available option. Because of their desirable characteristics multiple
access networks are now used even in environments where a wired pointtopoint network
could have been installed. The understanding of the operation of multiple access network
through their performance analysis is the focus of this book.
In many aspects broadcast multiple access networks are similar, or even identical, to point
topoint networks. The major difference lies in the way in which the data links are used by
the network nodes. This book concentrates on mechanisms for link access in multiaccess
communication systems including local area networks and radio networks. The text has a
mathematical orientation with emphasis on insight, that is, the analysis is mathematical in
nature yet the purpose is understanding the operation of the systems through their analy
sis. We have assumed acquaintance with probabilistic modeling of systems, some knowl
edge in stochastic processes and just a bit of elementary queueing systemsall on the level
of undergraduate studies. With this knowledge the reader should be able to follow all the
mathematical derivations.
While some of the material covered in this book appeared in other books, the vast body of
the text has appeared only in professional journals in their typical cryptic language and
inconsistent notation. Some of the material appears here for the ﬁrst time. Because of the
inconsistent notation used in the diverse exposition of the material a great emphasis is
placed on uniform notationidentical concepts have been assigned the same notation
throughout the book. This should make it ever so easier to understand the concepts and
compare derivations and results.
The subjects covered in the book were chosen judiciously. Each subsection presents a
communication system whose nature differs from the others in the system characteristics,
the purpose of the system, or the method of analysis. Via this approach we cover all types
of multiaccess systems known to date and most of the analytical methods used in their
analysis.
The introduction chapter, presents our way of classifying the multiple access protocols. It
is here that we present the concepts that dictate the order in which we address the protocol
in the rest of the book. The introduction also includes a thorough deﬁnition of the model
ii : PREFACE
that is used throughout the book in the analysis of the protocols. The reader is encouraged
to read this section before proceeding to the main text and to keep on referring to this
chapter while studying the material. The main body of the book resides in chapters 2, 3,
and 4. Chapter 2 addresses the traditional conﬂict free protocols. In chapter 3 we address
the random retransmission family of protocols that has become so popular in local area
and satellite networks. In chapter 4 we present the family of collision resolution protocols
that are algorithmically somewhat more complex but that exhibit many other desirable fea
tures. These chapters are to some extent independent from one another, in the sense that
one can be studied without the others and because different analytical techniques are used.
We do, however, feel that studying the subjects in the order presented contributes signiﬁ
cantly to the understanding of the material. Finally, in chapter 5 we scan brieﬂy other
major subjects that have been studied and published in the open literature but that is
beyond the scope of this book.
This book is aimed both at the student and the professional engineer. From a curriculum
standpoint the material here contains more than can be covered in a single semester of
studies. However, with sufﬁcient mathematical background an in depth course can be con
structed covering enough material so that the student could complete the rest of the mate
rial himself. A small set of problems and exercises is included at the end of every chapter.
These exercises are, in many cases, nontrivial and require a bit of time to solve; they are
meant to enhance the reader’s knowledge and train him in the techniques covered in the
text.
To enhance its use as a reference for professionals, the book points the reader to variations
and other systems through an extensive bibliography. One might expect the professional to
study the basic material as presented in the book and then follow the bibliography to an
analysis of a system that might be closer to the one he seeks.
Haifa, Israel Raphael Rom
June 1989 Moshe Sidi
CONTENTS
PREFACE .......................................................................................................................... i
CONTENTS ..................................................................................................................... iii
CHAPTER 1: INTRODUCTION ......................................................................................1
1.1. PROTOCOL CLASSIFICATION ....................................................................2
1.2. THE SYSTEM MODEL ..................................................................................5
CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS ..............................................9
2.1. FREQUENCY DIVISION MULTIPLE ACCESS ..........................................9
2.1.1. Delay Distribution...................................................................................................12
2.2. TIME DIVISION MULTIPLE ACCESS .......................................................12
2.2.1. FDMA  TDMA Comparison .................................................................................14
2.2.2. Message Delay Distribution....................................................................................15
2.3. GENERALIZED TDMA ...............................................................................20
2.3.1. Number of Packets at Allocated Slots  Distribution..............................................20
2.3.2. Expected Number of Packets at Allocated Slots.....................................................24
2.3.3. Message Delay Distribution....................................................................................26
2.3.4. Expected Message Delay........................................................................................29
2.4. DYNAMIC CONFLICTFREE PROTOCOLS ..............................................33
2.4.1. Expected Delay.......................................................................................................35
2.5. RELATED ANALYSIS .................................................................................38
EXERCISES ..................................................................................................... 40
APPENDIX A: Distribution of the Mod Function ........................................... 43
CHAPTER 3: ALOHA PROTOCOLS ............................................................................47
3.1. PURE ALOHA ..............................................................................................47
3.2. SLOTTED ALOHA .......................................................................................49
3.3. SLOTTED ALOHA  FINITE NUMBER OF USERS .................................53
SteadyState Probabilities .................................................................................................54
Throughput Analysis .........................................................................................................56
Expected Delay .................................................................................................................58
The Capture Phenomenon .................................................................................................60
3.4. (IN)STABILITY OF ALOHA PROTOCOLS ...............................................62
3.4.1. Analysis...................................................................................................................64
3.4.2. Stabilizing the Aloha System..................................................................................67
3.5. RELATED ANALYSIS .................................................................................71
EXERCISES ..................................................................................................... 74
CHAPTER 4: CARRIER SENSING PROTOCOLS .......................................................79
4.1. NONPERSISTENT CARRIER SENSE MULTIPLE ACCESS ....................80
Throughput Analysis .........................................................................................................80
4.2. 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS .........................83
Throughput Analysis .........................................................................................................84
4 : CONTENTS
Transmission Period Lengths ........................................................................................... 86
State Probabilities ............................................................................................................. 87
4.3. SLOTTED CARRIER SENSE MULTIPLE ACCESS ..................................89
Throughput of Slotted Nonpersistent CSMA ................................................................... 90
Throughput of Slotted 1PCSMA .................................................................................... 92
4.4. CARRIER SENSE MULTIPLE ACCESS WITH COLLISION
DETECTION ............................................................................................94
Throughput of Slotted Nonpersistent CSMA/CD ............................................................ 96
Throughput of Slotted 1Persistent CSMA/CD ............................................................... 99
4.5. RELATED ANALYSIS ...............................................................................101
EXERCISES ................................................................................................... 104
CHAPTER 5: COLLISION RESOLUTION PROTOCOLS .........................................107
5.1. THE BINARYTREE PROTOCOL .............................................................107
5.1.1. Moments of the Conditional Length of a CRI...................................................... 110
Bounds on the Moments ................................................................................................. 114
5.1.2. Stability Analysis ................................................................................................. 119
5.1.3. Bounds on Expected Packet Delay....................................................................... 121
5.2. ENHANCED PROTOCOLS ........................................................................123
5.2.1. The Modified BinaryTree Protocol..................................................................... 123
5.2.2. The Epoch Mechanism......................................................................................... 125
5.2.3. The Clipped BinaryTree Protocol ....................................................................... 129
5.3. LIMITED SENSING PROTOCOLS ...........................................................135
5.3.1. Throughput Analysis ............................................................................................ 139
5.4. RELATED ANALYSIS ...............................................................................140
EXERCISES ................................................................................................... 143
APPENDIX A: Moments of Collision Resolution Interval Length ................ 146
CHAPTER 6: ADDITIONAL TOPICS .........................................................................149
Multihop Networks ........................................................................................................ 149
Multistation Networks .................................................................................................... 150
Multichannel Systems .................................................................................................... 150
REFERENCES .............................................................................................................. 153
APPENDIX A: MATHEMATICAL FORMULAE AND BACKGROUND ............... 167
GLOSSARY OF NOTATION ....................................................................................... 171
CHAPTER 1
INTRODUCTION
Three major components characterize computer communication networks: switches, chan
nels, and protocols. The switches (or nodes) are the hardware entities that house the data
communication functions; the protocols are the sets of rules and agreements among the
communicating parties that dictate the behavior of the switches, and the channel is the
physical medium over which signals, representing data, travel from one switch to another.
Traditional networks make use of pointtopoint channels, that is, channels that are dedi
cated to an (ordered) pair of users. These channels, beyond being very economical, are
advantageous due to their noninterference feature namely, that transmission between a
pair of nodes has no effect on the transmission between another pair of nodes even if these
two pairs have a common node. Pointtopoint channels, however, require the topology to
be ﬁxed, mostly determined at network design time. Subsequent topological changes are
quite hard (and costly) to implement.
When pointtopoint channels are not economical, not available, or when dynamic topolo
gies are required broadcast channels can be used. Informally stated, a broadcast channel is
one in which more than a single receiver can potentially receive every transmitted mes
sage. Broadcast channels appear naturally in radio, satellite, and some local area networks.
This basic property has its advantages and disadvantages. If, indeed, a message is destined
to a large number of destinations then a broadcast channel is clearly superior. However, in
a typical case a message is destined to a single or a very small number of destinations and
wasteful processing results in all those switches for whom the message is not intended.
Moreover, transmissions over a broadcast channel interfere, in the sense that one transmis
sion coinciding in time with another may cause none of them to be received. In other
words, the success of a transmission between a pair of nodes is no longer independent of
other transmissions.
To make a transmission successful interference must be avoided or at least controlled. The
channel then becomes the shared resource whose allocation is critical for proper operation
of the network. This book focuses on access schemes to such channels known in the liter
ature as Multiple Access Protocols. These protocols are nothing but channel allocation
schemes that posses desirable performance characteristics. In terms of known networking
models, such as the OSI reference model, these protocols reside mostly within a special
layer called the Medium Access Control (MAC) layer. The MAC layer is between the Data
Link Control (DLC) layer and the Physical Layer.
The need for multiple access protocols arises not only in communications systems but also
in many other systems such as a computer system, a storage facility or a server of any
kind, where a resource is shared (and thus accessed) by a number of independent users. In
this book we mainly address a shared communications channel. To brieﬂy summarize the
2 CHAPTER 1: INTRODUCTION
environment in which we are interested we assume, in general, that (1) sending a message
to multiple users in a single transmission is an inherent capability, (2) users typically hear
one another, (3) we conﬁne ourselves to the Medium Access Control layer freeing us from
worrying about network wide functions such as routing and ﬂow control, and (4) interfer
ence is inherent, i.e., communication between one pair of nodes may inﬂuence the com
munication between other pairs. More precise deﬁnitions appear later in the introduction
and in the description of the various protocols.
1.1. PROTOCOL CLASSIFICATION
The multiple access protocols suggested and analyzed to date are too numerous to be all
mentioned here. We therefore classify these protocols and take samples of the various
classes to be analyzed in the text. In this description we consider the channel as the focal
point and refer to the nodes transmitting through the channel as its users.
There are various ways to classify multiple access protocols. Examples of such classiﬁca
tions appear in [KSY84] and [Sac88]. Our classiﬁcation is presented in Figure 1.1. First
and foremost, we are interested in noncentralized multiple access protocols. These are
protocols in which all nodes behave according to the same set of rules. In particular there
is no single node coordinating the activities of the others (whose protocol, by necessity,
differs from the rest). This also excludes, for example, all pollingtype access protocols
The classiﬁcation of Figure 1.1 attempts to exhibit the underlying balance and symmetry
behind existing multiple access protocols.
At the highest level of the classiﬁcation we distinguish between conﬂictfree and conten
tion protocols. Conﬂictfree protocols are those ensuring a transmission, whenever made,
is a successful one, that is, will not be interfered by another transmission. Conﬂictfree
transmission can be achieved by allocating the channel to the users either statically or
dynamically. The channel resources can be viewed, for this purpose, from a time, fre
quency, or mixed timefrequency standpoint. Hence, the channel can be divided by giving
the entire frequency range (bandwidth) to a single user for a fraction of the time as done in
Time Division Multiple Access (TDMA), or giving a fraction of the frequency range to
every user all of the time as done in Frequency Division Multiple Access (FDMA), or pro
viding every user a portion of the bandwidth for a fraction of the time as done in spread
spectrum based systems such as Code Division Multiple Access (CDMA).
To counter the static allocation, the dynamic one allocates the channel based on demand so
that a user who happens to be idle uses only little, if at all, of the channel resources, leav
ing the majority of its share to the other, more active users. Such an allocation can be done
by various reservation schemes in which the users ﬁrst announce their intent to transmit
and all those who have so announced will transmit before new users have a chance to
announce their intent to transmit. Another common scheme is referred to as token passing
in which a single (logical or physical) token is passed among the users permitting only the
token holder to transmit, thereby guaranteeing noninterference. The MiniSlotted Alternat
S
e
c
t
i
o
n
1
.
1
.
:
P
R
O
T
O
C
O
L
C
L
A
S
S
I
F
I
C
A
T
I
O
N
3
Dynamic Resolution
Multiple Access
Protocols
Contention Conﬂict Free
Static Resolution Dynamic Allocation Static Allocation
Time of
Arrival
Proba
bilistic
ID
Proba
bilistic
Reser
vation
Token
Passing
Time
and
Freq
Freq.
Based
Time
Based
FIGURE 1.1: Classiﬁcation of Multiple Access Protocols
4 CHAPTER 1: INTRODUCTION
ing Priority (MSAP) and the Broadcast Recognition Access Method (BRAM) are exam
ples of multiple access protocols that belong to this class.
Contention schemes differ in principle from conﬂictfree schemes since a transmitting
user is not guaranteed to be successful. The protocol must prescribe a way to resolve con
ﬂicts once they occur so all messages are eventually transmitted successfully. The resolu
tion process does consume resources and is one of the major differences among the
various contention protocols. If the probability of interference is small, such as might be
the case with bursty users, taking the chance of having to resolve the interference compen
sates for the resources that have to be expanded to ensure freedom of conﬂicts. Moreover,
in most conﬂictfree protocols, idle users do consume a portion of the channel resources;
this portion becomes major when the number of potential users in the system is very large
to the extent that conﬂictfree schemes are impractical. In contention schemes idle users
do not transmit and thus do not consume any portion of the channel resources.
When contentionbased multiple access protocols are used, the necessity arises to resolve
the conﬂicts, whenever they occur. As in the conﬂictfree case, here too, both static and
dynamic resolutions exist. Static resolution means that the actual behavior is not inﬂu
enced by the dynamics of the system. A static resolution can be based, for example, on
user ID’s or any other ﬁxed priority assignment, meaning that whenever a conﬂict arises
the ﬁrst user to ﬁnally transmit a message will be the one with, say, the smallest ID (this is
done in some treeresolution protocols). A static resolution can also be probabilistic,
meaning that the transmission schedule for the interfering users is chosen from a ﬁxed dis
tribution that is independent of the actual number of interfering users, as is done in Aloha
type protocols and the various versions of Carrier Sense Multiple Access (CSMA) proto
cols.
Dynamic resolution, namely taking advantage and tracking system changes is also possi
ble in contentionbased protocols. For example, resolution can be based on time of arrival
giving highest (or lowest) priority to the oldest message in the system. Alternatively reso
lution can be probabilistic but such that the statistics change dynamically according to the
extent of the interference. Estimating the multiplicity of the interfering packets, and the
exponential backoff scheme of the Ethernet standard fall into this category.
The main body of the text contains typical examples of multiple access protocols that are
analyzed and discussed. These examples were chosen judiciously to cover the above
classes of protocols. Each example presents a system or a protocol whose nature differs
from the others either in the system characteristics, or in the purpose of the system, or in
the method of analysis. Via this approach we cover all types of noncentralized multiple
access protocols known to date and most of the analytical methods used in their analysis.
Section 1.2.: THE SYSTEM MODEL 5
1.2. THE SYSTEM MODEL
The main body of the text contains typical examples of multiaccess protocols that are ana
lyzed and discussed. In the analysis we are interested mainly in throughput and delay
characteristics. We take the throughput of the channel to mean the aggregate average
amount of data that can be transported through the channel in a unit of time. In those cases
where only a single transmission can be successful at any time (as is typical in many sin
glehop systems) the throughput, as thus deﬁned, equals the fraction of time in which the
channel is engaged in the successful transmission of user data. In delay calculations we
generally consider the time from the moment a message is generated until it makes it suc
cessfully across the channel. Here one must distinguish between the user and the system
measures as it is possible that the average delay measured for the entire system does not
necessarily reﬂect the average delay experienced by any of the users. In “fair”, or homoge
neous systems we expect these to be almost identical. Two other criteria are also of inter
est: system stability and message storage requirement (buffer occupancy). The notion of
stability arises in this context because the protocol characteristics may be such that some
message generation rates, even smaller than the maximal transmission rate in the channel,
cannot be sustained by the system for a long time. Evaluation of those input rates for
which the system remains stable is therefore essential. We postpone further deﬁnitions to
the sections dealing directly with stability. Buffer occupancy is clearly an important per
formance measure since having to provide larger buffers generally translates into more
costly and complex implementation. Higher buffer occupancy usually also means longer
message delays and vice versa.
To analyze multiple access protocols one must make assumptions regarding the environ
ment in which they operate. Hence, in each and every protocol we must address the fol
lowing issues:
• Connectivity. In general, the ability of a node to hear the transmission of another node
depends on the transmission power used, on the distance between the two nodes, and on
the sensitivity of the receiver at the receiving node. In this text we assume a symmetric
connectivity pattern, that is, every node can successfully transmit to every node it can
hear. Basically, connectivity patterns can be classiﬁed into three categories known as
singlehop, dualhop and multihop topologies. In a singlehop topology all users hear
one another, and hence no routing of messages is required. Dualhop topologies are
those in which messages from a source to a destination do not have to pass more than
two hops, meaning that either the source and destination can communicate directly or
there exists a node that communicates directly with both the source and the destination.
This conﬁguration is peculiar in a broadcast channel context since the intermediary
node can be affected by the behavior of two other nodes that do not hear one another.
The multihop topology is the most general one in which beyond (and in addition to) the
problems encountered in the single and dualhop topologies one must address routing
issues that become complex if the topology is allowed to vary dynamically.
6 CHAPTER 1: INTRODUCTION
• Channel type. The channel is the medium through which data is transferred from its
source to its destination. In this text we deal with an errorless collision channel. Colli
sion is a situation in which, at the receiver, two or more transmissions overlap in time
wholly or partially. A collision channel is one in which all the colliding transmissions
are not received correctly and in most protocols have to be retransmitted. A channel is
errorless, if a single transmission heard at a node is always received correctly. Other
possible channels include the noisy channel in which errors may occur even if only a
single transmission is heard at a node and, furthermore, the channel may be such that
errors between successive transmissions are not independent. Another channel type is
the capture channel in which one or more of the colliding transmissions “captures” the
receiver and can be received correctly. Yet another case is a channel in which coding is
used so that even if transmissions collide the receiver can still decode some or all of the
transmitted information.
• Synchronism. Users are generally not assumed to be synchronized and are capable of
accessing and transmitting their data on the channel at any time. Another important
class of systems is that of.slotted systems in which a global clock exists that marks
equally long intervals of time called slots. In these systems transmissions of data start
only at slot boundaries. Other operations, such as determining activities on the channel
can be done at any time. Various degrees of synchronismare required in the slotted pro
tocols we consider.
• Feedback/Acknowledgment. Feedback is the information available to the users regard
ing activities on the channel at prior times. This information can be obtained by listen
ing to the channel, or by explicit acknowledgment messages sent by the receiving node.
For every protocol we assume that there exist some instants of time (typically slot
boundaries or end of transmissions) in which feedback information is available. Com
mon feedback information indicates whether a message was successfully transmitted,
or a collision took place, or the channel was idle. It is generally assumed that the feed
back mechanism does not consume channel resources, for example, by utilizing a dif
ferent channel or by being able to determine the feedback locally. Other feedback
variations include indication of the exact or the estimated number of colliding transmis
sions, or providing uncertain feedback (e.g., in the case of a noisy channel). Recently
nofeedback protocols have also been proposed.
• Message size. The basic unit of data generated by a user is a message. It is possible,
though, that due to its length, a message cannot be transmitted in a single transmission
and must therefore be broken into smaller units called packets each of which can be
transmitted in a single channel access. A message consists of an integral number of
packets although the number of packets in a message can vary randomly. Packet size is
measured by the time required to transmit the packet once access to the channel has
been granted. Typically, we assume all packets to be of equal size and variations
include randomly varying packets.
• Message generation. All users are statistically identical and generate new messages
according to a Poisson process. Variations include cases in which all users are not the
same and in particular one heavy user and many identical small ones. Few analyses can
be found in the literature accommodating nonPoisson generation processes.
Section 1.2.: THE SYSTEM MODEL 7
• User population. The number of users in the systemcan be ﬁnite or inﬁnite. Every user
is a different, generally independent entity. One interesting observation is that most
conﬂictfree protocols are useless if the user population increases beyond a certain
point. For such cases contention based protocols are the only possible solution.
• Buffering capability. Messages generated by the user are stored in a buffer. In a typical
analysis it is assumed that every user has a buffer for a single message and that it does
not generate new messages unless its buffer is empty. Other alternatives include more
buffering, both inﬁnite and ﬁnite, at each user.
A word about notation
Throughout the text we adopt a consistent notation as follows. A random variable is
denoted by a letter with a tilde, e.g., . For this random variable we denote by its
probability distribution function, by its probability density function, by the
Laplace transform of , and by its kth moment. If is a discrete random variable
then X(z) denotes its generating function. The expectation is denoted by or just x. In
general, a discrete stochastic process is denoted .
x
˜
F
x
˜
x ( )
f
x
˜
x ( ) F
x
˜
*
s ( )
f
x
˜
x ( ) x
˜
k
x
˜
x
x
˜
n
n 0 ≥ , { }
8 CHAPTER 1: INTRODUCTION
CHAPTER 2
CONFLICTFREE ACCESS PROTOCOLS
Conﬂictfree protocols are designed to ensure that a transmission, whenever made, is not
interfered by any other transmission and is therefore successful. This is achieved by allo
cating the channel to the users without any overlap between the portions of the channel
allocated to different users. An important advantage of conﬂictfree access protocols is the
ability to ensure fairness among users and the ability to control the packet delaya feature
that may be essential in realtime applications.
The ﬁrst three sections are devoted to static channel allocation strategies in which channel
allocation is predetermined (typically at network design time) and does not change during
the operation of the system. The two most well known protocols in this class are the Fre
quency Division Multiple Access (FDMA) in which a fraction of the frequency bandwidth
is allocated to every user all the time, and the Time Division Multiple Access (TDMA) in
which the entire bandwidth is used by each user for a fraction of the time.
For both the FDMA and the TDMA protocols no overhead, in the form of control mes
sages, is incurred. However, due to the static and ﬁxed assignment, parts of the channel
might be idle even though some users have data to transmit. Dynamic channel allocation
protocols attempt to overcome this drawback by changing the channel allocation based on
the current demands of the users.
2.1. FREQUENCY DIVISION MULTIPLE ACCESS
With Frequency Division Multiple Access (FDMA) the entire available frequency band is
divided into bands each of which serves a single user. Every user is therefore equipped
with a transmitter for a given, predetermined, frequency band, and a receiver for each band
(which can be implemented as a single receiver for the entire range with a bank of band
pass ﬁlters for the individual bands).
The main advantage of FDMA is its simplicityit does not require any coordination or
synchronization among the users since each can use its own frequency band without inter
ference. This, however, is also the cause of waste especially when the load is momentarily
uneven, since when one user is idle his share of the bandwidth cannot be used by other
users. It should be noted that if the users have uneven long term demands, it is possible to
divide the frequency range unevenly, i.e., proportional to the demands. FDMA is also not
ﬂexible; adding a new user to the network requires equipment modiﬁcation (such as addi
tional ﬁlters) in every other user. For more details the reader may consult any of the many
texts treating FDMA that have been published, e.g., by Stallings [Sta85] or the one by
Martin [Mar78].
10 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
To evaluate the performance of the FDMA protocol let us assume that the entire channel
can sustain a rate of R bits/sec which is equally divided among M usersR/M bits/sec for
each. Since the individual bands are disjoint, there is no interference among users’ trans
missions and the system can therefore be viewed as M independent queues (See Figure
2.1). Each of these queues has an individual input process governing the packet generation
process for that user. If the packet length is a random variable , then the service time
afforded to every packet is the random variable . To evaluate the throughput
of the individual user we note that every bit transmitted is a “good” bit and thus the indi
vidual throughput is the fraction of time the individual server is busy (i.e., nonempty sys
tem). The total throughput is M times the individual throughput while the average packet
delay can be obtained by applying Little’s result to the individual queue. In general, all
parameters relating to FDMA can be obtained by applying known results of the corre
sponding queue discipline.
Consider a typical user that generates packets according to a Poisson process with rate λ
packets/sec. and his buffering capabilities are not limited. The time required for the trans
mission of a packet is . Each node can therefore be viewed as an M/G/1 queue. Thus,
using the known system delay time formula for M/G/1 queueing systems we get that the
expected delay of a packet is (see Appendix)
where and .
λ
T
˜
1
λ
T
˜
2
λ
T
˜
Μ
FIGURE 2.1: FDMA System Model
P
˜
T
˜
MP
˜
R ⁄ =
T
˜
D x
λx
2
2 1 λx – ( )
 + T
λT
2
2 1 λT – ( )
 + = =
T E T
˜
[ ] = T
2
E T
˜ 2
[ ] =
Section 2.1.: FREQUENCY DIVISION MULTIPLE ACCESS 11
When all packets are of equal length consisting of P bits each then the transmission time
of every packet is (deterministically) equal to T= MP/R seconds. In this case the queueing
model corresponds to an M/D/1 queue in which and therefore
(2.1)
where .
For M/G/1 systems is the fraction of time the server is busy. In our case, there
fore, ρ equals the individual user’s throughput. Normalizing the expected delay given in
(2.1) by P/R, the time required to transmit a packet in a channel with rate R, and substitut
ing S for ρ we get the normalized expected delay, ,
which is the desired throughputdelay characteristic for FDMA with constant packet size.
Graphs depicting the throughput delay for various population sizes is depicted in Figure
2.2. Note that for a wide range the delay is rather insensitive to the throughput. For values
of throughput beyond 0.8 the delay increases quickly to values which cannot be tolerated .
T
2
T
2
=
D T
λT
2
2 1 λT – ( )
 + T 1
ρ
2 1 ρ – ( )
 +
MP
R
 1
ρ
2 1 ρ – ( )
 + = = =
ρ λT
·
∆
ρ λx =
D
ˆ
D
ˆ
D
P R ⁄
 1
S
2 1 S – ( )
 + M M
2 S –
2 1 S – ( )

M
2
 1
1
1 S –
 +
¸ ,
¸ _
= = = =
1
10
100
1000
10000
100000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
M=5
M=10
M=100
M=1000
FDMA
E
x
p
e
c
t
e
d
D
e
l
a
y
D
ˆ
(
)
Throughput S ( )
FIGURE 2.2: ThroughputDelay Characteristic for FDMA
12 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
2.1.1. Delay Distribution
The delay distribution of FDMA is also taken directly from M/G/1 results. Speciﬁcally, if
is a random variable representing the packet delay, then the Laplace transform of the
probability density function (pdf) of , , is given by (see Appendix)
where is the Laplace transform of the packet transmission time, i.e.,
. For the case of equally sized packets we have and there
fore
The expected delay can be obtained by taking the derivative of with respect to s at
s=0 and for equally sized packets the expression is given in (2.1). Higher moments can be
obtained by taking higher order derivatives.
2.2. TIME DIVISION MULTIPLE ACCESS
In the time division multiple access (TDMA) scheme the time axis is divided into time
slots, preassigned to the different users. Every user is allowed to transmit freely during the
slot assigned to it, that is, during the assigned slot the entire system resources are devoted
to that user. The slot assignments follow a predetermined pattern that repeats itself period
ically; each such period is called a cycle or a frame. In the most basic TDMA scheme
every user has exactly one slot in every frame (see Figure 2.3). More general TDMA
schemes in which several slots are assigned to one user within a frame, referred to as gen
eralized TDMA, are considered in the next section. Note that for proper operation of a
TDMA scheme, the users must be synchronized so that each one knows exactly when and
for how long he can transmit. Further details on TDMA schemes can be found in texts
such as [Kuo81,Sta85].
D
˜
D
˜
D
*
s ( ) E e
sD
˜
–
[ ] =
D
*
s ( ) X
*
s ( )
s 1 ρ – ( )
s λ – λX
*
s ( ) +
 =
X
*
s ( )
X
*
s ( ) E e
sT
˜
–
[ ] = X
*
s ( ) e
sT –
=
D
*
s ( )
s 1 ρ – ( )
λ s λ – ( )e
sT
+
 =
D
*
s ( )
FIGURE 2.3: TDMA Slot Allocation
t
Frame
Frame
Slot for
User
1 2 3 4 5 6 7 1 2 3 4 5 6 7
Section 2.2.: TIME DIVISION MULTIPLE ACCESS 13
To analyze the performance of a TDMA scheme consider a system composed of M users
each transmitting equally long packets of P bits each. If the total rate of transmission is R
bits/sec then the packet transmission time is T=P/R which is taken to be the slot size. The
duration of the entire cycle is therefore T
c
= MT. Assuming that the packet arrival pro
cesses of new packets to the different users are independent, it follows that the queueing
behavior of the queue at one user is independent of the queueing behavior of all others.
The reason is that a user transmits a packet, if he has any, every T
c
seconds, independently
of any event in any of the other queues of other users. Consequently, in the following we
concentrate on the characteristics of one user, and without loss of generality, assume that
the user transmits a packet, if he has any, at the ﬁrst slot of every frame.
Consider a typical packet generated by the user. The delay suffered by this packet has
three components: (1) the time between its generation and the end of the current frame, (2)
the queueing time to allow all the packets already queued to be transmitted and, (3) the
packet transmission time itself. Of these components the ﬁrst and the third are readily
known. Since all frames are of equal length, the average time between the packet genera
tion time and the end of the current frame is 0.5 T
c
. Packet transmission time is T.
To compute the queueing time (once the end of the current frame is reached) we observe
that the queue behaves exactly like one with deterministic service time of T
c
. If we assume
a Poisson arrival process of λ packets/sec for the user and that the number of packets that
can be stored in a queue is not bounded, then the queueing time is identical to that of the
queueing time in an M/D/1 queueing system in which the deterministic service time is
T
c
. We thus have that the expected queueing time of a packet, W
q
, is given by (see Appen
dix)
where ρ=λT
c
=λMP/R (note that this is the same value for ρ as in the FDMA case). The
total expected packet delay is therefore
.
As in the FDMA case, every bit transmitted should be counted in the throughput or, in
other words, the throughput equals the fraction of time the server is busy, which for an M/
D/1 queue equals ρ. Thus, as in the FDMA case, we have S = ρ leading to
,
and the normalized expected packet delay is obtained by dividing D by the time required
to transmit a packet,
x
W
q
ρ
2 1 ρ – ( )
x
ρ
2 1 ρ – ( )
T
c
ρ
2 1 ρ – ( )
MT = = =
D
1
2
 T
c
W
q
T + +
1
2
 MT
ρ
2 1 ρ – ( )
MT T + + T 1
M
2 1 ρ – ( )
 + = = =
D T 1
M
2 1 S – ( )
 + =
14 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
(2.2)
which is the desired throughput delay characteristic of a TDMA scheme. A family of
graphs of the expected packet delay versus the throughput for various values of Mis given
in Figure 2.4. They are, not surprisingly, similar to those of the FDMA (Figure 2.2).
2.2.1. FDMA  TDMA Comparison
Comparing the throughput delay characteristics of FDMA and TDMA we note that
.
We thus conclude that for (i.e., every meaningful case) the TDMAexpected delay is
always less than that of FDMA and the difference grows linearly with the number of users
and independent of the load! The difference stems from the fact that the actual transmis
sion of a packet in TDMA takes only a single slot while in FDMA it lasts the equivalent of
an entire frame. This difference is somewhat offset by the fact that a packet arriving to an
empty queue may have to wait until the proper slot when a TDMA scheme is employed,
whereas in FDMA transmission starts right away.
D
ˆ
D
T
 1
M
2 1 S – ( )
 + = =
E
x
p
e
c
t
e
d
D
e
l
a
y
D
ˆ
(
)
Throughput S ( )
1
10
100
1000
10000
100000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
M=5
M=10
M=100
M=1000
TDMA
FIGURE 2.4: ThroughputDelay Characteristic for TDMA
D
FDMA
D
TDMA
P
R

M
2
 1 – + D
TDMA
≥ =
M 2 ≥
Section 2.2.: TIME DIVISION MULTIPLE ACCESS 15
It must be remembered, however, that at high throughput the dominant factor in the
expected delay is proportional to 1/(1S) in both TDMAand FDMAand therefore the ratio
of the expected delays between the two schemes approaches unity when the load
increases.
As a practical matter, while FDMA performs slightly worse than TDMA, it is somewhat
easier to implement since it does not require any synchronization among users, which is
necessary to keep the TDMA users from transmitting in a slot which is not their own.
2.2.2. Message Delay Distribution
In the previous section we derived the expected packet delay in a TDMA system. In this
section we generalize the arrival process and demonstrate how one can, with fairly
straightforward queueing theory techniques, compute the distribution of the delay in such
a system. The analysis presented is essentially due to Lam [Lam77].
As before, the queueing behavior of one user is independent of the queueing behavior of
other users and therefore we consider a typical user in an M user system in which the slot
size T equals the duration of packet transmission. The user transmits a packet, if he has
any, in the ﬁrst slot of every frame. At each arrival epoch a new message arrives. A mes
sage consists of a random bulk of packets. Let L(z) be the generating function of , L
its mean and its second moment, i.e.,
We assume that messages arrive to the user according to a Poisson process at a rate of λ
messages/sec. Notice that the arrival process considered here is somewhat more general
than that considered in the previous section. If one takes L(z)=z, then each message con
sists of a single packet and the general arrival process we consider here degenerates to that
of the previous section. The buffering capabilities of the user are not limited.
The message delay is deﬁned as the time elapsing between the message arrival epoch
until after the transmission of the last packet of that message is completed. In this section
we derive the Laplace transform of the message delay distribution.
Consider an arbitrary “tagged” message arriving seconds after the beginning of the
(j+1)st frame (i.e., seconds before its end). Assume that our tagged message is the
st message arriving in that frame, that is messages arrived prior to it in the
same frame ( and are random variables and when is given, has a Poisson distri
bution with parameter ). Figure 2.5 shows the relation among these quantities.
L
˜
L
˜
L
2
L z ( ) Prob L
˜
l = [ ]z
l
l 1 =
∞
∑
=
L L' z ( )
z 1 =
l Prob L
˜
l = [ ] ⋅
l 1 =
∞
∑
= = L
2
L + L'' z ( )
z 1 =
l
2
Prob L
˜
l = [ ] ⋅
l 1 =
∞
∑
= =
D
˜
T
c
w
˜
–
w
˜
k
˜
1 + k 0 ≥
k
˜
w
˜
w
˜
k
˜
λ T
c
w
˜
– ( )
16 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
Denote by the number of packets awaiting transmission at the beginning of the (j+1)st
frame. Then the delay of the tagged message is given by
(2.3)
where the ﬁrst term represents the waiting time of the tagged message until the end of
the (j+1)st frame. The second term in (2.3) represents the time required to transmit all the
packets already queued in the user buffer upon the tagged message arrival. When mes
sages are transmitted in a FIFO manner, then this is the time the message will wait from
the end of the (j+1)st frame until its ﬁrst packet will be transmitted. This second term is
composed of two components: the ﬁrst corresponds to waiting for
transmission of packets present in the queue at the beginning of the (j+1)st frame. The
other component corresponds to waiting for transmission of packets that
arrived since the beginning of the (j+1)st frame until the arrival of the tagged message.
Finally, the third term in (2.3) represents the time required to transmit the tagged message
itself: seconds to transmit all packets of the tagged message, except the last
one that requires only T seconds.
The expression in (2.3) can be rewritten as
(2.4)
The three components in this expression are statistically independent of one another: the
ﬁrst contains quantities relating to previous frames, the second to the current frame up to
the arrival of the tagged message, and the last one to the tagged message itself. Thus,
FIGURE 2.5: TDMA Packet Delay Components
1 2 3 k
˜
q
˜
j
q
˜
j 1 +
¹ ¹ ' ¹ ¹
w
˜
Arrivals of
User 1
k
˜
1 + ( )st
Arrival of
User 1
T
c
jth Frame (j+1)st Frame_ (j+2)nd Frame
1 1
t
q
˜
j
D
˜
w
˜
max q
˜
j
1 – 0 , ( ) L
˜
i
i 1 =
k
˜
∑
+ T
c
L
˜
k
˜
1 +
1 – ( )T
c
T + [ ] + + =
w
˜
T
c
max q
˜
j
1 – 0 , ( ) [ ]
T
c
L
˜
i
i 1 =
k
˜
∑
L
˜
k
˜
1 +
1 – ( )T
c
D
˜
max q
˜
j
1 – 0 , ( )T
c
w
˜
T
c
L
˜
i
i 1 =
k
˜
∑
+ L
˜
k
˜
1 +
T
c
T T
c
– + [ ] + + =
Section 2.2.: TIME DIVISION MULTIPLE ACCESS 17
, the Laplace transformof the probability density function of is the product of the
Laplace transforms of these three components.
Let a
j
be the number of packets arriving during the jth frame. Because the number of mes
sages arriving in a frame depends only on the arrival process and on the frame length, and
because all frames are equally long we conclude that a
j
is independent of j. The relation
among the values of in consecutive frames is therefore given by
(2.5)
where ∆(i) equals 0 for i=0 and equals 1 elsewhere. The explanation of (2.5) is simple.
The packets awaiting transmission at the beginning of the (j+1)st frame are those packets
that were queued up at the beginning of the jth frame, less the packet (if there were any)
that has been transmitted in the ﬁrst slot of the jth frame. In addition, the packets that
arrived during the jth frame are also queued up at the beginning of the (j+1)st frame.
Let Q
j
(z) be the generating function of , i.e., . From (2.5) we obtain,
(2.6)
where we used the fact that the arrival process is independent of the queue size, hence is
independent of .
Let A(z) be the generating function of i.e., . The derivation of A(z) is sim
ple,
(2.7)
where we used the fact that the generating function of the number of packets in a message
is L(z).
We now turn to compute the quantity that appears in (2.6).
(2.8)
Combining (2.6) and (2.8) we obtain,
(2.9)
The chain is clearly a Markov chain (see Appendix). Assuming that
this Markov chain is ergodic (see below), the existence of steadystate (invariant) proba
D
*
s ( ) D
˜
q
˜
q
˜
j 1 +
q
˜
j
∆ q
˜
j
( ) – a
˜
+ =
q
˜
j
Q
j
z ( ) E z
q
˜
j
[ ] =
Q
j 1 +
z ( ) E z
q
˜
j 1 +
[ ] E z
q
˜
j
∆ q
˜
j
( ) – a
˜
+
[ ] E z
a
˜
[ ]E z
q
˜
j
∆ q
˜
j
( ) –
[ ] = = =
a
˜
q
˜
j
a
˜
A z ( ) E z
a
˜
[ ] =
A z ( ) E z
a
˜
[ ] E E z
a
˜
m messages arrive in frame [ ] { } = =
E L z ( )
m
[ ] { }
e
λT
c
–
λT
c
( )
m
m!
 L z ( ) [ ]
m
m 0 =
∞
∑
e
λT
c
L z ( ) 1 – [ ]
= = =
E z
q
˜
j
∆ q
˜
j
( ) – [ ]
E z
q˜
j
∆ q˜
j
( ) –
[ ] z
k ∆ k ( ) –
P q
˜
j
k = [ ]
k 0 =
∞
∑
P q
˜
j
0 = [ ]
1
z
  Q
j
z ( ) P q
˜
j
0 = [ ] – [ ] + = =
z
1 –
Q
j
z ( ) 1 z
1 –
– ( )P q
˜
j
0 = [ ] + =
Q
j 1 +
z ( ) A z ( ) z
1 –
Q
j
z ( ) 1 z
1 –
– ( )P q
˜
j
0 = [ ] + { } =
q
˜
j
j 1 2 … , , , = , { }
18 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
bilities for the queue size at the beginning of a frame is guaranteed. Let Q(z) be the gener
ating function of this invariant distribution, i.e., and let be
the probability that the queue will be empty at the beginning of a frame in steadystate.
Then from (2.9) we obtain,
or
The computation of is now easy; we use the normalization condition that
and obtain that (we used
L’Hopital’s rule in this calculation). Therefore, using (2.7) we obtain
(2.10)
The derivation of equation (2.10) holds if and only if ρ <1, which renders the q’s an
ergodic Markov chain and an altogether stable system.
We now turn to compute the Laplace transform of the probability density function of the
three components in equation (2.4). Starting with the ﬁrst component we note that
. From (2.8) we have,
(2.11)
We are interested in the steadystate behavior, hence we let in (2.11) to obtain,
(2.12)
where we used the expression for Q(z) from (2.10). Consequently, we get (by substituting
)
Substituting (2.12) and (2.7) into the last equation and denoting we get
(2.13)
To handle the second component of equation (2.4) we deﬁne and
thus we need to compute . This is done by computing the conditional transform
and then relaxing the conditions one by one as follows:
Q z ( ) Q
j
z ( )
j ∞ →
lim = P q
˜
0 = [ ]
Q z ( ) A z ( ) z
1 –
Q z ( ) 1 z
1 –
– ( )P q
˜
0 = [ ] + { } =
Q z ( ) A z ( )
P q
˜
0 = [ ] z 1 – ( )
z A z ( ) –
 =
P q
˜
0 = [ ]
Q z ( )
z 1 =
1 = P q
˜
0 = [ ] 1 A' z ( )
z 1 =
– 1 λLT
c
– 1 ρ –
·
∆
= =
Q z ( ) A z ( )
1 ρ – ( ) z 1 – ( )
z A z ( ) –
 e
λT
c
L z ( ) 1 – [ ]
1 ρ – ( ) z 1 – ( )
z e
λT
c
L z ( ) 1 – [ ]
–
 = =
max q
˜
j
1 – 0 , ( ) q
j
∆ q
˜
j
( ) – =
E z
q
˜
j
∆ q
˜
j
( ) –
[ ] z
1 –
Q
j
z ( ) 1 z
1 –
– ( )P q
˜
0 = [ ] + =
j ∞ →
E z
q
˜
∆ q
˜
( ) –
[ ] z
1 –
Q z ( ) 1 z
1 –
– ( )P q
˜
0 = [ ] +
1 ρ – ( ) z 1 – ( )
z A z ( ) –
 = =
z e
sT
c
–
=
E sT
c
q
˜
∆ q
˜
( ) – ( ) – [ ] exp [ ] E sT
c
– ( ) exp [ ]
q
˜
∆ q
˜
( ) –
[ ] E z
q
˜
∆ q
˜
( ) –
[ ]
z sT
c
– ( ) exp =
= =
L
*
s ( ) L e
sTc –
( )
·
∆
E sT
c
q
˜
∆ q
˜
( ) – ( ) – [ ] exp [ ]
1 ρ – ( ) 1 e
sT
c
–
– ( )
λT
c
L
*
s ( ) 1 – [ ] { } exp sT
c
– { } exp –
 =
y w
˜
T
c
L
˜
i
i 1 =
k
˜
∑
+
·
∆
E e
sy
˜
–
[ ]
Section 2.2.: TIME DIVISION MULTIPLE ACCESS 19
Noting that for a given w, is distributed according to a Poisson distribution with mean
λ(T
c
w) we now remove the condition on
Finally, relaxing the condition on and recalling that the latter is uniformly distributed
on [0, T
c
] we get
(2.14)
This last result can be obviously computed directly by
Finally, treating the third component of equation (2.4) we get
(2.15)
Putting together the results of equations (2.13), (2.14), and (2.15) we ﬁnally get
(2.16)
which is our desired result, i.e., the Laplace transform of the message delay.
To ﬁnd the expected message delay we take (the negative) derivative of with respect
to s computed at s=0 to obtain
E e
sy
˜
–
k
˜
k = w
˜
w = , [ ] E sw – sT
c
L
˜
i
i 1 =
k
∑
–
¸ ,
¸ _
exp e
sw –
E sT
c
L
˜
i
i 1 =
k
∑
–
¸ ,
¸ _
exp = =
e
sw –
E sT
c
L
˜
– ( ) exp [ ] [ ]
k
= e
sw –
( )L
*k
s ( ) =
k
˜
k
˜
E e
sy
˜
–
w
˜
w = [ ] E e
sw –
L
*k
˜
s ( ) w
˜
w = [ ] e
sw –
E L
*k
˜
s ( ) w
˜
w = [ ] = =
e
sw –
E z
k
˜
w
˜
w = [ ]
z L
*
s ( ) =
= e
sw –
e
λ T
c
w – ( ) z 1 – ( )
z L
*
s ( ) =
=
λT
c
L
*
s ( ) 1 – [ ] [ ] exp w s λ – λL
*
s ( ) + [ ] – [ ] exp =
w
˜
E e
sy
˜
–
[ ] E λT
c
L
*
s ( ) 1 – [ ] [ ] exp w
˜
s λ – λL
*
s ( ) + [ ] – [ ] exp [ ] =
λT
c
L
*
s ( ) 1 – [ ] [ ] exp E w
˜
s λ – λL
*
s ( ) + [ ] – [ ] exp [ ] =
λT
c
L
*
s ( ) 1 – [ ] [ ] exp sT
c
– [ ] exp –
T
c
s λ – λL
*
s ( ) + [ ]
  =
E e
sy
˜
–
[ ]
1
T
c
 e
sw –
L
*k
s ( )
λ T
c
w – ( ) [ ]
k
k!
e
λ T
c
w – ( ) –
k 0 =
∞
∑
w d
w 0 =
T
c
∫
=
E s L
˜
k
˜
1 +
T
c
T T
c
– + ( ) – [ ] exp [ ] e
s T
c
T – ( )
E e
sL
˜
T
c
–
[ ] e
s T
c
T – ( )
L
*
s ( ) = =
D
*
s ( )
1 ρ –
T
c

1 e
sT
c
–
–
s λ – λL
*
s ( ) +
 e
s T
c
T – ( )
L
*
s ( ) ⋅ ⋅ =
D
*
s ( )
20 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
,
and in a normalized form
which, for the case that each message consists of a single packet ( ), reduces
to the previously obtained result given in equation (2.2).
2.3. GENERALIZED TDMA
The allocation of a single slot within a frame to each user is reasonable in a homogeneous
system. However, when the communication requirements of the users in a system are
unequal, the channel will be utilized more efﬁciently if users with greater requirements
have more slots allocated within each frame than users with light trafﬁc. This is done in
the generalized TDMA scheme in which a user might be allocated more than one slot
within a frame, with arbitrary distances between successive allocated slots. An example is
depicted in Figure 2.6 where the frame consists of 7 slots and 4 users share the channel.
Slots 1,2 and 4 are allocated to one user, slots 3 and 5 to another user, and slots 6 and 7 one
to each of the other users. Consider a user to whom K slots are allocated in every frame
and let d(k)≥1 (1≤ k≤K) be the distance between the (k+1)modK allocated slot and the
kmodK allocated slot. Notice that . In each allocated slot, the user trans
mits one packet, if it has any. Without loss of generality we assume that the ﬁrst slot in a
frame belongs to our user. The analysis of the performance of a user in a generalized
TDMA scheme is more complicated than the analysis presented in the previous section for
regular TDMA scheme. This chapter is interesting mainly due to the mathematical tools
and techniques used.
2.3.1. Number of Packets at Allocated Slots  Distribution
Let be the number of packets awaiting transmission at the beginning of the kth allo
cated slot (1≤ k ≤ K) in the (j+1)st frame. We start by determining the generating function
of the steadystate distribution of (1≤ k ≤K). Steadystate distribution exists when
since λLT
c
is the expected number of packets arriving at the user during a frame and the
user can transmit at most K packets during a frame. In steadystate the user’s throughput is
.
D T
c
L
1
2
  –
¸ ,
¸ _
λT
c
2
L
2
2 1 λT
c
L – ( )
 T + + =
D
ˆ
D
T
 M L
1
2
 –
¸ ,
¸ _
ML
2
L

ρ
2 1 ρ – ( )
 ⋅ 1 + + = =
L L
2
1 = =
d k ( )
k 1 =
K
∑
T
c
=
q
˜
j
k ( )
q
˜
j
k ( )
ρ
λLT
c
K
 1 <
·
∆
S Kρ λLT
c
= =
Section 2.3.: GENERALIZED TDMA 21
From the operation of generalized TDMA we have that the number of packets awaiting
transmission at the beginning of the (k+1)st allocated slot (1 ≤ k ≤K1) in the (j+1)st
frame equals the number of packets awaiting transmission at the beginning of the kth allo
cated slot, less the packet (if there were any) that has been transmitted in the kth allocated
slot, plus the packets that arrived during d(k). In addition, the number of packets awaiting
transmission at the beginning of the ﬁrst allocated slot in the (j+1)st frame equals the
number of packets awaiting transmission at the beginning of the last (kth) allocated slot in
the jth frame, less the packet (if there were any) that has been transmitted in the Kth allo
cated slot, plus the packets that arrived during d(k). Therefore,
(2.17)
where (1≤ k ≤K) is the number of packets arriving to the user during d(k), and
if , if . Notice that does not depend on k.
Let be the steadystate generating function of . Then from (2.17) we have
(2.18)
where 1≤ k ≤ K). In (2.18) we used the fact that is
independent of . The derivation of (2.18) is identical to that of equation (2.8) in the
regular TDMA. From (2.18) we obtain
t
Frame Frame
Slot for
User
1 1 2 1 2 3 4 1 1 2 1 2 3 4
t
1 1 2 1 2 3 4 1 1 2 1 2 3 4
d(1) d(2) d(3)
FIGURE 2.6: GeneralizedTDMA Slot Allocation
q
˜
j
k 1 + ( ) q
˜
j
k ( ) ∆ q
˜
j
k ( ) ( ) a
˜
k ( ) + = = 1 k K 1 – ≤ ≤
q
˜
j 1 +
1 ( ) q
˜
j
K ( ) ∆ q
˜
j
K ( ) ( ) a
˜
K ( ) + = =
a
˜
k ( )
∆ q
˜
( ) 0 = q
˜
0 = ∆ q
˜
( ) 1 = q
˜
0 > a
˜
k ( )
Q
k
z ( ) q
˜
j
k ( )
Q
k 1 +
z ( ) A
k
z ( ) Q
k
0 ( ) Q
k
z ( ) Q
k
0 ( ) – [ ]z
1 –
+ { } = 1 k K 1 – ≤ ≤
Q
1
z ( ) A
K
z ( ) Q
K
0 ( ) Q
K
z ( ) Q
K
0 ( ) – [ ]z
1 –
+ { } =
A
k
z ( ) E z
a
˜
k
[ ] e
λd k ( ) L z ( ) 1 – [ ]
= = a
˜
k ( )
q
˜
j
k ( )
22 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
. (2.19)
We therefore have
. (2.20)
Had the boundary probabilities (1 ≤ k ≤ K) been known, the
generating functions (1 ≤ k ≤ K) would be completely determined. To complete the
calculation we must therefore compute these unknown probabilities which we do, using a
standard method (see for instance the book by Hayes [Hay84]). The method exploits the
fact that , being a generating function, must be analytic within the unit disk and thus
any zero of the denominator within the unit disk must also be a zero of the numerator.
The approach is to prove that there are exactly K zeroes of the denominator of equation
(2.20) within the unit disk, all of them distinct. Let their values be denoted by z
n
. These
values must also be zeroes of the numerator of equation (2.20) which results in K linear
equations in the K unknowns .
Consider the zeroes of the denominator of (2.20) within the unit disk. Any such zero,
,satisﬁes the equation
. (2.21)
We ﬁrst prove that each root of (2.21) is a simple root. If there were a multiple root z
n
, then
the derivative of the denominator of (2.20) with respect to z computed at z = z
n
would also
vanish, i.e.,
which when substituted into (2.21) yields
. (2.22)
Q
k
z ( ) Q
1
z ( )z
k 1 – ( ) –
A
m
z ( ) +
m 1 =
k 1 –
∏
=
1 z
1 –
– ( ) Q
v
0 ( )z
k v – 1 – ( ) –
A
m
z ( )
m v =
k 1 –
∏
v 1 =
k 1 –
∑
+
2 k K ≤ ≤
Q
1
z ( ) Q
1
z ( )z
K –
A
m
z ( )
m 1 =
K
∏
1 z
1 –
– ( ) Q
v
0 ( )z
K v – ( ) –
v 1 =
K
∑
A
m
z ( )
m v =
K
∏
+ =
Q
1
z ( )
z 1 – ( ) Q
u
0 ( )z
u 1 –
u 1 =
K
∑
A
m
z ( )
m u =
K
∏
z
K
A
m
z ( )
m 1 =
K
∏
–
 =
Q
k
0 ( ) q
˜
k ( ) 0 = [ ] Prob =
Q
k
z ( )
Q
1
z ( )
Q
k
0 ( )
z
n
1 ≤
z
n
K
A
m
z ( )
m 1 =
K
∏
e
λT
c
L z
n
( ) 1 – [ ]
= =
Kz
n
K 1 –
λT
c
L' z
n
( )e
λT
c
L z
n
( ) 1 – [ ]
=
K λT
c
L' z
n
( )z
n
=
Section 2.3.: GENERALIZED TDMA 23
We also have
. (2.23)
Equations (2.22) and (2.23) imply that K ≤ λT
c
L which contradicts the stability condition,
and therefore each root of (2.21) is a simple one.
Next we determine the number of roots of the denominator of (2.20) within the unit disk.
To do so we apply Rouche’s theorem.
Rouche’s Theorem: Given two functions f(z) and g(z) analytic in a region R, consider a
closed contour C in R; if on C we have and , then f(z) and f(z)+g(z)
have the same number of zeroes within C.
To apply this theorem we identify and . The region R is
the disk of radius 1+δ (i.e., ) for some δ>0. If δ is small enough, z
K
and
are both analytic in R since they are analytic in z ≤1. Also, because δ is
strictly positive we can ﬁnd some δ’ such that δ > δ’ > 0 so that z = 1+δ’ is an appropriate
contour for Rouche’s theorem. If δ is small enough, we can use a Taylor series expansion
to obtain
.
From the stability condition λLT
c
<K we see that on the ring z = 1+δ’ we have
, implying that and have the same number of
roots within z= 1+δ’. But z
K
has K roots in the unit circle (actually a root of multiplicity K
at the origin), and hence the denominator of (2.20) has K roots within the unit circle, that
are all distinct. One of these roots is z
K
=1. The other roots are denoted by .
We have already indicated that whenever the denominator of (2.20) vanishes within the
unit disk, its numerator must vanish too. We can thus substitute the values of z
n
(1≤n≤K1)
into the numerator of (2.20) and obtain the following K1 equations:
. (2.24)
An additional equation comes from the normalization condition , namely
(we use L’Hospital’s rule in (2.20))
. (2.25)
L' z
n
( ) l z
n
l 1 –
L
˜
l = [ ] Prob ⋅ ⋅
l 1 =
∞
∑
l L
˜
l = [ ] Prob ⋅
l 1 =
∞
∑
≤ L = =
f z ( ) 0 ≠ f z ( ) g z ( ) >
f z ( ) z
K
= g z ( ) e
λT
c
L z ( ) 1 – [ ]
– =
z 1 δ + <
e
λT
c
L z ( ) 1 – [ ]
z
k
1 δ' + ( )
K
1 Kδ' + ≈ =
e
λT
c
L z ( ) 1 – [ ]
1 λT
c
L δ' ( ) + ≈
z
k
e
λT
c
L z ( ) 1 – [ ]
< z
K
z
K
e
λT
c
L z ( ) 1 – [ ]
–
z
1
z
2
…z
K 1 –
, ,
Q
v
0 ( )z
n
v 1 –
A
m
z
n
( )
m v =
K
∏
v 1 =
K
∑
0 = 1 n K 1 – ≤ ≤
Q
1
z ( )
z 1 =
1 =
K λT
c
L – Q
i
0 ( )
i 1 =
K
∑
=
24 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
It is not difﬁcult to verify that the set of K equations (2.24)(2.25) has a unique solution.
For a complete description of the computation of the underlying determinant the interested
reader is referred to the book by Hayes [Hay84]. The solution of equations (2.24)(2.25)
determines (1 ≤ k ≤ K).
To summarize, the actual solution procedure is ﬁnding the roots of equation (2.21) within
the unit disk and then solving the set of equations (2.24) and (2.25). The solutions are then
substituted into equation(2.20). Solving (2.21) is, by all counts, the toughest part of the
procedure. One quite efﬁcient method to do it is due to Mueller [Mue56, CoB80]. This
method is particularly useful since it is iterative, does not require the evaluation of deriva
tives, obtains both real and complex roots even when these are not simple, and converges
almost quadratically in the vicinity of a root. In addition, the roots are computed in an
increasing absolute value order and therefore the roots within the unit disk are computed
ﬁrst. Another alternative for computing the boundary probabilities is to use Neuts’ theory
of matrix geometric computation [Neu81] as is described in [HoR87].
2.3.2. Expected Number of Packets at Allocated Slots
The expected number of packets at the beginning of an allocated slot in steadystate can be
computed by evaluating the derivative of with respect to z at z=1 (see (2.19) and
(2.20)). An alternative method (the one we employ here) is to use (2.17) directly. To that
end, we square, take expectations of both sides of (2.17) and let j → ∞. We obtain
Let and for 1 ≤ k ≤ K. With these notations and using the
independence between and and the identities
; (which stem from
the structure of the ∆(. ) function) we obtain
. (2.26)
Summing(2.26) for all we obtain
(2.27)
Using (2.17) we have
Q
k
0 ( )
Q
k
z ( )
E q
˜
2
k 1 + ( ) [ ] E q
˜
2
k ( ) [ ] E ∆ q
˜
k ( ) ( )
2
[ ] E a
˜
2
k ( ) [ ] 2E q
˜
k ( )a
˜
k ( ) [ ] + + + =
2E ∆ q
˜
k ( ) ( )a
˜
k ( ) [ ] – 2E q
˜
k ( )∆ q
˜
k ( ) ( ) [ ] –
1 k K 1 – ≤ ≤
E q
˜
2
1 ( ) [ ] E q
˜
2
K ( ) [ ] E ∆ q
˜
K ( ) ( )
2
[ ] E a
˜
2
K ( ) [ ] 2E q
˜
K ( )a
˜
K ( ) [ ] + + + =
2E q
˜
K ( )∆ q
˜
K ( ) ( ) [ ] – 2E ∆ q
˜
K ( ) ( )a
˜
K ( ) [ ] –
q k ( ) E q
˜
k ( ) [ ]
·
∆
a k ( ) E a
˜
k ( ) [ ]
·
∆
q
˜
a
˜
E ∆ q
˜
k ( ) ( )
2
[ ] E ∆ q
˜
k ( ) ( ) [ ] 1 Q
k
0 ( ) – = =
E q
˜
k ( )∆ q
˜
k ( ) ( ) [ ] E q
˜
k ( ) [ ] =
q
2
k 1 + ( ) q
2
k ( ) E a
˜
2
k ( ) [ ] 1 Q
k
0 ( ) – [ ] 1 2a k ( ) – [ ] + + =
2q k ( ) 1 a k ( ) – [ ] –
1 k K 1 – ≤ ≤
q
2
1 ( ) q
2
K ( ) E a
˜
2
K ( ) [ ] 1 Q
K
0 ( ) – [ ] 1 2a K ( ) – [ ] 2q K ( ) 1 a K ( ) – [ ] – + + =
k 1 2 …K , , =
2 q k ( ) 1 a k ( ) – [ ]
k 1 =
K
∑
E a
˜
2
k ( ) [ ]
k 1 =
K
∑
1 Q
k
0 ( ) – [ ] 1 2a k ( ) – [ ]
k 1 =
K
∑
+ =
Section 2.3.: GENERALIZED TDMA 25
(2.28)
where an empty sum vanishes. Substituting (2.28) into (2.27) we obtain
. (2.29)
Because the arrival is Poisson we have and
. Also, from (2.25) .
Therefore, we obtain from (2.29)
q k ( ) 1 a k ( ) – [ ]
k 1 =
K
∑
q 1 ( ) 1 a 1 ( ) – [ ] q k ( ) 1 a k ( ) – [ ]
k 2 =
K
∑
+ =
q 1 ( ) 1 a 1 ( ) – [ ] q k ( ) 1 a k 1 + ( ) – [ ]
k 1 =
K 1 –
∑
+ =
a k ( ) E ∆ q
˜
k ( ) ( ) [ ] – ( ) 1 a k 1 + ( ) – [ ]
k 1 =
K 1 –
∑
+
q 1 ( ) 2 a 1 ( ) – a 2 ( ) – [ ] a k ( ) E ∆ q
˜
k ( ) ( ) [ ] – ( ) 1 a k 1 + ( ) – [ ]
k 1 =
K 1 –
∑
+ =
q k ( ) 1 a k 1 + ( ) – [ ]
k 2 =
K 1 –
∑
+ … = =
q 1 ( ) K a k ( )
k 1 =
K
∑
–
¸ ,
¸ _
1 a k ( ) – [ ]
k 1 =
K
∑
a m ( ) E ∆ q
˜
m ( ) ( ) [ ] – ( )
m 1 =
k 1 –
∑
+ =
q 1 ( )
E a
˜
2
k ( ) [ ]
k 1 =
K
∑
1 Q
k
0 ( ) – [ ] 1 2a k ( ) – [ ]
k 1 =
K
∑
+
2 K a k ( )
k 1 =
K
∑
–
 =
2 1 a k ( ) – [ ] a m ( ) E ∆ q
˜
m ( ) ( ) [ ] – ( )
m 1 =
k 1 –
∑
k 1 =
K
∑
2 K a k ( )
k 1 =
K
∑
–
 –
a k ( ) E a
˜
k ( ) [ ] λLd k ( ) = =
E a
˜
2
k ( ) [ ] λL
2
d k ( ) λ
2
L
2
d
2
k ( ) + = 1 Q
k
0 ( ) – [ ]
k 1 =
K
∑
λLT
c
=
26 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
,
and ﬁnally from (2.17) we get
. (2.30)
Note that once the expected number of packets at the beginning of allocated slots is deter
mined, it is straight forward to compute the expected number of packets at the beginning
of an arbitrary slot (see the Exercises section).
2.3.3. Message Delay Distribution
Consider a tagged message arriving within d(k) (1 ≤ k ≤ K), seconds before the
beginning of the (k+1)st allocated slot and tag its last packet. The delay of the tagged mes
sage is the time elapsed from its arrival, until its last packet is transmitted, i.e., it is the
delay of the tagged packet. Thus, if is a random variable representing the total num
ber of packets that are to be transmitted before the tagged packet then the delay of the
tagged packet is plus the time needed to transmit the packets plus the time to
transmit the tagged packet itself (note that as in the regular TDMA depends on ).
Some insight into is appropriate. The quantity is actually the number of intervals
that have to elapse before the interval in which the tagged packet is transmitted. This num
ber can be (uniquely) decomposed into a number of complete frames and some leftover. In
other words we can write
where designates the number of complete frames of delay and designates the number
of intervals left over. As a matter of fact and depend only on and not on k itself.
We use the notation and (or and as a shorthand. Both and are
integervalued random variables and their distributions are derived in Appendix A at the
end of this section.
q 1 ( )
λLT
c
λ d k ( ) L
2
λL
2
d k ( ) 2L 1 Q
k
0 ( ) – [ ] – + ( )
k 1 =
K
∑
+
2 K λLT
c
– ( )
 =
2 1 λLd k ( ) – [ ] λLd m ( ) 1 – Q
m
0 ( ) + [ ]
m 1 =
k 1 –
∑
k 1 =
K
∑
2 K λLT
c
– ( )
 –
q k ( ) q 1 ( ) a m ( ) E ∆ q
˜
m ( ) ( ) [ ] – [ ]
m 1 =
k 1 –
∑
+ =
q 1 ( ) λLd m ( ) 1 – Q
m
0 ( ) + [ ]
m 1 =
k 1 –
∑
+ = 2 k K ≤ ≤
w
˜
k ( )
l
˜
k ( )
w
˜
k ( ) l
˜
k ( )
l
˜
k ( ) w
˜
k ( )
l
˜
k ( ) l
˜
k ( )
l
˜
k ( ) f
˜
k ( ) K ⋅ J
˜
k ( ) + = 0 J
˜
k ( ) K 1 – ≤ ≤
f
˜
J
˜
f
˜
J
˜
l
˜
k ( )
f
˜
k ( ) J
˜
k ( ) f
˜
J
˜
J
˜
k ( ) f
˜
k ( )
Section 2.3.: GENERALIZED TDMA 27
The delay of the tagged message after waiting the initial seconds is seconds
(representing the number of complete frames) plus the time to transmit the left over
packets which requires seconds if there are any packets left. In all cases,
the transmission time of the tagged packet is T. In summary, given the total delay is
where the summation wrapsaround from K to 1 when necessary. The above can be rewrit
ten (along with the relation ) as follows:
. (2.31)
For notational convenience, let us deﬁne
which then turns equation (2.31) into
. (2.32)
Note that , and hence depend on . Moving to the Laplace transform domain,
the above equation turns into
. (2.33)
We proceed by eliminating the condition on . Let be the generating function
of given . Continuing from equation (2.33) we get
w
˜
k ( ) f k ( )T
c
J
˜
k ( )
d u ( )
u k 1 + =
k J
˜
k ( ) +
∑
w
˜
k ( )
D
˜
k w
˜
k ( ) l
˜
k ( ) , ( )
w
˜
k ( ) f
˜
k ( )T
c
T + + J
˜
k ( ) 0 =
w
˜
k ( ) f
˜
k ( )T
c
d u ( )
u k 1 + =
k J
˜
k ( ) +
∑
T + + + 1 J
˜
k ( ) K 1 – ≤ ≤
¹
¹
¹
'
¹
¹
¹
=
f
˜
K 1 J
˜
– =
D
˜
k w
˜
k ( ) l
˜
k ( ) , ( ) w
˜
k ( ) T l
˜
k ( )
T
c
K
 J
˜
k ( )
T
c
K
 – d u ( )
u k 1 + =
k 1 J
˜
k ( ) + +
∑
d k 1 J
˜
k ( ) + + ( ) – + + + =
V J k , ( ) d u ( )
u k 1 + =
k 1 J k ( ) + +
∑
d k 1 J k ( ) + + ( ) –
·
∆
D
˜
k w
˜
k ( ) l
˜
k ( ) , ( ) w
˜
k ( ) T l
˜
k ( )
T
c
K
 J
˜
k ( )
T
c
K
 – V J
˜
k , ( ) + + + =
l
˜
k ( ) J
˜
w
˜
k ( )
D
k
*
s w
˜
k ( ) l
˜
k ( ) , ( ) E e
sD
˜
k w
˜
k ( ) l
˜
k ( ) , ( ) –
[ ] =
e
sT –
e
sw
˜
k ( ) –
e
sT
c
l
˜
k ( ) K ⁄ –
e
sT
c
J
˜
k ( ) K ⁄
e
sV J
˜
k , ( ) –
=
l
˜
l
k
z w
˜
k ( ) , ( )
l
˜
w
˜
k ( )
28 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
.
We recognize the bracketed term as the basic relation derived in Appendix A at the end of
this chapter, with z replaced by . Making the substitution we get
(2.34)
where is the unit root of order K.
The next step is to remove the condition on , but before doing so we must evaluate the
generating function of since it depends on . To do so we notice that given , the
total number of packets that are to be transmitted before the tagged packet, , is the
sum of three independent random variables: (i) the packets already waiting at the begin
ning of the kth allocated slot less one packet (if there were any) that is transmitted in the
kth allocated slot, i.e., generating function ; see
equation (2.8), (ii) Packets arriving from the beginning of the kth allocated slot until the
arrival of the tagged message (generating function ; see equation
(2.7), (iii) Packets of the tagged message, not including the tagged packet itself (generat
ing function ). Therefore,
(2.35)
By deﬁning
equation (2.35) can be written as
D
k
*
s w
˜
k ( ) ( ) e
sT –
e
sw
˜
k ( ) –
E e
sT
c
l
˜
k ( ) K ⁄ –
e
sT
c
J
˜
k ( ) K ⁄
e
sV J
˜
k , ( ) –
[ ] =
e
sT –
e
sw
˜
k ( ) –
e
sT
c
l K ⁄ –
e
sT
c
J
˜
k ( ) K ⁄
e
sV J
˜
k , ( ) –
l
˜
k ( ) l = [ ] Prob
l 0 =
∞
∑
=
e
sT –
e
sw˜ k ( ) –
e
sT
c
fK j + ( ) K ⁄ –
e
sT
c
j K ⁄ ( )
e
sV j k , ( ) –
l
˜
k ( ) fK j + = [ ] Prob
j 0 =
K 1 –
∑
f 0 =
∞
∑
=
e
sT –
e
sw
˜
k ( ) –
e
sV j k , ( ) –
e
sT
c
K ⁄ –
( )
fK
l
˜
k ( ) fK j + = [ ] Prob
f 0 =
∞
∑
j 0 =
K 1 –
∑
=
z
s
e
sT
c
k ⁄ –
·
∆
D
k
*
s w
˜
k ( ) ( ) e
sT –
e
sw˜ k ( ) –
e
sV j k , ( ) –
1
K
 z
s
β
m
( )
j –
l z
s
β
m
w
˜
k ( ) , ( )
m 0 =
K 1 –
∑
j 0 =
K 1 –
∑
=
1
K
e
sT –
e
sw
˜
k ( ) –
l z
s
β
m
w
˜
k ( ) , ( ) e
sV j k , ( ) –
z
s
β
m
( )
j –
j 0 =
K 1 –
∑
m 0 =
K 1 –
∑
=
β
m
e
i
2πm
K

=
w
˜
k ( )
l
˜
w
˜
k ( ) w
˜
k ( )
l
˜
k ( )
q
˜
k ( ) ∆ q
˜
k ( ) ( ) – 1 z
1 –
– ( )Q
k
0 ( ) z
1 –
Q
k
z ( ) + [ ]
e
λ d k ( ) w
˜
k ( ) – [ ] L z ( ) 1 – [ ]
L z ( )z
1 –
l
k
z w
˜
k ( ) , ( ) E z
l
˜
k ( )
w
˜
k ( ) [ ]
·
∆
1 z
1 –
– ( )Q
k
0 ( ) z
1 –
Q
k
z ( ) + [ ]e
λ d k ( ) w
˜
k ( ) – [ ] L z ( ) 1 – [ ]
L z ( )z
1 –
=
h
k
z ( ) 1 z
1 –
– ( )Q
k
0 ( ) z
1 –
Q
k
z ( ) + [ ]e
λd k ( ) L z ( ) 1 – [ ]
L z ( )z
1 –
·
∆
Q
k 1 +
z ( )L z ( )z
1 –
=
Section 2.3.: GENERALIZED TDMA 29
and when substituted into (2.34) we get
Since is uniformly distributed between 0 and d(k) we have from the above equation
Finally, the Laplace transform of the delay distribution is given by
(2.36)
since d(k)/T
c
is the probability that the tagged message will arrive within d(k).
2.3.4. Expected Message Delay
The expected delay of a message can be computed by evaluating the derivative of D*(s)
with respect to s at s=0 (see (2.36)). An alternative method (the one we employ here) is to
use (2.32) directly. To that end, we take expectation on (2.32) and obtain
. (2.37)
We now compute each of the terms in (2.37). Clearly,
From (2.35) and (2.30),
l
k
z w
˜
k ( ) , ( ) h
k
z ( )e
λw
˜
k ( ) 1 L z ( ) – [ ]
=
D
k
*
s w
˜
k ( ) ( )
1
K
e
sT –
e
sw
˜
k ( ) –
h
k
z
s
β
m
( )e
λw
˜
k ( ) 1 L z
s
β
m
( ) – [ ]
e
sV j k , ( ) –
z
s
β
m
( )
j –
j 0 =
K 1 –
∑
m 0 =
K 1 –
∑
=
1
K
e
sT –
h
k
z
s
β
m
( ) e
sV j k , ( ) –
z
s
β
m
( )
j –
j 0 =
K 1 –
∑
e
w
˜
k ( ) s λ 1 L z
s
β
m
( ) – [ ] – [ ] –
m 0 =
K 1 –
∑
=
w
˜
k ( )
D
k
*
s ( ) E e
s kD
˜
d –
[ ] E D
k
*
s w
˜
k ( ) ( ) [ ]
1
d k ( )
 D
k
*
s w ( ) w d
0
d k ( )
∫
= = =
1
K
e
sT –
h
k
z
s
β
m
( )
1
d k ( )
 e
w s λ 1 L z
s
β
m
( ) – [ ] – { } –
w d
0
d k ( )
∫
¸ ,
¸ _
e
sV j k , ( ) –
z
s
β
m
( )
j –
j 0 =
K 1 –
∑
m 0 =
K 1 –
∑
=
1
K
e
sT –
h
k
z
s
β
m
( )
1
d k ( )

1 e
d k ( ) s λ 1 L z
s
β
m
( ) – [ ] – { } –
–
s λ 1 L z
s
β
m
( ) – [ ] – { }
 e
sV j k , ( ) –
z
s
β
m
( )
j –
j 0 =
K 1 –
∑
m 0 =
K 1 –
∑
=
D
*
s ( )
d k ( )
T
c
D
k
*
s ( )
k 1 =
K
∑
=
dD k ( ) E w
˜
k ( ) [ ] T E l
˜
k ( ) [ ]
T
c
K
 E J
˜
k ( ) [ ]
T
c
K
 – E V J
˜
k , ( ) [ ] + + + = 1 k K ≤ ≤
E w
˜
k ( ) [ ]
1
2
 d k ( ) =
30 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
Using (2.35), the distribution of (from Appendix A at the end of this chapter) and the
deﬁnition of h
k
(z) we have
where . In a similar manner,
.
Considering that V(J,k)=0 for J=0, β
0
= 1, and c
0
=0, the above yields
.
Combining all terms we obtain
E l
˜
k ( ) [ ] q k ( ) 1 Q
k
0 ( ) – [ ] –
1
2
 λLd k ( ) L 1 – + + =
q 1 ( ) L 1 –
1
2
λLd k ( ) – λLd m ( ) 1 – Q
m
0 ( ) + [ ]
m 1 =
k
∑
+ + =
J
˜
E J
˜
k ( ) [ ] E E J
˜
k ( ) w k ( ) [ ] [ ] E
K 1 –
2

l
k
β
m
w
˜
k ( ) , ( )
1
1
β
m
 –

m 1 =
K 1 –
∑
– = =
K 1 –
2

E l
k
β
m
w
˜
k ( ) , ( ) [ ]
1
1
β
m
 –

m 1 =
K 1 –
∑
– =
K 1 –
2

h
k
β
m
( )
1
1
β
m
 –
E e
c
m
w˜ k ( )
[ ]
m 1 =
K 1 –
∑
– =
K 1 –
2

h
k
β
m
( )
1
1
β
m
 –

1
d k ( )
 ⋅ e
c
m
w
w d
0
d k ( )
∫
m 1 =
K 1 –
∑
– =
K 1 –
2

h
k
β
m
( )
1
1
β
m
 –

e
c
m
d k ( )
1 –
c
m
d k ( )
  ⋅
m 1 =
K 1 –
∑
– =
c
m
λ 1 L β
m
( ) – [ ]
·
∆
E V J
˜
k , ( ) [ ] E E V J
˜
k , ( ) w
˜
k ( ) [ ] [ ] E V J k , ( ) J
˜
J = w
˜
k ( ) [ ] Prob
J 0 =
K
∑
= =
E
1
K
 V J k , ( ) β
m
J –
l
k
β
m
w
˜
k ( ) , ( )
n 0 =
K 1 –
∑
J 0 =
K 1 –
∑
=
1
K
 V J k , ( )β
m
J –
h
k
β
m
( )E e
c
m
w
˜
k ( )
[ ]
m 0 =
K 1 –
∑
J 0 =
K 1 –
∑
=
E V J
˜
k , ( ) [ ]
1
K
 V J k , ( )
J 1 =
K 1 –
∑
1
K
 V J k , ( )β
m
J –
h
k
β
m
( )
e
c
m
d k ( )
1 –
c
m
d k ( )
 
J 1 =
K 1 –
∑
m 1 =
K 1 –
∑
+ =
Section 2.3.: GENERALIZED TDMA 31
.
Finally, the expected delay of a message is
. (2.38)
Figure 2.7 and Figure 2.8 contain some numerical results for a frame with 24 slots, four of
which are allocated to the user under consideration. We assume that an arriving message
D k ( )
1
2
d k ( ) T
T
c
K
 q 1 ( ) L 1 –
1
2
λLd k ( ) – λLd m ( ) 1 – Q
m
0 ( ) + [ ]
m 1 =
k
∑
+ +
¸ ,
¸ _
+ + =
T
c
K 1 – ( )
2K
 –
1
K
 V J k , ( )
J 1 =
K 1 –
∑
+
T
c
Kd k ( )
 h
k
β
m
( )
e
c
m
d k ( )
1 –
c
m
 
1
1
a
β
m
 –
  V J k , ( )β
m
J –
J 1 =
K 1 –
∑
+ ⋅
m 1 =
K 1 –
∑
+ +
D
d k ( )
T
c
D k ( )
k 1 =
K
∑
=
0
50
100
150
200
250
0 0.4 0.8 1.2 1.6 2.0 2.4 2.8 3.2 3.69 4.0
E
x
p
e
c
t
e
d
D
e
l
a
y
(
D
)
Throughput (S=Kρ)
K=4; T
c
= 24
d(1) = d(2) = d(3) = d(4) = 6
FIGURE 2.7: Throughput Delay for Generalized TDMA
L z ( ) z z
2
z
3
z
4
+ + + ( ) 4 ⁄ =
32 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
can contain one, two, three or four packets with equal probabilities, i.e.,
L(z)=(z+z
2
+z
3
+z
4
)/4 . In Figure 2.7 the allocated slots are evenly spaced, so that
d(1)=d(2)=d(3)=d(4)=6. Figure 2.8 contains the evenly spaced case as well as the contig
uous allocation in which the ﬁrst four slots are allocated to the user, i.e.,
d(1)=d(2)=d(3)=1, d(4)=21. We observe that for the arrival pattern under consideration,
the evenly space allocation is better, but the differences between the two allocations is
rather small. It is interesting to mention that in this case the evenly spaced allocation gives
the minimal expected delay and the contiguous allocation gives the maximal expected
delay for all values of arrival rate. Thus any other allocation results in expected delay that
is between the curves of Figure 2.8.
A natural question to ask is how to allocate the K slots available to a user in a frame, in
order to improve the performance. When the expected number of packets in the user’s
buffer is the performance measure, or equivalently, when expected packet delay is the
measure, Hofri and Rosberg showed [HoR87] that the best allocation is the uniform one,
namely, all the internal periods d(k) (1 ≤ k ≤ K) should be equal. Furthermore, this alloca
tion remains optimal for all arrival rates.
When the expected message delay is used as the performance measure, numerical experi
mentation leads us to believe, that the optimal allocation depends both on the arrival rate λ
10
15
20
25
30
35
40
45
50
0 0.4 0.8 1.2 1.6 2.0 2.4 2.8 3.2
E
x
p
e
c
t
e
d
D
e
l
a
y
(
D
)
Throughput (S=Kρ)
K=4; T
c
= 24
d(1) = d(2) = d(3) = d(4) = 6
d(1) = d(2) = d(3) = 1
d(4)=21
FIGURE 2.8: Throughput Delay for Generalized TDMA
L z ( ) z z
2
z
3
z
4
+ + + ( ) 4 ⁄ =
Section 2.4.: DYNAMIC CONFLICTFREE PROTOCOLS 33
and on the speciﬁc distribution of the message length. Whereas complete characterization
of the optimal allocation pattern is still an open question, the following captures some of
our observations.
When a message arrives at the user’s buffer, its delay is affected by the number of packets
ahead of it in the buffer. This number amounts to a number of whole frames plus some
leftover; The allocation of slots within a frame can affect only this leftover. Thus, for
heavy load (ρ → 1), the expected message delay is not very sensitive to changes in the
interallocation distances since the major portion of the delay is due to the large number of
whole frames a message must wait before its transmission starts. For light load (ρ → 0 or
equivalently λ → 0), the expected message delay might be sensitive to the allocation dis
tances. Let γ
i
(1 ≤ k ≤ K) be the probability that a message transmission requires i slots
beyond the number of whole frames, or in other words the probability that a message
length is imodK (clearly ). It can be shown (we leave the details as an exer
cise) that if γ
i
= γ
Ki+1
for i=1,2,..., K, then the expected delay is completely independent
of the interallocation distances and if γ
i
= γ
Ki+1
for i=2,3, ... , K1, then for γ
1
> γ
K
the
optimal allocation is the uniform (equidistant) one while for γ
1
<γ
K
the K slots should be
contiguous in order to minimize D
λ→ 0
.
2.4. DYNAMIC CONFLICTFREE PROTOCOLS
Static conﬂictfree protocols such as the FDMA and TDMA protocols do not utilize the
shared channel very efﬁciently, especially when the system is lightly loaded or when the
loads of different users are asymmetric. The static and ﬁxed assignment in these protocols,
cause the channel (or part of it) to be idle even though some users have data to transmit.
Dynamic channel allocation protocols are designed to overcome this drawback. With
dynamic allocation strategies, the channel allocation changes with time and is based on
current (and possibly changing) demands of the various users. The more responsive and
better usage of the channel achieved with dynamic protocols does not come for free: it
requires control overhead that is unnecessary with static protocols and consumes a portion
of the channel.
As an example of a protocol that belongs to the family of dynamic conﬂictfree protocols
we take the protocol known as Mini Slotted Alternating Priority (MSAP) protocol. It is
designed for a slotted system, i.e., one in which the time axis is divided into slots of equal
duration and where a user’s transmission is limited to within the slot (the TDMA systemis
also such a system).
To ensure freedom of transmission conﬂicts it is necessary to reach an agreement among
the users on who transmits in a given slot. This agreement entails collecting information as
to who are the ready users, i.e., those who request channel allocation, and an arbitration
algorithm by which one of these users is selected for transmission. This latter mechanism
is nothing but imposing a priority structure on the set of users each of which constitutes a
separate priority class. The MSAP protocol handles properly various such structures. The
presentation here follows that of Kleinrock and Scholl [KlS80].
γ
i
i 1 =
K
∑
1 =
34 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
Let the users be numbered sequentially 0,1,..., M1. The priority enforcement is based on
the observation that if in the most recent slot the channel was allocated to user i then it
must have been the one with the highest priority. Deﬁning the priority structure is thus the
determination of the transmission order after the transmission of some user. Given then,
that user i transmitted last we deﬁne the following priority structures:
• Fixed Priorities. Transmission order: 0, 1,..., M1.
• RoundRobin. Transmission order: i+1, i+2,..., i+M (user arithmetic is modulo M).
• Alternating Priorities. Transmission order: i, i+1,..., i+M1.
The ﬁxed priority structure implies that user i has always higher priority than user i+1,
thus it treats preferentially the smaller numbered users, and is somewhat “less fair”
(although desired in some cases). The roundrobin priority structure implies a channel that
is allocated to the users in a cyclic order and ensures that between any two transmissions
of any user all other users have a chance to transmit at least once. The alternating priorities
structure allows a user to whom the channel is allocated to transmit all the messages in its
buffer before the channel is allocated to another user; in other respects this structure is a
roundrobin one since the channel is allocated to the different users in a cyclic order.
Once the priority structure is decided upon the only issue left is to identify the one user
with the highest priority among those wishing to transmit. MSAP does this by means of
reservations as follows. Denote by τ the maximum system propagation delay, that is, the
longest time it takes for a signal emitted at one end of the network to reach the other end.
Let every slot consist of initial M1 reservation “minislots” each of duration τ, followed by
a data transmission period of duration T, followed by another minislot (see Figure 2.9).
Only those users wishing to transmit in a slot take any action; a user that does not wish to
transmit in a given slot remains quiet for the entire slot duration. Given that every user
wishing to transmit knows his own priority they behave as follows:
1 2
M

1
1 2
τ
T
Slot
Reservation
Minislots
Data
Transmission
t
FIGURE 2.9: Msap Slot Structure
Section 2.4.: DYNAMIC CONFLICTFREE PROTOCOLS 35
• If the user of the highest priority wishes to transmit in this slot then he starts immedi
ately. His transmission consists of an unmodulated carrier for a duration of M1 minis
lots followed by a message of duration T.
• A user of the ith priority ( ) wishing to transmit in this slot will do so only
if the ﬁrst i minislots are idle. In this case he will transmit M1i minislots of unmodu
lated carrier followed by a message of duration T.
The speciﬁc choice of the minislot duration ensures that when a given user transmits in a
minislot all other users know it by the end of that minislot allowing themto react appropri
ately. The additional minislot at the end allows the data signals to reach every user of the
network. This is needed to ensure that all start synchronized in the next slot, as required by
the reservation scheme.
The evaluation of throughput is fairly simple. Since transmission is conﬂictfree every
nonempty slot conveys useful data. However, the ﬁrst M1 minislots as well as the one
after data transmission are pure overhead and should not be counted in the throughput.
Thus if all slots are used, that is in the highest possible load circumstances, we get a chan
nel capacity (maximum throughput) of
where is a characteristic parameter of the system. It is evident that the capacity
increases both with the reduction of a and the number of users. For a typical value of
a=0.01 only a few tens of users can be tolerated before the capacity reduces below an
acceptable level.
2.4.1. Expected Delay
Consider the problem in a somewhat more general setting. Let the arrival processes of
packets at user i be Poisson with rate λ
i
( ). Let be the waiting time of a
packet at user i (with mean w
i
), and let be the transmission (service) time of a single
packet (with mean x
i
). Denote , and . In the following derivation we
assume a ﬁxed priority structure among the users; as we indicate later, some of the results
apply also to other priority structures. In a ﬁxed priority structure λ
i
can be interpreted as
the arrival rate of packets of the ith priority. We also assume that the buffering capabilities
of each user are not limited.
Consider a random “tagged” packet that joins user k. Upon its arrival there are packets
at the ith user already waiting. Let i be the number of packets that arrive at user i dur
ing the waiting time of the tagged packet. The waiting time of that packet is composed of
waiting for the currently transmitting user to complete his transmission (or if there is no
transmitting user the packet will wait until the channel is available for transmission), the
transmission time of all packets with equal or higher priority (anywhere in the system) that
1 i M 1 – ≤ ≤
S
max
T
T Mτ +

1
1 Ma +
 = =
a τ T ⁄
·
∆
1 i M 1 – ≤ ≤ w
˜
x
˜
ρ
i
λ
i
x
i
·
∆
ρ ρ
i ∑
=
L
˜
i
N
˜
i
36 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
are waiting upon his arrival, and the transmission time of all higher priority packets that
will arrive (anywhere in the system) before he starts transmission. Denote by the forced
idletime, that is an idle time imposed by the server (the channel) when it empties. With
the above deﬁnitions we have
(2.39)
The ﬁrst termin this equation is the average time before the next transmission starts. If the
system is nonempty, then ρ
i
is the probability that user i is transmitting, and the average
residual transmission time is . Similarly, a user encountering an empty
system (probability 1ρ) must wait the residual forced idletime. The second term denotes
the total transmission time of all packets of equal or higher priority that are waiting for
transmission. The third term is the total transmission time of all packets of higher priority
that will arrive while the tagged packet is waiting. Note that within a priority class FCFS
order prevails.
Straightforward application of Little’s formula yields and
which, when substituted into equation (2.39) yields
where . The above equation can be recursively evaluated to yield
The mean packet delay, averaged over all priority levels is which in general
cannot be evaluated in closed form. However, if the transmission times of all priority lev
els have the same distribution (that is is independent of i) then
(2.40)
t
˜
E w
˜
k
[ ] w
k
ρ
i
E x
˜
i
2
[ ]
2E x
˜
i
[ ]

i 0 =
M 1 –
∑
1 ρ – ( )
E t
˜
2
[ ]
2E t
˜
[ ]
 + = =
E x
˜
i
[ ]E L
˜
i
[ ]
i 0 =
k
∑
E x
˜
i
[ ]E N
˜
i
[ ]
i 0 =
k 1 –
∑
+ +
E x
i
˜ 2
[ ] 2E x
˜
i
[ ] ( ) ⁄
E L
˜
i
[ ] λ
i
w
i
= E N
˜
i
[ ] λ
i
w
k
=
w
k
ρ
i
E x
˜
i
2
[ ]
2E x
˜
i
[ ]

i 0 =
M 1 –
∑
1 ρ – ( )
E t
˜
2
[ ]
2E t
˜
[ ]
 ρ
i
w
i
i 0 =
k 1 –
∑
+ +
1 ρ
k ( )
–
 =
ρ
k ( )
ρ
i
i 0 =
k
∑
·
∆
w
k
ρ
i
E x
˜
i
2
[ ]
2E x
˜
i
[ ]

i 0 =
M 1 –
∑
1 ρ – ( )
E t
˜
2
[ ]
2E t
˜
[ ]
 +
1 ρ
k ( )
– ( ) 1 ρ
k 1 – ( )
– ( )
 =
λ
i
λ ⁄ ( )w
i ∑
x
˜
i
w E w
˜
[ ]
λ
i
λ
w
i
i 0 =
M 1 –
∑
ρ
E x
˜
2
[ ]
2E x
˜
[ ]
 1 ρ – ( )
E t
˜
2
[ ]
2E t
˜
[ ]
 +
1 ρ –
 = = =
Section 2.4.: DYNAMIC CONFLICTFREE PROTOCOLS 37
If, in addition is distributed in the same way as we get
(2.41)
For the MSAP case, all the are indeed the same and equal the slot size namely
. The quantity represents the forced idletime incurred when the
system becomes empty (a transmission can start only at slot boundaries) and is also deter
ministically equal to the slot size. Thus the conditions of equation (2.41) hold and when
substituted yields
with . The mean packet delay is thus
which ﬁnally yields
(2.42)
Note that since ρ is the fraction of periods in which transmission takes place and since Ma/
(1+Ma) of every slot is overhead, we conclude that S =ρ /(1+Ma), or ρ=(1+Ma)S. When
substituted into the last equation the throughput delay characteristic of MSAP results.
Although the preceding analysis was performed for ﬁxed priorities the ﬁnal results hold
for other priority structures as well. The priority conservation law [Kle76] states that for
any M/G/1 queue and any nonpreemptive workconserving queueing discipline (which
includes all those of concern here) .Thus, as surprising as it might
appear, the average delay given by equations (2.40)and (2.42) holds for all priority struc
tures mentioned in the beginning of this subsection. Higher moments are, as expected, dif
ferent.
t
˜
x
˜
E w
˜
[ ]
E x
˜
2
[ ]
2E x
˜
[ ]

1 ρ –
 =
x
˜
i
Mτ T + T 1 Ma + ( ) = t
˜
w
k
1 Ma + ( )T
2 1 ρ
k ( )
– ( ) 1 ρ
k 1 – ( )
– ( )
 =
ρ
k ( )
1 Ma + ( )T λ
i
i 0 =
k
∑
=
D
λ
i
λ
D
i
i 0 =
M 1 –
∑
λ
i
λ
 w
i
1 Ma + ( )T + [ ]
i 0 =
M 1 –
∑
E w
˜
[ ] 1 Ma + ( )T + = = =
1 Ma + ( )T 1
1
2 1 ρ – ( )
 + =
D
ˆ
D
T
 1 Ma + ( ) 1
1
2 1 ρ – ( )
 + = =
ρ
i
w
i ∑
Const =
38 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
2.5. RELATED ANALYSIS
Conﬂictfree access protocols, especially the static ones, are the oldest and most popular
protocols around. This is the reason for the large volume of analyses of such protocols. We
suggest here several of those analyses along with some other conﬂictfree protocols.
FDMA and TDMA
A good treatment of TDMA and FDMA analysis can be found in [Hay84]. A sample path
comparison between FDMA and TDMA schemes is carried out in [Rub79] where it is
shown that TDMA scheme is better than the FDMA scheme not only on the average. A
TDMA scheme in which the packets of each user are serviced according to a priority rule
is analyzed in [MoR84]. The question of optimal allocation of slots to the users in the gen
eralized TDMA scheme is addressed in [ItR84] where the throughput of the system is
maximized (assuming single buffers for each user) and in [HoR87] where the expected
packetdelay in the system is minimized.
Code Division Multiple Access
Both FDMA and TDMA does not allow any time overlap of transmissions. A conﬂictfree
scheme that does allow overlap of transmission both in the frequency and the time domain
is the.code division multiple access (CDMA) [Pur87]. The conﬂictfree property of
CDMA is achieved by using orthogonal signals in conjunction with matched ﬁlters in the
corresponding receivers. Interconnecting all users in the system requires that matched ﬁl
ters corresponding to all signals to be available at all receivers. The use of multiple orthog
onal signals increases the bandwidth required for transmission. Yet, CDMA allows the
coexistence of several systems in the same frequency bands, as long as different signals
are used in different systems.
Reservation Protocols
The MSAP protocol presented in the text is a representative of an entire family of protocol
that guarantee conﬂictfree transmission by way of reservation. All these protocols have a
sequence of bits precede serving to reserve or announce upcoming transmissions (this is
known as the reservation preamble). In MSAP there are M1 such bits for every packet
transmitted. We mention here those reservation protocols which basically do not involve
contention for the reservation itself; such protocols are discussed in Section 3.5.
An improvement to the MSAP protocol is the bitmap protocol described in [Tan81]. The
idea behind this protocol is to use a single reservation preamble to schedule more than a
single transmission. This is done by utilizing the fact that all participating nodes are aware
of the reservations made in the preamble. The bitmap protocol requires synchronization
among the users that is somewhat more sophisticated than the MSAP protocol, but the
overhead paid per transmitted packet is less than the overhead in the MSAP protocol.
Section 2.5.: RELATED ANALYSIS 39
Another variation of a reservation protocol has been described in [Rob75]. There, every
user can make a reservation in every minislot of the reservation preamble, and if the reser
vation remains uncontended that reserving user will transmit. If there is a collision in the
reservation minislot all users but the “owner” of that minislot will abstain from transmis
sion. Altogether, this is a standard TDMA with idle slots made available to be grabbed by
others.
One of the most efﬁcient reservation protocol is the Broadcast Recognition Access
Method [CFL79]. This is essentially a combination between the bitmap and the MSAP
protocols. As in the MSAP protocol, a reservation preamble serves to reserve the channel
for a single user but unlike the MSAP, the reservation preamble does not necessarily con
tain all M1 minislots. The idea is that users start their transmission with a staggered delay
not before they ensure that another transmission is not ongoing (the paper [KlS80] also
refers to a similar scheme). Under heavy load BRAM reduces to regular TDMA.
40 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
EXERCISES
Problem 1.
Assume that a portion y of every transmitted packet is overhead (e.g., address, sync bits,
etc.).
1. What will be the throughput delay characteristic of an FDMA channel?
2. What will be the throughput delay characteristic of a TDMA channel?
Problem 2.
Derive the Laplace transformof the message delay in FDMA in which every message con
tains a random number of packets. Compare the expected message delay with that of
TDMA.
Problem 3.
Compare the ﬁrst two moments of the distribution of the queueing time of FDMA with
that of TDMA (Note: the queueing time does not include the actual transmission time).
Problem 4.
Derive the steadystate distribution and the ﬁrst two moments of the number of messages
in a TDMA system where L(z) is the generating function of the number of packets in a
message.
Problem 5.
Consider a TDMA systemin which a user is assigned the ﬁrst two slots of every frame but
a message transmission will start only at the ﬁrst slot in the frame. Assume a Poisson mes
sage arrival process with rate λ messages/second and the number of packets in a message
distributed according to the generating function L(z). We are interested in the message
delay distribution at the user.
1. Deﬁne an appropriate set of random variables and write the equation for the delay
of an arbitrary (“tagged”) message.
2. Find the generating function of the number of frames required to transmit a message
containing packets.
3. Compute the Laplace transform of the message delay and derive from it the average
message delay.
D
˜
f
˜
L
˜
Section : EXERCISES 41
Problem 6.
For the generalized TDMA scheme derive the generating function of the number of pack
ets at the beginning of an arbitrary slot. Compute also the expected number of packets at
the beginning of an arbitrary slot.
Problem 7.
Show that when the interallocation of slots in the generalized TDMA scheme is uniform,
i.e., d(k) = T
c
/K for , then the Laplace transform of the delay distribution in
(2.37) reduces to that in (2.16) (see Sections 2.2. and 2.3.).
Problem 8.
This problem addresses the optimal allocation of slots (i.e., choosing d(k)) in the general
ized TDMA protocol when the load is very light, i.e., or equivalently . Let
where γ
i
( ) is the probability that a message transmission requires i slots beyond
the number of whole frames, or in other words the probability that a message length is
imodK (clearly ).
1. Show that it is sufﬁcient to minimize F in order to minimize the expected delay under
the light load circumstances
2. Show that the expression for F reduces to
Note fromthe above expression that when γ
i
= γ
K^i+1
for i=1,2,..., K, then the expected
delay is completely independent of the interallocation distances.
3. Let γ
i
= γ
K^i+1
for i=2,3,..., K1. What is the optimal allocation when γ
1
>γ
K
and when
γ
1
<γ
K
?
Problem 9.
Referring to the delay analysis of the MSAP protocol. Prove that
1 k K ≤ ≤
ρ 0 → λ 0 →
F 1
1
T
c
 γ
i
1
2
 d
2
j ( )
j 1 =
K
∑
d j ( )d i ( )
l j 1 + =
j i 1 – +
∑
j 1 =
K
∑
+
i 1 =
K
∑
+ =
1 i K ≤ ≤
γ
i
i 1 =
K
∑
1 =
F 1
1
2
 T
c
1
T
c
 d l ( )d l k + ( ) γ
i
γ
K i – 1 +
– ( )
i k 1 + =
K
∑
k 1 =
K 1 –
∑
l 1 =
K
∑
+ + =
w
k
ρ
i
E x
˜
i
2
[ ]
2E x
˜
i
[ ]
 1 ρ – ( )
E t
˜
2
[ ]
2E t
˜
[ ]
 +
i 0 =
M 1 –
∑
1 ρ
k ( )
– ( ) 1 ρ
k 1 – ( )
– ( )
 =
42 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
(with ) is a solution of the general equation, and that if all have the
same distribution then
ρ
k ( )
ρ
i
i 0 =
k
∑
= x
˜
i
E w
˜
[ ]
λ
i
λ
w
i
i 0 =
M 1 –
∑
ρ
E x
˜
2
[ ]
2E x
˜
[ ]
 1 ρ – ( )
E t
˜
2
[ ]
2E t
˜
[ ]
 +
1 ρ
k ( )
– ( ) 1 ρ
k 1 – ( )
– ( )
 = =
Section : APPENDIX A Distribution of the Mod Function 43
APPENDIX A
Distribution of the Mod Function
Let be a nonnegative integer valued random variable with a known distribution and a
generating function
l(z), and let K be a known positive integer constant. The quantity can be uniquely
decomposed into
.
In another form this can be written as and . We
would like to compute the distributions of and from that of .
Let β
m
be the unit roots of order K namely, . These roots obey
Our most basic relation is derived as follows:
(2.43)
By setting z=1 in equation (2.43) we get
which gives us the distribution of . From this J(z), the generating function of , can be
computed as follows:
l
˜
l
˜
l
˜
f
˜
K ⋅ J
˜
+ = 0 J
˜
K 1 – ≤ ≤
J
˜
l
˜
modK = f
˜
l
˜
K ⁄ l
˜
J
˜
– ( ) K ⁄ = =
J
˜
f
˜
l
˜
β
m
e
i
2πm
K

=
1
K
 β
n
m 0 =
K 1 –
∑
1 ∆ nmodK ( ) –
1 K divides n
0 Otherwise
¹
'
¹
= =
1
K
 zβ
m
( )
n –
l zβ
m
( )
m 0 =
K 1 –
∑
1
K
 zβ
m
( )
n –
l
˜
l = [ ] zβ
m
( )
l
Prob
l 0 =
∞
∑
m 0 =
K 1 –
∑
=
1
K
 l
˜
l = [ ]z
l n –
β
m
( )
l n –
m 0 =
K 1 –
∑
Prob
l 0 =
∞
∑
=
l
˜
l = [ ]z
l n –
1 ∆ l n – ( )modK ( ) – [ ] Prob
l 0 =
∞
∑
=
l
˜
f K ⋅ n + = [ ] Prob
f 0 =
∞
∑
=
J
˜
n = [ ] Prob l
˜
f K ⋅ n + = [ ] Prob
f 0 =
∞
∑
1
K
 β
m
n –
l β
m
( )
m 0 =
K 1 –
∑
= = 0 n K 1 – ≤ ≤
J
˜
J
˜
44 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
where in the step before last we used the fact that . Overall, we thus have
Taking the derivative at z=1 yields the expectation
We turn now to calculate the generating function of . Clearly
and thus
We note that the bracketed term in the summation appears in equation (2.43) when z
1/K
is
substituted for z. Hence
J
˜
n = [ ]z
n
Prob
n 0 =
K 1 –
∑
1
K
 β
m
n –
l β
m
( )
m 0 =
K 1 –
∑
z
n
n 0 =
K 1 –
∑
1
K

z
β
m

¸ ,
¸ _
n
n 0 =
K 1 –
∑
l β
m
( )
m 0 =
K 1 –
∑
= =
1
K

1
z
β
m

¸ ,
¸ _
K
–
1
z
β
m
 –
  l β
m
( )
m 0 =
K 1 –
∑
1
K

1 z
K
–
1
z
β
m
 –
  l β
m
( )
m 0 =
K 1 –
∑
1 z
K
–
K

l β
m
( )
1
z
β
m
 –
 
m 0 =
K 1 –
∑
= = =
β
m
K
1 =
J z ( )
1 z
K
–
K

l β
m
( )
1
z
β
m
 –
 
m 0 =
K 1 –
∑
1 z
K
–
K 1 z – ( )

1 z
K
–
K

l β
m
( )
1
z
β
m
 –
 
m 1 =
K 1 –
∑
+ = =
E J
˜
[ ]
K 1 –
2

l β
m
( )
1
1
β
m
 –
 
m 1 =
K 1 –
∑
+ =
f
˜
f
˜
f = [ ] Prob l
˜
f K ⋅ n + = [ ] Prob
n 0 =
K 1 –
∑
= f 0 ≥
z ( ) f
˜
f = [ ] Prob z
f
f 0 =
∞
∑
l
˜
f K ⋅ n + = [ ] Prob
n 0 =
K 1 –
∑
z
f
f 0 =
∞
∑
= = =
l
˜
f K ⋅ n + = [ ] Prob z
f
f 0 =
∞
∑
n 0 =
K 1 –
∑
=
Section : APPENDIX A Distribution of the Mod Function 45
Calculating the expected value of can be done by taking the derivative of the above
equation at z=1 or using the direct approach, i.e.,
.
F z ( )
1
K
 z
1 K /
β
m
( )
n –
l z
1 K /
β
m
( )
m 0 =
K 1 –
∑
n 0 =
K 1 –
∑
=
1
K
 z
1 K /
β
m
( )
n –
n 0 =
K 1 –
∑
l z
1 K /
β
m
( )
m 0 =
K 1 –
∑
1
K

1 z
1 K /
β
m
( )
K –
–
1 z
1 K /
β
m
( )
1 –
–
l z
1 K /
β
m
( )
m 0 =
K 1 –
∑
= =
1
K

1 z
1 –
–
1 z
1 K /
β
m
( )
1 –
–
 l z
1 K /
β
m
( )
m 0 =
K 1 –
∑
=
1 z
1 –
–
K

l z
1 K /
β
m
( )
1 z
1 K /
β
m
( )
1 –
–

m 0 =
K 1 –
∑
=
f
˜
E f
˜
[ ]
E l
˜
[ ] E J
˜
[ ] –
K
 
1
K
E l
˜
[ ]
K 1 –
2K
 –
1
K

l β
m
( )
1
1
β
m
 –

m 1 =
K 1 –
∑
+ = =
46 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
CHAPTER 3
ALOHA PROTOCOLS
The Aloha family of protocols is probably the richest family of multiple access protocols.
Its popularity is due ﬁrst of all to seniority, as it is the ﬁrst random access technique intro
duced. Second, many of these protocols are so simple that their implementation is straight
forward. Many local area networks of today implement some sophisticated variants of this
family’s protocols.
With the conﬂictfree protocols that were discussed in Chapter 2, every scheduled trans
mission is guaranteed to succeed. The Aloha family of protocols belongs to the conten
tiontype or randomretransmission protocols in which the success of a transmission is not
guaranteed in advance. The reason is that whenever two or more users are transmitting on
the shared channel simultaneously, a collision occurs and the data cannot be received cor
rectly. This being the case, packets may have to be transmitted and retransmitted until
eventually they are correctly received. Transmission scheduling is therefore the focal con
cern of contentiontype protocols.
Because of the great popularity of Aloha protocols, analyses have been carried out for a
very large number of variations. The variations present different protocols for transmission
and retransmission schedules as well as adaptation to different circumstances and channel
features. This chapter covers a few of these variations.
3.1. PURE ALOHA
The pure Aloha protocol is the basic protocol in the family of the Aloha protocols. It con
siders a singlehop systemwith an inﬁnite population generating packets of equal length T
according to a Poisson process with rate λ packets/sec. The channel is errorfree without
capture: whenever a transmission of a packet does not interfere with any other packet
transmission, the transmitted packet is received correctly while if two or more packet
transmissions overlap in time, a collision is caused and none of the colliding packets is
received correctly and they have to be retransmitted. The users whose packets collide with
one another are called the colliding users. At the end of every transmission each user
knows whether its transmission was successful or a collision took place.
The pure Aloha protocol is very simple [Abr70]. It states that a newly generated packet is
transmitted immediately hoping for no interference by others. Should the transmission be
unsuccessful, every colliding user, independently of the others, schedules its retransmis
sion to a random time in the future. This randomness is required to ensure that the same
set of packets does not continue to collide indeﬁnitely. A simple example of the operation
of the protocol is depicted in Figure 3.1 where the arrows indicate arrival instants, success
ful transmissions are indicated by blank rectangles and collided packets are hatched.
48 CHAPTER 3: ALOHA PROTOCOLS
Since the population is inﬁnite each packet can be considered as if it belongs to a different
user. Hence, each newly arrived packet can be assigned to an idle user i.e., one that does
not have a packet to retransmit. This allows us to interchange the roles of users and pack
ets and consider only the points in time when packet transmission attempts are made.
Observing the channel over time we deﬁne a point process consisting of scheduling points,
i.e., the points in which packets are scheduled for transmission. The scheduling points
include both the generation times of new packets and the retransmission times of previ
ously collided packets. Let the rate of the scheduling points be g packets/sec. The parame
ter g is referred to as the offered load to the channel. Clearly, since not all packets are
successful on their ﬁrst attempted transmission, g>λ.
The exact characterization of the scheduling points process is extremely complicated. To
overcome this complexity it is assumed that this process is a Poisson process (with rate g,
of course). This assumption can, however, be a good approximation at best (as has indeed
been shown by simulation). The reason is that a Poisson process implies independence
between events in nonoverlapping intervals, which cannot be the case here because of the
dependence between the interval containing the original transmission and the interval con
taining a retransmission of the same packet. It can be shown, however, that if the retrans
mission schedule is chosen uniformly froman arbitrarily large interval then the number of
scheduling points in any interval approaches a Poisson distribution. The Poisson assump
tion is used because it makes the analysis of Alohatype systems tractable and predicts
successfully their maximal throughput.
Pure Aloha is a singlehop system. Hence, the throughput is the fraction of time the chan
nel carries useful information, namely noncolliding packets. The channel capacity is the
highest value of arrival rate λ for which the rate of departure (throughput) equals the total
arrival rate (but see the discussion of stability in Section 3.4.).
Vulnerable
Period
T
tT
t+T
t
Retransmissions
FIGURE 3.1: Pure aloha Packet Timing
time
Section 3.2.: SLOTTED ALOHA 49
Consider a packet (new or old) scheduled for transmission at some time t (see Figure 3.1).
This packet will be successful if no other packet is scheduled for transmission in the inter
val (tT,t+T) (this period of 2T is called the vulnerable period). The probability of this hap
pening, that is, the probability of success, is that no packet is scheduled in an interval of
length 2T and since scheduling is Poisson we have
Now, packets are scheduled at a rate of g per second of which only a fraction P
suc
are suc
cessful. Thus, the rate of successfully transmitted packets is gP
suc
. When a packet is suc
cessful the channel carries useful information for a period of T seconds; in any other case
it carries no useful information at all. Using the deﬁnition that the throughput is the frac
tion of time that useful information is carried on the channel we get
which gives the channel throughput as a function of the offered load. Deﬁning to
be the normalized offered load to the channel, i.e., the rate (per packet transmission time)
packets are transmitted on the channel, we have
The relation between S and G is depicted in Figure 3.2, which is typical to many Aloha
type protocols. At G=1/2, S takes on its maximal value of . This value is
referred to as the capacity of the pure Aloha channel.
We recall that for a systemto be stable the long termrate of input must equal the long term
rate of output meaning that stability requires S= λT. Larger values of λ clearly cannot
result in stable operation. Note however, that even for smaller values of λ there are two
values of G to which it correspondsone larger and one smaller than 1/2. The smaller one
is (conditionally) stable while the other one is conditionally unstable, meaning that if the
offered load increases beyond that point the system will continue to drift to higher load
and lower throughput. Thus, without additional measures of control, the stable throughput
of pure Aloha is 0. We return to the stability issue in Section 3.4. (It is appropriate to add
that this theoretical instability is rarely a severe problem in real systems, where the long
term load including, of course, the “off hours” load, is fairly small although temporary
problems may occur).
3.2. SLOTTED ALOHA
The slotted Aloha variation of the Aloha protocol is simply that of pure Aloha with a slot
ted channel. The slot size equals Tthe duration of packet transmission. Users are
restricted to start transmission of packets only at slot boundaries. Thus, the vulnerable
period is reduced to a single slot. In other words, a slot will be successful if and only if
exactly one packet was scheduled for transmission sometime during the previous slot. The
P
suc
e
2gT –
=
S gTe
2gT –
=
G gT
·
∆
S Ge
2G –
=
1 2e ( ) ⁄ 0.18 ≈
50 CHAPTER 3: ALOHA PROTOCOLS
throughput is therefore the fraction of slots (or probability) in which a single packet is
scheduled for transmission. Because the process composed of newly generated and
retransmitted packets is Poisson we conclude that
or using the deﬁnition of the normalized offered load G = gT
This relation is very similar to that of pure Aloha, except of increased throughput (see Fig
ure 3.2). Channel capacity is and is achieved at G=1.
We would like to introduce an additional method of calculating throughput which will be
useful later and which can be easily demonstrated in the slotted Aloha case.
When observing the channel over time we notice a cyclic behavior of busy and idle peri
ods (see Figure 3.3 in which uparrows point at arrival instants, blank rectangles refer to
successful slots, and hatched rectangles denote colliding packets, i.e., unsuccessful slots).
A busy period is a succession of slots in which transmission takes place (successful or
not). The idle period is deﬁned as the interval between two busy period. The starting times
S gTe
gT –
=
S Ge
G –
=
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.001 0.01 0.1 1 10 100
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
Slotted Aloha
Pure Aloha
FIGURE 3.2: ThroughputLoad of Pure and Slotted Aloha
1 e ⁄ 0.36 ≈
Section 3.2.: SLOTTED ALOHA 51
of every cycle (just before the start of the busy period) deﬁne renewal points. In fact, these
are points of a regenerative process, since the system is memoryless in the sense that its
behavior in a given cycle does not depend on the behavior in any previous cycle. As such,
the expected fraction of time the system is in a given state equals the expected fraction of
time during a single cycle that the system is in that state (see Appendix).
Let be a random variable describing the number of slots in the idle period. The random
variable must be strictly positive since there must be at least one empty slot in an idle
period. The probability that the idle period consists of a single slot is the probability that
there were some packets scheduled during that slot that will be transmitted in the next slot.
Thus,
The probability that the idle period lasts exactly two slots is the probability of the event
that no packets were scheduled in the ﬁrst slot and some were scheduled in the second (to
be transmitted in the third slot). Thus,
In general, the length of the idle period is seen to be geometrically distributed, namely
yielding an average length (measured in slots) of
Busy Period
Idle
Period
Cycle
FIGURE 3.3: SlottedAloha Packet Timing
I
˜
I
˜
P I
˜
1 = [ ] P Some packets scheduled in first slot [ ] =
1 P No packets scheduled in first slot [ ] – = 1 e
gT –
– =
P I
˜
2 = [ ] e
gT –
1 e
gT –
– ( ) ⋅ =
P I
˜
k = [ ] e
gT –
( )
k 1 –
1 e
gT –
– ( ) ⋅ = k 1 2 … , , =
I
1
1 e
gT –
–
 =
52 CHAPTER 3: ALOHA PROTOCOLS
Similarly, let us deﬁne as the number of slots in the busy period. Clearly . An
argument similar to that used in calculating the distribution of the idle period leads to the
derivation of the distribution of . For the busy period to be k>0 slots long, packets must
be scheduled for transmission in each and every one of the ﬁrst k1 slots and none sched
uled in the kth. This leads to
yielding an expected value of
Since not all the slots in the busy period are successful, let denote the number of useful,
or successful slots in a cycle and let U be its expected value. The probability that a given
slot in the busy period is successful is
which is the probability of a single arrival in a slot given that we are in the busy period
(i.e., some arrivals do occur). Thus, given that the duration of the busy period is slots we
have
and hence
from which we get
Now, the throughput is the expected fraction of slots within a cycle in which successful
transmission takes place. If denotes the number of slots in a cycle then
B
˜
B
˜
0 >
B
˜
P B
˜
k = [ ] 1 e
gT –
– ( )
k 1 –
e
gT –
= k 1 2 … , , =
B
1
e
gT –
 =
U
˜
gTe
gT –
1 e
gT –
–

B
˜
P U
˜
k = B
˜
[ ]
B
˜
k
¸ ,
¸ _
gTe
gT –
1 e
gT –
–

¸ ,
¸ _
k
1
gTe
gT –
1 e
gT –
–
 –
¸ ,
¸ _
B
˜
k –
= 0 k B
˜
≤ ≤
E U
˜
B
˜
[ ] B
˜
=
gTe
gT –
1 e
gT –
–

U E U
˜
[ ] E E U
˜
B
˜
[ ] [ ] B
gTe
gT –
1 e
gT –
–
 = = =
C
˜
S
E U
˜
[ ]
E C
˜
[ ]

U
B I +

B
gTe
gT –
1 e
gT –
–

B I +

1
e
gT –

gTe
gT –
1 e
gT –
–
 ⋅
1
e
gT –

1
1 e
gT –
–
 +
 gTe
gT –
= = = = =
Section 3.3.: SLOTTED ALOHA  FINITE NUMBER OF USERS 53
which is the result obtained previously. The technique of using regenerative processes as
done above is an important tool in deriving the throughput of more complicated protocols.
3.3. SLOTTED ALOHA  FINITE NUMBER OF USERS
To make the previous model more realistic we analyze here the case of an Aloha system
with a ﬁnite number of users. The analysis of this model enables us to derive packet delays
which we were unable to do in the previous model. The following analysis is based on that
of Kleinrock and Lam [KlL75]. We consider a case in which slotted Aloha is used by a
group of M users each with a single packet buffer (this less general case makes the com
parison with the inﬁnite population cases more meaningful since there too every user had
only a single buffer). All packets are of the same size, requiring T seconds for transmis
sion, which is also the slotduration.
To gain insight into the relation between transmission of new packets and retransmission
of old ones we build the following packetscheduling model (referred to as the linear feed
back model). Let every user be in one of two statesthinking and backlogged. In the think
ing state the user does not have a packet in its buffer and does not participate in any
scheduling activities. When in this state, the user generates a packet in every slot with
probability σ and does not generate a packet in a slot with probability 1σ; packet genera
tion is independent of any other activity. The preceding means that packet generation is an
independent process distributed geometrically with mean 1/σ. Once a packet is generated
its transmission is attempted immediately, that is, in the next slot. If the transmission was
successful the user remains in the thinking state and the packet generating process starts
anew. If packet transmission was unsuccessful the user moves to the backlogged state and
schedules the retransmission of the packet according to an independent geometric distribu
tion with parameter ν. In other words, in every slot the user will retransmit the packet with
probability ν and will refrain fromdoing so with probability 1ν. While in the backlogged
state the user does not generate any new packets. When the packet is ﬁnally successfully
transmitted the user moves back to the thinking state.
Let the slots of the system be numbered sequentially k=0,1,... and let denote the
number of backlogged users at the beginning of the kth slot. The random variable is
referred to as the state of the system. The number of backlogged users at the beginning of
the k+1st slot depends on the number of backlogged users at the beginning of the kth slot
and the number of users that moved from state to state within the slot. Since statetransi
tion of the users is independent of the activities in any previous slot the process
is a Markov chain. Because the number of backlogged users can
not exceed M this chain is ﬁnite; thus, if all states communicate (as we subsequently indi
cate) this Markov chain is also ergodic, meaning that steadystate distribution exists.
The transition diagram for the system is shown in Figure 3.4. “Upward” transitions are
possible between every state and all the highernumbered states, since collision of any
number of packets is possible. “Downward” transitions are possible only to the adjacent
N
˜
k ( )
N
˜
k ( )
N
˜
k ( ) k 1 2 … , , = ( ) , { }
54 CHAPTER 3: ALOHA PROTOCOLS
state since only one packet can be successfully transmitted in a slot, at which time the
backlog is reduced by unity. Note also the missing transition from state 0 to state 1 which
is clear since if all users were thinking and a single user generated and transmitted a
packet he could not cause a collision and become backlogged. The fact that all states com
municate is evident from the diagram.
SteadyState Probabilities
For analysis purposes we introduce the following notation (see Appendix). Let π
i
be the
steadystate probability of the system being in state i, that is
. Further, let p
ij
be the steadystate transition probability,
i.e., . Finally, denote by P the matrix whose
elements are p
ij
and by the row vector whose elements are π
i
. From the above argumen
tation it follows that the steadystate probability vector is the solution to the ﬁnite set of
linear equations
to which the existence of a unique solution is guaranteed. We must therefore construct the
matrix P and derive the desired solution.
FIGURE 3.4: State Transitions of Finite Population Aloha
0 1 2 i M
From
0
1
i1
To
M
i+2
i+1
To i1
From i+1
i
π
i
lim
k ∞ →
Prob N
˜
k ( ) i = [ ] =
p
ij
lim
k ∞ →
Prob N
˜
k ( ) j N
˜
k 1 – ( ) i = = [ ] =
π
π πP = π
i
i
∑
1 =
Section 3.3.: SLOTTED ALOHA  FINITE NUMBER OF USERS 55
Since the retransmission process of every user is an independent geometric process the
probability that i out of the j backlogged users will schedule a retransmission in a given
slot is binomially distributed, namely
(3.1)
In a similar manner, we obtain for the thinking users
(3.2)
since when j users are backlogged, Mj users are thinking.
The matrix P can be constructed by applying equations (3.1) and (3.2) as follows.
Clearly a transition from state i to state j<i1 is impossible implying that p
ij
=0 for those
cases. Consider the transition from state i to state i1. This indicates a reduction in the
backlog which is possible only if a single backlogged packet was transmitted (and no new
packet was generated, of course).
The transition from state i to the very same state can come about from two distinct rea
sons. The ﬁrst results from the circumstance in which no new packet was generated (and
transmitted) while several backlogged users attempted transmission. The transmitting
users clearly collide and remain in the backlog; because no transmission of new packets
was attempted the backlog did not change. A special case of this latter one is when the slot
remains idleneither a transmission of a new packet is attempted nor is a retransmission
attempted. The second reason for this transition results from a situation in which none of
the backlogged users attempt retransmission and a single thinking user transmits. In this
case the thinking user succeeds and therefore remains in the thinking state leaving the sys
tem in the same state. The above can be summarized by the union of the two independent
events: “No backlogged user succeeds and no thinking user attempts” and “No backlogged
user attempts and a single thinking user attempts”.
The next transition to consider is from state i to state i+1. Since the backlog increased, a
collision must have taken place. Furthermore, since the backlog increased by unity,
exactly one thinking user has attempted together with at least one attempt from the back
log.
The last case is the transition from state i to a state j>i+1. Here the backlog increased by
two or more meaning that ji thinking users generated packets and, of course, collided.
The activity of the backlogged users is immaterial in this case since the collision is gener
ated by the thinking users alone.
Prob i backlogged users in a slot  j in backlog [ ]
j
i
¸ ,
¸ _
ν
i
1 ν – ( )
j i –
=
Prob i thinking users transmit in a slot  j in backlog [ ]
M j –
i
¸ ,
¸ _
σ
i
1 σ – ( )
M j – i –
=
56 CHAPTER 3: ALOHA PROTOCOLS
The above can be summarized in the following formulae (where the bracketed terms cor
respond to the events described in the preceding explanation):
(3.3)
It can be easily veriﬁed that as required. Also, note that p
01
turns out to be
identically zero; this result is correct and expected since it takes at least two colliding
packets to increase the backlog and because no users were backlogged before, it is impos
sible to have a single backlogged user at the end of the slot.
Solving the set of equations for the above matrix cannot be done in closed form.
However, the special structure of the P matrixhaving nonzero elements in the upper right
triangle and in the ﬁrst subdiagonalallows fairly easy computation.
Consider the homogeneous set of equations x(IP)=0 where I is the identity matrix. This is
the same set of equations rewritten with x replacing . It can be shown that the rank of I
P is M1 which means, among others, that nontrivial solutions exist and are all collinear.
To ﬁnd one of these solutions assume x
0
=1. The ﬁrst equation, namely
yields . Fromthe second equation we have
, from which x
2
can be calculated. Proceeding simi
larly, every step involves a simple computation and results in determining the value of an
additional x
i
. Having done this Mtimes and calculated the values of all x
i
we then compute
The value of thus computed is colinear with x and therefore solves the original set of
equations. The π
i
clearly sum up to 1. This is therefore our desired solution.
Throughput Analysis
To evaluate the throughput of the system consider the epochs at the beginning of every
slot. Since the activity within a given slot is independent of the activity in any previous slot
these epochs are renewal points. Hence, the long term fraction of time the channel carries
useful informationthe throughputequals the expected fraction of slots containing useful
transmission. If we denote by P
suc
the probability of a successful slot then
p
ij
0 j i 1 – <
iν 1 ν – ( )
i 1 –
[ ] 1 σ – ( )
M i –
j i 1 – =
1 iν 1 ν – ( )
i 1 –
– [ ] 1 σ – ( )
M i –
M i – ( )σ 1 σ – ( )
M i – 1 –
[ ] 1 ν – ( )
i
+ j i =
M i – ( )σ 1 σ – ( )
M i – 1 –
[ ] 1 1 ν – ( )
i
– [ ] j i 1 + =
M i –
j i –
¸ ,
¸ _
σ
j i –
1 σ – ( )
M j –
j i 1 + >
¹
¹
¹
¹
¹
'
¹
¹
¹
¹
¹
=
p
ij
j
∑
1 =
π πP =
π
x
0
1 p
00
– ( ) x
1
p
10
– 0 = x
1
1 p
00
– ( ) p
10
⁄ =
x
0
p
01
– x
1
1 p
11
– ( ) x
2
p
21
– + 0 =
π
i
x
i
x
j
j
∑
  =
π
Section 3.3.: SLOTTED ALOHA  FINITE NUMBER OF USERS 57
For a slot to be successful only a single transmission must take place within it. This means
that either all backlogged users remain silent and a single new user transmits, or a single
backlogged user transmits while no new packet is generated. Given that there are i back
logged users this can be stated as
. (3.4)
The total throughput is therefore
(3.5)
Note that since all users are statistically identical, the individual throughput is given by the
value of S from the last equation divided by M.
As a special case, consider a situation in which we do not distinguish between backlogged
packets and new packets, i.e., we set ν·σ. Substituting this into equation (3.4) yields
indicating that P
suc
(i) is independent of i. This result is, of course, not surprising since if
we cease to distinguish between backlogged and thinking users we cannot expect the
probability of success to depend on the number of backlogged users. Moreover, because
P
suc
(i) is independent of i we obtain from equation (3.5) a closed form expression for the
throughput, namely
(3.6)
Let us continue a bit with this line of thought, i.e., not distinguishing the backlogged from
the thinking users. In previous sections we denoted by G the total, system wide, average
number of transmissions per slot; in our case this equals Mσ. Substituting this value into
the throughput equation above yields
Under these circumstances, letting M increase to inﬁnity we ﬁnd that in the limit
, a result identical to the one derived in Section 3.1.3 for the inﬁnite population
slotted Aloha scheme. We might conclude, therefore, that the inﬁnite population model is
indeed in some sense the limit of the ﬁnite population model if backlogged users are not
S P
suc
=
P
suc
i ( ) Prob successful slot  i users in backlog [ ]
·
∆
=
1 ν – ( )
i
M i – ( )σ 1 σ – ( )
M i – 1 –
iν 1 ν – ( )
i 1 –
1 σ – ( )
M i –
+ =
S P
suc
E P
suc
i ( ) [ ] P
suc
i ( )π
i
i 0 =
M
∑
= = =
P
suc
i ( ) Mσ 1 σ – ( )
M 1 –
=
E P
suc
i ( ) [ ] Mσ 1 σ – ( )
M 1 –
=
S G 1
G
M
 –
¸ ,
¸ _
M 1 –
=
S Ge
G –
=
58 CHAPTER 3: ALOHA PROTOCOLS
distinguished fromthe thinking ones and if the number of users is increased under the con
straint that the total average arrival rate remains ﬁnite.
Expected Delay
The previous derivation considered the throughput from the departure standpoint since
P
suc
is the average rate of packet departure from the system. If the system is to be stable
then this rate must equal the average rate of new packet generation. Now, when the system
is in state i there are Mi thinking users each generating packets in every slot with proba
bility σ. Thus, the average rate of new packet generation when in state i is (Mi)σ. Taking
expectation yields
(3.7)
where s the average number of backlogged users.
Denote by b the average rate at which packets (actually, users with packets) join the back
log; then according to Little’s formula, the average amount of time spent in the backlog is
the ratio of the average number of backlogged users to the average rate of joining or .
Not all packets going through the system are backloggedthe lucky ones make it the ﬁrst
time. Since b is the rate of packets joining the backlog and S (the throughput) is the rate of
packets leaving the system then a fraction (Sb)/S of the packets are never backlogged.
These packets suffer a delay of 1 slot only. All the others (whose fraction is b/S) suffer the
backlog delay mentioned above plus the one slot in which their transmission is successful.
Measured in slots (i.e., normalized to the packet transmission time) the average delay is
.
Using the value of from equation (3.7) ﬁnally yields
(3.8)
With the value of S taken from equation (3.5). This last equation is the desired throughput
delay relation. It should be noted that this representation is parametric since σ inﬂuences
the value of S. Throughput delay characteristics for several parameter choices are depicted
in Figure 3.5. Each of the curves in the ﬁgure represents one value of ν with σ varying
from 0 to 1 along the curve. Thus, the throughput ﬁrst increases with σ until capacity is
achieved (for that value of ν); thereafter the throughput decreases with increasing load.
The delay, as intuitively expected, increases monotonically with σ.
Consider again the special case in which σ·ν, for which the throughput is given in equa
tion (3.6). Substituting this into equation (3.8) yields
S E M i – ( )σ [ ] M i – ( )σπ
i ∑
M N – ( )σ = = =
N
N b ⁄
D
ˆ
S b –
S
 1 ⋅
b
S

N
b
 1 +
¸ ,
¸ _
⋅ + =
N
D
ˆ
1
1
σ
 –
M
S
 + =
Section 3.3.: SLOTTED ALOHA  FINITE NUMBER OF USERS 59
E
x
p
e
c
t
e
d
D
e
l
a
y
D
ˆ
(
)
1
10
100
1000
10000
100000
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Throughput (S)
ν=0.3 ν=0.4 ν=0.5 ν=0.6 ν=0.7
Finite Population
Slotted Aloha
M=10
1
10
100
1000
10000
100000
0 0.05 0.1 0.15 0.2 0.25 0.3
Throughput (S)
ν=0.3 ν=0.4 ν=0.2
Finite Population
Slotted Aloha
M=25
E
x
p
e
c
t
e
d
D
e
l
a
y
D
ˆ
(
)
FIGURE 3.5: ThroughputDelay of Finite Population Slotted Aloha
(a) 10 Users
(b) 25 Users
60 CHAPTER 3: ALOHA PROTOCOLS
Two interesting observations can be made regarding this last result. First, keeping the
product Mσ constant and increasing M shows an ever increasing delay. That is, the model
we have developed cannot be used to evaluate the delay for the inﬁnite population case.
This is not a surprising fact and is due to the instability of the inﬁnitepopulation Aloha
systems, a subject we shall discuss in more detail in Section 3.4. The second interesting
observation relates to the expected delay when σ tends to zero. Taking the limit we ﬁnd
that , a result that may look surprising at ﬁrst. When σ is very small,
hardly ever will a collision result, and in most cases therefore the delay will be a single
slotthat of the transmission itself. However, in the rare case of a collision the colliding
users become of course backlogged, and remain in this state for a very long time since the
average waiting time for a backlogged packet is inversely proportional to σ. Putting it all
together we ﬁnd most packets having a delay of unity, very few packets having extremely
large delays, yielding a combined average delay of M slots.
The Capture Phenomenon
There is a curious phenomenon in ﬁnite population Aloha systems that appears in certain
situations when the retransmission probability is small. We shall introduce this phenome
non through the equations themselves. Consider a situation in which σ accepts nonnegligi
ble values (although not necessarily very close to 1) and ν accepts very small values, in
fact such that Mν<<1. These assumptions in fact mean that eventually most users will
become backlogged since the rate of exit from the backlog is very small. One can there
fore safely assume that most of the time either M or M1 backlogged users will be
observed. Using these assumptions, and approximating (1 ν)
i
as 1iν equations (3.3)
become
and all other values of p
ij
vanish. Solving this set of equations yields
.
The corresponding values of the probability of success are
.
Putting all these together we obtain the following expression for the throughput
D
ˆ
1
1 1 σ – ( )
M 1 –
–
σ 1 σ – ( )
M 1 –
 + =
D
ˆ
σ 0 → ( ) M →
p
M 1 – M 1 – ,
1 M 1 – ( )νσ – = p
M 1 – M ,
M 1 – ( )νσ =
p
M M 1 – ,
Mν = p
M M ,
1 Mν – =
π
M 1 –
M
M M 1 – ( )σ +
 = π
M
M 1 – ( )σ
M M 1 – ( )σ +
 =
P
suc
M 1 – ( ) σ M 1 – ( )ν M 2 – ( )νσ + + = P
suc
M ( ) Mν =
Section 3.3.: SLOTTED ALOHA  FINITE NUMBER OF USERS 61
.
This expression is interesting because it indicates that the throughput increases with the
load σ. Furthermore, substituting these values into equation (3.8) yields
indicating that the average delay is actually decreasing with increasing load! Furthermore,
both the throughput and the average delay do not depend on ν hence these systems exhibit
identical performance for various values of ν. The phenomenon is demonstrated in Figure
3.6 where the throughputdelay curve for a system with M=10 is shown (this ﬁgure
depicts the result of equation (3.8)). One can clearly observe that as long as σ is small
delay increases with throughput as one might generally expect. At higher values of σ we
note a decrease of the delay while the throughput increases, which is rather strange and is
indicative of a curious phenomenon. Also note that the graphs for the various values of ν
coincide for larger values of σ, demonstrating our previous observation that systemperfor
mance under these conditions is almost independent of ν.
To understand the nature of this phenomenon, recall that the retransmission rate is very
small and that backlogged users remain in that state for a fairly long time. Since the rate of
S M
σ M 1 – ( )ν M 3 – ( )νσ + +
M M 1 – ( )σ +

Mσ
M M 1 – ( )σ +
 ≈ =
D
ˆ
1 M 1 – ( )
1 σ +
σ
  + M
M 1 –
σ
 + = =
E
x
p
e
c
t
e
d
D
e
l
a
y
D
ˆ
(
)
Throughput (S)
10
20
30
40
50
60
70
80
90
100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
ν=0.01
ν=0.001
ν=0.1
ν=0.05
Finite Population
Slotted Aloha
M=10
FIGURE 3.6: ThroughputDelay of ﬁnite population Aloha under Capture
62 CHAPTER 3: ALOHA PROTOCOLS
new packet generation σ is nonnegligible, eventually all users will become backlogged.
With this situation in place consider a user that retransmits its packet and (usually) suc
ceeds. This user will generate another packet after a relatively short time and, having gen
erated a new packet, is very likely to succeed again since the probability that its
transmission will be interfered with that of a backlogged user is slight. This chain of
events continues for quite a while until the “unfortunate” event that a retransmission of a
backlogged user is scheduled at which time either a collision occurs immediately, or there
are two thinking users that are likely to collide. When this happens all users are back
logged again, remaining so for quite a while, since retransmission rate is small. Eventu
ally, however, another user retransmits and usually succeeds and the previous scenario
repeats itself. What actually happens is that a random user captures the channel for a
while, i.e., transmits successfully a succession of packets. Performance clearly improves
with increasing σ because the more packets generated in between two retransmissions the
better the throughput. When σ becomes sufﬁciently high, the throughput approaches 0.5
because the capturing user transmits in one slot and generates a new packet in the next and
so forth.
This capture effect appears in all ﬁnitepopulation Alohalike protocols such as all the
variants of the carriersensing protocols, with and without collision detection (see, for
example, [ShH82]).
3.4. (IN)STABILITY OF ALOHA PROTOCOLS
An underlying assumption in the analysis of the Aloha protocol is that the total arrival pro
cess of new and retransmitted (due to collisions) packets is a Poisson process. There is no
justiﬁcation to this assumption except that it simpliﬁes the analysis of the Aloha protocol.
We recall that under this assumption, the throughput of the slotted Aloha protocol is pre
dicted to be
(3.9)
where T is the (ﬁxed) transmission time of a packet, g is the expected number of new and
retransmitted packets per second and G=gT.
Another assumption that is implicitly used in the analysis of the Aloha protocol is a stabil
ity assumption, namely, that the number of backlogged users with packets awaiting to be
retransmitted is not steadily growing. In other words, it is assumed that packets are enter
ing and leaving the system at the same rate. We ﬁrst give an intuitive reasoning why this
assumption is false and then prove it rigorously.
Consider Figure 3.2 where relation (3.9) is depicted. Assume that the arrival rate of new
packets is λ packets per slot and assume that ( is the maximum throughput
predicted for the slotted Aloha protocol). If equilibrium between arrival and departure
rates prevails, then the rate of the total trafﬁc on the channel (new and retransmitted pack
S gTe
gT –
Ge
G –
= =
λ e
1 –
< e
1 –
Section 3.4.: (IN)STABILITY OF ALOHA PROTOCOLS 63
ets) G=gT will be G
1
=g
1
T, as is shown in Figure 3.7. This is of course an “average” rate,
and, over any ﬁxed interval of time, the actual rate will ﬂuctuate around this mean. If the
actual trafﬁc rate moves a little above G
1
, the actual throughput increases a little above λ.
Thus, packets leave the system faster than they arrive, which causes the actual trafﬁc rate
to decrease back to G
1
. If the actual trafﬁc rate moves a little below G
1
, the actual through
put decreases a little below λ. Thus, packets leave the system slower than they arrive,
which causes the actual trafﬁc rate to increase back to G
1
. Consequently, the point
(S,G)=(λ,G
1
) is a conditionally stable point, namely, it is stable under small variations in
G
1
. However, if a large variation (and this will happen with probability one) causes the
actual trafﬁc to exceed G
2
in Figure 3.7 then the actual throughput decreases below λ.
Thus, packets leave the system at a slower rate than they enter, which causes further
increase in the actual trafﬁc rate, a further decrease in actual throughput, etc. The system
never returns to the point (λ,G
1
), but rather drifts relentlessly toward the catastrophic,
unconditionally stable, point . We conclude that the maximum stable
throughput of the Aloha protocol is zero.
In the following, this result is established in a more precise and formal manner. The fol
lowing is based on the analysis by Fayolle et. al. [FLB74]. We ﬁrst describe the concrete
model that is used in the analysis.
S G , ( ) 0 ∞ , ( ) =
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.001 0.01 0.1 1 10 100
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
Slotted Aloha
FIGURE 3.7: Demonstrating the Instability of Slotted Aloha
λ
G
2
G
1
64 CHAPTER 3: ALOHA PROTOCOLS
3.4.1. Analysis
Consider the slotted Aloha system with an inﬁnite population of users, so that each new
packet that arrives to the system is associated with a new user. Let be the number of
newpackets that are generated (arrive) during the kth slot. These packets are transmitted in
slot k+1. It is assumed that is a sequence of independent and
identically distributed (i.i.d.) random variables with a common distribution
(3.10)
and mean λ packets/slot . Notice that if the arrival process of new packets
is Poisson, .
Let be the number of backlogged users at the beginning of slot k (k=0,1,2,...). Back
logged users have packets that collided, and that have to be retransmitted. We assume that
. The retransmission delay of a packet is assumed to be geometric, namely, a
user that is backlogged at the beginning of slot k, retransmits its packet during slot k with
probability ν, independently of any other event in the system. From these deﬁnitions we
have:
(3.11)
As in the case of a ﬁnite number of users (see Section 3.3.) the number of backlogged
users is referred to as the state of the system. The number of backlogged users at the
beginning of the k+1st slot depends on the number of backlogged users at the beginning of
the kth slot (one of them might transmit a packet successfully) and the number of new
packets that arrived within the slot. Since the process of new arrivals is independent of the
activities in any previous slot, the process is a Markov chain.
Unlike the case of a ﬁnite number of users, the chain is not
ﬁnite, so it is not obvious whether or not the chain is ergodic.
Let denote the probability that . Let us list all possible transitions that will
lead into the state in which there are n backlogged users at the beginning of slot k+1, i.e.,
in parenthesis we indicate the probabilities of the corresponding events):
1. There were n backlogged users at the beginning of slot k ( ), none of them trans
mitted ( ) and a single new packet was transmitted (a
1
). The single new packet is
successfully transmitted and therefore the number of backlogged users is unchanged.
2. There were n backlogged users at the beginning of slot k ( ), at least two of them
transmitted ( ) and hence collided, and no new packet was transmitted
(a
0
). The number of backlogged users is unchanged.
A
˜
k ( )
A
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
Prob A
˜
k ( ) i = [ ] Prob i new packets arrive at slot k [ ] a
i
= = i 0 ≥
λ ia
i
i 1 =
∞
∑
=
a
i
λ
i
e
i –
( ) i! ⁄ =
N
˜
k ( )
N
˜
0 ( ) 0 =
b
i
n ( ) Prob i backlogged users transmit in slot k  n in backlog [ ]
·
∆
Prob i backlogged users transmit in slot k  N
˜
k ( ) n = [ ]
n
i
¸ ,
¸ _
ν
i
1 ν – ( )
n i –
= =
N
˜
k ( )
N
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
N
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
π
n
k ( ) N
˜
k ( ) n =
N
˜
k 1 + ( ) n =
π
n
k ( )
b
0
n ( )
π
n
k ( )
1 b
0
n ( ) – b
1
n ( ) –
Section 3.4.: (IN)STABILITY OF ALOHA PROTOCOLS 65
3. There were n backlogged users at the beginning of slot k ( ), none of them trans
mitted ( ) and no new packet was transmitted (a
0
). This corresponds to an idle slot
and the number of backlogged users is unchanged.
4. There were n+1 backlogged users at the beginning of slot k ( ), exactly one of
them transmitted ( ) and no new packet was transmitted (a
0
). In this case, the
transmission of the backlogged user was successful, and therefore the number of back
logged users decreases by one.
5. There were n1 backlogged users at the beginning of slot k( ), at least one of
them transmitted ( ) and a single new packet was transmitted (a
1
). In this
case a collision occurs and the user with the new packet joins the backlogged users.
6. There were nj backlogged users at the beginning of slot k ( ) and at
least two new packets were transmitted a
j
. In this case a collision occurs and
all users with the new packets join the backlogged users.
Summarizing the above, the following balance equation can be written:
(3.12)
Notice that (3.12) is valid for all if one adopts the convention that for
i<0.
The Markov chain is aperiodic and irreducible. It is ergodic if an
invariant probability distribution exists satisfying (3.12) such that
π
n
> 0 for all n, and . Assuming the latter limit exists,
we obtain from (3.12):
(3.13)
Deﬁne
, (3.14)
and sum (3.13) for n=0,1,..., N
π
n
k ( )
b
0
n ( )
π
n 1 +
k ( )
b
1
n 1 + ( )
π
n 1 –
k ( )
1 b
0
n 1 – ( ) –
2 j n ≤ ≤ π
n 1 –
k ( )
2 j n ≤ ≤
π
n
k 1 + ( ) π
n
k ( )b
0
n ( )a
1
π
n
k ( ) 1 b
0
n ( ) – b
1
n ( ) – [ ]a
0
π
n
k ( )b
0
n ( )a
0
+ + =
π
n 1 +
k ( )b
1
n 1 + ( )a
0
π
n 1 –
k ( ) 1 b
0
n 1 – ( ) – [ ]a
1
π
n j –
k ( )a
j
j 2 =
n
∑
+ + +
n 0 ≥ π
i
k ( ) 0 =
N
˜
k ( ) k 0 1 2 … , , , = , { }
π
n
k 0 1 2 … , , , = , { }
π
n
n 0 =
∞
∑
1 = π
n
lim
k ∞ →
π
n
k ( ) =
π
n
π
n
b
0
n ( )a
1
b
1
n ( )a
0
– [ ] π
n 1 +
b
1
n 1 + ( )a
0
π
n 1 –
b
0
n 1 – ( )a
1
– π
n j –
a
j
j 0 =
n
∑
+ + =
P
N
π
n
n 0 =
N
∑
=
66 CHAPTER 3: ALOHA PROTOCOLS
(3.15)
Using the fact that b
1
(0)=0 we have:
(3.16)
or
(3.17)
where the ﬁrst inequality above is due to the fact that P
N
does not decrease as N increases,
and the second inequality is due to the fact that .
From (3.17) we have:
(3.18)
or (we use (3.11))
(3.19)
for any . The inequality in (3.19) implies that the ratio increases without
limit as . Therefore, the sum exists only if for all ﬁnite
values of N; otherwise, is divergent, which cannot be the case when the π
n
,
deﬁne a probability distribution. Thus, the Markov chain repre
senting the number of backlogged users is not ergodic, and the Aloha protocol is not sta
ble. Furthermore, we will see that the throughput of the system is zero.
Let be the conditional probability that one packet is successfully transmitted during
the kth slot, given that . The throughput of the system is then
P
N
π
n
n 0 =
N
∑
π
n
b
0
n ( )a
1
b
1
n ( )a
0
– [ ]
n 0 =
N
∑
π
n 1 +
b
1
n 1 + ( )a
0
n 0 =
N
∑
+ = =
π
n 1 –
b
0
n 1 – ( )a
1
n 0 =
N
∑
– π
n j –
a
j
j 0 =
n
∑
n 0 =
N
∑
+
π
n
b
0
n ( )a
1
b
1
n ( )a
0
– [ ]
n 0 =
N
∑
π
n 1 +
b
1
n 1 + ( )a
0
n 0 =
N
∑
+ =
π
n 1 –
b
0
n 1 – ( )a
1
n 0 =
N
∑
– a
j
π
n j –
n j =
N
∑
j 0 =
N
∑
+
π
N
b
0
N ( )a
1
π
0
b
1
0 ( )a
0
– π
N 1 +
b
1
N 1 + ( )a
0
a
j
P
N j –
j 0 =
N
∑
+ + =
P
N
π
N
b
0
N ( )a
1
π
N 1 +
b
1
N 1 + ( )a
0
a
j
P
N j –
j 0 =
N
∑
+ + =
P
N
1 a
0
– ( ) π
N
b
0
N ( )a
1
π
N 1 +
b
1
N 1 + ( )a
0
a
j
P
N j –
j 1 =
N
∑
+ + =
π
N
b
0
N ( )a
1
π
N 1 +
b
1
N 1 + ( )a
0
P
N 1 –
a
j
j 1 =
N
∑
+ + ≤
π
N
b
0
N ( )a
1
π
N 1 +
b
1
N 1 + ( )a
0
P
N 1 –
1 a
0
– ( ) + + ≤
a
j
j 1 =
N
∑
a
j
j 1 =
∞
∑
≤ 1 a
0
– =
π
N
1 a
0
– ( ) P
N
P
N 1 –
– ( ) 1 a
0
– ( ) π
N
b
0
N ( )a
1
π
N 1 +
b
1
N 1 + ( )a
0
+ ≤ =
π
N 1 +
π
N

1 a
o
– b
0
N ( )a
1
–
b
1
N 1 + ( )a
0
 ≥
1 a
0
– 1 ν – ( )
N
a
1
–
N 1 + ( )ν 1 ν – ( )
N
a
0
 =
N 0 ≥ π
N 1 +
π
N
⁄
N ∞ → P
∞
lim
N ∞ →
P
N
·
∆
π
n
0 =
P
∞
n 0 ≥
N
˜
k ( ) k 0 1 2 … , , , = , { }
S
n
k ( )
N
˜
k ( ) n =
Section 3.4.: (IN)STABILITY OF ALOHA PROTOCOLS 67
(3.20)
A packet will be successfully transmitted only if exactly one backlogged user is transmit
ting and no new packet is transmitted or no backlogged user is transmitting and exactly
one new packet is transmitted. Hence,
(3.21)
( does not depend on k). Therefore,
(3.22)
where we used our conclusion from (3.19) that π
n
=0 for any ﬁnite n.
The facts that the Markov chain is not ergodic and that the
throughput of the systemis zero, indicate that the number of backlogged users will eventu
ally grow to inﬁnity, no packets will be successfully transmitted, and the expected delay of
a packet will be inﬁnite.
3.4.2. Stabilizing the Aloha System
From the intuitive arguments given at the beginning of this chapter and from the analysis
presented in the previous section, it is clear that the Aloha system (with inﬁnitely many
users) cannot be stable for a policy of retransmission of collided packets that does not take
into account (somehow) the system state. The schemes presented thus far use ﬁxed
retransmission probabilities, meaning that the retransmission policy is independent of sys
tem state, rendering these schemes unstable. In order to stabilize the system, the retrans
mission probabilities must somehow adapt in accordance with the state of the system.
Assuming that some coordination among the backlogged users is possible prior to each
slot, it is not difﬁcult to develop retransmission policies that stabilize the Aloha system.
For instance, consider the following threshold policy: At most θ (the threshold) of the
backlogged users at the beginning of slot k, retransmit their packets during slot k, each of
themwith probability ν independently of the other users. In other words, when the number
of backlogged users does not exceed θ, each of them retransmits its packet with probabil
ity ν. If the number exceeds θ, a subset of size θ is chosen from the backlogged users and
each user in this subset retransmits its packet with probability ν. All other backlogged
users remain silent during that slot.
The implementation of the threshold policy is not speciﬁed but is clearly not simple; It is
not clear how the number of backlogged users would be known, and even if it is known, it
S S
n
k ( )π
n
k ( )
n 0 =
∞
∑
k ∞ →
lim =
S
n
k ( ) b
1
n ( )a
0
b
0
n ( )a
1
+ =
S
n
k ( )
S b
1
n ( )a
0
b
0
n ( )a
1
+ [ ]π
n
k ( )
n 0 =
∞
∑
k ∞ →
lim b
1
n ( )a
0
b
0
n ( )a
1
+ [ ]π
n
n 0 =
∞
∑
= =
b
1
n ( )a
0
b
0
n ( )a
1
+ [ ]π
n
n 0 =
N
∑
N ∞ →
lim 0 = =
N
˜
k ( ) k 0 1 2 … , , , = , { }
68 CHAPTER 3: ALOHA PROTOCOLS
is not clear how the backlogged users coordinate to choose the subset of size θ. Neverthe
less, it is instructive to see why this policy stabilizes the Aloha system for a certain range
of arrival rate.
To prove the stabilizing properties of the threshold policy, we use a lemma due to Pakes
[Pak69] that is often useful in proving ergodicity of homogeneous Markov chains.
Pakes’ Lemma:
Let be an irreducible, aperiodic homogeneous Markov chain whose
state space is the set of nonnegative integers. The following two conditions are sufﬁcient
for the Markov chain to be ergodic:
1. ,
2. .
It is clear that the Markov chain is irreducible, aperiodic and
homogeneous for the threshold policy. Assume that and i < θ. In this case we
have only to showthat condition (a) of Pakes’ Lemma ( plays the role of ) holds, since
cannot increase to inﬁnity in this case. Given that we have
(3.23)
The explanation of (3.23) is similar to that of (3.12). From (3.23) we obtain
(3.24)
Hence, condition (a) holds if .
Consider now the case that and . As in (3.23) we have:
Z
˜
k
k 0 1 2 … , , , = , { }
E Z
˜
k 1 +
Z
˜
k
– Z
˜
k
i = ( ) [ ] ∞ < i ∀
supE Z
˜
k 1 +
Z
˜
k
– Z
˜
k
i = ( ) [ ]
i ∞ →
lim 0 <
N
˜
k ( ) k 0 1 2 … , , , = , { }
N
˜
k ( ) i =
N
˜
Z
˜
N
˜
k ( ) N
˜
k ( ) i =
N
˜
k 1 + ( )
i 1 – with probability b
1
i ( )a
0
i with probability 1 b
1
i ( ) – [ ]a
0
b
0
i ( )a
1
+
i 1 + with probability 1 b
0
i ( ) – [ ]a
1
i j + j 2 ≥ , with probability a
j
=
E N
˜
k 1 + ( ) N
˜
k ( ) – N
˜
k ( ) i = ( ) [ ] E N
˜
k 1 + ( ) N
˜
k ( ) i = ( ) [ ] E N
˜
k ( ) N
˜
k ( ) i = ( ) [ ] – =
i 1 – ( )b
1
i ( )a
0
i 1 b
1
i ( ) – [ ]a
0
b
0
i ( )a
1
+ { } + =
i 1 + ( ) 1 b
0
i ( ) – [ ]a
1
i j + ( )a
j
j 2 =
∞
∑
i – + +
i j + ( )a
j
j 2 =
∞
∑
i – b
1
i ( )a
0
– b
0
i ( )a
1
– = λ b
1
i ( )a
0
– b
0
i ( )a
1
– =
λ ∞ <
N
˜
k ( ) i = i θ ≥
Section 3.4.: (IN)STABILITY OF ALOHA PROTOCOLS 69
(3.25)
The above stems fromthe fact that at most θ users are transmitting when the threshold pol
icy is employed. From (3.25) we have
(3.26)
Hence, both conditions (a) and (b) hold if
(3.27)
Therefore, if λ satisﬁes (3.27), the system is stable.
In summary, we saw that by limiting the number of backlogged users that contend for the
channel, it is possible to stabilize the Aloha system. Moreover, we derived a condition on
the arrival rate, (3.27), that guarantees stability.
Let us now consider another stable policy for the Aloha system that requires only the
knowledge of the number of backlogged users at the beginning of each slot. This policy is
based on adaptively controlling the retransmission probabilities. Speciﬁcally, assuming
that the number of backlogged users is known, we allow the retransmission probability ν
to be a function of that number. Denoting this function by ν(n) where n is the number of
backlogged users we have (see (3.11)):
(3.28)
For this retransmission policy, equation (3.23) holds for all i and therefore (see (3.24)):
Consequently, conditions (a) and (b) of Pakes’ Lemma will hold if
. (3.29)
Let us now quantify this value of λ. Deﬁne
N
˜
k 1 + ( )
i 1 – with probability b
1
θ ( )a
0
i with probability 1 b
1
θ ( ) – [ ]a
0
b
0
θ ( )a
1
+
i 1 + with probability 1 b
0
θ ( ) – [ ]a
1
i j + j 2 ≥ , with probability a
j
=
E N
˜
k 1 + ( ) N
˜
k ( ) – N
˜
k ( ) i = ( ) [ ]
i 1 – ( )b
1
θ ( )a
0
i 1 b
1
θ ( ) – [ ]a
0
b
0
θ ( )a
1
+ { } + =
i 1 + ( ) 1 b
0
θ ( ) – [ ]a
1
i j + ( )a
j
j 2 =
∞
∑
i – + +
λ b
1
θ ( )a
0
– b
0
θ ( )a
1
– =
λ b
1
θ ( )a
0
b
0
θ ( )a
1
+ <
b
i
n ( )
n
i
¸ ,
¸ _
ν n ( ) [ ]
i
1 ν n ( ) – [ ]
n i –
=
E N
˜
k 1 + ( ) N
˜
k ( ) – N
˜
k ( ) i = ( ) [ ] λ b
1
i ( )a
0
– b
0
i ( )a
1
– = i ∀
λ limsup
n ∞ →
b
1
n ( )a
0
b
0
n ( )a
1
+ [ ] <
70 CHAPTER 3: ALOHA PROTOCOLS
. (3.30)
By differentiating with respect to ν and setting the result equal to zero, one observes
that is maximized for
The maximum value of is
Taking the limit as (see equation (3.29)) we see that the system will be stable for
It is interesting to note that if the arrival process is Poisson, a
1
/a
0
= log a
0
= λ and there
fore the system would be stable if , exactly as was predicted for constant retrans
mission probabilities. However, one should not forget that the Aloha systemwith constant
retransmission probabilities is unstable for any arrival rate. The stabilizing policy pre
sented above, namely retransmitting with probabilities that are (approximately) inversely
proportional to the number of backlogged users, requires the knowledge of this number
and it is not clear how the users would know this number.
There exist retransmission policies that do not require the knowledge of the exact number
of backlogged users. These policies are based on updating the retransmission probabilities
recursively in each slot, according to what happened during that slot. The general structure
of these policies is:
, (3.31)
namely, the retransmission probability (of a backlogged user) in slot k +1 is some function
f of the retransmission probability in the previous slot and of the event that occurred in slot
k. In essence, all such policies (namely, all the functions f that are used) increase the
retransmission probability when an idle slot occurs and decrease it when a collision
occurs. Examples of such policies and their analysis can be found in the work by Hajek
and Van Loon [HaL82]. The policies of the form (3.31) were proved to yield maximal sta
ble throughput of at most .
To summarize, the virtue of the Aloha protocol is its simplicity. However, the simple pro
tocol yields an unstable system. The protocols that stabilize the Aloha system are no
longer as simple as the original protocol, and yet, they only guarantee throughput of at
most . The reason for this low throughput is that in all the stabilizing policies dis
S
n
ν ( ) b
1
n ( )a
0
b
0
n ( )a
1
+
·
∆
1 ν n ( ) – [ ]
n
a
1
nν n ( ) 1 ν n ( ) – [ ]
n i –
a
0
+ =
S
n
ν ( )
S
n
ν ( )
ν
*
n ( )
a
0
a
1
–
na
0
a
1
–
 =
S
n
ν ( )
S
*
ν
*
( ) a
0
n 1 –
n a
1
a
0
⁄ –

n 1 –
=
n ∞ →
λ e
a
0
log
a
1
a
0
 1 – +
<
λ e
1 –
<
ν
k 1 +
f ν
k
feedback of slot k , ( ) =
e
1 –
e
1 –
Section 3.5.: RELATED ANALYSIS 71
cussed in this chapter (excluding the threshold policy that is not practical), all backlogged
users are using the same retransmission probability. In the next chapter we shall see that if
the decisions of users whether to transmit or not are based both upon their own history of
retransmissions and the feedback history, higher stable throughput can be obtained.
3.5. RELATED ANALYSIS
Numerous variations of the environment under which the Aloha protocol operates have
been addressed in the literature. We considered a very small part of these variations; slot
ted and nonslotted system; ﬁnite and inﬁnite population; ﬁxed length packets and Poisson
arrivals. Several books such as [Kl75, Tann81, Hay84, Tas86, HaO86, BeG87] cover part
of these and other variations. In the following we list a few of the variations that we did
not describe. In addition, we refer to some papers in which performance measures, other
than the throughput, have been computed for the Aloha protocol.
Variablelength packets
Abramson [Abr77] studied the performance of the inﬁnite population pure Aloha system
with two different possible packet lengths. Ferguson [Fer77b] and Bellini and Borgonovo
[BeB80] considered a system with an arbitrary packet length distribution. It is interesting
to note that it was shown in [BeB80] that constant length packets yield the maximum
throughput over all packet length distributions (see Exercises).
Arbitrary interarrival distribution
Sant [San80] studied the performance of the pure Aloha system where packet interarrival
times are statistically independent and identically distributed (i.i.d.), but not necessarily
exponential.
Capture
The assumption that whenever two or more packets overlap at the receiver, all packets are
lost, is overly pessimistic. In radio systems the receiver might correctly receive a packet
despite the fact that it is timeoverlapping with other transmitted packets. This phenome
non is known as capture and it can happen as a result of various characteristics of radio
systems. Most studies [Abr77, Met76, Sha84, Lee87] considered power capture, namely
the phenomenon whereby the strongest of several transmitted signals is correctly received
at the receiver. Thus, if a single high powered packet is transmitted then it is correctly
received regardless of other transmissions and hence the utilization of the channel
increases. Some other works (for instance, Davis and Gronemeyer [DaG80]) studied the
effect of delay capture (the receiver captures a packet since it arrived a short time before
any other packet that is transmitted during the same slot) in slotted Aloha protocol (see
Exercises).
72 CHAPTER 3: ALOHA PROTOCOLS
Buffered users
In some practical systems the users can provide buffer space for queueing of exogeneous
packets that arrive at the user. The case of a single buffer per user has been considered in
Section 3.1.3. When a ﬁnite number of buffers are available at each user, the analysis pro
ceeds along the same lines, although the number of states increases dramatically. When
the buffering capability at each user is not limited, one faces a very complicated queueing
problem due to the strong interaction among the various queues of the system. Speciﬁ
cally, the success probability of a transmission of a certain user depends on the activity of
the other users that have packets to transmit. Exact analysis of a twouser system has been
carried by Sidi and Segall [SiS83]. Approximate analysis of an Muser system has been
carried out by Saadawi and Ephremides [SaE81] and by Sidi and Segall [SiS83] for a sym
metric system, and by Ephremides and Zhu [EpZ87] for a nonsymmetric system. Bounds
for the expected queue lengths have been derived by Szpankowski [Szp86]. Sufﬁcient con
ditions for ergodicity of the system have been provided by Tsybakov and Mikhailov
[TsM79] and Tsybakov and Bakirov [TsB88].
Reservation and adaptive protocols
Reservation schemes are designed to have the advantages of both the Aloha and the
TDMA approaches. The operation of such schemes is discussed in Section 2.5, where
only conﬂictfree reservation schemes are discussed. An immediate extension is, of
course, to use a reservation scheme with contention, i.e., that users contend during a reser
vation period an those who succeed in making reservations transmit without interference.
These schemes derive their efﬁciency from the fact that reservation periods are shorter
than transmission periods by several orders of magnitude.
In the category of reservation schemes fall the works by Binder [Bin75] that requires
knowledge of the number of users (or an upper bound thereof), and the works by Crowther
et. al. [CRW73] and Roberts [Rob75] that do not require this knowledge. Approximate
analysis of a reservation Aloha protocol can be found in [Lam80]. Additional variations of
reservation protocols and their analysis can be found in [ToK76, TaI84, TsC86].
Another kind of protocols that are designed to operate as the Aloha protocol in light load
and according to a TDMA scheme in heavy load are known as adaptive protocols. The urn
scheme [KlY78] is an example of such protocol.
Delay and interdeparture times
Ferguson presented an approximate analysis of the delay in the Aloha protocol [Fer75,
Fer77b] and compared it to TDMA [Fer77a]. Exact packet delay and interdeparture time
distribution for a ﬁnite population slotted Aloha system with a single buffer per each user
has been derived in [Tob82b]. The interdeparture process of the Aloha protocols has been
studied in [TaK85b].
Section 3.5.: RELATED ANALYSIS 73
Stability
Instability issues of the Aloha protocol were ﬁrst identiﬁed in [CaH75] and [LaK75].
Later, similar issues were identiﬁed for the CSMA family of protocols in [ToK77]. Other
papers that discussed these issues are [Jen80, MeL83, OnN85].
Stable protocols for the Aloha system of the form (3.31) have been suggested in several
studies. For instance, Kelly [Kel85] proposed an additive rule for determining ν
k+1
(as
opposed to the multiplicative rule suggested in [HaL82]). Another additive rule known as
the pseudoBayesian rule has been suggested in [Riv87] and analyzed in [Tsi87].
The operation of the stable algorithms depends very strongly on the feedback information
that is obtained in each slot (see (3.31)). Therefore, if the feedback information is not reli
able, the tuning of the algorithms should be adjusted as is discussed in [MeK85].
74 CHAPTER 3: ALOHA PROTOCOLS
EXERCISES
Problem 1. (Busy period of pure Aloha)
Consider a pure Aloha system with Poisson offered load g packets/sec and packets of
equal size 1 (T=1). Denote by the length of an unsuccessful transmission period.
1. Let (i=1,2,...) be the ith interarrival time between packets arriving within a single
unsuccessful transmission period. Find , the distribution of , its pdf , and
its Laplace transform .
2. Let be the number of transmissions in an unsuccessful transmission period. Find
.
3. By conditioning on ﬁnd the Laplace transformof the probability density func
tion of . Compute the expectation and variance of .
4. Using the result of part (3) ﬁnd the average length of a transmission period (successful
or not) and the throughput of the system.
Problem 2. (Slotted Aloha with acknowledgments)
Consider a system that employs the slotted Aloha protocol. After each successful trans
mission the receiving station sends an acknowledgment packet indicating the transmission
was successful. Acknowledgment packets are transmitted on the same channel used for
data packets and their length is the same as that of a data packet. A collision between a
data packet and an acknowledgment packet destroys both of them. Consequently, the col
lided data packets and the data packet that has been acknowledged by the collided
acknowledgment packet have to be retransmitted. In the following use the standard Pois
son assumptions.
1. Find the relation between the throughput S and the offered load G in this system. What
is the maximal throughput of this system?
2. Draw S as a function of G for this system and for slotted Aloha without acknowledg
ments on the same ﬁgure.
Problem 3. (Slotted Aloha with time capture)
Consider a slotted Aloha systemwith a large number of terminals transmitting to a central
station. Typically, when two or more packets are transmitted concurrently all are lost.
However, if the ﬁrst packet arrives at the station θ seconds before any other packet in that
slot, the receiver of the station “locks on” the packet and can receive it correctly. This phe
nomenon is called time capture.
The terminals are evenly distributed around the station so that in terms of propagation time
the most remote terminal is τ seconds away from the station. The slot size is T sec and
F
˜
A
˜
i
F
A
t ( ) A
˜
i
f
A
t ( )
F
A
*
s ( )
L
˜
L z ( ) E z
L
˜
[ ] =
L
˜
F
F
*
s ( )
F
˜
F
˜
Section : EXERCISES 75
packets arrive for transmission on the channel according to a Poisson process with average
g packets per second.
In the following we would like to compute the probability of capture and the channel
throughput. Let t
1
be the time of the ﬁrst arrival in the station in the given slot and let t
2
be
the time of the second arrival in the station in the given slot. (In questions (1) and (2)
below we compute the quantities conditioned on the event that k packets are ready to
transmit at the beginning of the slot).
1. Deﬁne t = t
2
t
1
. Find the pdf conditioned on k).
2. What is the probability of capture in a slot (conditioned on k)?
3. What is the (unconditional) probability of capture in a slot?
4. What is the throughput S of the system?
5. Show that regardless of θ, maximum throughput is achieved for gT>1. Explain this
result.
Problem 4. (Slotted Aloha with power capture [Met76])
This problem deals with a different model for the capture phenomenon. As explained in
Problem 3.3, a capture means receiving correctly a packet even when other packets are
transmitted during the same time. The model used here is typical to power capture,
namely, that some nodes transmit with higher power than others.
Assume that the population of users in the system is divided into K classes all using the
slotted Aloha protocol. If a single node of class i (2≤ i ≤K) is transmitting simultaneously
with any number of users of classes 1,2,..., i1, then the transmitted packet of node i is cap
tured and thus successfully received. All other nodes that transmitted during the slot have
to retransmit their packets at some later time.
Let S
i
denote the rate of generation of new packets (per slot) by nodes of class i (1≤ i ≤K)
and let G
i
denote the total rate of transmitted packets (per slot) by nodes of class i. Use the
standard Poisson assumptions.
1. Let be the total throughput of the systems. Determine the maximal
throughput for K=2.
2. Determine the maximal throughput for any K. How should the G
i
’s (1≤ i ≤K) be chosen
in order to obtain the maximal throughput? What happens when K→ ∞?
3. For K=2, determine the possible values of the couple (S
1
, S
2
). Draw your result.
Problem 5. (Pure Aloha with variable length packets [BeB80])
Consider a network whose nodes transmit (arbitrarily distributed) variable length packets
according to the pure Aloha protocol.
f
t
w ( )
S S
i
j 1 =
K
∑
=
76 CHAPTER 3: ALOHA PROTOCOLS
Observe that long packets are more likely to collide than short packets. Therefore, the
length distribution of packets that are successfully transmitted is different from the length
distribution of an arbitrary packet transmitted in the channel.
Let be a random variable representing the length of packets (the time required to trans
mit a packet) generated by the nodes (this random variable represents also the length of
packets that are successfully transmitted). Let X
0
(x) be its distribution function, x
0
(x) its
density function, X
0
*
(s) its Laplace transform, and x
0
its mean. In a similar manner we
deﬁne , X
c
(x), x
c
(x), X
c
*
(x) and x
c
for the packets transmitted on the channel (new pack
ets and those that are retransmitted because of collisions).
Assume that new packets are generated according to Poisson process with mean λ
0
and
the total trafﬁc on the channel is also Poisson with mean λ
c
. Let S = λ
0
x
0
and G = λ
c
x
c
.
1. Prove that the probability for successful transmission of a packet whose length
is (the packet is from the total trafﬁc on the channel) is given by:
Note that depends only on x
c
and not on the distribution of .
2. Prove the following relationships:
3. Prove that the throughput is given by:
where ; .
4. Let λ
0
(x) = λ
0
x
0
(x) and . Prove that
5. Find the relation between S and G in the following three special cases:
 Constant packet length: with probability 1).
x
˜
0
x
˜
c
P
suc
x ( )
x
˜
c
x =
P
suc
x ( ) exp λ
c
x
c
x + ( ) – { } =
P
suc
x ( ) x
˜
c
X
c
*
s ( )
X
0
*
s λ
c
– ( )
X
0
*
λ
c
– ( )
 =
X
0
*
s ( )
X
c
*
s λ
c
+ ( )
X
c
*
λ
c
( )
 =
S
Ge
G –
x
c
 X
c
*(1)
G
x
c
 ( )
x
0
Ge
G –
X
0
*(1)
G –
x
c
 ( )
 = =
X
0
*(1)
s d
d
X
0
*
s ( ) =
s d
d
X
c
*
s ( )
Λ
0
*
e
sx –
λ
0
x ( ) x d
0
–
∞
∫
=
Λ
o
*
s ( ) λ
c
X
c
*
s λ
c
+ ( )exp λ –
c
x
c
{ } =
S
Λ
o
*
s ( ) d
s d

s 0 =
– =
x
˜
0
x
0
=
Section : EXERCISES 77
 Dual packet size: ; There are packets of two
types  those with constant length and those with constant length ( ).
The probability that a packet is of the ﬁrst type is α.
 Exponential packet length: .
6. Prove that the throughput of a pure Aloha system is maximized when the packet length
distribution is deterministic i.e., all packets have the same constant length.
Problem 6. (Aloha with K channels)
Consider a systemconsisting of K separate slotted channels (with slot boundaries synchro
nized). The system operates according to the slotted Aloha protocol with the speciﬁc
channel chosen according to some rule.
1. Find the throughput of such a system for the inﬁnite population case if the channels are
selected uniformly at random.
2. For the ﬁnite population case ﬁnd (and compare) the throughput delay characteristics if
 The users are divided in advance into K equally sized groups each assigned one
channel.
 Each user selects randomly and uniformly one of the K channels and performs all its
activity on that channel.
 Every user and for every packet transmission (or retransmission) selects randomly,
uniformly, and independent of the past the channel over which this speciﬁc packet is
to be transmitted.
 Every user upon generating a new packet selects randomly and uniformly one chan
nel to be used for transmission and all retransmissions of that packet.
x
0
αδ x x
1
– ( ) 1 α – ( )δ x x
2
– ( ) + =
x
1
x
2
x
1
x
2
<
x
0
µe
µx –
=
78 CHAPTER 3: ALOHA PROTOCOLS
CHAPTER 4
CARRIER SENSING PROTOCOLS
The Aloha schemes, described in the previous section, exhibited fairly poor performance
which can be attributed to the “impolite” behavior of the users namely, whenever one has a
packet to transmit he does so without consideration of others. It does not take much to
realize that even little consideration can beneﬁt all. Consider a behavior that we generi
cally characterize as “listen before talk”, that is, every user before attempting any trans
mission listens whether somebody else is already using the channel. If this is the case the
user will refrain from transmission to the beneﬁt of all; his packet will clearly not be suc
cessful if transmitted and, further, disturbing another user will cause the currently trans
mitted packet to be retransmitted, possibly disturbing yet another packet.
The process of listening to the channel is not that demanding. Every user is equipped with
a receiver anyway. Moreover, to detect another user’s transmission does not require receiv
ing the information; it sufﬁces to sense the carrier that is present when signals are trans
mitted. The carrier sensing family of protocols is characterized by sensing the carrier and
deciding according to it whether another transmission is ongoing.
Carrier sensing does not, however, relieve us from collisions. Suppose the channel has
been idle for a while and two users concurrently generate a packet. Each will sense the
channel, discover it is idle, and transmit the packet to result in collision. “Concurrently”
here does not really mean at the very same time; if one user starts transmitting it takes
some time for the signal to propagate and arrive at the other user. Hence “concurrently”
actually means within a time window of duration equal to signal propagation time. This
latter quantity becomes therefore a crucial parameter in the performance of these proto
cols.
All the carrier sense multiple access (CSMA) protocols share the same philosophy: when
a user generates a new packet the channel is sensed and if found idle the packet is trans
mitted without further ado. When a collision takes place every transmitting user resched
ules a retransmission of the collided packet to some other time in the future (chosen with
some randomization) at which time the same operation is repeated. The variations on the
CSMA scheme are due to the behavior of users that wish to transmit and ﬁnd (by sensing)
the channel busy. Most of these variations were introduced and ﬁrst analyzed in a series of
papers by Tobagi and Kleinrock [KlT75, ToK75, ToK77].
For more detail the reader is referred to any of the many books and surveys that deal with
various aspects of CSMA protocols such as Tannenbaum’s [Tan81], Stallings’ [Sta85], or
Clark Pogran and Reed’s [CPR78].
80 CHAPTER 4: CARRIER SENSING PROTOCOLS
4.1. NONPERSISTENT CARRIER SENSE MULTIPLE ACCESS
In the nonpersistent versions of CSMA (NPCSMA) a user that generated a packet and
found the channel to be busy refrains from transmitting the packet and behaves exactly as
if its packet collided, i.e, it schedules (randomly) the retransmission of the packet to some
time in the future. The following analysis is based on Kleinrock and Tobagi [KlT75].
Throughput Analysis
To evaluate the performance of NPCSMA let us adopt a model similar to that used in the
evaluation of the performance of the Aloha protocol. We assume an inﬁnite population of
users aggregately generating packets according to a Poisson process with parameter λ. All
packets are of the same length and require T seconds of transmission. When observing the
channel, packets (new and retransmitted) arrive according to a Poisson process with
parameter g packets/sec.
In addition to the assumptions of the model used to analyze the Aloha protocol, the model
used for CSMA deals also with system conﬁguration which is manifested by a propaga
tion delay among users. Denote by τ the maximum propagation delay in the system (mea
sured in seconds) and deﬁne to be the normalized propagation time. We assume
that all users are “τ seconds apart” that is, τ is the propagation delay between every pair of
users. With this assumption the following analysis provides a lower bound to the actual
performance.
Consider an idle channel and a user scheduling a transmission at some time t (see Figure
4.1). This user senses the channel, starts transmitting at time t and does so for T seconds;
once he is done it will take τ additional seconds before the packet arrives at the destina
tion. This transmission therefore causes the channel to be busy for a period of T+ τ sec
onds. If, at time t’> t+τ another user scheduled a packet for transmission, that user would
sense the channel busy and refrain from transmission. If, however, some other user sched
uled a packet for transmission during the period [t,t+τ], that user would sense the channel
idle, transmit its packet, and cause a collision. The initial period of the ﬁrst τ seconds of
transmission is called the vulnerable period since only within this period is a packet vul
nerable to interference. Figure 4.1(b) depicts a situation in which a packet transmission
starting at time t is interfered by two other transmissions that start in the interval [t,t+τ]. In
the case of a collision the channel will therefore be busy for some (random) duration
between T+τ and T+2τ. This period in which transmission takes place is referred to as the
transmission period (TP). In the case of NPCSMAthe transmission period coincides with
the busy period. Having completed a transmission period the channel will be idle for some
time until the next packet transmission is scheduled.
We therefore observe along the time axis a succession of cycles each consisting of a trans
mission period followed by an idle period (see Figure 4.1(a)). Because packet scheduling
is memoryless, the times in which these cycles start are renewal points. As we did before,
a τ T ⁄
·
∆
Section 4.1.: NONPERSISTENT CARRIER SENSE MULTIPLE ACCESS 81
we denote by the duration of the busy (transmission) period, and by B its mean. Let
be the time duration within the transmission period in which a successful packet is being
transmitted (mean U), and let the duration of the idle period (with mean I). The cycle
length is clearly and the throughput is given by S = U / (B+I). We now derive these
quantities.
Consider ﬁrst the idle period. Its duration is the same as the duration between the end of
packet transmission and the arrival of the next packet. Because packet scheduling is mem
oryless we get
B
˜
B
˜
I
˜
I
˜
T τ
Busy
Period
Idle
Period
Cycle Cycle Cycle Cycle
time
t τ +
Y
˜
Vulnerable
Period
t
t+T
t T τ Y
˜
+ + +
t+T+τ
B
˜
(a) Cycle Structure
(b) Unsuccessful Transmission Period
FIGURE 4.1: Non Persistent CSMA Packet Timing
B
˜
U
˜
I
˜
B
˜
I
˜
+
F
I
x ( ) Prob I
˜
x ≤ [ ] 1 Prob I
˜
x > [ ] – = =
1 P No packet scheduling during x [ ] – 1 e
gx –
– = =
82 CHAPTER 4: CARRIER SENSING PROTOCOLS
which means that is exponentially distributed with mean
The expected useful time U can also be easily computed. When a packet is successful the
channel carries useful information for a duration of T secondsthe time it takes to transmit
the packet; in the unsuccessful case no useful information is carried at all or, in other
words
If denotes the probability that a transmitted packet is successful then
.
The probability of a successful transmission, is the probability that no packet is
scheduled during the vulnerable period [t, t+ τ]. Hence,
and thus
.
To compute B, the average transmission period duration, let be a random variable such
that denotes the time at which the last interfering packet was scheduled within a
transmission period that started at time t (see Figure 4.1(b)). Clearly, and for a suc
cessful transmission period . Using this notation the duration of the transmission
period is
The period is characterized by the fact that no other packet is scheduled for transmis
sion during the period for otherwise the packet that is transmitted at
would not have been the last packet to be transmitted in . Thus, the probabil
ity distribution function of is
The above relation holds for . For negative values of y the probability distribution
function vanishes and for values greater than τ it equals unity. It is important to notice that
has a discontinuity at y=0 which means that care must be taken when the probabil
ity density function is derived. Denoting by δ(t) the Dirac impulse function we get
I
I
1
g
 =
U
T Successful Period
0 Unsuccessful Period
¹
'
¹
=
P
suc
U E U
˜
[ ] T P
suc
⋅ 0 1 P
suc
– ( ) ⋅ + TP
suc
= = =
P
suc
P
suc
Prob No arrival in the period t t τ + , [ ] [ ] e
gτ –
= =
U Te
gτ –
=
Y
˜
t Y
˜
+
Y
˜
τ <
Y
˜
0 =
B
˜
T τ Y
˜
+ + =
Y
˜
t Y
˜
+ t τ + , [ ] t Y
˜
+
t Y
˜
+ t τ + , [ ]
Y
˜
F
Y
y ( ) Prob Y
˜
y ≤ [ ] Prob No packet arrival during τ y – [ ] e
g τ y – ( ) –
= = =
0 y τ ≤ ≤
F
Y
y ( )
Section 4.2.: 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS 83
(4.1)
from which
(4.2)
and ﬁnally
Putting all these results together we get
which is the desired relation we were seeking.
As before we normalize the quantities with respect to the packet transmission time. To that
end, let G denote the average scheduling rate of packet measured in packets per packet
transmission time; in other words G=gT. With our previous deﬁnition of the normalized
propagation time a we get
.
A sketch of the throughput versus normalized offered load G for various values of the nor
malized propagation time a is shown in Figure 4.2. These graphs have the same shape as
those for the Aloha system except for the evidently improved throughput. As is expected,
the lower a the better the performance. In fact, the extreme case of yields a
throughput of G/(1+G) which does not decrease to zero with increasing load. We remark
also that having the same characteristic shape as the Aloha protocol means that NPCSMA
(as the other protocols in this family) suffer from the same instability problems from
which Aloha suffers.
4.2. 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS
With nonpersistent CSMA there are situations in which the channel is idle although one or
more users have packets to transmit. The 1persistent CSMA (1PCSMA) is an alternative
to nonpersistent CSMA that avoids such situations. This is achieved by applying the fol
lowing rule: A user that senses the channel and ﬁnds it busy, persists to wait and transmits
as soon as the channel becomes idle. Consequently, the channel is always used if there is a
user with a packet.
f
Y
y ( ) e
gτ –
δ y ( ) ge
g τ y – ( ) –
+ =
E Y
˜
[ ] τ
1 e
gτ –
–
g
 – =
B E T τ Y
˜
+ + [ ] T 2τ
1 e
gτ –
–
g
 – + = =
S
U
B I +

Te
gτ –
T 2τ
1 e
gτ –
–
g
 –
1
g
 + +

gTe
gτ –
g T 2τ + ( ) e
gτ –
+
 = = =
S
Ge
aG –
G 1 2a + ( ) e
aG –
+
 =
a 0 →
84 CHAPTER 4: CARRIER SENSING PROTOCOLS
The performance of the 1persistent CSMA scheme was ﬁrst analyzed by Kleinrock and
Tobagi [KlT75]. The analysis presented in the following is considerably simpler and is
based on Shoraby et. al. [SMV87].
Throughput Analysis
We adopt the same model as the one used in the analysis of nonpersistent CSMA. When
observing the channel over time, one sees a sequence of cycles, each consisting of an idle
period (no packet is scheduled for transmission during this period), followed by a busy
period that consists of several successive transmission periods (see Figure 4.3). All users
that sense the channel busy in some transmission period, transmit their scheduled packets
at the beginning of the successive transmission period. If no packet is scheduled for trans
mission during some transmission period, then an idle period begins as soon as this trans
mission period ends.
Notice that a transmission period starts either with the transmission of a single packet (call
it type 1 transmission period), or with the transmission of at least two packets (call it type
2 transmission period). Atransmission period that follows an idle period is always a type 1
transmission period. The type of a transmission period that follows another transmission
period depends on the number of those persistent users waiting for the current transmis
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.001 0.01 0.1 1 10 100 1000
T
h
r
o
u
g
h
p
u
t
(
S
)
Nonpersistent CSMA
a=0.001
a=0.01
a=0.1
a=1.0
a=0
FIGURE 4.2: ThroughputLoad of Nonpersistent CSMA
Offered Load (G)
Section 4.2.: 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS 85
sion to end. For consistency, an idle period is also viewed as a transmission period that
starts with no transmitted packets (call it type 0 transmission period). This is depicted in
Figure 4.3.
Deﬁne the state of the system at the beginning of a transmission period to be the type of
that transmission period. These states (0,1 and 2) correspond to a threestate Markov chain
embedded at the beginning of the transmission periods. The knowledge of the systemstate
at the beginning of some transmission period (together with the scheduling points of pack
ets during this transmission period) is sufﬁcient to determine the system state at the begin
ning of the successive transmission period. The possible transitions among the three states
of the embedded Markov chain are depicted in Figure 4.4.
t τ +
Y
˜
t
t+T
t T τ Y
˜
+ + +
t+T+τ
T Y
˜
+
τ Y
˜
–
t Y
˜
+
I
˜
T τ
TP
Type 1
TP
Type 2
Cycle Cycle Cycle
time
TP
Type 1
TP
Type 1
B
˜
B
˜ I
˜
Busy
Period
Idle
Period
Busy
Period
(a) Cycle Structure
(b) Unsuccessful Transmission Period
FIGURE 4.3: 1Persistent CSMA Timing
86 CHAPTER 4: CARRIER SENSING PROTOCOLS
During a type 0 transmission period no packets are transmitted and during type 2 transmis
sion periods two or more packets are transmitted and collide. Consequently, only type 1
transmission periods may result in a successful transmission. Yet, for a type 1 transmission
period to be successful, it is necessary that no packets arrive during its ﬁrst τ seconds that
constitute its vulnerable period (the probability of the latter event is .
Let i=0,1,2 be the stationary probability of being in state i, namely that the systemis in
a transmission period of type i. Let i=0,1,2 be a random variable representing the
length of type i transmission period and let . Since the length of a packet is T
seconds, and from the same renewal arguments we used before the throughput is given by
(4.3)
To compute the throughput we still have to compute and T
i
, i=0,1,2, which we do next.
Transmission Period Lengths
The nature of the idle period in 1persistent CSMA is identical to that of nonpersistent
CSMA, i.e., exponentially distributed with mean 1/g, hence . Regarding the
random variables and the important observation is that they have the same distri
bution. The reason is that the length of a transmission period with either a single packet or
with two or more packets, is determined only by the time of arrival of the last packet (if
any) within the vulnerable period (the ﬁrst τ seconds of the transmission period), and does
not depend at all on the type of the transmission period.
0
2
1
P
20
P
22
P
12
P
21
P
10
P
11
FIGURE 4.4: State Transitions of 1PCSMA
P
01
e
gτ –
π
i
T
˜
i
E T
˜
i
[ ] T
i
=
S
Tπ
1
e
gτ –
π
i
T
i
i 0 =
2
∑
 =
π
i
T
0
1 g ⁄ =
T
˜
1
T
˜
2
Section 4.2.: 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS 87
The computation of and is identical to that of in the nonpersistent CSMA (Fig
ure 4.1(b)). Let a transmission period start at time t and let be a random variable repre
senting the time (after t) of the last packet that arrived during the vulnerable period [t,t+τ]
of a transmission period that started at time t ( if no packets arrive during [t,t+τ]).
Then
(4.4)
We already derived the probability distribution function and probability density function
of and found (see equations (4.1) and (4.2))
(4.5)
(4.6)
Combining (4.4) and (4.6) we obtain:
(4.7)
State Probabilities
From the state diagram in Figure 4.4 we have,
(4.8)
When a type 1 or a type 2 transmission period starts, the type of the next transmission
period is determined (only) by those packets scheduled for transmission after the trans
mission period begins. Speciﬁcally, if no packets arrive within the transmission period, the
next transmission period will be of type 0. If a single packet arrives within the transmis
sion period, at least τ seconds after it begins, the next transmission period will be of type
1. Finally, if at lease two packets arrive within the transmission period, at least τ seconds
after it begins, the next transmission period will be of type 2. Therefore,
. (4.9)
Using (4.8) and (4.9) we have,
(4.10)
T
1
T
2
B
Y
˜
Y
˜
0 =
T
˜
1
T
˜
2
T τ Y
˜
+ + = =
Y
˜
f
Y
y ( ) e
gτ –
δ y ( ) ge
g τ y – ( ) –
+ = 0 y ≤ τ ≤
E Y
˜
[ ] τ
1 e
gτ –
–
g
 – =
T
1
T
2
T τ E Y
˜
[ ] + + T 2τ
1 e
gτ –
–
g
 – + = = =
π
0
π
1
P
10
π
2
P
20
+ =
π
1
P
12
π
2
P
21
P
20
+ ( ) =
π
0
π
1
π
2
+ + 1 =
P
1 j
P
2 j
= j 0 1 2 , , =
π
0
P
10
1 P
10
+
 = π
1
P
10
P
11
+
1 P
10
+
 = π
2
1 P –
10
P
11
–
1 P
10
+
 =
88 CHAPTER 4: CARRIER SENSING PROTOCOLS
Assume that a type 1 transmission period starts at time t. Conditioning on , the next
transmission period will be of type 0 (namely, an idle period) only if no packet is sched
uled for transmission after time t+τ and before the end of the type 1 transmission period
(namely, time t+y+T+τ). The probability of this event is . Unconditioning, we
obtain
(4.11)
In a similar manner we obtain
(4.12)
Combining (4.3), (4.7), (4.10), (4.11), and (4.12) we obtain the throughput
or in a normalized form:
This relation is depicted in Figure 4.5. While, generally, these graphs have the same form
as those of the nonpersistent CSMA, performance is less than expected.
Recall that the 1persistent CSMA was devised in an attempt to improve the performance
of the nonpersistent CSMA by reducing the extent of the idle periods. This attempt is, evi
dently, not quite successful since for high load the nonpersistent CSMA outperforms the
1persistent CSMA. In particular note that,
and thus, in the best possible case, when ; the maximumfor S in this case is
obtained for . For low load, however, 1persistent CSMA shows a slightly better
throughput and improved performance.
Y y =
e
g T y + ( ) –
P
10
e
g T y + ( ) –
f
Y
˜
y ( ) y d
0
τ
∫
e
g T y + ( ) –
e
gτ –
δ y ( ) ge
g τ y – ( ) –
+ ( ) y d
0
τ
∫
= =
1 gτ + ( )e
g T τ + ( ) –
=
P
11
g T y + ( )e
g T y + ( ) –
f
Y
˜
y ( ) y d
0
τ
∫
g T y + ( )e
g T y + ( ) –
e
gτ –
δ y ( ) ge
g τ y – ( ) –
+ ( ) y d
0
τ
∫
= =
ge
g T τ + ( ) –
T gτ T
τ
2
 +
¸ ,
¸ _
+ =
S
gTe
g T 2τ + ( ) –
1 gT gτ 1 gT gτ 2 ⁄ + + ( ) + + [ ]
g T 2τ + ( ) 1 e
gτ –
– ( ) – 1 gτ + ( )e
g T τ + ( ) –
+
 =
S
Ge
G 1 2a + ( ) –
1 G aG 1 G aG 2 ⁄ + + ( ) + + [ ]
G 1 2a + ( ) 1 e
Ga –
– ( ) – 1 Ga + ( )e
G 1 a + ( ) –
+
 =
S
a 0 →
G 1 G + ( )
1 Ge
G
+
 =
S 0 → G ∞ →
G 1.03 ≈
Section 4.3.: SLOTTED CARRIER SENSE MULTIPLE ACCESS 89
4.3. SLOTTED CARRIER SENSE MULTIPLE ACCESS
Consider an environment similar to that described for the CSMA protocols except for a
slotted time axis. Let the slot size equal the maximum propagation delay τ which means
that any transmission starting at the beginning of a slot reaches (and could be sensed by)
each and every user by the end of that slot. These slots are sometimes referred to as mini
slots since they are shorter than the time required to transmit a packet. As in every slotted
system users are restricted to start transmissions only at minislot boundaries. We assume
that carrier sensing can be done in zero time (we may assume that τ includes the propaga
tion delay as well as the carrier sensing time). All packets are of the same length and
require T seconds for transmission. We also assume that the packet size T is an integer
multiple of the propagation delay τ and denote by a the ratio between τ and T (1/a is there
fore an integer).
Users behave as follows. When a packet is scheduled for transmission at a given time the
user waits to the beginning of the next minislot at which time it senses the channel and if
idle transmits its packet for T seconds, i.e., 1/a minislots (the packet occupies the channel
one more minislot before all other users have received it). If the channel is sensed busy,
then the corresponding CSMA protocol is applied, namely, for Nonpersistent CSMA the
packet is rescheduled to some randomly chosen time in the future, and for 1PCSMA the
0
0.2
0.4
0.6
0.8
1
0.001 0.01 0.1 1 10
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
1Persistent CSMA
FIGURE 4.5: ThroughputLoad of 1Persistent CSMA
a=0
a=0.1
a=1.0
90 CHAPTER 4: CARRIER SENSING PROTOCOLS
user waits until the channel becomes idle and then starts transmission. In both cases col
lided packets are retransmitted at some random time in the future.
Throughput of Slotted Nonpersistent CSMA
We adopt a similar approach to that taken in the corresponding unslotted systems (see Sec
tions 3.2.1 and 3.2.2). Observing the channel we see that a busy period consists of con
secutive transmission periods. The idle period is the time elapsed between every two
successive busy periods (see Figure 4.6). By our deﬁnition, the length of an idle period is
at least one minislot.
For the idle period to be one minislot long means that there is at least one arrival in the
ﬁrst minislot of the idle period. For it to be two minislots long means that there are no
arrivals in its ﬁrst minislot and there is at least one arrival in its second minislot. Con
tinuing this reasoning and considering the Poisson scheduling process we have,
so,
(4.13)
An outcome of the deﬁnition of the model is the fact that both successful and unsuccessful
transmission periods last T+τ seconds (see Figure 4.6). A collision occurs if two or more
packets arrive within the same minislot and are scheduled for transmission in the next
minislot. A busy period will contain k transmission periods if there is at least one arrival
B
˜
I
˜
T
τ
Idle
Period
Idle
Period
Successful
Transmission
Period
Unsuccessful
Transmission
Period
FIGURE 4.6: Slotted Nonpersistent CSMA Packet Timing
P I
˜
kτ = [ ] e
gτ –
( )
k 1 –
1 e
gτ –
– ( ) = k 1 2 … , , =
I
τ
1 e
gτ –
–
 =
Section 4.3.: SLOTTED CARRIER SENSE MULTIPLE ACCESS 91
in the last minislot of each of the ﬁrst k1 transmission periods, and no arrival in the last
minislot of the kth transmission period. Thus,
so,
Following a similar approach to that used in the slotted Aloha case we deﬁne a cycle as the
period consisting of a busy period followed by an idle period and denote by the amount
of time within a cycle during which the channel carries useful information. When a trans
mission period is successful the channel carries useful information for T seconds, while it
carries no useful information in unsuccessful transmission periods. Since the number of
transmission periods during is , we have
where is the probability of a successful transmission period. We have
The division by the probability of “some arrivals” is noteworthy. It is necessary because
we are computing the probability of a single arrival in the last minislot of the preceding
transmission period knowing that there was at least one arrival, since a transmission period
has been initiated.
Putting all these together we get
Using a=τ/T and G=gT we have:
Prob B
˜
k T τ + ( ) = [ ] 1 e
gτ –
– ( )
k 1 –
e
gτ –
= k 1 2 … , , =
B
T τ +
e
gτ –
 =
U
˜
B
˜
B
˜
T τ + ( ) ⁄
E U
˜
[ ] T
B
T τ +
P
suc
=
P
suc
P
suc
Successful Transmission Period [ ] Prob =
single arrival in last minislot
before the transmission period
some arrivals Prob =
Single arrival in last minislot [ ] Prob
Some arrivals in last minislot [ ] Prob
  =
gτe
gτ –
1 e
g – τ
–
 =
S
U
B I +

T
B
T τ +
P
suc
T τ +
e
gτ –

τ
1 e
gτ –
–
 +

Tgτe
gτ –
T τ Te
gτ –
– +
 = = =
S
aGe
aG –
1 a e
aG –
– +
 =
92 CHAPTER 4: CARRIER SENSING PROTOCOLS
When a is very small we obtain:
which is identical to the unslotted case when .
Throughput of Slotted 1PCSMA
The analysis of the slotted 1PCSMA is similar to that of slotted Nonpersistent CSMA.
The mean of the idle period is given by (4.13). The distribution of the busy period is
since a busy period will contain k transmission periods if at least one packet arrives in each
of the ﬁrst k1 transmission periods (as opposed to minislots in the nonpersistent case),
and no packet arrives in the kth transmission period. So,
The probability of success in the ﬁrst transmission period in a busy period, , is differ
ent from the success probability in any other transmission period within the busy period,
. For the ﬁrst transmission period in a busy period to be successful we need the last
minislot of the idle period to contain exactly one arrival (notice that we know there is at
least one arrival there, since it is the last minislot of the idle period). Hence,
For any transmission period in a busy period, other than the ﬁrst, to be successful we must
have exactly one arrival during the previous transmission period, i.e.,
The channel carries useful information only during successful transmission periods. The
probability of success of the ﬁrst transmission period in a busy period is and there
fore is the expected amount of time the channel carries useful information during
these periods. The expected number of transmission periods (other than the ﬁrst) in a busy
S
a 0 →
G
1 G +
 =
a 0 →
B
˜
k T τ + ( ) = [ ] Prob 1 e
g T τ + ( ) –
– ( )
k 1 –
e
g T τ + ( ) –
=
B
T τ +
e
g T τ + ( ) –
 =
P
suc
1
P
suc
2
P
suc
1
Successful transmission in first
transmission period of a busy period
Prob =
Single transmission in a minislot at leat one arrival [ ] Prob =
gτe
gτ –
1 e
gτ –
–
 =
P
suc
2
Successful transmission in nonfirst period in a busy period [ ] Prob =
Single arrival in a transmission period at least one arrival [ ] Prob =
g T τ + ( )e
g T τ + ( ) –
1 e
g T τ + ( ) –
–
 =
P
suc
1
T P
suc
1
⋅
Section 4.3.: SLOTTED CARRIER SENSE MULTIPLE ACCESS 93
period is , since each transmission period lasts T+τ seconds. The probabil
ity of success in each of these transmission periods is and therefore
is the expected amount of time the channel carries useful
information during these periods. In summary, the expected amount of time within a cycle
that the channel carries useful information is
.
The throughput is therefore given by
or in a normalized form
This relation is depicted in Figure 4.7
For the case of a very small minislot size we have
B T τ + ( ) ⁄ 1 –
P
suc
2
B T τ + ( ) ⁄ 1 – [ ] T P
suc
2
⋅ ⋅
U T P
suc
1
B T τ + ( ) –
T τ +
P
suc
2
+ =
S
U
B I +

gTe
g T τ + ( ) –
T τ Te
gτ –
– + [ ]
T τ + ( ) 1 e
gτ –
– [ ] τe
g T τ + ( ) –
+
 = =
S
Ge
1 a + ( )G –
1 a e
aG –
– + [ ]
1 a + ( ) 1 e
aG –
– [ ] ae
1 a + ( )G –
+
  =
CSMA
a=0.01
0
0.2
0.4
0.6
0.8
1
0.001 0.01 0.1 1 10 100 1000
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
Slotted
Nonpersistent
Nonpersistent
Slotted
1Persistent
FIGURE 4.7: ThroughputLoad of Slotted 1Persistent and Nonpersistent CSMA
94 CHAPTER 4: CARRIER SENSING PROTOCOLS
Comparing these graphs with those of the corresponding unslotted systems we note, as
expected, a slightly better performance of the slotted systems. Practically speaking, the
very small gain achieved is probably not worth the cost of keeping the users synchronized.
From a theoretical standpoint the close performance means that the slotted system can
serve as an approximation of the unslotted one. This is advantageous since analysis of
slotted systems is often much simpler than the unslotted ones.
4.4. CARRIER SENSE MULTIPLE ACCESS WITH COLLISION
DETECTION
The Aloha family of protocols suffers fromthe inherent interference of concurrently trans
mitted packets, that is, whenever the transmission of two or more packets overlap in time,
even a bit, all are lost and must be retransmitted. The pure Aloha protocol suffers most, as
no precautions to reduce collisions are taken. CSMA reduces the level of interference
caused by overlapping packets by allowing users to sense the carrier due to other users’
transmissions, and inhibit transmission when the channel is in use and a collision is inevi
table. CSMA protocols appear to be the best possible solution since their performance
depends only on the endtoend propagation delaya quantity that is not alterable (except
by a different topological design). To further improve performance, a new avenue must
therefore be sought.
Throughput, our measure of performance, is the ratio between the expected useful time
spent in a cycle to the cycle duration itself. To improve the throughput we must therefore
reduce the cycle length, an observation that is the foundation of the protocols described in
this section. As we have seen, a cycle is composed of a transmission period followed by an
idle period. Shortening the idle period is possible by means of 1persistent protocols
which, unfortunately, perform poorly under most loads. Finding the way to shorten the
busy period is therefore our only recourse. Clearly, the duration of successful transmission
periods should not be changed for this is the time the channel is used best. Hence, perfor
mance can be improved by shortening the duration of unsuccessful transmission periods,
as we now explain.
Beside the ability to sense carrier, some local area networks (such as Ethernet) have an
additional feature, namely that users can detect interference among several transmissions
(including their own) while transmission is in progress and abort transmission of their col
lided packets. If this can be done sufﬁciently fast then the duration of an unsuccessful
transmission would be shorter than that of a successful one which is the effect we were
looking for. Together with carrier sensing this produces a variation of CSMA that is
known as CSMA/CD (Carrier Sense Multiple Access with Collision Detection).
S
a 0 →
Ge
G –
1 G + [ ]
G e
G –
+
 =
Section 4.4.: CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION 95
The operation of all CSMA/CDprotocols is identical to the operation of the corresponding
CSMA protocols, except that if a collision is detected during transmission, the transmis
sion is aborted and the packet is scheduled for transmission at some later time.
In all CSMA protocols, a transmission that is initiated when the channel is idle reaches all
users after at most one endtoend propagation delay, τ. Beyond this time, the channel will
be sensed busy. The spacetime diagramof Figure 4.8 captures this situation. In this ﬁgure
we consider two users A and B, the propagation between whom is τ. Suppose that user A
starts transmission at time t
0
when the channel is idle, then its transmission reaches user B
at t
0
+τ. Suppose, further, that B initiates a transmission at time t
1
<t
0
+τ (when B still
senses an idle channel). It takes τ
cd
for a user to detect the collision, so that at time t
0
+τ+τ
cd
user B positively determines the collision. In many local area networks such as
Ethernet, every user upon detection of a collision initiates a consensus reenforcement pro
cedure, which is manifested by jamming the channel with a collision signal for a duration
of τ
cr
to ensure that all network users indeed determine that a collision took place. Thus, at
t
0
+τ+τ
cd
+τ
cr
user B completed the consensus reenforcement procedure which reaches
user A at t
0
+2τ+τ
cd
+τ
cr
. From user A’s standpoint this transmission period lasted
.
By similar calculation, user B completes this transmission period at time t
1
+γ. The chan
nel is therefore busy for a period of t
1
+γ t
0
. In the worst case user B starts transmission
just prior to the arrival of A’s packet, i.e., at time t
1
=t
0
+τ; hence in the worst case, in an
t
0
τ + t
1
t
0
τ + τ
cd
τ
cr
+ + t
1
γ + t
0
T τ + +
t
1
τ +
t
1
τ + τ
cd
+
t
0
γ +
t
0
T +
A
B
Distance
Unsuccessful Transmission Period
Successful Transmission Period
time
t
0
FIGURE 4.8: Collision detection Timing
t
0
τ + τ
cd
+
t
1
τ + τ
cd
τ
cr
+ +
γ 2τ τ
cd
τ
cr
+ + =
96 CHAPTER 4: CARRIER SENSING PROTOCOLS
unsuccessful transmission period the channel remains busy for a duration of γ+τ. Denoting
by the length of the transmission period we have
In the following we analyze the slotted versions of CSMA/CD, namely, it is assumed that
time is quantized into minislots of length τ seconds and that all users are synchronized so
that transmissions can begin only at the start of a minislot. Thus, when a packet is sched
uled for transmission during some minislot, the user with that packet waits until the end
of that minislot, senses the channel, and follows the corresponding version of the CSMA/
CD protocol. In addition, we assume that both γ and T (the transmission time of a packet)
are integer multiples of τ. Thus takes on only values that are certain integer multiples of
τ. The analysis is based on the work by Tobagi and Hunt [ToH80].
Throughput of Slotted Nonpersistent CSMA/CD
With the nonpersistent CSMA/CD time alternates between busy periods (that contain both
successful and unsuccessful transmission periods) and idle periods. A cycle is a busy
period followed by an idle period (see Figure 4.9).
We denote, as before, the length of the busy period by , the length of the idle period by
and the useful time in a cycle by The distribution of the idle period is identical to that
computed for slotted nonpersistent CSMA, i.e.,
X
˜
X
˜
T τ + Successful transmission period
γ τ + Unsuccessful transmission period
¹
'
¹
=
X
˜
γ
τ
T
Successful
Transmission
Period
Idle
Period
Idle
Period
Unsuccessful
Transmission
Period
Busy Period
time
FIGURE 4.9: Slotted Nonpersistent CSMA/CD Packet Timing
B
˜
I
˜
U
˜
Section 4.4.: CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION 97
(4.14)
so the expected length of the idle period is
(4.15)
The probability that a certain transmission in a busy period is successful is the probability
that the transmission period contains exactly one packet (given that it contains at least one
packet), i.e., the probability of a single arrival in a minislot (given that there was at least
one arrival):
(4.16)
Each transmission period that contains a successful transmission is of length T+τ seconds
while a transmission period with an unsuccessful transmission is of length γ+τ seconds. A
busy period will contain l transmission periods if there was at least one arrival in the last
minislot of each of the ﬁrst l1 transmission periods, and no arrival in the last minislot of
the lth transmission period. Therefore, the probability that the busy period contains exactly
l ( ) transmission periods is and the average number of transmis
sion periods within the busy period is . In addition, we have that the probability
distribution of the length of the busy period is
where l corresponds to the total number of transmission periods in the busy period and k
corresponds to the successful transmission periods. Therefore,
. (4.17)
We now turn to compute . Every successful transmission period contributes T
to while unsuccessful transmission periods do not contribute anything. Thus,
I
˜
kτ = ( ) Prob e
gτ –
( )
k 1 –
1 e
gτ –
– ( ) = k 1 2 … , , =
I
τ
1 e
gτ –
–
 =
P
suc
Single tranmission at least one transmission [ ] Prob
gτe
gτ –
1 e
gτ –
–
 = =
l 1 ≥ e
gτ –
1 e
gτ –
– ( )
l 1 –
1 e
gτ –
⁄
B
˜
k T τ + ( ) l k – ( ) γ τ + ( ) + = [ ] Prob
e
gτ –
1 e
gτ –
– ( )
l 1 –
l
k
¸ ,
¸ _
P
suc
k
1 P
suc
– ( )
l k –
= l 1 2 … , , = k 0 1 … l , , , = ,
B k T τ + ( ) l k – ( ) γ τ + ( ) + [ ] k T τ + ( ) l k – ( ) γ τ + ( ) + [ ] Prob
k 0 =
l
∑
l 1 =
∞
∑
=
P
suc
T τ + ( ) 1 P
suc
– ( ) γ τ + ( ) +
e
gτ –
 =
U E U
˜
[ ] =
U
˜
U
˜
kT = ( ) Prob k successful transmission periods in a busy period [ ] Prob =
B
˜
k T τ + ( ) l k – ( ) γ τ + ( ) + = [ ] Prob
l k =
∞
∑
=
98 CHAPTER 4: CARRIER SENSING PROTOCOLS
from which
. (4.18)
Combining (4.16), (4.17), and (4.18) we compute the throughput:
. (4.19)
In a normalized form:
(4.20)
where γ’ is the ratio between γ and the transmission time of a packet (γ’ =γ/T). Notice that
when γ’=1 the result in (4.20) is identical to slotted nonpersistent CSMA. Figure 4.10
depicts the throughputload characteristics of the nonpersistent CSMA with collision
detection. The improvement in performance is readily apparent.
U kT U
˜
kT = [ ] Prob
k 0 =
∞
∑
T
e
gτ –
P
suc
= =
S
U
B I +

gτTe
gτ –
gτTe
gτ –
1 e
gτ –
– ( ) gτe
gτ –
– [ ]γ τ + +
 = =
S
aGe
aG –
aGe
aG –
1 e
aG –
– aGe
aG –
– ( )γ' a + +
 =
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.001 0.01 0.1 1 10 100 1000
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
a=0
a=0.01
a=0.1
a=1.0
FIGURE 4.10: ThroughputLoad of Slotted Nonpersistent CSMA/CD
γ’=2a
Section 4.4.: CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION 99
Throughput of Slotted 1Persistent CSMA/CD
With the 1persistent CSMA/CDthe time also alternates between busy periods (containing
successful and unsuccessful transmission periods) and idle periods, and a cycle is a busy
period followed by an idle period (see Figure 4.11). Notice that here a success or failure of
a transmission period in the busy period depends (only) on the length of the preceding
transmission period, except for the ﬁrst transmission period that depends on arrivals on the
preceding minislot. Denoting by the duration of the ith transmission period in the busy
period, then the duration of the i+1st transmission period depends only on . This is so
since the type of the ith transmission period (success or collision) is determined by the
number of arrivals during the previous transmission period which, in turn, depends only
on its duration. Hence, given that a transmission period is of length x, the length of the
remainder of the busy period is a function of x, and its average is denoted by B(x). Simi
larly, given that a transmission period is of length x, the average time the channel is carry
ing successful transmissions in the remainder of the busy period is denoted by U(x). Let
a
i
(x) be the probability of i arrivals during a period of length x. Under the Poisson assump
tion .
The quantity B(x) is given by:
(4.21)
γ
τ
T
Successful
Transmission
Period
Idle
Period
Idle
Period
Unsuccessful
Transmission
Period
Busy Period
time
FIGURE 4.11: Slotted 1Persistent CSMA/CD Packet Timing
x
˜
i
x
˜
i
a
i
x ( ) gx ( )
i
e
gx –
( ) i! ⁄ =
B x ( )
a
1
x ( )
1 a
0
x ( ) –
 T τ 1 a
0
T τ + ( ) – [ ]B T τ + ( ) + + [ ] =
1
a
1
x ( )
1 a
0
x ( ) –
 – γ τ 1 a
0
γ τ + ( ) – [ ]B γ τ + ( ) + + [ ] +
100 CHAPTER 4: CARRIER SENSING PROTOCOLS
The ﬁrst term in (4.21) corresponds to a successful transmission of the single packet that
arrives during x, in which case the remainder of the busy period will be of length T+τ (the
length of a successful transmission period). In addition, if there is at least one arrival
within T+τ (probability 1a
0
(T+τ)), the remainder of the busy period is of length B(T+τ).
The second term in (4.21) corresponds to an unsuccessful transmission due to arrivals dur
ing x, in which case the remainder of the busy period will be of length γ+τ (the length of
an unsuccessful transmission period). In addition, if there is at least one arrival within γ+τ,
an additional length of B(γ+ τ) is the remainder of the busy period. The expected duration
of the entire busy period is B(τ) since x, the argument of can be interpreted as an
arrival period for the next transmission period and clearly the arrival period for the entire
busy period is the ﬁrst minislot before it started. Observing (4.21) we notice that substi
tuting τ for x in (4.21) is not quite enough since values of appear on the right hand
side as well. This is overcome by setting x=T+τ and x =γ+τ in (4.21) and obtaining two
equations with two unknowns B(T+τ) and B(γ+τ) which can be solved easily. Having
determined these values, B(x) can be determined for any x, in particular x=τ, to yield the
expected length of a busy period.
In a similar manner, U(x) is given by
. (4.22)
The explanation of (4.22) is similar to that of (4.21). Again, using (4.22) with x=T+τ and
x=γ+τ one obtains two equations with two unknowns U(T+τ) and U(γ+τ). The average
time during a cycle that the channel is carrying successful transmissions, U, is given by
U(τ), expressed in terms of U(T+τ) and U(γ+τ).
The average length of an idle period is (see (4.15), and the throughput is
given by
(4.23)
While the above evaluation of the throughput does not result in a closed form, computa
tion is straight forward. Figure 4.12 shows the throughput load characteristic of the 1per
sistent CSMA with collision detection. As opposed to the nonpersistent case, a rather
dramatic change is seen here. For comparison Figure 4.13 shows the characteristics of the
slotted nonpersistent and 1persistent CSMA with and without collision detection, all for
a=0.01. Superiority of the collision detection mechanism is evident. Moreover, the “gap”
in performance between the nonpersistent and the 1persistent CSMA when collision
detection is used has narrowed down. Because of its better performance in low load and
B
.
( )
B
.
( )
U x ( )
a
1
x ( )
1 a
0
x ( ) –
 T 1 a
0
T τ + ( ) – [ ]U T τ + ( ) + [ ] =
1
a
1
x ( )
1 a
0
x ( ) –
 – 1 a
0
γ τ + ( ) – U γ τ + ( ) [ ] +
τ 1 e
gτ –
– ( ) ⁄
S
U
B I +

U τ ( )
B τ ( )
τ
1 e
gτ –
–
 +
 = =
Section 4.5.: RELATED ANALYSIS 101
because its delay characteristic is favorable the 1persistent CSMA with collision detec
tion is so popular in local area networks.
4.5. RELATED ANALYSIS
Carrier sensing has become extremely popular in recent year for one major reasonlocal
area networks (LANs). This is a result of the ease of implementing collision detection in
broadcast LANs. The direct outcome is an enormous amount of research and analysis of
these protocols under all types of circumstances. In fact, the amount of published material
is so large that it is impossible to cover it all or even classify it properly. In the next few
paragraphs we point the reader at some relevant additional work on the subject. For a
broader survey the reader is referred to [Tob80].
Variablelength packets
The performance of CSMA and CSMA/CD protocols with two different possible packet
lengths has been studied in [ToH80]. Batch packet arrivals are considered in [Hey82] for
the CSMA protocol and in [Hey86] the constant length packets are allowed to be grouped
into messages of random size (with a geometric distribution) and the performance of
CSMA/CD is studied.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.01 0.1 1 10 100 1000 10000
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
a=1.0
a=0.1
a=0.01
a=0.001
FIGURE 4.12: ThroughputLoad of Slotted 1Persistent CSMA/CD
γ’=2a
102 CHAPTER 4: CARRIER SENSING PROTOCOLS
Buffered users
The performance of CSMA and CSMA/CD with ﬁnite number of users having ﬁnite or
inﬁnite buffering capabilities has been considered in several works. A twouser system
with inﬁnite buffers is analyzed in [TaK85a], a system with ﬁnite number of buffers per
user is studied in [ApP86] and approximate analysis of a system with inﬁnite buffers is
presented in [TTH88].
Delay and interdeparture times
Numerous papers studied the throughput delay characteristics of CSMA and CSMA/CD
protocols. For instance, Coyle and Liu [CoL83] treated a ﬁnite population as did Tobagi
and Hunt in [ToH80]. Packet delay was analyzed in [CoL85, BeC88] using the matrix
geometric approach [Neu81] and interdeparture time (both distributions and moments)
was derived by Tobagi [Tob82b].
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.001 0.01 0.1 1 10 100 1000
T
h
r
o
u
g
h
p
u
t
(
S
)
Offered Load (G)
1Persistent
1Persistent/CD
Nonpersistent
Nonpersistent/CD
Slotted CSMA
a=0.01
γ’=.02
FIGURE 4.13: Comparison of ThroughputLoad of Slotted systems with a=0.01
Nonpersistent without collision detection,
1persistent without collision detection
Nonpersistent with collision detection
1persistent with collision detection
Section 4.5.: RELATED ANALYSIS 103
Ordered users
Most CSMAtype local networks are implemented using a coax as the transmission
means. As such, the attachment of the users to the cable introduces an inherent order,
which, if properly used, can improve performance substantially. Such attempts were done
by Tobagi and Rom [ToR80, RoT81], by Limb and Flores [LiF82], and by Tobagi and
Fine [FiT84]. It was shown by Rom [Rom84] that users can identify themselves their ordi
nal number on the network.
Performance improvement
In an attempt to improve the performance of CSMA and CSMA/CD protocols Meditch
and Lea [MeL83] derive some optimized versions (keeping stability in mind), Takagi and
Yamada [TaY83] proposed to resolve collisions deterministically while Molle and Klein
rock [MoK85] proposed using virtual time to resolve collisions. Other versions of virtual
time CSMA have been considered in [ZhR87, CuM88]. Attempts to incorporate priority
structures are presented in [RoT81, Tob82a, KiK83].
Collision detection in radio systems
Implementing collision detection in local area networks is relatively simple since the
transmitted and received signals are of the same order of magnitude. In radio systems the
received signal is considerably weak compared with the transmitted signal and therefore
collision detection cannot be implemented via a simple comparison. An idea of how to
implement collision detection in radio systems is described in [Rom86] (see Exercises).
104 CHAPTER 4: CARRIER SENSING PROTOCOLS
EXERCISES
Problem 1.
Find the throughput of nonpersistent and 1persistent CSMA for a→0 and g→∞. Explain
the difference. Find the throughput of a slotted 1persistent CSMA for a→0 and compare
with the result for an unslotted system.
Problem 2.
For Nonpersistent CSMA show that
1. Increasing a uniformly decreases the throughput.
2. There is a single load for which the channel attains its capacity.
Problem 3. (CSMA with a heavy user [ScK79]
Consider a network containing one central computer and a large (read: inﬁnite) number of
terminals all operating as follows. The terminals each have a single packet buffer and com
municate using the slotted ALOHA scheme. The computer has an inﬁnite buffer and uses
a modiﬁed CSMA (see below). All packets are of equal length T.
To increase the total throughput the slot size is set to T+2a and the terminal packets carry
a preamble of length a (where a is the propagation delay in the system). That is, a terminal
packet transmission consists of a seconds of carrier followed by T seconds of information
(and, of course, a more seconds to ‘clean’ the channel). The computer, at the beginning of
the slot where transmission is attempted, listens to the channel and will transmit a packet
of duration T only if the channel is sensed idle. (Note that the computer has in fact a lower
priority since it defers transmission to an ongoing terminal transmission).
Let λ
1
and λ
2
be the Poisson arrival rate (in packets/second) of the computer and the ter
minals respectively and let g be the combined offered load of the terminals. Deﬁne, as
usual, the partial throughputs S
1
= λ
1
T, S
2
=λ
2
T, and the total throughput S = S
1
+S
2
. (For
convenience deﬁne Λ
i
=λ
i
(T+2a)).
1. Find the throughputs S
1
, S
2
, and S.
2. When is the throughput maximal? What is the throughput in this case?
3. Let denote the service time of a central computer’s packet. The random variable is
the time from the moment the channel is ﬁrst sensed for that packet until its transmis
sion is complete. Find and .
4. From the results of part (3) compute the average delay of a computer packet. What is
the average delay under the conditions of part (2)?
x
˜
x
˜
E x
˜
[ ] E x
˜
2
[ ]
Section : EXERCISES 105
Problem 4. (Mixed mode CSMA)
Consider the following version of slotted CSMA: Whenever a packet is scheduled for
transmission, the corresponding user senses the channel. If the channel is idle the packet is
transmitted. If the channel is busy, a coin is ﬂipped; with probability p the packet is sched
uled for transmission at some later time (nonpersistent) and with probability 1p the user
waits until the channel becomes idle and then transmits the packet (1persistent).
1. Assume a=0 and use the standard Poisson assumption to determine the relation
between S and G as a function of p. Check your results for p=0 and p=1.
2. Repeat part (a) for . How should p be chosen to maximize the throughput?
Problem 5. (CSMA with a noisy channel)
Consider a slotted nonpersistent CSMA system with a noisy channel. Each slot is noisy
with probability p
e
which is independent of the system and of the noise in previous slots.
When a slot is noisy while a packet is being transmitted, that packet is destroyed. In addi
tion, a user arriving in a noisy slot within an idle period thinks (erroneously) that the chan
nel is busy and behaves accordingly.
To analyze the system we assume an inﬁnite number of users collectively forming a Pois
son arrival process with average g packets per slot. Let the slot size equal the endtoend
propagation delay τ and let all the packets be of equal length T (assumed to occupy an
integer number of slots). For the analysis we deﬁne an embedded Markov chain in the
beginning and end of each transmission period.
1. What is the probability that a given slot in the idle period is not the last slot of that
period.
2. Compute the throughput of the system. Verify your answer for the case p
e
= 0.
Problem 6. (Collision detection in radio systems [Rom86])
This problem deals with a collision detection scheme usable also in radio network. The
scheme is essentially a nonpersistent CSMA with the following exception. Each transmit
ting user pauses during transmission and senses the channel. If the channel is sensed idle
transmission proceeds as usual. If the channel is sensed busy the user concludes that his
packet collided and will not transmit the entire packet. However, the user will not abort
transmission immediately but rather will continue transmitting for some period of time
and then abort. The period from the start of transmission to the time of abortion is called
the collision detection interval.
The analysis of such a system is based on a slotted nonpersistent CSMA. Let there be inﬁ
nitely many users collectively generating (on the channel) a Poisson distributed offered
load with mean g packets/second. The slot size τ equals the endtoend propagation delay
in the system. All packets are of equal length T and occupy an integer number of slots. The
a 0 ≠
106 CHAPTER 4: CARRIER SENSING PROTOCOLS
collision detection interval is R=rτ (where r is an integer) and a transmitting user will
pause for one slot randomly and uniformly chosen among the r slots of transmission start
ing with the second slot.
1. Is it necessary for R to be identical for all users?
2. Deﬁne an embedded Markov chain with which you intend to analyze the system.
3. Watching the channel we observe long and short transmission periods. What is the
probability of a long transmission period? What is the probability of a short transmis
sion period?
4. What is the probability of a successful transmission period?
5. Derive an expression for the channel throughput.
6. Let r
opt
be that r that maximizes the throughput. Why does such an r exist? What sys
tem parameters does it depend on? What is the value of r for low load.
CHAPTER 5
COLLISION RESOLUTION PROTOCOLS
We have seen that the original Aloha protocol is inherently unstable in the absence of
some external control. If we look into the philosophy behind the Aloha protocol, we notice
that there is no sincere attempt to resolve collisions among packets as soon as they occur.
Instead, the attempts to resolve collisions are always deferred to the future, with the hope
that things will then work out, somehow, but they never do.
In this chapter we introduce and analyze multi access protocols with a different philoso
phy. In these protocols, called Collision Resolution Protocols (CRP), the efforts are con
centrated on resolving collisions as soon as they occur. Moreover, in most versions of
these protocols, new packets that arrive to the system are inhibited from being transmitted
while the resolution of collisions is in progress. This ensures that if the rate of arrival of
new packets to the system is smaller than the rate at which collisions can be resolved (the
maximal rate of departing packets), then the system is stable.
The basic idea behind these protocols is to exploit in a more sophisticated manner the
feedback information that is available to the users in order to control the retransmission
process, so that collisions are resolved more efﬁciently and without chaotic events.
The underlying model and the assumptions used here are identical to those assumed for
the slotted Aloha protocol. The channel is slotted and the users can transmit packets
(whose length is one slot) only at beginning of slots. New packets arrive to the system
according to a Poisson process with rate λ packets/slot. If two or more packets are trans
mitted in a slot, a collision occurs and the packets involved in the collision have to be
retransmitted. At the end of each slot the users of the system know what happened in the
slot, namely, whether the slot was idle (no packet was transmitted), or contained a success
ful transmission (exactly one packet was transmitted) or there was a collision (at least two
packets were transmitted). This is known as the ternary feedback model. For some proto
cols it sufﬁces for the users to know whether the slot contained a collision or not. The lat
ter is referred to as binary feedback model.
5.1. THE BINARYTREE PROTOCOL
The most basic collision resolution protocol is called the binarytree CRP (or binarytree
protocol) and was proposed almost concurrently by Capetanakis [Cap79], Tsybakov and
Mikhailov[TsM78], and Hayes[Hay78]. According to this protocol when a collision
occurs, in slot k say, all users that are not involved in the collision wait until the collision is
resolved. The users involved in the collision split randomly into two subsets, by (for
instance) each ﬂipping a coin. The users in the ﬁrst subset, those that ﬂipped 0, retransmit
in slot k+1 while those that ﬂipped 1 wait until all those that ﬂipped 0 transmit success
fully their packets. If slot k+1 is either idle or contains a successful transmission, the users
108 CHAPTER5: COLLISIONRESOLUTION
of the second subset (those that ﬂipped 1) retransmit in slot k+2. If slot k+1 contains
another collision, then the procedure is repeated, i.e., the users whose packets collided in
slot k+1 (the “colliding users”) ﬂip a coin again and operate according to the outcome of
the coin ﬂipping, and so on. We refer to a user having a packet that collided (at least once)
as a backlogged user.
The above explanation shows that the protocol is speciﬁed by a recursion: a group of col
liding packets is split into two subgroups each of which is subjected to the same procedure
as the original group. This recursion will be manifested later when these protocols are ana
lyzed. But even at the description level the recursive operation of the protocol is best
described by a binarytree (see Figure 5.1) in which every vertex corresponds to a time
slot. The root of the tree corresponds to the slot of the original collision. Each vertex of the
tree also designates a subset (perhaps empty) of backlogged users. Vertices whose subsets
contain at least two users (labeled “≥ 2”) indicate collisions and have two outgoing
branches, corresponding to the splitting of the subset into two new subsets. Vertices corre
sponding to empty subsets (labeled “0”), or subsets containing one user (labeled “1”) are
leaves of the tree and indicate an idle and a successful slot, respectively.
To further understand the operation of the protocol we consider in detail the example
depicted in Figure 5.1. A collision occurs in slot 1. At this point it is neither known how
many users nor who are the users that collided in this slot. Each of the colliding users ﬂip
a coin and those that ﬂipped 0 transmit in slot 2. By the rules of the protocol no newly
arrived packet is transmitted while the resolution of a collision is in progress, so that only
users that collided in slot 1 and ﬂipped 1 transmit in slot 2. Another collision occurs in slot
2 and the users involved in that collision ﬂip a coin again. In this example, all the colliding
users of slot 2 ﬂipped 1 and therefore slot 3 is idle. The users that ﬂipped 1 in slot 2 trans
mit again in slot 4, resulting in another collision and forcing the users involved in it to ﬂip
a coin once more. One user ﬂips 0 and transmits (successfully) in slot 5 causing all users
that ﬂipped 1 in slot 4 to transmit in slot 6. In this example there is one such user and
therefore slot 6 is a successful one. Now that the collision among all users that ﬂipped 0 in
slot 1 has been resolved, the users that ﬂipped 1 in that slot transmit (in slot 7). Another
0 1
1
1
1
0
1 2 3 4 5 6 7 8 9 10 11 12
FIGURE 5.1: Example of the BinaryTree Protocol Operation
≥2 ≥2 ≥2
≥2 ≥2
0
1
0
1
0
1
0
1
0
1
Collision resolution Interval (CRI)
Section 5.1.: THE BINARYTREE PROTOCOL 109
collision occurs, and the users involved in it ﬂip a coin. Another collision is observed in
slot 8, meaning that at least two users ﬂipped 0 in slot 7. The users that collided in slot 8
ﬂip a coin and, as it happens, there is a single user that ﬂipped 0 and it transmits (success
fully) in slot 9. Then, in slot 10, transmit the users that ﬂipped 1 in slot 8. There is only one
such user, and his transmission is, of course, successful. Finally, the users that ﬂipped 1 in
slot 7 must transmit in slot 11. In this example there is no such user, hence slot 11 is idle,
completing the resolution of the collision that occurred in slot 7 and, at the same time, the
one in the ﬁrst slot. Observing again Figure 5.1 we see that the order of transmission cor
responds exactly to the traversal of the tree.
It is clear from this example that each user can construct the binarytree shown in Figure
5.1 by following the feedback signals corresponding to each slot. Users that are not
involved in the collision, can also follow the binarytree and thus know exactly when the
collision is resolved. In the same manner, each backlogged user can keep track of his own
position on that tree (while a collision is being resolved), and thus can determine when to
transmit his packet. For the correct operation of the binarytree protocol, the binary feed
back sufﬁces, i.e., users do not have to distinguish idle slots from successful ones.
We say that a collision is resolved when the users of the system know that all packets
involved in the collision have been transmitted successfully. The time interval starting
with the original collision (if any) and ending when this collision is resolved is called col
lision resolution interval (CRI). In the above example the length of the CRI is 11 slots.
The binarytree protocol dictates how to resolve collisions once they occur. To complete
the description of the protocol, we need to specify when newly generated packets are
transmitted for the ﬁrst time or, in other words, to specify the ﬁrsttime transmission rule.
One alternative, that which we assumed all along (known as the obviousaccess scheme),
is that new packets are inhibited frombeing transmitted while a resolution of a collision is
in progress. That is, packets that arrive to the system while a resolution of a collision is in
progress, wait until the collision is resolved, at which time they are transmitted. In the
example of Figure 5.1 all new packets arriving to the system during slots 1 through 11 are
transmitted for the ﬁrst time in slot 12.
The operation of the binarytree protocol can also be described in terms of a stack (see
Figure 5.2). This is, in fact, a standard description of tree traversal by a stack representa
tion. In each slot the stack is popped, and all the packets that were at the top of the stack
are transmitted. In case of a collision, the stack is pushed with the users that ﬂip 1 and then
pushed again with those that ﬂip 0. The users that ﬂip 0 remain therefore at the top of the
stack to be popped and transmitted in the next slot. In case of a successful transmission or
an idle slot no further operations are done on the stack. Clearly then, when the stack emp
ties the collision is resolved, the CRI is over and newly arrived packets (if any) are pushed
onto the top of the stack and operation proceeds as before.
Performance analysis of binarytree CRP can be done either with the tree or stack repre
sentations reaching, of course, the same results. In this book we conﬁne ourselves to the
110 CHAPTER5: COLLISIONRESOLUTION
tree approach. For analysis via the stack representation the interested reader is referred to
the work by Fayolle et. al. [FFH85]. In the following we ﬁrst compute the moments of the
time required to resolve a collision among n packets and obtain tight bounds for these
moments. These results are then used to derive the stability condition for this system.
Finally, we showthat when the systemis stable, the expected delay of a packet is bounded.
The analysis in this section is based on Massey [Mas81].
5.1.1. Moments of the Conditional Length of a CRI
Assume that at some given slot n packets collide. To resolve the collision each participat
ing user ﬂips a coin and proceeds correspondingly. Clearly, the number of slots required to
resolve such a collision is a random variable. Denote therefore by the length (in slots)
of a CRI given that it starts (in its ﬁrst slot) with a collision among n packets, and let
. The quantity B
n
plays a crucial role in the analysis of the binarytree CRP.
Loosely speaking, the ratio n/B
n
represents the “effective service rate” of packets in a CRI
that starts with the transmission of n packets, since n packets are transmitted successfully
during B
n
slots. One would expect that if the arrival rate of new packets is smaller than the
“effective service rate” even when the systemis highly loaded (n is large), then the system
is stable. This statement is made rigorous in Section 5.1.2.. But ﬁrst we compute the
moments of .
When no packet, or a single packet, is transmitted in the ﬁrst slot of a CRI then the CRI
lasts exactly one slot, hence
(5.1)
FIGURE 5.2: BinaryTree Protocol Stack Representation
Transmission Level
Level 0
Level 1
Level 2
Level 3
Flip 0
Flip 1
Idle or
Success
Collision
B
˜
n
B
n
E B
˜
n
[ ] =
B
˜
n
B
˜
0
B
˜
1
1 = =
Section 5.1.: THE BINARYTREE PROTOCOL 111
When n ≥ 2, there is a collision in the ﬁrst slot of the CRI and the random coin ﬂipping
takes place. Let p be the probability that a user ﬂips 0 whenever it ﬂips the binary coin.
Then, the probability that exactly i of the n colliding users ﬂip 0 (and hence transmit in the
next slot) is
(5.2)
Given that i users ﬂipped 0, the length of the CRI is
(5.3)
The 1 corresponds to the slot of the initial collision among the n users. Then it takes
slots to resolve the collision among the i users that ﬂipped 0. Finally, it takes i addi
tional slots to resolve the collision among the ni users that ﬂipped 1. From the above rela
tion the ﬁrst moment of can be recursively derived as follows. First, from equation
(5.1) we have
(5.4)
Then, from equation (5.3)
leading to
(5.5)
In this last equation B
n
appears on both sides of the equation; solving for B
n
we obtain the
recursion
(5.6)
where equation (5.4) provides the initial values.
Q
i
n ( )
n
i
¸ ,
¸ _
p
i
1 p – ( )
n i –
= 0 i n ≤ ≤
B
˜
n i
1 B
˜
i
B
˜
n i –
+ + = n 2 ≥
B
˜
i
B
˜
n i –
B
˜
n
B
0
B
1
1 = =
B
n i
E B
˜
n i
[ ] 1 B
i
B
n i –
+ + = = n 2 ≥
B
n
1 Q
i
n ( ) B
i
B
n i –
+ ( )
i 0 =
n
∑
+ = n 2 ≥
B
n
1 Q
i
n ( )B
i
i 0 =
n 1 –
∑
Q
i
n ( )B
n i –
i 1 =
n
∑
+ +
1 Q
0
n ( ) – Q
n
n ( ) –
 =
1 Q
i
n ( ) Q
n i –
n ( ) + [ ]B
i
i 0 =
n 1 –
∑
+
1 Q
0
n ( ) – Q
n
n ( ) –
 = n 2 ≥
112 CHAPTER5: COLLISIONRESOLUTION
Computation of a similar nature can be done for any moment desired. However, a more
general approach is that of the generating function . To compute this gen
erating function notice that the random variables and are independent and thus
or
(5.7)
In equation (5.7) G
n
(z) appears on both sides so, as we did before, solving for G
n
(z) we
obtain:
and the initial conditions of equation (5.1) translate into
(5.8)
allowing recursive computation of G
n
(z). From this generating function the moments of
can be computed recursively by taking derivatives at z=1. Taking the ﬁrst derivative at
z=1 leads, obviously, to the result of equation (5.6). We now proceed to calculate the sec
ond moment.
Let V
n
be the second moment of i.e., . From (5.1) we immediately have
To compute the second moment for higher values of n we differentiate equation (5.7)
twice with respect to z and obtain
(5.9)
Substituting z=1 in (5.9) and using the facts that and we
obtain
G
n
z ( ) E z
B
˜
n
[ ]
·
∆
B
˜
i
B
˜
n i –
G
n
z ( ) E z
B
˜
n
[ ] Q
i
n ( )E z
1 B
˜
i
B
˜
n i –
+ +
[ ]
i 0 =
n
∑
Q
i
n ( )zE z
B
˜
i
[ ]E z
B
˜
n i –
[ ]
i 0 =
n
∑
= = = n 2 ≥
G
n
z ( ) z Q
i
n ( )G
i
z ( )G
n i –
z ( )
i 0 =
n
∑
= n 2 ≥
G
n
z ( )
z Q
i
n ( )G
i
z ( )G
n i –
z ( )
i 1 =
n 1 –
∑
1 z
2
Q
0
n ( ) Q
n
n ( ) + [ ] –
 = n 2 ≥
G
0
z ( ) G
1
z ( ) z = =
B
˜
n
B
˜
n
V
n
E B
˜
n
2
[ ] =
V
0
V
1
1 = =
G
˙˙
n
z ( ) 2 Q
i
n ( )G
˙
i
z ( )G
n i –
z ( )
i 0 =
n
∑
2 Q
i
n ( )G
i
z ( )G
˙
n i –
z ( )
i 0 =
n
∑
2 Q
i
n ( )G
˙
i
z ( )G
˙
n i –
z ( )
i 0 =
n
∑
+ + =
Q
i
n ( )G
˙˙
i
z ( )G
n i –
z ( )
i 0 =
n
∑
Q
i
n ( )G
i
z ( )G
˙˙
n i –
z ( )
i 0 =
n
∑
+ +
G
˙
n
1 ( ) B
n
= G
˙˙
n
1 ( ) V
n
B
n
– =
Section 5.1.: THE BINARYTREE PROTOCOL 113
(5.10)
where we used (5.5). Solving for V
n
we obtain the following recursion:
(5.11)
The recursive nature of equations (5.5) and (5.11) are sometimes inconvenient to work
with and a direct expression might be preferred. Indeed, it is possible to obtain closed
form expressions for the moments of . The derivations of these expressions are quite
lengthy and for the interested reader are given in Appendix A at the end of this chapter.
The resulting expressions are:
(5.12)
where
It is interesting to investigate the behavior of B
n
as a function of p. Differentiating (5.12)
twice with respect to p we note that independent of n, at the ﬁrst derivative van
ishes while the second derivative is positive. In fact, is the only real value for
which the ﬁrst derivative vanishes. We conclude therefore, that B
n
is minimized for
for all n. Table 1 contains some of the values of B
n
and V
n
when
along with values of the “effective service rate” n/B
n
. Judging by the values of n/B
n
it is
interesting to note that the protocol resolves collisions among a small number of packets
V
n
B
n
– 2 Q
i
n ( ) B
i
B
n i –
B
i
B
n i –
+ + ( )
i 0 =
n
∑
Q
i
n ( ) V
i
B –
i
V
n i –
B –
n i –
+ ( )
i 0 =
n
∑
+ =
B
n
1 – 2 Q
i
n ( )B
i
B
n i –
i 0 =
n
∑
Q
i
n ( ) V
i
V
n i –
+ ( )
i 0 =
n
∑
+ + = n 2 ≥
V
n
2B
n
1 – 2 Q
i
n ( )B
i
B
n i –
i 0 =
n
∑
Q
i
n ( ) Q
n i –
n ( ) + [ ]V
i
i 0 =
n 1 –
∑
+ +
1 Q
0
n ( ) – Q
n
n ( ) –
 = n 2 ≥
B
˜
n
B
n
1
n
k
¸ ,
¸ _
2 k 1 – ( ) 1 – ( )
k
1 p
k
– 1 p – ( )
k
– [ ]

k 2 =
n
∑
+ = n 2 ≥
V
n
1
2n!
n k – ( )! 1 p
k
– 1 p – ( )
k
– [ ]
 B
k
*
B
i
*
B
k i –
*
i 0 =
k
∑
+
k 2 =
n
∑
+ =
n
k
¸ ,
¸ _
4 k 1 – ( ) 1 – ( )
k
1 p
k
– 1 p – ( )
k
– [ ]

k 2 =
n
∑
+ n 2 ≥
B
0
*
1 = B
1
*
0 = B
k
*
n
k
¸ ,
¸ _
2 k 1 – ( ) 1 – ( )
k
k! 1 p
k
– 1 p – ( )
k
– [ ]

k 2 =
n
∑
= k 2 ≥
p 1 2 ⁄ =
p 1 2 ⁄ =
p 1 2 ⁄ = p 1 2 ⁄ =
114 CHAPTER5: COLLISIONRESOLUTION
more efﬁciently than collisions among a large number of packets. We return to this topic in
Section 5.2.2.
Bounds on the Moments
In the previous section the ﬁrst two moments of the conditional length of a CRI are com
puted. In this section we derive upper bounds for these moments. We are seeking an upper
bound for B
n
of the form
(5.13)
for some arbitrary m and with some α
m
>0. The motivation for a bound of this form is that
it guarantees a strictly positive “effective service rate” (n/B
n
) for large n. If indeed a bound
of that form can be found then one would be able to write
and hence the effective service rate is guaranteed to be larger than 1/α
m
. This also moti
vates looking for the smallest α
m
for which the bound holds.
The approach to determine α
m
is as follows. We ﬁx m (m ≥2) and choose some α
m
so that
B
n
≤α
m
n 1 for n= m (any α
m
≥ (B
m
+1)/m is feasible). Next, we assume that (5.13) holds
Table 1: The ﬁrst and second moments of for .
n 1 2 3 4 5
B
n
1.0000 5.0000 7.6667 10.5238 13.4191
n/B
n
1.0000 0.4000 0.3913 0.3801 0.3726
V
n
1.0000 33.000 68.555 124.28 197.00
n 6 7 8 9 10
B
n
16.3131 19.2010 22.0854 24.9691 27.8532
n/B
n
0.3678 0.3646 0.3622 0.3604 0.3590
V
n
286.42 392.36 514.82 653.89 809.63
n 11 12 13 14 15
B
n
30.7382 33.6238 36.5097 39.3955 42.2813
n/B
n
0.3579 0.3569 0.3561 0.3554 0.3548
V
n
982.05 1171.1 1376.9 1599.3 1838.4
B
˜
n
p 1 2 ⁄ =
B
n
α
m
n 1 – ≤ n m ≥
n
B
n

1
α
m

1
α
m
B
n
 + ≥ n m ≥
Section 5.1.: THE BINARYTREE PROTOCOL 115
up to j1, i.e., B
n
≤ α
m
n 1 for n= m,m+1,..., j1 and by induction establish the validity of
(5.13) for j.
The point of departure is equation (5.6), i.e.,
Applying the induction hypothesis we obtain
where we used the facts that , and
. Therefore,
(5.14)
It thus follows that (5.13) holds if we choose α
m
so that the summation in (5.14) is non
positive for all j>m, i.e., such that
or,
B
j
1 Q
0
j ( ) – Q
j
j ( ) – [ ] 1 Q
i
j ( ) Q
j i –
j ( ) + [ ]B
i
i 0 =
j 1 –
∑
+ =
1 Q
i
j ( ) Q
j i –
j ( ) + [ ]B
i
i 0 =
m 1 –
∑
Q
i
j ( ) Q
j i –
j ( ) + [ ]B
i
i m =
j 1 –
∑
+ + =
B
j
1 Q
0
j ( ) – Q
j
j ( ) – [ ] 1 Q
i
j ( ) Q
j i –
j ( ) + [ ]B
i
i 0 =
m 1 –
∑
Q
i
j ( ) Q
j i –
j ( ) + [ ] α
m
i 1 – ( )
i m =
j 1 –
∑
+ + ≤
1 Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
α
m
i – 1 + ( )
i 0 =
m 1 –
∑
+ =
Q
i
j ( ) Q
j i –
j ( ) + [ ] α
m
i 1 – ( )
i 0 =
j
∑
Q
0
j ( ) Q
j
j ( ) + [ ] α
m
j 1 – ( ) – +
1 Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
α
m
i – 1 + ( )
i 0 =
m 1 –
∑
α
m
j 2 – Q
0
j ( ) Q
j
j ( ) + [ ] α
m
j 1 – ( ) – + + =
Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
α
m
i – 1 + ( )
i 0 =
m 1 –
∑
1 Q
0
j ( ) – Q
j
j ( ) – [ ] α
m
j 1 – ( ) + =
Q
i
j ( )
i 0 =
j
∑
1 = iQ
i
j ( )
i 0 =
j
∑
jp =
iQ
j i –
j ( )
i 0 =
j
∑
j 1 p – ( ) =
B
j
α
m
j 1 – ( )
Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
α
m
i – 1 + ( )
i 0 =
m 1 –
∑
1 Q
0
j ( ) – Q
j
j ( ) –
 + ≤
Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
α
m
i – 1 + ( )
i 0 =
m 1 –
∑
0 ≤ j m >
116 CHAPTER5: COLLISIONRESOLUTION
(5.15)
Having computed previously the values of B
i
, the right hand side of equation (5.15) can
also be computed for any desired value of m. Table 2 depicts some of the values of the
expression on the right side of (5.15) for . Recall that we started with α
m
that
satisﬁes . Therefore, if we choose α
m
as the maximum between (B
m
+1)/m
and the following supremum:
(5.16)
the induction step follows and hence (5.13) holds.
To summarize, for a given m, after computing B
i
for i <m (using (5.6) or (5.12)), one can
compute the supremum in (5.16) and hence determine α
m
as the maximum between that
supremum and (B
m
+1)/m.
We recall that B
n
is minimized for . Hence, the best bound of the form (5.13) is
obtained when . For example, choosing say, m=6, we see from Table 2 that the
supremum is 2.886 when , thus
Table 2: Values of right side of (5.15) for .
3 4 5 6 7 8 9 10 11
2 2.667 2.500 2.400 2.333 2.286 2.250 2.222 2.200 2.182
3 2.875 2.880 2.889 2.898 2.907 2.914 2.920 2.926
4 2.885 2.889 2.892 2.894 2.895 2.896 2.896
5 2.886 2.886 2.887 2.887 2.886 2.886
6 2.886 2.886 2.886 2.885 2.885
7 2.886 2.886 2.886 2.885
8 2.886 2.886 2.886
9 2.886 2.886
α
m
Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
1 + ( )
i 0 =
m 1 –
∑
i Q
i
j ( ) Q
j i –
j ( ) + [ ]
i 0 =
m 1 –
∑
 ≥ j m >
p 1 2 ⁄ =
B
m
α
m
m 1 – ≤
Q
i
j ( ) Q
j i –
j ( ) + [ ] B
i
1 + ( )
i 0 =
m 1 –
∑
i Q
i
j ( ) Q
j i –
j ( ) + [ ]
i 0 =
m 1 –
∑

¹ ¹
¹ ¹
¹ ¹
' ;
¹ ¹
¹ ¹
¹ ¹
j m >
sup
p 1 2 ⁄ =
j
m
p 1 2 ⁄ =
p 1 2 ⁄ =
p 1 2 ⁄ =
Section 5.1.: THE BINARYTREE PROTOCOL 117
(5.17)
Although inequality (5.17) does not hold for n<6, it is very easy to bound B
n
for all n ≥0
(by using (5.17) and Table 1) by
(5.18)
We conclude that the “effective service rate”, n/ B
n
, for large n is 1/2.886 ≅ 0.346 and thus
the system is expected to be stable for arrival rates smaller than 0.346. This effective rate
is smaller than e
1
the maximum throughput of the slotted Aloha protocol; we shall
shortly present improved versions of the binarytree protocol that yield much better per
formance. In subsequent computations of system parameters and performance we shall
take the value α
m
= 2.886.
Using the above methodology, it is possible to develop a bound on the second moment of
the conditional length of a CRI, V
n
, of the form
(5.19)
where α
m
is the same one used to bound B
n
and is determined by the procedure described
above. This bound will be required in developing an upper bound on the expected delay of
a packet.
We ﬁrst check the validity of (5.19) for n=m. For , m =6, and α
m
=2.886 we see
from Table 1 that (5.19) holds. Next, we assume that (5.19) holds for all values of n up to
j1 i.e., for n=m,m+1,..., j1 and by induction establish the validity of
(5.19) for j. The point of departure is equation (5.11) that can be rewritten for j≥2 (using
(5.5) as
(5.20)
Using (5.13) we have
B
n
2.886n 1 – ≤ n 6 ≥
B
n
2.886n 1 + ≤ n 0 ≥
V
n
α
m
2
n
2
1 + ≤ n m ≥
p 1 2 ⁄ =
V
n
α
m
2
n 1 + =
V
j
1 2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
+ + ( )
i 0 =
j
∑
Q
i
j ( ) Q
j i –
j ( ) + [ ]V
i
i 0 =
j 1 –
∑
+ +
1 Q
0
j ( ) – Q
j
j ( ) –
 =
118 CHAPTER5: COLLISIONRESOLUTION
(5.21)
Similarly, assuming that up to we have
(5.22)
Substituting (5.21) and (5.22) in (5.20) we obtain:
Therefore, for the induction hypothesis to hold we require that
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
+ + ( )
i 0 =
j
∑
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
+ + ( )
i 0 =
m 1 –
∑
≤
2 Q
i
j ( ) α
m
i 1 – α
m
j i – ( ) 1 – α
m
i 1 – ( ) α
m
j i – ( ) 1 – ( ) + + [ ]
i m =
j
∑
+
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
+ + ( )
i 0 =
m 1 –
∑
2 Q
i
j ( ) α
m
2
i j i – ( ) 1 – [ ]
i m =
j
∑
+ =
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
α
m
2
i j i – ( ) – 1 + + + [ ]
i 0 =
m 1 –
∑
2 Q
i
j ( ) α
m
2
i j i – ( ) 1 – [ ]
i 0 =
j
∑
+ =
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
α
m
2
i j i – ( ) – 1 + + + [ ]
i 0 =
m 1 –
∑
2α
m
2
j
2
p jp 1 p – ( ) – j
2
p
2
– [ ] 2 – + =
V
i
α
m
2
i
2
1 + ≤ i j 1 – ≤
Q
i
j ( ) Q
j i –
j ( ) – [ ]V
i
i 0 =
j 1 –
∑
Q
i
j ( ) Q
j i –
j ( ) – [ ]V
i
i 0 =
m 1 –
∑
Q
i
j ( ) Q
j i –
j ( ) – [ ] α
m
2
i
2
1 + ( )
i m =
j 1 –
∑
+ ≤
Q
i
j ( ) Q
j i –
j ( ) – [ ] V
i
α
m
2
i
2
– 1 – ( )
i 0 =
m 1 –
∑
=
Q
i
j ( ) Q
j i –
j ( ) + [ ] α
m
2
i
2
1 + ( )
i 0 =
j
∑
Q
0
j ( ) Q
j
j ( ) – [ ] α
m
2
j
2
1 + ( ) – +
Q
i
j ( ) Q
j i –
j ( ) – [ ] V
i
α
m
2
i
2
– 1 – ( )
i 0 =
m 1 –
∑
=
α
m
2
2 jp 1 p – ( ) j
2
p
2
j
2
1 p – ( )
2
+ + [ ] 2 Q
0
j ( ) Q
j
j ( ) + [ ] α
m
2
j
2
1 + ( ) – + +
V
j
α
m
2
j
2
1
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
α
m
2
i j i – ( ) – 1 + + + [ ]
i 0 =
m 1 –
∑
1 Q
0
j ( ) – Q
j
j ( ) –
  + + ≤
Q
i
j ( ) Q
j i –
j ( ) + [ ] V
i
α
m
2
i
2
– 1 – ( )
i 0 =
m 1 –
∑
1 Q
0
j ( ) – Q
j
j ( ) –
 +
Section 5.1.: THE BINARYTREE PROTOCOL 119
(5.23)
The correctness of (5.23) for , m=6 and α
m
=2.886 can be checked directly.
By analogous arguments to those we used to establish upper bounds of the ﬁrst two
moments of , it is possible to establish lower bounds on these moments. For instance, it
can be shown that (see [Mas81])
(5.24)
and therefore one can show that the system is unstable for arrival rate that is larger than 1/
2.881 ≅ 0.347.
5.1.2. Stability Analysis
One of the most important properties of the binarytree CRP is its stable behavior to which
we alluded in previous sections. We are now ready to prove this claim. When the binary
tree CRP is executed then along the time axis we observe a sequence of collision resolu
tion intervals. Let be an integervalued randomvariable that represents the length (in
slots) of the kth CRI. When the obvious access scheme is employed, the chain
forms a Markov chain because the length of the k+1st CRI is
determined by the number of packets transmitted in its ﬁrst slot. This number equals the
number of packets arriving during the kth CRI and depends only on the length of the kth
CRI. The system is said to be stable if the Markov chain is
ergodic. (We shall see in the next section that when the system is stable, the expected
delay of a packet is ﬁnite).
To obtain sufﬁcient conditions for which the Markov chain is
ergodic, we use again Pakes’ Lemma (see Section 3.4.). We ﬁrst notice that the chain is
irreducible, aperiodic and homogeneous. To be ergodic, it is sufﬁcient that the chain fulﬁll
the following two conditions:
1. ;
2. .
We start by computing the following conditional expectation:
2 Q
i
j ( ) B
i
B
j i –
B
i
B
j i –
α
m
2
i j i – ( ) – 1 + + + [ ]
i 0 =
m 1 –
∑
Q
i
j ( ) Q
j i –
j ( ) + [ ] V
i
α
m
2
i
2
– 1 – ( )
i 0 =
m 1 –
∑
+ 0 ≤ j m >
p 1 2 ⁄ =
B
˜
n
B
n
2.881n 1 – ≥ n 6 ≥
B
˜
k ( )
B
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
B
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
B
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
E B
˜
k 1 + ( ) B
˜
k ( ) – ( ) B
˜
k ( ) i = ( ) [ ] ∞ < i ∀
E B
˜
k 1 + ( ) B
˜
k ( ) – ( ) B
˜
k ( ) i = ( ) [ ]
i ∞ →
limsup 0 <
120 CHAPTER5: COLLISIONRESOLUTION
(5.25)
where B
n
is the expected length (in slots) of a CRI given that it started with a collision
among n packets and λ is the expected number of packets that arrive to the system in a
slot. In (5.25) we used the fact that the distribution of the length of a CRI, given that it
starts with the transmission of n packets, does not depend on the length of the previous
CRI, and that the arrival process is Poisson.
In the previous section we have shown that B
n
is ﬁnite for n ≥0 and, moreover, is bounded
by
(5.26)
where α = 2.886. Substituting this bound in (5.25) we have
Therefore
. (5.27)
From (5.27) we see that conditions (a) and (b) of Pakes’ Lemma hold if
.
It follows then, that a sufﬁcient condition for stability of the system is that λ, the arrival
rate of new packets, be less than 1/α
m
. Consequently, the system is stable for arrival rates
that are smaller than 0.346 (packets per slot).
E B
˜
k 1 + ( ) B
˜
k ( ) i = [ ]
E B
˜
k 1 + ( ) B
˜
k ( ) i = number of arrivals in kth CRI n = , [ ]
λi ( )
n
e
λi –
n!

n 0 =
∞
∑
=
E B
˜
k 1 + ( ) B
˜
k ( ) i = n packets transmitted at start of B
˜
k 1 + ( ) , [ ]
λi ( )
n
e
λi –
n!

n 0 =
∞
∑
=
E B
˜
k 1 + ( ) n packets transmitted at start of B
˜
k 1 + ( ) [ ]
λi ( )
n
e
λi –
n!

n 0 =
∞
∑
=
B
n
λi ( )
n
e
λi –
n!

n 0 =
∞
∑
=
B
n
αn 1 + ≤ n 0 ≥
E B
˜
k 1 + ( ) B
˜
k ( ) i = [ ] αn 1 + ( )
λi ( )
n
e
λi –
n!

n 0 =
∞
∑
≤ αλi 1 + =
E B
˜
k 1 + ( ) B
˜
k ( ) – B
˜
k ( ) i = [ ] E B
˜
k 1 + ( ) B
˜
k ( ) i = [ ] E B
˜
k ( ) B
˜
k ( ) i = [ ] – =
E B
˜
k 1 + ( ) B
˜
k ( ) i = [ ] i – = αλ 1 – ( )i 1 + ≤
λ
1
α
 <
Section 5.1.: THE BINARYTREE PROTOCOL 121
5.1.3. Bounds on Expected Packet Delay
Let be the delay of a randomly chosen packet (a tagged packet), namely, the difference
between its arrival time to the system and the time it is transmitted successfully. The pur
pose of this section is to show that the expected delay of a randomly chosen
packet is ﬁnite when λ < α
1
where α
1
= 0.346.
We already proved that the Markov chain is ergodic when λ<α
1
.
Let be the number of packets transmitted at the beginning of the kth CRI. We now
show that the Markov chain is also ergodic when λ < α
1
. Since
the arrival process of new packets is Poisson we have
(5.28)
and therefore,
Testing conditions (a) and (b) of Pakes’ Lemma for the chain we
conclude that it is ergodic when λ < α
1
.
Being ergodic Markov chains, steadystate distribution of and exist. Let
denote the length of a CRI in steadystate and let B and be the ﬁrst and second
moments of , respectively, namely, and . Similarly, let
denote the number of packets transmitted at the beginning of a CRI in steadystate and let
A and be the ﬁrst and second moments of , namely, and .
From (5.28) we see that A=λ B. In addition we have that
which implies
(5.29)
Let denote the length of the CRI in progress when the tagged packet arrives at the
system and let be the length of the subsequent CRI during which the tagged packet
departs from the system. Then
(5.30)
since at the earliest the tagged packet arrives at the beginning of the interval whose length
is and at the latest it will be transmitted successfully at the end of the next interval
whose length is .
Let be the number of packets transmitted in steadystate at the beginning of the CRI
in which the packet leaves the system. By deﬁnition,
D
˜
D E D
˜
[ ] =
B
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
A
˜
k ( )
A
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
E A
˜
k 1 + ( ) A
˜
k ( ) i = [ ] λB
i
=
E A
˜
k 1 + ( ) A
˜
k ( ) – A
˜
k ( ) i = ( ) [ ] λB
i
i – λα 1 – ( )i λ – ≤ =
A
˜
k ( ) k 0 1 2 … , , , = ( ) , { }
B
˜
k ( ) A
˜
k ( ) B
˜
B
2
B
˜
B E B
˜
[ ] = B
2
E B
˜ 2
[ ] = A
˜
A
2
A
˜
A E A
˜
[ ] = A
2
E A
˜ 2
[ ] =
E A
˜
k 1 + ( ) ( )
2
B
˜
k ( ) l = [ ] λl λl ( )
2
+ =
A
2
λB λ
2
B
2
+ A λ
2
B
2
+ = =
B
˜ a ( )
B
˜ d ( )
D E B
˜ a ( )
[ ] E B
˜ d ( )
[ ] + ≤
B
˜ a ( )
B
˜ d ( )
A
˜ d ( )
122 CHAPTER5: COLLISIONRESOLUTION
and therefore with the help of (5.18),
where α= 2.886. Unconditioning with respect to , we obtain
(5.31)
Similarly to (5.28) we have
(5.32)
Combining (5.30), (5.31) and (5.32) we have
Lastly, the residual life theorem states that (see Appendix)
and since B≥1 we have
(5.33)
and therefore we only have to bound in order to bound D.
We have that
(5.34)
The above along with (5.29) implies that
and since λ < α
1
we have
E B
˜ d ( )
A
˜ d ( )
n = [ ] B
n
=
E B
˜ d ( )
A
˜ d ( )
n = [ ] αn 1 + ≤
A
˜ d ( )
E B
˜ d ( )
[ ] αE A
˜ d ( )
[ ] 1 + ≤
E A
˜ d ( )
[ ] λE B
˜ a ( )
[ ] =
D 1 λα + ( )E B
˜ a ( )
[ ] 1 + ≤
E B
˜ a ( )
[ ]
B
2
B
 =
D 1 λα + ( )B
2
1 + ≤
B
2
B
2
E B
˜
( )
2
[ ] E B
˜
( )
2
A
˜
n = ( ) [ ]Prob A
˜
n = [ ]
n 0 =
∞
∑
= =
V
n
Prob A
˜
n = [ ]
n 0 =
∞
∑
α
2
n
2
1 + ( )Prob A
˜
n = [ ]
n 0 =
∞
∑
≤ = α
2
A
2
1 + =
B
2
α
2
A λ
2
B
2
+ ( ) 1 + ≤
B
2
α
2
A 1 +
1 λ
2
α
2
–
 ≤
Section 5.2.: ENHANCED PROTOCOLS 123
Substituting the above result in (5.33) we obtain
(5.35)
where we used the fact that A=λ B.
Similarly to (5.34) and using (5.18) we have
or
(5.36)
Thus, substituting (5.36) in (5.35) we obtain
Therefore, we showed that when the system is stable, the expected delay of a packet is
ﬁnite and an explicit upper bound for this quantity is given above.
5.2. ENHANCED PROTOCOLS
The performance of the binarytree protocol can be improved in two ways. The ﬁrst is to
speed up the collision resolution process by avoiding certain, avoidable, collisions. The
second is based on the observation (see Table 4.1) that collisions among a small number of
packets are resolved more efﬁciently than collisions among a large number of packets
(compare the ratio n/B
n
for small n and for large n). Therefore, if most CRIs start with a
small number of packets, the performance of the protocol is expected to improve. These
ideas are the basis of the protocols presented in this section.
5.2.1. The Modiﬁed BinaryTree Protocol
Consider again the example depicted in Figure 5.1. In slots 2 and 3 a collision is followed
by an idle slot. This implies that in slot 2 all users (and there were at least two of them)
ﬂipped 1. The binarytree protocol dictates that these users must transmit in slot 4,
although it is obvious that this will generate a collision that can be avoided. The modiﬁed
binarytree protocol is due to Massey [Mas81] and eliminates such avoidable collisions by
D 1 λα + ( )
α
2
A 1 +
1 λ
2
α
2
–
 1 + ≤
α
2
A 1 +
1 λα –
 1 +
α
2
λB 1 +
1 λα –
 1 + = =
B E B
˜
[ ] E B
˜
A
˜
n = [ ]Prob A
˜
n = [ ]
n 0 =
∞
∑
B
n
Prob A
˜
n = [ ]
n 0 =
∞
∑
= = =
αn 1 + ( )Prob A
˜
n = [ ]
n 0 =
∞
∑
≤ αA 1 + αλB 1 + = =
B
1
1 λα –
 ≤
D
α
2
λ 1 λα – +
1 λα – ( )
2
 1 + ≤
124 CHAPTER5: COLLISIONRESOLUTION
letting the users that ﬂipped 1 in slot 2 in the example above, ﬂip coins before transmitting
is slot 4. Consequently, the slot in which a avoidable collision would occur is skipped and
the evolution of the protocol for the same example of Section 5.1. is depicted in Figure
5.3. Except for eliminating these avoidable collisions the modiﬁed binarytree protocol
evolves exactly as the binarytree protocol. Note that the correct operation of the modiﬁed
binarytree protocol requires ternary feedback, i.e., the users have to be able to distinguish
between idle and successful slots.
The analysis of the modiﬁed binarytree protocol is essentially the same as that of the
binarytree protocol. We have
and given that a CRI starts with a collision of n (n≥2) packets and that i users ﬂip 0, the
conditional length of the CRI is given by
which accommodates for the saving of one slot when no users ﬂip 0 (i=0).
The procedure of determining the expected length of a CRI given that it starts with a colli
sion among n users is entirely analogous to the derivation in the previous section. The
equation analogous to (5.6) is,
with the initial values B
0
=B
1
=1.
0 1
1
1
1
0
1 2 3 4 5 6 7 8 9 10 11 12
≥2 ≥2 ≥2
≥2 ≥2
0
1
0
1
0
1
0
1
0
1
Collision resolution Interval
FIGURE 5.3: Example of the Modiﬁed BinaryTree Protocol Operation
B
˜
0
B
˜
1
1 = =
B
˜
n i
1 B
˜
i
B
˜
n i –
+ + 1 i n ≤ ≤
1 B
˜
n
+ i 0 =
¹
¹
'
¹
¹
=
B
n
1 Q
0
n ( ) – Q
i
n ( ) Q
n i –
n ( ) + [ ]B
i
i 0 =
n 1 –
∑
+
1 Q
0
n ( ) – Q
n
n ( ) –
  = n 2 ≥
Section 5.2.: ENHANCED PROTOCOLS 125
The equation analogous to (5.12) is,
(5.37)
In this case, however, the B
n
are not minimized for p=1/2. Moreover, there is no single
value of p that minimizes all the B
n
. If we choose p=1/2, we can establish an upper bound
on B
n
of the form (5.13) with α
m
=2.664, while if we use p=0.4175, we can establish an
upper bound on B
n
with α
m
=2.623. This implies that when the modiﬁed binarytree proto
col is employed with fair coins, then the system is stable for arrival rates up to 1/2.664 ≅
0.375 while if biased coins are used, then the system is stable for arrival rates up to 1/
2.623 ≅ 0.381 which is higher than the maximal throughput for the slotted Aloha
protocol.
5.2.2. The Epoch Mechanism
From Table 1 we see that1/B
1
=1, 2/B
2
=0.4, 3/B
3
=0.3913 and when n is large (5.17) and
(5.24) imply that n/ B
n
≅ 0.346. The conclusion is that the binarytree protocol resolves
collisions among a small number of packets more efﬁciently than among a large number
of packets. When obvious access is employed, it is very likely that a CRI will start with a
collision among a large number of packets when the previous CRI was long. When the
systemoperates near its maximal throughput most CRIs are long, hence, collisions among
a large number of packets have to be resolved frequently, yielding non efﬁcient operation.
Ideally, if it were possible to start each CRI with the transmission of exactly one packet,
the throughput of the system would have been 1. Since this is not possible, one should try
to design the system so that in most cases a CRI starts with the transmission of about one
packet. There are several ways to achieve this goal by determining the ﬁrsttime transmis
sion rule, i.e., when packets are transmitted for the ﬁrst time. One way, suggested by
Capetanakis [Cap79], is to have an estimate on the number of packets that arrived in the
previous CRI and divide them into smaller groups, each having an expected number of
packets on the order of one and handling each group separately. Another way, known as
the epoch mechanism has been suggested by Gallager [Gal78] and by Tsybakov and
Mikhailov [TsM80], and is described next.
Consider the arrivals of packets to the system and divide the time axis into consecutive
epochs (called the arrival epochs), each of length ∆ slots (∆ is not necessarily an integer).
The ith arrival epoch is the time interval ( i∆,(i+1)∆ ). Packets that arrive during the ith
arrival epoch are transmitted for the ﬁrst time in the ﬁrst slot after the collision among
packets that arrived during the (i1)st arrival epoch is resolved. The parameter ∆ is chosen
to optimize the performance of the system. The operation of the epoch mechanism is illus
trated in Figure 5.4. On the channel we observe a sequence of collision resolution intervals
each corresponding to arrivals during some time interval on the arrival axis. If we number
these collision resolution intervals sequentially then in the ith CRI all packets (if any) that
B
n
1
n
k
¸ ,
¸ _
1 – ( )
k
k 1 p + ( ) 1 – p
k
– [ ]
1 p
k
– 1 p – ( )
k
– [ ]

k 2 =
n
∑
+ = n 2 ≥
e
1 –
126 CHAPTER5: COLLISIONRESOLUTION
arrive during the ith epoch are successfully transmitted. All the packets arriving in the 0th
epoch, i.e., in the period (0, ∆) are transmitted in the ﬁrst slot of CRI0; they collide, and a
resolution process starts (see Figure 5.4). In the meantime, newly arrived packets wait and
when CRI0 ends all packets belonging to the ﬁrst epoch, i.e., those arriving in the period
(∆, 2∆), are transmitted, and so on. An interesting phenomenon occurs at the end of CRI2
in our example, since the CRI2 ended before the third epoch. There are two options at this
point (corresponding to two different protocols): one could shorten the third arrival epoch
to match the end of CRI2 or, as is shown in the ﬁgure (and analyzed in this section), enter
a waiting period lasting from the end of CRI2 to the end of the third epoch. It turns out
that throughput in both cases is the same although the average delay in the latter method is
slightly higher (see Huang and Berger [HuB85].
When the epoch mechanism is employed as the ﬁrsttime transmission rule, it is possible
to describe the binarytree protocol via interval splitting. Whenever the transmission of
packets that arrive during some interval results in a collision, the interval is split in two.
The nodes that arrived in the left part correspond to users that would have ﬂipped 0 in the
binarytree protocol and the nodes that arrived in the right part correspond to users that
would have ﬂipped 1. In this way, it is guaranteed that packets are transmitted in the order
they arrive (FCFS). The example of Section 5.1. is depicted again in Figure 5.5 to demon
strate the interval splitting procedure. Slot number 1 is the ﬁrst slot of a CRI whose corre
sponding arrival epoch is the period (a,g) in the ﬁgure; thus all packets that arrived during
(a,g) are transmitted in slot 1. Since a collision occurred one needs to split the group but
instead of ﬂipping a coin we split the interval in two halves: all those users that arrived
during (a,d) behave as if they ﬂipped 0 while all those that arrived during (d,g) behave as if
they ﬂipped 1. This results in another collision in slot 2, requiring further splitting of the
interval (a,d) so that in slot 3 transmit those users that arrived during (a,b)none in our
example. This causes packets that arrived during (b,d) to be transmitted in slot 4, and so
on. Note that the probability p that a user ﬂips 0 corresponds here to the interval splitting
FIGURE 5.4: Example of the Epoch Mechanism
¹ ¹ ; ¹ ¹ ¹ ¹ ; ¹ ¹ ¹ ¹ ; ¹ ¹ ¹ ¹ ; ¹ ¹ ¹ ¹ ; ¹ ¹
0 ∆ 2∆ 3∆ 4∆ 5∆
CRI0 CRI1 CRI2 CRI3 CRI4
Arrival
Axis
Channel
Axis
}
Waiting
Period
Section 5.2.: ENHANCED PROTOCOLS 127
ratio. Thus, when p=1/2 the interval is halved, and when p=0.3 the left part of the split
interval is 0.3 times the length of the original interval.
We now turn to evaluate the performance of this protocol. Since arrival process is Poisson
the arrival points in every given interval are uniformly distributed. Thus, splitting an inter
val is completely equivalent to ﬂipping a coin. This means that the values B
n
and V
n
are
identical to those of the binarytree CRP. The main difference lies in the region of stable
throughput.
When the epoch mechanismis employed as the ﬁrsttime transmission rule, there is no sta
tistical dependencies among the corresponding collision resolution intervals. If
denotes the number of new packets that are transmitted at the beginning of the kth CRI,
then , , , ... is a sequence of independent and identically distributed (i.i.d.)
randomvariables. Since the length of the kth CRI is completely determined by the number
of packets transmitted in its ﬁrst slot, we conclude that the sequence , , , ...
is also a sequence of independent and identically distributed random variables. Let and
denote an arbitrary pair and . Since the arrival process is Poisson, we have:
Also,
0 1
1
1
1
0
1 2 3 4 5 6 7 8 9 10 11 12
≥2 ≥2 ≥2
≥2 ≥2 ≥2
(a,g) (a,d) (a,b) (b,d) (b,c) (c,d) (d,g) (d,f) (d,e) (e,f) (f,g) (g,k)
FIGURE 5.5: IntervalSplitting Procedure for the Epoch Mechanism
Channel
Axis
Collision resolution Interval (CRI) CRI
a
b c d e f
g
h i j
k
∆ ∆
A
˜
k ( )
A
˜
0 ( ) A
˜
1 ( ) A
˜
2 ( )
B
˜
0 ( ) B
˜
1 ( ) B
˜
2 ( )
A
˜
B
˜
A
˜
k ( ) B
˜
k ( )
Prob A
˜
n = [ ]
λ∆ ( )
n
e
λ∆ –
n!
 =
128 CHAPTER5: COLLISIONRESOLUTION
from which we obtain,
where B
n
and V
n
are given by (5.6) and (5.11), respectively.
The system can be viewed as a (discretetime) queueing system in which packets that
arrive during the interval (i ∆, (i+1)∆) are served in the ith CRI. The total service time of
these arrivals has a ﬁrst and a second moment B and , respectively. In order for the
“server” not to fall behind the arrivals we need that
(5.38)
in other words, the time it takes, on the average, to successfully transmit all packets that
arrive in a period of duration ∆ must be less than ∆. The quantity B∆ is the expected
change in the time backlog of the system, namely the expected change in the difference
between the current time and the time of the last epoch whose packets were transmitted
successfully. When condition (5.38) holds, the system is stable, and if is ﬁnite, the
expected delay of a packet is ﬁnite.
Condition (5.38) can be written as:
and rewritten as:
(5.39)
where . The function f(z) is depicted in Figure 5.6 for various values of p (recall
that p is the probability that a user will ﬂip 0 and it is equivalent to the splitting ratio of the
interval when a collision is observed). For p=1/2 the function f(z) is maximized for
z
*
=1.15 and the maximum value is 0.429. Hence, the system is stable for arrival rates
λ<0.429. The maximal value of f(z) is smaller for other values of p. The fact that
Prob B
˜
i = [ ] Prob B
˜
n
i = A
˜
n = [ ]Prob A
˜
n = [ ]
n 0 =
∞
∑
=
Prob B
˜
n
i = [ ]Prob A
˜
n = [ ]
n 0 =
∞
∑
= i 1 ≥
B E B
˜
[ ] B
n
Prob A
˜
n = [ ]
n 0 =
∞
∑
= = B
2
E B
˜ 2
[ ] V
n
Prob A
˜
n = [ ]
n 0 =
∞
∑
= =
B
2
B ∆ <
B
2
B
n
λ∆ ( )
n
e
λ∆ –
n!

n 0 =
∞
∑
∆ <
λ
λ∆
B
n
λ∆ ( )
n
e
λ∆ –
n!

n 0 =
∞
∑
 <
z
B
n
z
n
e
z –
n!

n 0 =
∞
∑
 f z ( )
·
∆
=
z λ∆
·
∆
Section 5.2.: ENHANCED PROTOCOLS 129
z
*
=λ∆
*
=1.15 is not surprising. It conforms with the intuition that most of the CRIs should
start with the transmission of a single packet. The fact that z
*
is slightly higher than 1 is
due to the waste incurred by idle slots. The length ∆
*
that should be chosen to obtain the
maximal throughput is ∆
*
=1.15/0.429=2.68. FromFigure 5.6 we observe that the function
f(z) is not very sensitive to small changes in z, especially for values of z larger than z
*
. The
conclusion is that slightly longer epochs (slightly larger ∆) will not cause the maximal
achievable throughput to deteriorate by a large amount.
The description and the analysis above corresponds to use of the epoch mechanism and
resolving collisions as done in the binarytree protocol. When the epoch mechanism is
used with the modiﬁed binarytree protocol resolution method the analysis is the same,
except that in (5.39) one should use the values of B
n
that correspond to this protocol (equa
tion (5.37). The results are as follows: When p=1/2 the system is stable for λ<0.462 and
when p=0.4175 the system is stable for input rates up to λ<0.468.
5.2.3. The Clipped BinaryTree Protocol
In the previous analysis we made the point that to improve performance a CRI should start
with the transmission of about one packet. However, that idea was not fully exploited by
the above enhancements. To see why, assume that the epoch mechanism is employed and
consider what happens when a collision is followed by another collision (see slots 1 and 2
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0 1 2 3 4 5
f
(
z
)
z = λ ∆
p=0.5
p=0.6, 0.4
P=0.7, 0.3
FIGURE 5.6: Determining Permissible Arrival Rates for the Epoch Mechanism
130 CHAPTER5: COLLISIONRESOLUTION
in Figure 5.5. The collision in the ﬁrst slot is among the packets that arrived during an
interval of length ∆. The users whose packets collided are divided into two groups by split
ting the interval into two parts (for explanation purposes we assume p=1/2 so the length of
either of the two parts equals ∆/2). The packets that arrived in the left part are then trans
mitted in slot 2 and collide again. By repeated interval splitting all packets in the left part
are eventually successfully transmitted (slot 6 in our example). At this point the packets
that arrived during the right part of the original interval are transmitted. It is at this point
that we have not done as best we can. The underlying observation is that there is no infor
mation regarding the number of packets in the part we attempt to resolve, yet the expected
number of arrivals during that period is different from the desired quantity, namely,
slightly larger than one. Indeed, if the distribution of the number of packets in that part is
as that of new packets i.e., Poisson with parameter λ, then the transmission of packets in
that part is identical to starting a CRI with half the optimal expected number of packets. To
remedy this, it would be better to choose a new interval of length ∆ and let packets that
arrived in this interval transmit. Protocols based on this observation have been suggested
by Gallager [Gal80] and Tsybakov and Mikhailov [TsM80].
To incorporate this strategy into the protocol we adopt the rule that whenever a collision is
followed by two successive successful transmissions, a new epoch of length ∆ from the
arrival axis is enabled, that is, the packets that arrived in that interval are transmitted. The
protocol that results from such an operation is called the clipped binarytree protocol,
since part of the binarytree is clipped and not enabled (see Figure 5.7). One might add
that the clipping idea can be used with the original binarytree protocol (Section 5.1.) as
well as the modiﬁed one (Section 5.2.1.), i.e., the one that avoids transmission of packets
that are guaranteed to collide (the latter is called the modiﬁed clipped binary tree proto
col). The example of Section 5.1. is depicted in Figure 5.7 when the clipped binarytree
protocol is employed. Note that in slot 7 a new CRI is started, corresponding to the arrival
epoch (d,i), rather than enabling the interval (d,f) as would be the case in the regular epoch
mechanism. In the modiﬁed clipped binarytree protocol the collision in slot 3 will be
avoided (skipped) as was the case in the example depicted in Figure 5.3.
The argument that leads to the enhancement described above is based on the assertion that
the distribution of the number of packets in the right part of the interval is the same as that
of the newly arrived ones i.e., those that arrived in the part of the arrival axis that was
never explored. This assertion is not as trivial as it ﬁrst appears, since after receiving the
feedback indicating that a collision took place, it is not clear that the distribution of the
right part remains Poisson with the same parameter as before. To show this property, let
be the number of packets involved in the ﬁrst collision and let and be the number of
packets in the left part and the right part of the interval, respectively. We need to show that
if it is known that and then the distribution of is Poisson. We
compute
x
˜
x
˜
l
x
˜
r
x
˜
x
˜
l
x
˜
r
+ 2 ≥ = x
˜
l
2 ≥ x
˜
r
Section 5.2.: ENHANCED PROTOCOLS 131
where we used the fact that and are independent, a result stemming from the known
property of the Poisson process that the number of events in nonoverlapping intervals are
independent. Generally, receiving the collision feedback, results in a nonPoisson distribu
tion of the number of colliding packets; but, receiving the additional feedback of a colli
sion in the left part of the interval means that the number of packets that arrived in the
right part is greater or equal to zeroinformation we had to start withmeaning that we
have the same Poisson distribution.
The analysis of the clipped binarytree protocol is similar to that of the binarytree proto
col. We present the analysis for the modiﬁed clipped binarytree (see Section 5.2.1.) in
which deﬁnite collisions, i.e., those that can be predicted beforehand, are avoided. Denote,
as before, by the length of the collision resolution interval that starts with a collision of
n packets and by the length of the collision resolution interval that starts with a colli
sion of n packets, i of which arrived in its left part, and by B
n
and their respective
expected values. For n=0 and n=1 we have
1
1
≥2
0 1
1
1 2 3 4 5 6 7 8 9 10 11 12
≥2 ≥2 ≥2
≥2 ≥2 0
(a,g) (a,d) (a,b) (b,d) (b,c) (c,d) (d,i) (d,g) (d,f) (d,e) (e,f) (f,j)
Channel
Axis
CRI
CRI
a
b c d e f
g
h i j
k
∆
∆
∆
Arrival
Axis
CRI
FIGURE 5.7: Example of the ClippedBinaryTree Protocol Operation
Prob x
˜
r
i = x
˜
l
x
˜
r
+ 2 ≥ x
˜
l
2 ≥ , [ ] Prob x
˜
r
i = x
˜
r
0 ≥ x
˜
l
2 ≥ , [ ] =
Prob x
˜
r
i = x
˜
l
2 ≥ [ ] =
λ∆ 2 ⁄ ( )
i
e
λ∆ 2 ⁄ –
i!
 =
x
˜
r
x
˜
l
B
˜
n
B
˜
n i
B
˜
n i
B
˜
0
B
˜
1
1 = =
132 CHAPTER5: COLLISIONRESOLUTION
Consider a CRI that starts with a collision of n (n≥2) packets, i of which arrived in the left
part and ni in the right part (0≤i≤n). If i=0 then the length of the CRI will include the
original collision slot followed by the need to still resolve all n packets. If i=1 then the CRI
includes the original collision slot, the successful slot for the left part, and then the number
of slots needed to resolve the remaining n1 packets. Finally, when 2≤i≤n the CRI
includes the original collision slot and then the number of slots needed to resolve the i
packets that arrived in the left part. Note that in the latter case (that is unique to the clipped
tree protocol) the resolution of the remaining ni packets is not a part of the current CRI.
The conditional length of a CRI for the clipped binarytree protocol can therefore be sum
marized by:
and correspondingly the expected values are
(5.40)
Denoting, as before, by Q
i
(n) the probability that in an interval containing n arrivals i
occurred in its left part, we can write, based on equation (5.40), an expression for B
n
:
or
(5.41)
The quantity B
n
can be computed recursively from (5.41) with the initial values B
0
=B
1
=1.
B
˜
n i
1 B
˜
i
+ 2 i n ≤ ≤
2 B
˜
n 1 –
+ i 1 =
1 B
˜
n
+ i 0 =
¹
¹
¹
'
¹
¹
¹
=
B
n i
1 B
i
+ 2 i n ≤ ≤
2 B
n 1 –
+ i 1 =
1 B
n
+ i 0 =
¹
¹
¹
'
¹
¹
¹
=
B
n
E B
˜
n i
[ ] B
n i
Q
i
n ( )
i 0 =
n
∑
= =
1 Q
0
n ( )B
n
Q
1
n ( ) 1 B
n 1 –
+ ( ) Q
i
n ( )B
i
i 2 =
n
∑
+ + + = n 2 ≥
B
n
1 Q
1
n ( ) 1 B
n 1 –
+ ( ) Q
i
n ( )B
i
i 2 =
n 1 –
∑
+ +
1 Q
0
n ( ) – Q
n
n ( ) –
  = n 2 ≥
Section 5.2.: ENHANCED PROTOCOLS 133
With the clipped binarytree protocol not all packets that collide in the ﬁrst slot of a CRI
are successfully transmitted during that CRI (since part of the tree is clipped). For
Instance, in the example depicted in Figure 5.7 during the CRI that starts with a collision
among four packets, only two packets are transmitted successfully (the rest will be part of
the next CRI). To evaluate the performance of the protocol, one must compute the rate of
successful transmissions during a CRI. To that end let be a random variable represent
ing the number of packets that are successfully transmitted during a CRI given that it
started with the transmission of n packets and be the same variable conditioned on
having i packets in the left part, and by U
n
and their respective expected values. For
n=0 and n=1 we have
and for n≥2, similarly to equation (5.40), we have
(5.42)
Leading to
or
(5.43)
and U
n
can be computed recursively from (5.43) with the initial values U
0
=0 and U
1
=1.
Values of B
n
and U
n
are given in Table 3 for p=1/2. It is interesting to note the slow growth
of B
n
with n compared to the linear growth of B
n
with n when the binarytree protocol is
employed (see Table 1). In addition, observe that the expected number of packets transmit
ted successfully during a CRI is almost a constant for n≥3.
The expected length of a CRI, B, and the expected number of packets that are successfully
transmitted in a CRI, U, are given by
U
˜
n
U
˜
n i
U
˜
n i
U
0
0 = U
1
1 =
U
n i
U
i
2 i n ≤ ≤
1 U
n 1 –
+ i 1 =
U
n
i 0 =
¹
¹
¹
'
¹
¹
¹
=
U
n
Q
0
n ( )U
n
Q
1
n ( ) 1 U
n 1 –
+ ( ) Q
i
n ( )U
i
i 2 =
n
∑
+ + = n 2 ≥
U
n
Q
1
n ( ) 1 U
n 1 –
+ ( ) Q
i
n ( )U
i
i 2 =
n 1 –
∑
+
1 Q
0
n ( ) – Q
n
n ( ) –
 = n 2 ≥
134 CHAPTER5: COLLISIONRESOLUTION
(5.44)
The expected number of packets transmitted in the ﬁrst slot of a CRI is λ∆ and therefore
the fraction of packets successfully transmitted during a CRI is U/λ∆. The Poisson process
has an interesting property that given a number of arrivals in an interval, the arrival points
are uniformly distributed in the interval. Consequently, if a fraction U/λ∆ of the packets
are successfully transmitted it means that U/λ∆ is also the fraction of the interval resolved.
Hence, (U/ λ∆)∆=U/ λ is, on the average, the portion of the resolved interval. On the other
hand, on the average, it takes B slots to resolve a collision. Thus, for the system to remain
stable, it must be able to resolve collisions at least at the rate in which time progresses or,
which, upon substitution of equation (5.44) leads to
(5.45)
where we have substituted . The right handside is a function of z (with parameter
p). For a given p this function can be plotted (the curves obtained are similar to those in
Table 3: The ﬁrst moment of and for p=1/2.)
n 1 2 3 4 5
B
n
1.0000 4.0000 5.8333 6.4762 6.6698
U
n
1.0000 2.000 2.5000 2.5714 2.5238
n 6 7 8 9 10
B
n
6.8363 7.0286 7.2180 7.3894 7.5406
U
n
2.4977 2.4958 2.5008 2.5052 2.5075
n 11 12 13 14 15
B
n
7.6741 7.7937 7.9027 8.0035 8.0980
U
n
2.5079 2.5073 2.5064 2.5055 2.5049
B
˜
n
U
˜
n
B E B
n
[ ] B
n
λ∆ ( )
n
e
λ∆ –
n!

n 0 =
∞
∑
= =
U E U
n
[ ] U
n
λ∆ ( )
n
e
λ∆ –
n!

n 0 =
∞
∑
= =
B
U
λ
 <
λ
U
n
λ∆ ( )
n
e
λ∆ –
n!

n 0 =
∞
∑
B
n
λ∆ ( )
n
e
λ∆ –
n!

n 0 =
∞
∑
 <
U
n
z
n
e
z –
n!

n 0 =
∞
∑
B
n
z
n
e
z –
n!

n 0 =
∞
∑
 =
z λ∆
·
∆
Section 5.3.: LIMITED SENSING PROTOCOLS 135
Figure 5.6) and its maximal value found. For p=1/2 the right side of (5.45) is maximized
for z=1.26 and the maximum value is 0.487 packets/sec. Thus, the system is stable for
arrival rates λ<0.487 and the length ∆ that should be chosen is ∆=1.26/0.487=2.60 slots.
When the splitting probability p is optimized it is possible to slightly increase the through
put to 0.4877 as was demonstrated by Mosley and Humblet [MoH85].
The analysis presented above was carried out for the modiﬁed clipped binarytree proto
col, i.e., avoidable collisions are eliminated. When the modiﬁcation is not used, the only
change in the analysis is the addition of Q
0
(n) to the numerator of equation (5.41), and the
resulting maximal throughput is 0.449.
5.3. LIMITED SENSING PROTOCOLS
The protocols described so far require every user to monitor the channel feedback at all
times, even if that user has no packet to transmit. This is necessary because a newly gener
ated packet can be transmitted for the ﬁrst time only at the end of the current CRI, requir
ing every user to positively determine the end of the CRI. This kind of feedback
monitoring is known as fullsensing. This mode of operation is quite impractical because a
user that crashed can never again join the systemand, furthermore, a user that due to some
fault did not receive properly all signals may actually disturb the others and decrease the
efﬁciency of the protocol. Thus, it is desirable that users monitor the feedback signals only
during limited periods, preferably after having generated a packet for transmission and
until the packet is transmitted successfully. This kind of monitoring is known as limited
sensing, and protocols with such monitoring are referred to as protocols operating in a
limited sensing environment.
Several collision resolution protocols have been devised for the limitedsensing environ
ment. The simplest, albeit not the most efﬁcient one, is the freeaccess protocol analyzed
by Mathys and Flajolet [MaF83] and by Fayolle et. al. [FFH85]. In this protocol, new
packets are transmitted as soon as possible, i.e., at the beginning of the slot subsequent to
their arrival time; thereafter, a user that transmitted a new packet monitors the feedback
signals and continues operating as if he were an “old” user. This protocol works in con
junction with both the basic binarytree protocol (see Section 5.1.) or the modiﬁed binary
tree protocol (see Section 5.2.1.). The maximal throughput of this limited sensing protocol
when the modiﬁed binarytree protocol is employed is 0.360.
To date, the most efﬁcient protocol for a limitedsensing environment is the one intro
duced by Humblet [Humb86] and Georgiadis and PapantoniKazakos [GeK87] which is
essentially an adaptation of the (fullsensing) modiﬁed clipped binarytree protocol
described in Section 5.2.3.. As mentioned earlier, a crucial feature required for the correct
operation of the full sensing protocols is the ability of all users to determine the end of a
CRI. This is necessary so that users with new packets know exactly when they may trans
mit for the ﬁrst time in a manner that would not interfere with an ongoing resolution of a
collision. The major change involved in the adaptation of the clipped binarytree protocol
136 CHAPTER5: COLLISIONRESOLUTION
to the limited sensing environment is therefore controlling the extent of CRIs so that their
end remains uniquely and easily detectable. As before, newly generated packets are not
considered for transmission until the ongoing collision resolution is done (and all waiting
users receive sufﬁcient feedback to detect the end of a CRI). A collision feedback obvi
ously indicates that a CRI is in progress; the difﬁculty is to realize whether or not a colli
sion resolution is in progress when a series of idle and successful slots is observed, and
then to detect the end of a CRI.
Going back to the clipped binarytree protocol described in Section 5.2.3., we recall that
the end of a CRI that started with a collision is characterized by two consecutive success
ful slots. In addition, there are CRIs that consist of a single slot, either an idle one or a suc
cessful one. Thus, if we have a successful slot followed by either an idle slot or another
successful slot, we are assured that a CRI just ended. We take this event to be the endof
CRI marker, that is, a successful slot followed immediately by an idle slot or another suc
cessful slot denotes the end of the CRI. Such marking is, however, not enough. Observing
the channel once the system becomes idle with no user having a new packet to transmit,
reveals a sequence of idle slots. A user generating a new packet at this state will wait for
ever for the endofCRI marker. To overcome this potential deadlock, the endofCRI
marker is augmented to include an event consisting of R+1 idle slots, where R is a globally
known protocol parameter. Thus, a user that observes R+1 consecutive idle slots con
cludes that a CRI ended. This indication of the end of a CRI is correct only if R+1 consec
utive idle slots never occur during an actual resolution of a collision. Yet, in the modiﬁed
clipped binarytree protocol, such event is possible since the protocol dictates to avoid def
inite collisions, i.e., those that are guaranteed to occur. Consequently, in the limited sens
ing environment, we avoid deﬁnite collisions at most R1 times in succession. Users that
participate in a collision resolution and observe R consecutive idle slots, retransmit in the
next slot to cause a collision, and continue regularly thereafter. In summary, a user that
generates a new packet at some slot will be able to decide whether or not a CRI is in
progress within at most R+1 slots and if he ﬁnds out that a CRI is in progress, he will be
able to determine its end.
There is a clear performance tradeoff in the choice of R. A large value for R may cause a
packet arriving to an idle system to wait quite long before it can determine that the system
is idle, but it will eliminate more deﬁnite collisions. A small value for R requires fairly
often to retransmit a packet unnecessarily, only to announce an ongoing CRI, but the delay
incurred by a new packet upon arrival to an idle system will be small. Observe that the
case R=1 corresponds to elimination of the modiﬁcation introduced in Section 5.2.1.,
while when R is very large, the protocol is similar to the clipped binarytree protocol.
To complete the speciﬁcation of the protocol, we must deﬁne the behavior of the users
involved in the resolution process. One of the rules has already been deﬁned, and requires
a participating user to retransmit his packet after having sensed R consecutive idle slots.
The other rules deﬁne the exact transmission schedules and are similar to the rules of the
clipped binarytree protocol, namely, upon detection of the end of a CRI, a new epoch is
Section 5.3.: LIMITED SENSING PROTOCOLS 137
chosen and packets that arrived during this epoch are transmitted; upon a collision the
epoch is split in two, and so forth. To describe how new epochs are chosen and how
epochs are split we note that at each point in time all users having packets awaiting trans
mission, both the newly arrived and the old, untransmitted, ones can be divided into three
classes. Class A contains those users that cannot decide thus far whether or not some CRI
is in progress. Class B contains those users that deﬁnitely know a CRI is in progress (they
heard a collision feedback) but do not know when it started. Class C contains those users
that know a CRI is in progress as well as the time it started. While all users in class C can
simultaneously decide which epoch will be chosen for transmission, those in classes A and
B cannot. The beginning of the CRI is not known to the users since they start monitoring
the channel only upon a packet arrival meaning that the only common time reference is the
end of the CRI (which is known to all class C users). Since it is desirable to select an initial
epoch whose length is optimal we select it so that its end coincides with the latest class C
slot. In other words, the protocol in the limited sensing environment evolves in such a way
that the epochs for transmission are chosen so that the most recent arrivals belonging to
class C attempt transmission ﬁrst. In this sense the protocol is a lastcome ﬁrstserved pro
tocol.
Figure 5.8 presents seven snapshots, taken at seven consecutive slots, of a system with
R=2. Each snapshot is taken at the time marked CT (Current Time) and depicts the time
axis starting at some time t
0
, a moment for which all previously arrived packets have
already left. The uparrows indicate arrival instants of packets, the numbers are packet
numbers used for explanation, and the feedback for the slot starting at CT is shown above
the axis (`0’=idle, `1’=success, ` ≥2’=collision). Such a diagram is often called the arrival
timeaxis diagram.
Figure 5.8(a) depicts the initial situation and shows the arrival times of the various classes.
All but those users that generated packets in the last slot before CT belong to class C while
those that arrived in the last slot belong to class A since they have not sensed a collision
and cannot decide, based on a single slot, whether some activity is going on. At this time,
a portion of the rightmost part of the class C time is enabled namely, packets that arrive
during this transmission interval (marked TI in the ﬁgure) are transmitted, and a collision
between packets 2 and 3 occurs. Users that previously belonged to class A have now
sensed a collision and therefore belong to class B. The same applies to packet 4 that
arrived during the most recent slot. This is depicted in snapshot (b).
At this point in time the rightmost half of the previous TI is enabled, and since it contains
no packets an idle slot occurs and therefore it is concluded that the left part contains at
least two packets and its rightmost half is enabled, as illustrated in Figure 5.8(c). Another
idle slot occurs and it is concluded that the corresponding left part contains at least two
packets. However, having sensed two idle slots in a system with R=2 the entire left part of
the previous TI is enabled, causing a (deﬁnite) collision as is illustrated in Figure 5.8(d).
Note that the two idle slots, depicted in snapshots (b) and (c) do not increase the extent of
class B users since two idle slots leave all users arriving during these slots in uncertainty
138 CHAPTER5: COLLISIONRESOLUTION
FIGURE 5.8: Enabled Intervals in Limited Sensing
A
C
TI
C
B
TI
C
B C
A B
TI
C
A
B
TI
C
TI
C
A
TI
C
1 2 3 4
t
0
1 2 3 4
t
0
1 2 3 4
t
0
1 2 3 4
t
0
1 2 3 4
t
0
1 2 4
t
0
1 4
t
0
CT
≥2
B
0
CT
CT
0
≥2
TI
1
CT
CT
A CT
1
TI
CT
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Section 5.3.: LIMITED SENSING PROTOCOLS 139
they have not sensed a collision and cannot determine whether or not the system is idle.
Having sensed a collision (snapshot (d)) all class A users become class B users.
After the collision of snapshot (d), operation continues as before. The previously enabled
interval is halved and packet 3 is transmitted successfully (snapshot (e)) after which
packet 2 is transmitted successfully (snapshot (f)). At this point two consecutive success
ful slots took place and an endofCRI marker is detected. All users but those arriving in
the most recent slot join class C while those arriving in the most recent slot join class B
(having sensed a single success they know the system is not idle). Note that the extent of
time covered by class C users is not contiguousit is separated by a period for which all
arrivals, if any, where already transmitted successfully.
Snapshot (g) illustrates the start of the next CRI. It starts by enabling a transmission inter
val that contains a portion of the time covered by class C users which consists of a small
interval after t
0
and the entire period during which the most recent CRI took place. Note
that in general, intervals corresponding to different classes do not overlap, and whenever
an interval is resolved, class B is empty. As we have seen the subset of the arrival axis that
contains packets from class C may consist of disjoint subintervals, but these will eventu
ally be resolved if the system is to be stable.
5.3.1. Throughput Analysis
The analysis of the above protocol is similar to the analysis presented in Section 5.2.3. for
the clipped binarytree protocol, except that one must accommodate the rule that R+1 slots
will not appear within a CRI. We ﬁrst derive the evolution of the length of a CRI that
starts with n users. For the cases n=0 and n=1 there is no change and hence
.
A CRI that starts with n≥2 packets, starts with a collision slot followed by 0 or more idle
slots (belonging to the rightmost subintervals without arrivals) after which a nonidle slot
must occur. The number of transmitting users in this slot can be any number between 1
and n, depending on the speciﬁc arrival times. Given that a CRI starts with a collision of n
(n≥2) packets, followed by exactly l consecutive idle slots, and having i packets in the
interval that is enabled after the l idle slots, then the length of a CRI for the above protocol
is
(5.46)
B
˜
n
B
˜
0
B
˜
1
1 = =
B
˜
n i
1 l B
˜
i
+ + 2 i n ≤ ≤ 0 l R 1 – ≤ ≤ ,
2 l B
˜
n 1 –
+ + i 1 0 l R 1 – ≤ ≤ , =
1 R B
˜
n
+ + l R =
¹
¹
¹
'
¹
¹
¹
=
140 CHAPTER5: COLLISIONRESOLUTION
where is the length of a CRI that started with the transmission of n packets given that
there were l consecutive idle slots at the beginning of a CRI, and when l < R, i users
belonged to the right part of the interval. These equations are similar to those of the modi
ﬁed clipped binarytree protocol with the exception that after R idle slots the original n
packets are retransmitted and collide, as is indicated by the third component of equation
(5.46). Recalling that Q
in
is the probability of i arrivals occurring during the right portion
of the an interval containing n arrivals (see equation (5.2)) we obtain from (5.46) in the
same manner as before
or after some algebra we have for n≥2 that
and B
n
is computed recursively from the above equation with the initial values B
0
= B
1
=1.
The derivation of the expected number of packets transmitted successfully during a CRI is
identical to that presented for the clipped binarytree (see Section 5.2.3. equations (5.42)
and (5.41)). The maximal throughput of the protocol is given by equation (5.44). When
R=1 the maximal throughput is 0.449the same throughput obtained for the clipped
binarytree protocol (without the modiﬁcation). When R is large (for that matter R=5 is
already large enough), the maximal throughput is 0.487, the same as the throughput for the
fullsensing environment.
5.4. RELATED ANALYSIS
Numerous variations of the environment under which collision resolution protocols oper
ate have been addressed in the literature and excellent surveys on the subject have been
written by Gallager [Gal85] and Tsybakov [Tsy85]. Chapter 4 in Bertsekas and Gallager’s
book [BeG87] is also an excellent source on collision resolution protocols. We considered
but a very small number of these variations: slotted time, inﬁnite population, Poisson
arrivals and reliable ternary feedback. In the following we list a few of the other variations.
Bounds on throughput
Considerable effort has been spent on ﬁnding upper bounds to the maximum throughput
that can be achieved in an inﬁnite population model with Poisson arrivals and ternary feed
B
˜
n i
B
n
1 Q
0
n ( ) [ ]
R
R B
n
+ ( ) + =
Q
0
n ( ) [ ]
l
Q
1
n ( ) 1 l B
n i –
+ + ( ) Q
i
n ( ) l B
i
+ ( )
i 2 =
n
∑
+
l 0 =
R 1 –
∑
= n 2 ≥
B
n
1 Q
0
n ( ) – 1 Q
0
n ( ) [ ]
R
– [ ] Q
0
n ( ) Q
1
n ( ) 1 B
n i –
+ ( ) Q
i
n ( )B
i
i 2 =
n 1 –
∑
+ + +
1 Q
0
n ( ) [ ]
R
– [ ] 1 Q
0
n ( ) – Q
n
n ( ) – [ ]
 =
Section 5.4.: RELATED ANALYSIS 141
back [Pip81, Mol82, MiT81, CrH82, TML83, BeZ88]. The best upper bound known to
date is 0.568 and is due to Tsybakov and Likhanov [TsL88].
Feedback types
In the text we considered mainly the ternary feedback distinguishing among idle, success
ful and collision slots, and the binary feedback that informs the users only whether or not
there was a conﬂict in the slot. There are two other binary feedback types that can be con
sidered: (i) Something/Nothing feedback that informs the users whether or not a slot was
idle; (ii) Success/Failure feedback that informs the users whether or not a slot contained
exactly one packet. Collision resolution protocols for these binary feedback channels have
been studied in [MeB84].
In some cases it might be possible to increase the amount of feedback detail by using an
extra control channel for reservation [HuB85, ToV87], or by indicating the exact number
of users involved in a collision. The latter information can be obtained by using energy or
power detectors and this kind of feedback is termed known multiplicity feedback [Tsy80]
and [GeP82].
When a packet is transmitted successfully it is possible to use the contents of the packet in
the feedback in order to improve the performance of the CRP. Aprotocol that uses an extra
bit that can be read only when a packet is transmitted successfully is presented in [KeS88].
Multiple access protocols without feedback have been considered in [TsL83, MaM85].
Noise errors, erasures and captures
Practical multipleaccess communication systems are prone to various types of errors. The
most common are the noise errors that are intrinsic in any physical radio channel. Such
errors cause the feedback to indicate a collision although no user or a single user transmit
ted. Another type of errors are erasures. These errors correspond to situations in which
one or several nodes are transmitting, but the feedback detected by the users indicates that
the slot was idle. Reasons for erasures in practical systems are mobile users that may occa
sionally be hidden (for example, because of physical obstacles) or because of fading prob
lems. The capture phenomenon can also be considered as erroneous operation of the
system, corresponding to the ability to receive a packet successfully although more than
one packet is transmitted at the same time. Collision resolution protocols that operate in
presence of noise errors, erasures and captures have been studied in [VvT83, SiC85,
CiS85, CiS87, CKS88].
Group testing
Group testing, a branch of applied statistics, addresses the problem of classifying items of
some population as either defective or nondefective. It has been discovered that group
142 CHAPTER5: COLLISIONRESOLUTION
testing ideas and algorithms can be applied in the design of protocols, similar to the colli
sion resolution protocols, for random access communication. The basic idea is that a
defective item in the group testing problem corresponds to an active user in the communi
cation problem and a nondefective item corresponds to an idle user. Collision resolution
protocols based on group testing ideas have been developed in [BMT84] for a homoge
neous population of users and in [KuS88] for a nonhomogeneous population of users.
General arrival processes
In the analysis of the collision resolution protocols it has been assumed that the arrival
process of newpackets to the systemis a stationary Poisson process. Furthermore, some of
the parameters of the protocols were carefully tuned, based on the Poisson assumption, to
yield the best performance (the epoch length ∆, for instance). It is not difﬁcult to realize,
though, that the performance of some of the algorithms is not sensitive to the speciﬁc
arrival process. For instance, the maximal throughput of the basic binarytree protocol
with the obvious access scheme is 0.346, independently of the arrival process. Collision
resolution protocols yielding high throughputs for general arrival processes (even if their
statistics are unknown) were developed in [GFL87] and [CiS88].
Delay analysis
Several delay analyses of collision resolution protocols appear in the literature. The
expected packet delay of the binarytree protocol has been derived in [TsM78, MaF85,
FFH85]. Bounds on the expected packet delay of the clipped binarytree with the epoch
mechanism have been obtained in [TsM80, GMP87]. Other variations of delay analysis
appear in [HuB85, PMV87].
Section : EXERCISES 143
EXERCISES
Problem 1.
For the binarytree algorithm we found that (see (5.12))
For p =1/2 prove:
1.
2.
Problem 2.
Prove equation (5.37) and determine the corresponding equation for V
n
in this case.
Problem 3. (Noise errors)
Assume that the binarytree protocol with the epoch mechanismis employed. Assume that
the channel is noisy, so that an idle slot is interpreted as a collision with probability π
0,C
and a slot that contains a single transmission is interpreted as a collision with probability
π
1,C
(in the latter event the user that transmitted has to retransmit his packet again accord
ing to the protocol). All error events are independent of each other and of the system state.
1. Compute B
0
, B
1
and write recursive relations for computing B
n
, for this system.
What are the conditions on the error probabilities to insure that these quantities are
ﬁnite?
2. Write an expression for computing the throughput.
3. Let
Prove that


 The optimal z that maximizes the throughput (z
*
) does not depend on π
1,C
. How
would you explain this phenomenon?
4. For π
0,C
= π
1,C
= 1/4 determine z
*
and the maximal throughput.
B
n
1
n
k ¸ ,
¸ _
2k 1 – ( ) 1 – ( )
k
1 p
k
– 1 p – ( )
k
–

k 2 =
n
∑
+ = n 2 ≥
B
n
3
n
k ¸ ,
¸ _
2k 1 – ( ) 1 – ( )
k
2
k 1 –
1 –

k 2 =
n
∑
+ = n 2 ≥
B
n
B
n 1 –
2 n 1 – ( ) 2
k –
1 2
k –
– ( )
n 2 –
k 1 =
∞
∑
+ = n 3 ≥
n 2 ≥
B z ( ) B
n
z
n
e
z –
n!

n 0 =
∞
∑
·
∆
B 0 ( ) B
0
=
B z ( ) d
z d

z 0 =
B
1
B
0
– =
B z ( ) 2B z 2 ⁄ ( ) – 1 1 z + ( ) 1 B
0
+ ( )e
z –
– =
144 CHAPTER5: COLLISIONRESOLUTION
Problem 4. (Power capture [CKS88])
This problem deals with the capture phenomenon employing a model similar to that of
Problem 3.4. As explained there a capture means receiving correctly a packet even when
other packets are transmitted during the same time. The model used is typical to power
capture, namely, that some nodes transmit with higher power than others.
When users are transmitting in a slot, a capture occurs with probability π
n,1
, namely,
one of the n transmitted packets is successfully received and the other n1 packets have to
be retransmitted.
Assume that the users are executing the binarytree protocol with the epoch mechanism.
Also assume that when a transmission succeeds, the users are informed which packet was
received correctly. The latter implies that the users transmitting during a slot that con
tained a capture, know of that event (but other users are not aware that a capture occurred).
We therefore assume that in case of success all users (including those that transmitted and
failed in that slot) behave as if the slot was successful. The question then is when will the
users whose packets failed due to a capture retransmit their packets. Consider the follow
ing alternatives: (i) They transmit immediately in the subsequent slot after the capture
occurs (persist scheme). (ii) They wait until the current CRI ends and retransmit in the ﬁrst
slot of the subsequent CRI (wait scheme).
Let be the number of new packets transmitted at the beginning of the kth CRI and let
be the number of packets that due to captures are not transmitted successfully during
the kth CRI and hence are transmitted at the beginning of the (k+1)st CRI. Let
and for , let .
1. Write recursive equations for P
n
(l) for the wait and the persist schemes. Indicate the
order in which P
n
(l) should be computed.
2. Let . How would you compute from
P
n
(l)?
3. Let B
n
be the expected length of a CRI that starts with the transmission of n packets.
Write recursive equations for computing B
n
, for the wait and the persist schemes.
4. Write an expression for the throughput of the protocol as a function of B
n
, and
P
y
(l) the probability that in steadystate. How would you compute P
y
(l)?
Problem 5. (Known Multiplicity [Tsy80, GeP82])
Consider a system in which at the end of each slot the users are informed of the exact
number of users that transmitted during the slot. Assume that the epoch mechanism is
used and devise a collision resolution protocol for this system. Compute the maximal
throughput of your protocol.
n 2 ≥
A
˜
k
Y
˜
k
X
˜
k
A
˜
k
Y
˜
k 1 –
+ = n 0 ≥ 0 l n ≤ ≤ P
n
l ( ) Y
˜
k
l = X
˜
k
n = [ ] Prob =
p n
2
n
1
( ) Y
˜
k
n
2
= ( ) Y
˜
k 1 –
n
1
= [ ] Prob = p n
2
n
1
( )
n 0 ≥
n 0 ≥
Y
˜
k
l =
Section : EXERCISES 145
Problem 6. (Erasures, lost packets)
Consider a system that uses the binarytree CRP with the epoch mechanism. Assume that
whenever n ( ) packets are transmitted, there is a possibility that all these packets will
be erased, i.e., they will be lost and the feedback signal will indicate all users that the slot
was empty. Denote by π
n
the probability of this event. Assume that lost packets are never
retransmitted.
1. Write expressions for the expected length of a CRI that starts with the transmission of n
packets.
2. Write expressions for the expected number of packets that are transmitted successfully
during a CRI that starts with the transmission of n packets.
3. Write expressions for the expected number of packets that are lost during a CRI that
starts with the transmission of n packets.
4. Derive the throughput of this system when the arrival rate is λ and the epoch length is
∆. What is the rate (packets/slot) in which packets are lost in this case.
5. Let π
1
=0.5; π
i
=0 i>1. How should λ ∆ be chosen in order to maximize the throughput.
How would you compute the optimal epoch length in this case.
6. Repeat (1)(5) when the modiﬁed binarytree CRP is used.
n 1 ≥
146 CHAPTER5: COLLISIONRESOLUTION
APPENDIX A
Moments of Collision Resolution Interval Length
In this appendix we derive a closed form expression for the moments of . The reader
must bear with the lengthy (yet straight forward) algebraic manipulation involved in this
derivation.
We ﬁrst demonstrate the method to obtain a closed form expression for B
n
. Deﬁne the
exponential generating function of B
n
by
. (5.47)
Multiplying equation (5.5) by z
n
/n! and summing both sides for we obtain
.
Using (5.47) we have
or (using (5.1))
B
˜
n
B z ( ) B
n
z
n
n!
 
n 0 =
∞
∑
·
∆
n 2 ≥
B
n
z
n
n!
 
n 2 =
∞
∑
z
n
n!
 
n 2 =
∞
∑
z
n
n!
  Q
i
n ( ) B
i
B
n i –
+ ( )
i 0 =
n
∑
n 2 =
∞
∑
+ =
B z ( ) z – 1 – e
z
z – 1 –
z
n
n!
  Q
i
n ( ) B
i
B
n i –
+ ( )
i 0 =
n
∑
n 0 =
∞
∑
+ =
z
n
n!
  Q
i
n ( ) B
i
B
n i –
+ ( )
i 0 =
n
∑
n 0 =
1
∑
–
B z ( ) e
z
z
n
n!
  Q
i
n ( ) B
i
B
n i –
+ ( )
i 0 =
n
∑
n 0 =
∞
∑
2 2z 1 p – p + ( ) + [ ] – + =
e
z
2 1 z + ( ) –
z
n
i! n i – ( )!
 p
i
1 p – ( )
n i –
B
i
B
n i –
+ ( )
i 0 =
n
∑
n 0 =
∞
∑
+ =
e
z
2 1 z + ( ) –
z
n
i! n i – ( )!
 p
i
1 p – ( )
n i –
B
i
B
n i –
+ ( )
n i =
∞
∑
i 0 =
∞
∑
+ =
e
z
2 1 z + ( ) –
z 1 p – ( ) [ ]
n i –
n i – ( )!

zp [ ]
i
i!
 B
i
B
n i –
+ ( )
n i =
∞
∑
i 0 =
∞
∑
+ =
Section : APPENDIX A Moments of Collision Resolution Interval Length 147
therefore,
. (5.48)
It is convenient to deﬁne
(5.49)
which transforms (5.48) to
(5.50)
Expanding both sides of (5.50) to a Taylor series around z=0, letting
and equating the coefﬁcients of z
k
on both sides of (5.50) we get for .
or
. (5.51)
From the deﬁnitions (5.47) and (5.49) we obtain
from which, by equating corresponding coefﬁcients of z
n
, we obtain:
.
Finally, since and , and using (5.51) we obtain
e
z
2 1 z + ( ) –
zp [ ]
i
i!
B
i
i 0 =
∞
∑
z 1 p – ( ) [ ]
n i –
n i – ( )!

n i =
∞
∑
+ =
zp [ ]
i
i!
B
i
z 1 p – ( ) [ ]
n i –
n i – ( )!
B
n i –
n i =
∞
∑
i 0 =
∞
∑
+
e
z
2 1 z + ( ) – B zp ( )e
z 1 p – ( )
B z 1 p – ( ) ( )e
zp
+ + =
e
z –
B z ( ) 1 2e
z –
1 z + ( ) – B zp ( )e
zp –
B z 1 p – ( ) ( )e
z 1 p – ( ) –
+ + =
B
*
z ( ) e
z –
B z ( )
·
∆
B
*
z ( ) B
*
zp ( ) – B
*
z 1 p – ( ) ( ) – 1 2e
z –
1 z + ( ) – =
B
*
z ( ) B
k
*
z
k
k 0 =
∞
∑
=
k 2 ≥
B
k
*
p
k
B
k
*
– 1 p – ( )
k
B
k
*
– 2
1 – ( )
k
k!
 – 2
1 – ( )
k 1 –
k 1 – ( )!
 –
2 k 1 – ( ) 1 – ( )
k
k!
 = =
B
k
*
2 k 1 – ( ) 1 – ( )
k
k! 1 p
k
– 1 p – ( )
k
– [ ]
 = k 2 ≥
B
n
z
n
n!
 
n 0 =
∞
∑
B z ( ) e
z
B
*
z ( )
z
j
j!
 B
k
*
z
k
k 0 =
∞
∑
j 0 =
∞
∑
B
k
*
j!
z
k j +
j 0 =
∞
∑
k 0 =
∞
∑
= = = =
B
n
n!

B
k
*
n k – ( )!

k 0 =
n
∑
=
B
0
*
1 = B
1
*
0 =
148 CHAPTER5: COLLISIONRESOLUTION
(5.52)
which is the closed form expression for B
n
we were seeking.
In almost the same manner, a closed form expression for V
n
can be obtained. Let
Then from (5.10) we obtain
and with a deﬁnition we get
from which
.
B
n
1
n
k ¸ ,
¸ _
2 k 1 – ( ) 1 – ( )
k
1 p
k
– 1 p – ( )
k
– [ ]

k 2 =
n
∑
+ = n 2 ≥
V z ( ) V
n
z
n
n!
 
n 0 =
∞
∑
·
∆
V z ( ) 2B z ( ) e
z
– 4 1 z + ( ) – 2B zp ( )B z 1 p – ( ) ( ) + =
V zp ( )e
z 1 p – ( )
V z 1 p – ( ) ( )e
zp
+ +
V
*
z ( ) e
z –
V z ( ) =
V
*
z ( ) V
*
zp ( ) – V
*
z 1 p – ( ) ( ) – 2B
*
z ( ) 1 – 4e
z –
1 z + ( ) – 2B
*
zp ( )B
*
z 1 p – ( ) ( ) + =
V
n
1
2n!
n k – ( )! 1 p
k
– 1 p – ( )
k
– [ ]
 B
k
*
B
i
*
B
k i –
*
i 0 =
k
∑
+
k 2 =
n
∑
+ =
n
k ¸ ,
¸ _
4 k 1 – ( ) 1 – ( )
k
1 p
k
– 1 p – ( )
k
– [ ] [ ]

k 2 =
n
∑
+ n 2 ≥
CHAPTER 6
ADDITIONAL TOPICS
The ﬁeld of multipleaccess systems is much too broad to be contained in a single book.
Although we treated in depth many fundamental protocols and systems, we were able to
uncover just the tip of the iceberg. Many important and interesting subjects in this ﬁeld
were not discussed in the book, either because they are beyond the scope we planned for
the book or because they are still in a formative and fragmentary stage of research. This
section is devoted to short descriptions of several of these subjects.
Multihop Networks
The basic topology assumed throughout this book is the singlehop topology in which all
users hear one another, or there is a common receiver that can hear all transmissions in the
network. An important topology, known as a.i “multihop” network, is characterized by the
feature that each user is in reception range of only a subset of the users and similarly the
transmission of a user is heard by a subset of all users. A transmission is successful only if
it is the only transmission currently being heard by the.i “receiving” node. The term “mul
tihop” alludes to the need of packets to hop over several intermediate users in order to
arrive at their destinations, since not all pairs of users communicate directly. This gives
rise to routing issues i.e., which of the receiving users should forward a received packet
towards its destination. The multihop topology allows for concurrent successful transmis
sions, provided each receiver receives only a single transmission at a time; this property is
utilized to increase the capacity of the total network by an approach called spatial reuse
Multihop networks appear naturally in radio networks with low powered transmitters or in
interconnected local area networks. The fact that each receiver hears a subset of transmit
ters rather than all of them renders the analysis of multihop systems far more complex
than that of singlehop systems. Yet, some of the basic protocols such as the pure and the
slotted Aloha are still applicable in multihop systems. Carrier sensing can also be used,
although it will not be as effective as in singlehop systems since a transmission cannot be
sensed by all users. To improve the effectiveness of carrier sensing, the idea of busytone
can be used (see ToK75, SiS81, BrT85, CiR86]). The application of collision resolution or
controlled Aloha protocols require substantial revisions of the protocol since it is difﬁcult
to obtain the correct feedback information at the end of each slot in a multihop environ
ment.
A detailed description of multihop packet radio technology appears in [KGB78]. An
extensive survey of recent developments in the analysis of multihop systems appears in
[Tob87]. Other relevant papers are [BoK80, SiK83, TaK85b, BKM87, KlS87, ShK87,
KBC87, PYS87].
150 CHAPTER 6: ADDITIONAL TOPICS
Multistation Networks
The multihop topology assumes that every user in the network can and should serve as a
repeater for packets it receives from its neighbors and that needs to be forwarded to their
destinations. In many applications of packetradio networks, the population of users is not
homogeneous: some users are more powerful than others, some are not mobile, etc. In
addition, some inherent hierarchy may exist among the users of the network. In such sys
tems it is natural to build a backbone network of stations that is responsible for the routing
and other network functions.
In the multistation model the users of the network are originators of the data that are trans
mitted through a shared channel to the stations. The stations may be either the ﬁnal desti
nations for some packets sent by the users or can act as relays for other packets, by
forwarding them to their respective destinations (other stations or users). The network
operates as follows: a packet that is generated at some user, is forwarded to a station via
the shared channel by employing some multiaccess protocol. The station then forwards it
to some other station through the backbone network of stations to be ﬁnally transmitted on
the stationtouser channel to its destination (cellular phone systems use this approach).
The advantages of the multistation conﬁguration over the singlehop one are that the
former allows for lower power transmitters at the users, results in better utilization of the
common radio channel due to spatial reuse and allows distribution of control among sev
eral stations. It is also advantageous over the multihop conﬁguration because it simpliﬁes
both the design and the analysis of the network, it simpliﬁes nodal protocols, and is also
adequate for a large number of naturally structured hierarchical networks.
Multistation networks in which a TDMA scheme is used are considered in [RoS89].
Alohatype protocols in this environment are studied in [SiC88, CiR86], and collision res
olution protocols in [BaS88].
Multichannel Systems
The networks considered in this book contain a single shared channel used by the users for
communication. Many studies that consider multichannel networks. These networks are
characterized by the ability of the users to communicate via several different, noninterfer
ing, communication channels at different bands. To keep the same level of connectivity as
in the single channel environment, the users should be equipped with several receivers.
The main advantage of using multiple channels is the reduced interference level among
the users. There are two ways in which interference is reduced compared to the single
channel environment. The ﬁrst is obviousless users use each frequency band. The second
is characteristic to carrier sensing systems; since the total available bandwidth is ﬁxed,
each frequency band in the multichannel systemis narrower and the transmission time of a
packet is longer. The propagation delay is constant and hence the ratio between the propa
Section : ADDITIONAL TOPICS 151
gation delay and the packet transmission time becomes smaller, yielding better perfor
mance.
Multichannel systems for various Alohatype protocols are discussed in [MaR83, Tod85,
Kim85, MaB87, Kim87, ShK87, ChG88].
152 CHAPTER 6: ADDITIONAL TOPICS
REFERENCES
Abr70. N. Abramson, “The ALOHA System  Another Alternative for
Computer Communications,” pp. 281285 in Proc. of the Fall Joint
Computer Conference, (1970).
Abr77. N. Abramson, “The Throughput of Packet Broadcasting Channels,”
IEEE Trans. on Communications, COM25(1) pp. 117128 (Janu
ary 1977).
ApP86. T.K. Apostolopoulos and E.N. Protonotarios, “Queueing Analysis
of Buffered CSMA/CD Protocols,” IEEE Trans. on Communica
tions, COM34(9) pp. 898905 (September 1986).
BaS88. A. BarDavid and M. Sidi, “Collision Resolution Algorithms in
MultiStation PacketRadio Networks,” pp. 385400 in PERFOR
MANCE’87, Brussels (December 1987).
BeB80. S. Bellini and P. Borgonovo, “On the throughput of an ALOHA
channel with variable length packets,” IEEE Trans. Communica
tions, COM28(11) pp. 19321935 (November 1980).
BeC88. S.L. Beuerman and E.J. Coyle, “The Delay Characteristics of
CSMA/CD Networks,” IEEE Trans. on Communications, COM
36(5) pp. 553563 (May 1988).
BeG87. D. Bertsekas and R. Gallager, Data Networks, Prentice Hall, Inc.,
NewJersey (1987).
BeT88. T. Berger and T.S. Tszan, “An Improved Upper Bound for the
Capacity of a Channel with Multiple Random Access,” Problemy
Peredachi Informatsii, 21(4) pp. 8387 (January 1985).
BeZ88. T. Berger and R.Z. Zhu, “Upper Bound for the Capacity of a Ran
domMultiple Access System,” Problemy Peredachi Informatsii, 17
pp. 9095 (January 1988).
Bin75. R. Binder, “A Dynamic Packet Switching System for Satellite
Broadcast Channels,” pp. 41.141.5 in Proc. of ICC’75, San Fran
cisco, California (1975).
BKM87. R.R. Boorstyn, A. Kershenbaum, B. Maglaris, and V. Sahin,
“Throughput Analysis in Multihop CSMA Packet Radio Net
works,” IEEE Trans. on Communications, COM35(3) pp. 267274
(March 1987).
BMT84. T. Berger, N. Mehravari, D. Towsley, and J. Wolf, “Random Multi
pleAccess Communication and Group Testing,” IEEE Trans. on
Communications, COM32(7) pp. 769779 (July 1984).
154 : REFERENCES
BoK80. R.R. Boorstyn and A. Kershenbaum, “Throughput Analysis of Mul
tihop Packet Radio,” pp. 13.6.113.6.6 in Proc. of ICC’80, Seattle,
Washington (1980).
BrT85. J.M. Brazio and F.A. Tobagi, “Throughput Analysis of Spread
Spectrum Multihop Packet Radio Networks,” pp. 256265 in Pro
ceeding of IEEE INFOCOM’85, Washington, D.C. (March 1985).
CaH75. A.B. Carleial and M.E. Hellman, “Bistable Behavior of ALOHA
type Systems,” IEEE Trans. on Communications, COM23(4) pp.
401410 (April 1975).
Cap77. J.I. Capetanakis, “The Multiple Access Broadcast Channel: Proto
col and Capacity Considerations,” in Ph.D. Dissertation, Depart
ment of Electrical Engineering, MIT (August 1977).
Cap79. J.I. Capetanakis, “Tree Algorithm for Packet Broadcast Channels,”
IEEE Trans. on Information Theory, IT25(5) pp. 505515 (Septem
ber 1979).
CFL79. I. Chlamtac, W.R. Franta, and K.D. Levin, “BRAM: The Broadcast
Recognizing Access Mode,” IEEE Trans. on Communications,
COM27(8) pp. 11831189 (August 1979).
ChG88. I. Chlamtac and A. Ganz, “Channel Allocation Protocols in Fre
quencyTime Controlled High Speed Networks,” IEEE Trans. on
Communications, COM36(4) pp. 430440 (April 1988).
CiR86. I. Cidon and R. Rom, “Carrier Sense Access in a Two Interfering
Channels Environment,” Computer Networks and ISDN Systems,
12(1) pp. 110 (August 1986).
CiS85. I. Cidon and M. Sidi, “The Effect of Capture on CollisionResolu
tion Algorithms,” IEEE Trans. Communications, COM33(4) pp.
317324 (April 1985).
CiS87. I. Cidon and M. Sidi, “Erasures and Noise in Multiple Access Algo
rithms,” IEEE Trans. on Information Theory, IT33(1) pp. 132143
(January 1987).
CiS88. I. Cidon and M. Sidi, “Conﬂict Multiplicity Estimation and Batch
Resolution Algorithms,” IEEE Trans. on Information Theory, IT
34(1) pp. 101110 (January 1988).
CKS88. I. Cidon, H. Kodesh, and M. Sidi, “Erasure, Capture and Random
Power Level Selection in MultipleAccess Systems,” IEEE Trans.
Communications, COM36(3) pp. 263271 (March 1988).
CoB80. S.D. Conte and C. de Boor, Elementary Numerical Analysis: An
Algorithmic Approach (3rd ed.), McGraw Hill (1980).
CoL83. E.J. Coyle and B. Liu, “Finite Population CSMACD Networks,”
IEEE Trans. on Communications, COM31(11) pp. 12471251
(November 1983).
Section : REFERENCES 155
CoL85. E.J. Coyle and B. Liu, “A Matrix Representation of CSMA/CD
Networks,” IEEE Trans. on Communications, COM33(1) pp. 53
64 (January 1985).
CPR78. D.D. Clark, K.T. Pogran, and D.P. Reed, “Introduction to Local
Area Networks,” Proc. of the IEEE, 66(11) pp. 14971517 (Novem
ber 1978).
CrH82. R. Cruz and B. Hajek, “ANewUpper Bound to the Throughput of a
MultiAccess Broadcast Channel,” IEEE Trans. on Information
Theory, IT28 pp. 402405 (May 1982).
CRW73. W. Crowther, R. Rettberg, D. Walden, S. Ornstein, and F. Heart, “A
System for Broadcast Communication: ReservationALOHA,” pp.
371374 in Proc. of the 6th Hawaii International Conference on
Systems Sciences, Honolulu, Hawaii (1973).
CuM88. G.A. Cunningham and J.S. Meditch, “Distributed Retransmission
Controls for Slotted, Nonpersistent, and Virtual Time CSMA,”
IEEE Trans. on Communications, COM36(6) pp. 685691 (June
1988).
DaG80. D.H. Davis and S.A. Gronemeyer, “Performance of Slotted
ALOHA Random Access with Delay Capture and Randomized
Time of Arrival,” IEEE Trans. on Communications, COM28(5)
pp. 703710 (May 1980).
EpZ87. A. Ephremides and R.Z. Zhu, “Delay Analysis of Interacting
Queues with an Approximate Model,” IEEE Trans. on Communica
tions, COM35(2) pp. 194201 (February 1987).
Fer75. M.J. Ferguson, “On the Control, Stability, and Waiting Time in a
Slotted ALOHA RandomAccess System,” IEEE Trans. on Com
munications, COM23(11) pp. 13061311 (November 1975).
Fer77a. M.J. Ferguson, “ABound and Approximation of Delay Distribution
for FixedLength Packets in an Unslotted ALOHA Channel and a
Comparison with Time Division Multiplexing (TDM),” IEEE
Trans. on Communications, COM25(1) pp. 136139 (January
1977).
Fer77b. M.J. Ferguson, “An Approximate Analysis of Delay for Fixed and
Variable Length Packets in an Unslotted ALOHA Channel,” IEEE
Trans. on Communications, COM25(7) pp. 644654 (July 1977).
FFH85. G. Fayolle, P. Flajolet, M. Hofri, and P. Jacquet, “Analysis of a
Stack Algorithm for Random MultipleAccess Communication,”
IEEE Trans. Information Theory, IT31(2) pp. 244254 (March
1985).
FiT84. M. Fine and F.A. Tobagi, “Demand Assignment Multiple Access
Schemes in Broadcast Bus Local Area Networks,” IEEE Trans.
Computers, C33(12) pp. 113059 (December 1984).
156 : REFERENCES
FLB74. G. Fayolle, J. Labetoulle, D. Bastin, and E. Gelenbe, “The Stability
Problem of Broadcast Packet Switching Computer networks,” Acta
Informatica, 4(1) pp. 4953 (1974).
Gal78. R.G. Gallager, “Conﬂict Resolution in Random Access Broadcast
Networks,” pp. 7476 in Proc. AFOSR Workshop Communication
Theory and Applications, Provincetown (September 1978).
Gal85. R.G. Gallager, “A Perspective on Multiaccess Channels,” IEEE
Trans. on Information Theory, IT31(2) pp. 124142 (March 1985).
Gek87. L. Georgiadis and P. PapantoniKazakos, “A 0.487 Throughput
Limited sensing Algorithm,” IEEE Trans. on Information Theory,
IT33(2) pp. 233237 (March 1987).
GeP82. L. Georgiadis and P. PapantonyKazakos, “A Collision Resolution
Protocol for Random Access Channels with Energy Detectors,”
IEEE Trans. on Communications, COM30(11) pp. 24132420
(November 1982).
GGFL87. A.G. Greenberg, P. Flajolet, and R.E. Ladner, “Estimating the Mul
tiplicities of Conﬂicts to Speed Their Resolution in Multiple Access
Channels,” Journal of the ACM, 34(2) pp. 289325 (April 1987).
GMP87. L. Georgiadis, L.F. Merakos, and P. PapantoniKazakos, “AMethod
for the Delay Analysis of Random MultipleAccess Algorithms
whose Delay Process is Regenerative,” IEEE Journal on Selected
Areas in Communications, SAC5(6) pp. 10511062 (July 1987).
HaL82. B. Hajek and T. Van Loon, “Decentralized Dynamic Control of a
Multiaccess Broadcast Channel,” IEEE Trans. on Automatic Con
trol, AC27 pp. 559569 (June 1982).
HaO86. J.L. Hammond and P.J.P. O’Reilly, Performance Analysis of Local
Computer Networks, AddisonWesley Publishing Company (1986).
HaS79. L.W. Hansen and M. Schwartz, “An AssignedSlot ListenBefore
Transmission Protocol for a Multiaccess Data Channel,” IEEE
Trans. on Communications, COM27(6) pp. 846856 (June 1979).
Hay78. J.F. Hayes, “An Adaptive Technique for Local Distribution,” IEEE
Trans. on Communications, COM26(8) pp. 11781186 (August
1978).
Hay84. J.F. Hayes, Modeling and Analysis of Computer Communications
Networks, Plenum Press, New York (1984).
Hey82. D.P. Heyman, “An Analysis of the CarrierSense MultipleAccess
Protocol,” Bell System Technical Journal, 61 pp. 20232051 (Octo
ber 1982).
Hey86. D.P. Heyman, “The Effects of Random Message Sizes on the Per
formance of the CSMA/CD Protocol,” IEEE Trans. on Communica
tions, COM34(6) pp. 547553 (June 1986).
Section : REFERENCES 157
HoR87. M. Hofri and Z. Rosberg, “Packet Delay Under the Golden Ration
Weighted TDMPolicy in a Multiple Access Channel,” IEEE Trans.
on Information Theory, IT33(3) pp. 341349 (1987).
HuB85. J.C. Huang and T. Berger, “Delay Analysis of IntervalSearching
Contention Resolution Algorithms,” IEEE Trans. on Information
Theory, IT31(2) pp. 264273 (March 1985).
HuB86. J.C. Huang and T. Berger, “Delay Analysis of 0.487 Contention
Resolution Algorithms,” IEEE Trans. on Communications, COM
34(9) pp. 916926 (September 1986).
Hum86. P.A. Humblet, “On the Throughput of Channel Access Algorithms
with Limited Sensing,” IEEE Trans. Communications, COM34(4)
pp. 345347 (April 1986).
ItR84. A. Itai and Z. Rosberg, “A Golden Ratio Control Policy for a Multi
pleAccess Channel,” IEEE Trans. Automatic Control, AC29(8)
pp. 712718 (August 1984).
Jen80. Y.C. Jenq, “On the Stability of Slotted ALOHA Systems,” IEEE
Trans. Communications, COM28(11) pp. 19361939 (November
1980).
KBC87. A. Kershenbaum, R. Boorstyn, and M.S. Chen, “An Algorithm for
Evaluation of Throughput in Multihop Packet Radio Networks with
Complex Topologies,” IEEE Journal on Selected Areas in Commu
nications, SAC5(6) pp. 10031012 (July 1987).
Kel85. F.P Kelly, “Stochastic Models of Computer Communication Sys
tems,” Journal of the Royal Statistical Society, 47(1) (1985).
KeS88. I. Kessler and M. Sidi, “Mixing Collision Resolution Algorithms
Exploiting Information of Successful Messages,” IEEE Trans. on
Information Theory, IT34(3) pp. 531536 (May 1988).
KGB78. R.E. Kahn, A.A. Gronemeyer, J. Burchﬁel, and R.C. Kunzelman,
“Advances in packet radio technology,” Proceedings of the IEEE,
66(11) pp. 14681496 (November 1978).
KiK83. W.M. Kiesel and P.J. Kuehn, “A new CSMACD protocol for local
area networks with dynamic priorities and low collision probabil
ity,” IEEE Journal on Selected Areas in Communications, SAC
1(5) pp. 869876 (November 1983).
Kim85. G. Kimura, “An Analysis of the Multichannel CSMA/CD Protocol
by Nonslotted Model,” Trans. of the Japanese Inst. Electronics and
Communication Engineering, J68B (Part B)(12) pp. 13411348
(December 1985).
Kim87. G. Kimura, “An Analysis of the MultiChannel CSMA/CDProtocol
by Nonslotted Model,” Trans. of the Japanese Inst. Electronics and
Communication Engineering, 70(5) pp. 7885 (May 1987).
158 : REFERENCES
Kle76. L. Kleinrock, Queueing Systems (Vols. I, II), J. Wiley (1975, 1976).
Kle75. L. Kleinrock and S.S. Lam, “Packet Switching in a Multiaccess
Broadcast Channel: Performance Evaluation,” IEEE Trans. on
Communications, COM23(4) pp. 410423 (April 1975).
KlS80. L. Kleinrock and M. Scholl, “Packet Switching in Radio Channels:
New ConﬂictFree Multiple Access Schemes,” IEEE Trans. on
Communications, COM28(7) pp. 10151029 (July 1980).
KlS87. L. Kleinrock and J. Silvester, “Spatial Reuse in Multihop Packet
Radio Networks,” Proceedings of the IEEE, 75(1) pp. 156167
(January 1987).
KlT75. L. Kleinrock and F.A. Tobagi, “Packet Switching in Radio Chan
nels: Part I  Carrier Sense MultipleAccess Modes and Their
Throughput Delay Characteristics,” IEEE Trans. on Communica
tions, COM23(12) pp. 14001416 (December 1975).
KlY78. L. Kleinrock and Y. Yemini, “An optimal adaptive scheme for mul
tiple access broadcast communication,” in Proc. of ICC’78, Tor
onto, Canada (1978).
KSY88. J.F. Kurose, M. Schwartz, and Y. Yemini, “Controlling window pro
tocols for timeconstrained communication in multiple access net
works,” IEEE Trans. on Communications, COM36(1) pp. 4149
(January 1988).
Kuo81. F.F. Kuo, Protocols and Techniques for Data Communication Net
works, Prentice Hall, New Jersey (1981).
KuS88. D. Kurtz and M. Sidi, “Multiple Access Algorithms via Group Test
ing for Heterogeneous Population of Users,” IEEE Trans. Commu
nications, COM36(12) pp. 13161323 (December 1988).
LaK75. S.S. Lam and L. Kleinrock, “Packet Switching in a Multiaccess
Broadcast Channel: Dynamic Control Procedures,” IEEE Trans. on
Communications, COM23(9) pp. 891904 (September 1975).
Lam77. S.S. Lam, “Delay Analysis of a Time Division Multiple Access
(TDMA) Channel,” IEEE Trans. on Communications, COM
25(12) pp. 14891494 (December 1977).
Lam80. S.S. Lam, “Packet Broadcast Networksa Performance Analysis of
The RALOHA Protocol,” IEEE Trans. on Computers, C29(7) pp.
596603 (July 1980).
Lee87. C.C. Lee, “Random Signal Levels for Channel Access in Packet
Broadcast networks,” IEEE Journal on Selected Areas in Communi
cations, SAC5(6) pp. 10261034 (July 1987).
LeP87. J.S. Lehnert and M.B. Pursley, “Error Probabilities for Binary
DirectSequence SpreadSpectrum Communications with Random
Section : REFERENCES 159
Signature Sequence,” IEEE Trans. on Communications, COM
35(1) pp. 8798 (January 1987).
LiF82. J.O. Limb and C. Flores, “Description of Fasnet, a Unidirectional
Local Area Communication Network,” Bell System Technical Jour
nal, 61 (Part 1)(7) (September 1982).
MaB87. M.A. Marsan and M. Bruscagin, “Multichannel ALOHA Networks
with Reduced Connections,” pp. 268275 in IEEE INFOCOM’87,
San Francisco, CA (April 1987).
MaF83. P. Mathys and P. Flajolet, “Qary Collision Resolution Algorithms
in Randomaccess Systems with Free or Blocked ChannelAccess,”
IEEE Trans. Information Theory, IT31(2) pp. 217243 (March
1985).
MaM85. J.L. Massey and P. Mathys, “The Collision Channel Without Feed
back,” IEEE Trans. on Information Theory, IT31(2) pp. 192204
(March 1985).
Mar78. J. Martin, Communication Satellite Systems, PrenticeHall, New
Jersey (1978).
MaR83. M.A. Marsan and D. Rofﬁnella, “Multichannel Local Area Net
work Protocols,” IEEE Journal on Selected Areas in Communica
tions, SAC1(5) pp. 885897 (November 1983).
Mas81. J.L. Massey, “Collision Resolution Algorithms and Random
Access Communications,” pp. 73137 in MultiUser Communica
tions Systems (CISM Courses and Lectures Series), ed. G. Longo,
SpringerVerlag, New York (1981). (Also in UCLA Technical
Report UCLAENG8016, April 1980)
MeB84. N. Mehravari and T. Berger, “Poisson MultipleAccess Contention
with Binary Feedback,” IEEE Trans. Information Theory, IT30(5)
pp. 745751 (September 1984).
MeK85. L. Merakos and D. Kazakos, “On Retransmission Control Policies
in MultipleAccess Communication Networks,” IEEE Trans. Auto
matic Control, AC30(2) pp. 109117 (February 1985).
MeL83. J.S. Meditch and C.T.A. Lea, “Stability and Optimization of the
CSMA and CSMA/CD Channels,” IEEE Trans. on Communica
tions, COM31(6) pp. 763774 (June 1983).
Met73. A R. Metcalfe, "SteadyState Analysis of a Slotted and Controlled
ALOHA System with Blocking," pp. 375380 in Proc. 6th Hawaii
International Conference on System Sciences, Honolulu, Hawaii
(January 1973).
Met76. J.J. Metzner, “On Improving Utilization in ALOHA Networks,”
IEEE Trans. on Communications, COM24(4) pp. 447448 (April
1976).
160 : REFERENCES
MiT81. V.A. Mikhailov and B.S. Tsybakov, “Upper Bound for the Capacity
of a Random Multiple Access System,” Problemy Peredachi Infor
matsii, 17(1) pp. 9095 (January 1981).
MoH85. J. Mosley and P.A. Humblet, “A Class of Efﬁcient contention Reso
lution Algorithms for Multiple Access Channels,” IEEE Trans. on
Communications, COM33(2) pp. 145151 (February 1985).
MoK85. M.L. Molle and L. Kleinrock, “Virtual Time CSMA: Why Two
Clocks are Better Than One,” IEEE Trans. on Communications,
COM33(6) pp. 919933 (June 1985).
Mol82. M.L. Molle, “On the Capacity of Inﬁnite Population Multiple
Access Protocols,” IEEE Trans. Information Theory, IT28(3) pp.
396401 (May 1982).
MoR84. L.F.M. De Moraes and I. Rubin, “Message Delays for a TDMA
Scheme under a Nonpreemptive Priority Discipline,” IEEE Trans.
on Communications, COM32(5) pp. 5838 (May 1984).
Mue56. D.E. Mueller, “A Method of Solving Algebraic Equations Using an
Automatic computer,” Mathematical Tables and Other Aids to
Computation (MTAC), 10 pp. 208215 (1956).
Neu81. M.F. Neuts, MatrixGeometric Solutions in Stochastic Models: An
Algorithmic Approach, The John Hopkins Press (1981).
OnN85. Y. Onozato and S. Noguchi, “On the Thrashing Cusp in Slotted
ALOHASystems,” IEEE Trans. on Communications, COM33(11)
pp. 11711182 (November 1985).
Pak69. A.G. Pakes, “Some Conditions of Ergodicity and Recurrence of
Markov Chains,” Operations Research, 17 (1969).
Pip81. N. Pippenger, “Bounds on the Performance Of Protocols for a Mul
tipleAccess Broadcast Channel,” IEEE Trans. Information Theory,
IT27(2) pp. 145151 (March 1981).
PMV87. G.C. Polyzos, M.L. Molle, and A.N. Venetsanopoulos, “Perfor
mance Analysis of Finite Nonhomogeneous Population Tree Con
ﬂict Resolution Algorithms using Constant Size Window Access,”
IEEE Trans. on Communications, COM35(11) pp. 11241138
(November 1987).
Pur77. M.B. Pursley, “Performance Evaluation for PhaseCoded Spread
Spectrum MultipleAccess Communication (part I: System analy
sis),” IEEE Trans. on Communications, COM25(8) pp. 795799
(August 1977).
Pur87. M.B. Pursley, “The Role of Spread Spectrum in Packet Radio Net
works,” Proceedings of IEEE, (1) pp. 11634 (January 1987).
PYS87. E. Pinsky, Y. Yemini, and M. Sidi, “The canonical approximation in
the performance analysis of packet radio networks,” pp. 140162 in
Section : REFERENCES 161
Current advances in distributed computing and communications,
Ed. Y. Yemini, Computer Science Press, (1987).
Riv87. R.L. Rivest, “Network Control by Bayesian Broadcast,” IEEE
Trans. on Information Theory, IT33(3) pp. 323328 (May 1987).
Rob72. L.G. Roberts, “Dynamic Allocation of Satellite Capacity Through
Packet Reservation,” in Computer communication networks, ed.
R.L. Grimsdale and F.F. Kuo, Noordhoff Internat Publishing,
Groningen, Netherlands (1975).
Rom84. R. Rom, “Ordering Subscribers on Cable Networks,” ACM Trans
action on Computer Systems, 2(4) pp. 322334 (November 1984).
Rom86. R. Rom, “Collision Detection in Radio Channels,” pp. 235249 in
Local Area and Multiple Access Networks, Computer Science
Press, (1986).
Ros72. S.M. Ross, Introduction to Probability Models, Academic Press,
New York (1972).
RoS89. Z. Rosberg and M. Sidi, “TDM Policies in Multistation Packet
Radio Networks,” IEEE Trans. on Communications, COM37(1)
pp. 3138 (January 1989).
RoT81. R. Rom and F.A. Tobagi, “MessageBased Priority Functions in
Local Multiaccess Communication Systems,” Computer Networks,
5(4) pp. 273286 (July 1981).
Rub79. I. Rubin, “Message Delays in FDMA and TDMA Communication
Channels,” IEEE Trans. on Communications, COM27(5) pp. 769
777 (May 1979).
Rub78. I. Rubin, "Group RandomAccess Disciplines for MultiAccess
Broadcast Channels," IEEE Trans. on Information Theory IT24(5)
pp. 578592 (September 1978).
Sac88. S.R. Sachs, "Alternative Local Area Network Access Protocols,"
IEEE Communications Magazine (26)(3) pp. 2545 (March 1988).
SaE81. T.N. Saadawi and A. Ephremides, “Analysis, Stability and Optimi
zation of Slotted ALOHAwith a Finite Number of Buffered Users,”
IEEE Trans. on Automatic Control, AC26(3) pp. 680689 (June
1981).
San80. D. Sant, “Throughput of unslotted ALOHA channels with arbitrary
packet interarrival time distributions,” IEEE Trans. Communica
tions, COM28(8 (part 2)) pp. 14221425 (August 1980).
ScK79. M. Scholl and L. Kleinrock, “On a Mixed Mode Multiple Access
Scheme for PacketSwitched Radio Channels,” IEEE Trans. on
Communications, COM27(6) pp. 906911 (June 1979).
162 : REFERENCES
Sha84. N. Shacham, “ThroughputDelay Performance of PacketSwitching
MultipleAccess Channel with Power Capture,” Performance Eval
uation, 4(3) pp. 153170 (August 1984).
ShH82. N. Shacham and V.B. Hunt, “Performance Evaluation of the
CSMACD 1Persistent Channel Access Protocol in Common
Channel Local Networks,” pp. 401414 in Proc. of the International
Symposium on Local Computer Networks, IFIP TC6, Florence,
Italy (April 1982).
ShK87. N. Shacham and P.J.B. King, “Architectures and Performance of
Multichannel Multihop Packet Radio Networks,” IEEE Journal
Selected Areas Communications, SAC5(6) pp. 10131025 (20 July
1987).
SiC85. M. Sidi and I. Cidon, “Splitting Protocols in Presence of Capture,”
IEEE Trans. Information Theory, IT31 pp. 295301 (March 1985).
SiC88. M. Sidi and I. Cidon, “A MultiStation PacketRadio Network,”
Performance Evaluation, 8(1) pp. 6572 (February 1988).
SiK83. J.A. Silvester and L. Kleinrock, “On the Capacity of Multihop Slot
ted ALOHA Networks with Regular Structure,” IEEE Trans. on
Communications, COM31(8) pp. 974982 (August 1983).
SiS81. M. Sidi and A. Segall, “A BusyTone MultipleAccess Type
Scheme for PacketRadio Networks,” pp. 110 in The International
Conference on Performance of Data Communication Systems and
their Applications, Paris, France (1416 September 1981).
SiS83. M. Sidi and A. Segall, “Two Interfering Queues in PacketRadio
Networks,” IEEE Trans. on Communications, COM31(1) pp. 123
129 (January 1983).
SMV87. K. Sohraby, M.L. Molle, and A.N. Venetsanopoulos, “Comments
on ‘Throughput Analysis for Persistent CSMA Systems’,” IEEE
Trans. on Communications, COM35(2) pp. 240243 (February
1987).
Sta85. W. Stallings, Data and Computer Communications, Macmillan
Inc., New York (1985).
Szp86. W. Szpankowski, “Bounds for queue lengths in a contention packet
broadcast system,” IEEE Trans. Communications, COM34(11) pp.
11321140 (November 1986).
TaI84. S. Tasaka and Y. Ishibashi, “A Reservation Protocol for Satellite
Packet communicationa Performance Analysis and Stability Con
siderations,” IEEE Trans. on Communications, COM32(8) pp.
9207 (August 1984).
Section : REFERENCES 163
TaK85a. H. Takagi and L. Kleinrock, “Mean Packet Queueing Delay in a
Buffered TwoUser CSMA/CD system,” IEEE Trans. on Communi
cations, COM33(10) pp. 11361139 (October 1985).
TaK85b. H. Takagi and L. Kleinrock, “Throughput Analysis for Persistent
CSMA Systems,” IEEE Trans. on Communications, COM33(7)
pp. 627638 (July 1985). (Corrected February 1987)
H. Takagi and L. Kleinrock, “Output Processes in Contention
Packet Broadcasting Systems,” IEEE Trans. on Communications,
COM33(11) pp. 11911199 (November 1985).
TaK857. H. Takagi and L. Kleinrock, “Correction to ‘Throughput Analysis
for Persistent CSMA Systems’,” IEEE Trans. on Communications,
COM35(2) pp. 243245 (February 1987).
Tan81. A.S. Tannenbaum, Computer Networks, Prentice Hall, Inc., New
Jersey (1981).
Tas86. S. Tasaka, Performance Analysis of Multiple Access Protocols, MIT
Press, Cambridge, Mass. (1986).
TaY86. A. Takagi and S. Yamada, “CSMA/CD with Deterministic Conten
tion Resolution,” IEEE Journal on Selected Areas in Communica
tions, SAC1(5) pp. 877884 (November 1983).
TML83. B.S. Tsybakov, V.A. Mikhailov, and N.B. Likhanov, “Bounds for
Packet Transmission Rate in a RandomMultipleAccess System,”
Probl. Information Transmission, 19(1) pp. 5068 (JanuaryMarch
1983).
Tob80. F.A. Tobagi, “Multiaccess Protocols in Packet Communication Sys
tems,” IEEE Trans. on Communications, COM28(4) pp. 468488,
(April 1980).
Tob82a. F.A. Tobagi, “Carrier Sense Multiple Access With MessageBased
Priority Functions,” IEEE Trans. on Communications, 30(1 pt 2,)
pp. 185200 (January 1982).
Tob82b. F.A. Tobagi, “Distributions of Packet Delay and Interdeparture
Time in Slotted ALOHA and Carrier Sense Multiple Access,” Jour
nal of the ACM, 29(4) pp. 907927 (October 1982).
Tob87. F.A. Tobagi, “Modeling and Performance Analysis of Multihop
Packet Radio Networks,” Proc. of the IEEE, 75(1) pp. 135155
(January 1987).
Tod85. T.D. Todd, “Throughput in Slotted Multichannel CSMA/CD Sys
tems,” pp. 276280 in GLOBECOM’85, New Orleans, LA (Decem
ber 1985).
ToH80. F.A. Tobagi and V.B. Hunt, “Performance Analysis of Carrier Sense
Multiple Access with Collision Detection,” Computer Networks,
4(5) pp. 245259 (October/November 1980).
164 : REFERENCES
ToK75. F.A. Tobagi and L. Kleinrock, “Packet Switching in Radio Chan
nels: Part II  The Hidden Terminal Problemin Carrier Sense Multi
pleAccess and the Busy Tone Solution,” IEEE Trans. on
Communications, COM23(12) pp. 14171433 (December 1975).
ToK76. F.A. Tobagi and L. Kleinrock, “Packet Switching in Radio Chan
nels: Part III  Polling and (Dynamic) SplitChannel Reservation
MultipleAccess,” IEEE Trans. on Communications, COM24(8)
pp. 832845 (August 1976).
ToK77. F.A. Tobagi and L. Kleinrock, “Packet Switching in Radio Chan
nels: Part IV  Stability Considerations and Dynamic Control in
Carrier Sense MultipleAccess,” IEEE Trans. on Communications,
COM25(10) pp. 11031119 (October 1977).
ToR80. F.A. Tobagi and R. Rom, “Efﬁcient Round Robin and Priority
Schemes in Unidirectional Broadcast Systems,” in Proc. of the
IFIPWG 6.4 Local Area Networks Workshop, Zurich (August
1980).
ToV82. D. Towsley and G. Venkatesh, “Window RandomAccess Protocols
for Local Computer Networks,” IEEE Trans. on Computers, C
31(8) pp. 715722 (August 1982).
ToV87. D. Towsley and P.O. Vales, “Announced arrival random access pro
tocols,” IEEE Trans. on Communications, COM35(5) pp. 513521
(May 1987).
TsB88. B.S. Tsybakov and V.L. Bakirov, “Stability Analysis of a Packet
Switching Network and Its Application to Asynchronous Aloha
Radio Networks,” Problemy Peredachi Informatsii, 24(2) pp. 139
151 (October 1988).
TsC86. D. Tsai and J.F. Chang, “Performance Study of an Adaptive Reser
vation Multiple Access Technique for Data Transmissions,” IEEE
Trans. on Communications, COM34(7) pp. 725727 (July 1986).
Tsi87. J.N. Tsitsiklis, “Analysis of a Multiaccess Control Scheme,” IEEE
Trans. on Automatic Control, AC32(11) pp. 10171020 (November
1987).
TsL83. B.S. Tsybakov and N.B. Likhanov, “Packet Switching in a Channel
Without Feedback,” Probl. Information Transmission, 19 pp. 6984
(AprilJune 1983).
TsL88. B.S. Tsybakov and N.B. Likhanov, “Upper Bound on the Capacity
of a Random Multiple Access System,” Problemy Peredachi Infor
matsii, 23(3) pp. 224236 (January 1988).
TsM78. B.S. Tsybakov and V.A. Mikhailov, “Free Synchronous Packet
Access in a Broadcast Channel with Feedback,” Probl. Information
Transmission, 14(4) pp. 259280 (OctoberDecember 1978).
Section : REFERENCES 165
TsM79. B.S. Tsybakov and V.A. Mikhailov, “Ergodicity of a slotted
ALOHA system,” Probl. Information Transmission, 15(4) pp. 301
312 (OctoberDecember 1979).
Tsm80. B.S. Tsybakov and V.A. Mikhailov, “Random Multiple Packet
Access: PartandTry Algorithm,” Probl. Information Transmission,
16(4) pp. 305317 (OctoberDecember 1980).
Tsy80. B.S. Tsybakov, “Resolution of a Conﬂict of Known Multiplicity,”
Prob. Information Transmission, 16(2) pp. 134144 (AprilJune
1980).
Tsy85. B.S. Tsybakov, “Survey of USSR Contributions to MultipleAccess
Communications,” IEEE Trans. on Information Theory, IT31(2)
pp. 143165 (March 1985).
TTH88. T. Takin, Y. Takahashi, and T. Hasegawa, “An Approximate Analy
sis of a Buffered CSMA/CD,” IEEE Trans. on Communications,
COM36(8) pp. 932941 (August 1985).
VvT83. N.D. Vvedenskaya and B.S. Tsybakov, “Random Multiple Access
of Packets to a Channel With Errors,” Prob. Information Transmis
sion, 19(2) pp. 131147 (AprilJune 1983).
ZhR87. W. Zhao and K. Ramamritham, “Virtual Time CSMA Protocols for
Hard RealTime Communication,” IEEE Trans. Software Engineer
ing, SE13(8) pp. 938952 (August 1987).
166 : REFERENCES
APPENDIX A
MATHEMATICAL FORMULAE AND BACKGROUND
This appendix summarizes some of the important properties and results regarding queue
ing and Markov processes that are used in the text. This is only a list; the reader is
expected to be acquainted with the items on the list (and with stochastic processes in gen
eral) to the extent that he/she understands them and knows how to make use of them. The
material here is based on textbooks by Ross [Ros72] and Kleinrock [Kle76].
In this appendix as well as throughout the text we adopt a consistent notation as follows. A
random variable is denoted by a letter with a tilde, e.g., . For this random variable we
denote by its probability distribution function, by its probability density func
tion, by the Laplace transformof , and by its kth moment. If is a discrete
random variable then X(z) denotes its generating function. The expectation is denoted by
or just x. In general, a discrete stochastic process is denoted .
Markov Chains
Consider a ﬁnite or countable set and a stochastic process
in which designates the state of the process. We say that the process is in state j at
time n if . For conciseness we consider the states as being the set of integers, i.e.,
. Such a stochastic process is a Markov chain if
that is, the probability that at time n the process is in state j depends only on its state at
time n1 and not on prior history. The quantities are called the onestep transition
probabilities of the process at time n. When the transition probabilities are time indepen
dent, i.e., for all n, the chain is called homogeneous. The mstep transition prob
ability of a homogeneous Markov chain is deﬁned as
and is the probability of transitioning fromstate i to state j in exactly msteps. The onestep
probabilities can be arranged in a matrix P called the transition matrix.
Two states of a Markov chain are said to communicate if and only if there is a positive
probability that the process ever be in state j after having been in state i, and vice versa. In
fact all communicating states form a class of states. A Markov chain having but one class
of states is called irreducible. There is a variety of other ways to characterize states in
Markov chains, notably periodicity and ergodicity (whose deﬁnition we leave out); in this
textbook we are interested only in irreducible, aperiodic and homogeneous chains. Since
x
˜
F
x
˜
x ( ) f
x
˜
x ( )
F
x
˜
*
s ( ) f
x
˜
x ( ) x
˜
k
x
˜
x x
˜
n
n 0 ≥ , { }
E E
0
E
1
… , , { } = x
˜
n
n 0 ≥ , { }
x
˜
n
E ∈
x
˜
n
E
j
=
E
j
j =
x
˜
n
j = x
˜
n 1 –
i = x
˜
n 2 –
i
n 2 –
= … x
˜
0
i
0
= , , , [ ] Prob x
˜
n
j = x
˜
n 1 –
i = [ ] Prob p
ij
n
·
∆
=
p
ij
n
p
ij
n
p
ij
=
p
ij
m ( )
x
˜
m n +
m n + = x
˜
n
i = [ ] Prob =
168 : APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND
ergodicity plays an important role in the analyses the proof of ergodicity is included in the
text in the appropriate places.
The result most frequently used in the text stems from the following theorem:
For an irreducible ergodic Markov chain the limit
exists and the values π
j
are the unique nonnegative solutions of the
set of equations
Several remarks and corollaries result from the above theorem. First, the set of equations
can be written in matrix form as
where is the row vector of the values π
i
. This notation is especially useful when the
number of states is ﬁnite as the tools of linear algebra can be put to work. It can also be
shown that if the set of equations has a solution such that then the chain is
ergodic. The probabilities π
i
are (interchangeably) referred to in the literature as limiting
probabilities, steadystate probabilities, stationary probabilities, or invariant probabilities.
In general the term “steadystate” refers to the operation of the process after a long time,
i.e., for large values of n. The limiting probability π
i
is the steady state probability that the
process is in state i; it is also the proportion of time that the process stays in state i (the lat
ter remains true for periodic chains).
Recurrent Markov chains are members of another family of stochastic processes known as
regenerative processes. This special family of processes possesses the property that there
exist times t
0
, t
1
,... such that the behavior of the process after time t
i+1
is a repetition, in a
probabilistic sense, of the behavior of the process after time t
i
. Referring to the time
between two regeneration points as a cycle, we have that
This relation is used extensively in the textbook when a (regenerative) system is modeled
as having two statesuseful and uselessthe ratio of which is a good measure of efﬁ
ciency.
π
j
p
ij
n ( )
n ∞ →
lim
·
∆
π
j
π
i
p
ij
i
∑
=
1 π
i
i
∑
=
π πP =
π
π
i ∑
∞ <
Proportion of time in state j
Exected time in state j in a cycle
Expected cycle length
 =
Section : APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND 169
Residual Life
Consider a stochastic (renewal) process that marks time instants on the time axis in a way
that the length of the marked intervals, denoted , are independent and identically
distributed (i.i.d.) according to a common distribution (or density and
expected value . At some random time t, while the process is ongoing, it is
sampled and we are interested in the distribution and moments of the residual time i.e., the
time until the next marked point.
If denotes the residual time then:
Where F*( . ) is the Laplace transform of the corresponding probability density function.
The age of the process, i.e., the time from the beginning of the interval to the sampled
point has the same distribution as .
The M/G/1 Queue
Consider a queueing system in which arrivals occur according to a Poisson process with
parameter λ and in which the service rendered to the customersis distributed accord
ing to a distribution B(t). In such a queueing systemthe number of customers in the system
as seen by an outside observer equals that seen by an arriving customer which equals that
seen by a departing customer. With this in mind we make the following notation:
b(t)  Probability density function of the service time.
B*(s)  Laplace transform of b(t).
 Load factor
 Steady state number of customers in queue
Q(z)  Generating function of
 Time spent in the system (delay time)
D  Average delay time
 Queueing time (time spent in queue)
W  Average queueing time.)l
The following holds for an M/G/1 queueing system:
x
˜
n
n 0 ≥
F
x
˜
x ( ) f
x
˜
x ( )
E x [ ] x =
y
˜
f
y
˜
y ( )
1 F
x
˜
y ( ) –
x
  =
F
Y
˜
*
s ( )
1 F
x˜
*
s ( ) –
sx
 =
E y
˜
[ ]
x
˜
2
2x
 = E y
˜
[ ]
2
x
˜
3
3x
 =
y
˜
x
˜
ρ λx =
q
˜
q
˜
D
˜
W
˜
170 : APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND
Q z ( ) B
*
λ λz – ( )
1 ρ – ( ) 1 z – ( )
B
*
λ λz – ( ) z –
 =
E q
˜
[ ] ρ
λ
2
x
˜
2
2 91 ρ – ( )
 + =
W
*
s ( )
1 ρ – ( )s
s λ – λB
*
s ( ) +
 =
W
λx
˜
2
2 1 ρ – ( )
 =
D
*
s ( ) B
*
s ( )
1 ρ – ( )s
s λ – λB
*
s ( ) +
 =
D x W + x
λx
˜
2
2 1 ρ – ( )
 + = =
GLOSSARY OF NOTATION
Symbol Meaning (Form of Usage)
a Arrival per slot (a
i
, )
Normalized endtoend delay
A Number of arrivals ( , A(z), A
k
(z))
B Busy period, length of CRI ( , B, B
n
, B
ni
)
b Backlog departure rate (b, b
i
(n))
C Cycle length ( )
c Constant (c, c
n
)
D Delay ( , D, , D*(s), D(k))
d Distance between assignments (Generalized TDMA) (d(k))
E Expectation (E[ . ])
F General function (usually distribution) (F( . ))
f General function (f( . ))
G Normalized channel (offered) load
Generating function (G(z), G
n
(z))
g Channel (offered) load
I Idle period ( , I)
L Number of packets in a message ( , L, , L(z), L*(s))
M Number of users in the system
N Population size ( , N, )
P Packet size (P, )
Probabilities (P
suc
, P
n
)
p probability (p, p
i
, p
ij
)
P Transition matrix
Q Generating function of q (Q(z), Q
k
(z))
Probabilities (Q
i
(n))
q Number of packets in queue ( , , q(k))
R Channel transmission rate
S Throughput (S, S
n
, S
n(
k))
s Laplace variable
T Slot size
Packet length (time)
Transmission period ( , , T, T
i
, T
c
)
t General time ( , t)
U Useful (successful) time in a cycle ( , U , U
n
)
a
˜
A
˜
B
˜
C
˜
D
˜
D
ˆ
I
˜
L
˜
L
˜ 2
N
˜
N
˜
k
P
˜
q
˜
q
˜
j
T
˜
T
˜
i
t
˜
U
˜
U
˜
n
172 : GLOSSARY OF NOTATION
V Second moment of CRI length (V
n
, V(z))
W Waiting time
x General variable (x, x
i
, x
l
, x
r
)
General service time ( , )
z Generating function variable
α CRI length bound (α
m
)
β Root of unity (β
m
)
δ Impulse function (δ( . ))
General (bounding) number
∆ Step function (∆ ( . ))
CRI epoch length
γ Collision detection time (CSMA/CD)
λ Arrival rate
ν General probability
π Invariant probabilities (π
i
), Probability vector
ρ load factor
σ General probability
τ Minislot duration, endtoend propagation delay
Symbol Meaning (Form of Usage)
x
˜
x
PREFACE
Computer communication networks have come of age. Today, there is hardly any professional, particularly in engineering, that has not been the user of such a network. This proliferation requires the thorough understanding of the behavior of networks by those who are responsible for their operation as well as by those whose task it is to design such networks. This is probably the reason for the large number of books, monographs, and articles treating relevant issues, problems, and solutions. Among all computer network architectures, those based on broadcast multiaccess channels stand out in their uniqueness. These networks appear naturally in environments requiring user mobility where the use of any ﬁxed wiring is impossible and a wireless channel is the only available option. Because of their desirable characteristics multipleaccess networks are now used even in environments where a wired pointtopoint network could have been installed. The understanding of the operation of multiple access network through their performance analysis is the focus of this book. In many aspects broadcast multiple access networks are similar, or even identical, to pointtopoint networks. The major difference lies in the way in which the data links are used by the network nodes. This book concentrates on mechanisms for link access in multiaccess communication systems including local area networks and radio networks. The text has a mathematical orientation with emphasis on insight, that is, the analysis is mathematical in nature yet the purpose is understanding the operation of the systems through their analysis. We have assumed acquaintance with probabilistic modeling of systems, some knowledge in stochastic processes and just a bit of elementary queueing systemsall on the level of undergraduate studies. With this knowledge the reader should be able to follow all the mathematical derivations. While some of the material covered in this book appeared in other books, the vast body of the text has appeared only in professional journals in their typical cryptic language and inconsistent notation. Some of the material appears here for the ﬁrst time. Because of the inconsistent notation used in the diverse exposition of the material a great emphasis is placed on uniform notationidentical concepts have been assigned the same notation throughout the book. This should make it ever so easier to understand the concepts and compare derivations and results. The subjects covered in the book were chosen judiciously. Each subsection presents a communication system whose nature differs from the others in the system characteristics, the purpose of the system, or the method of analysis. Via this approach we cover all types of multiaccess systems known to date and most of the analytical methods used in their analysis. The introduction chapter, presents our way of classifying the multiple access protocols. It is here that we present the concepts that dictate the order in which we address the protocol in the rest of the book. The introduction also includes a thorough deﬁnition of the model
ii
: PREFACE
that is used throughout the book in the analysis of the protocols. The reader is encouraged to read this section before proceeding to the main text and to keep on referring to this chapter while studying the material. The main body of the book resides in chapters 2, 3, and 4. Chapter 2 addresses the traditional conﬂict free protocols. In chapter 3 we address the random retransmission family of protocols that has become so popular in local area and satellite networks. In chapter 4 we present the family of collision resolution protocols that are algorithmically somewhat more complex but that exhibit many other desirable features. These chapters are to some extent independent from one another, in the sense that one can be studied without the others and because different analytical techniques are used. We do, however, feel that studying the subjects in the order presented contributes signiﬁcantly to the understanding of the material. Finally, in chapter 5 we scan brieﬂy other major subjects that have been studied and published in the open literature but that is beyond the scope of this book. This book is aimed both at the student and the professional engineer. From a curriculum standpoint the material here contains more than can be covered in a single semester of studies. However, with sufﬁcient mathematical background an in depth course can be constructed covering enough material so that the student could complete the rest of the material himself. A small set of problems and exercises is included at the end of every chapter. These exercises are, in many cases, nontrivial and require a bit of time to solve; they are meant to enhance the reader’s knowledge and train him in the techniques covered in the text. To enhance its use as a reference for professionals, the book points the reader to variations and other systems through an extensive bibliography. One might expect the professional to study the basic material as presented in the book and then follow the bibliography to an analysis of a system that might be closer to the one he seeks. Haifa, Israel June 1989 Raphael Rom Moshe Sidi
.....2 1............ GENERALIZED TDMA .... RELATED ANALYSIS ......................................1.............60 3....3..............................................................................................................2...........5............ 40 APPENDIX A: Distribution of the Mod Function ..........................................................................................................................2...............................................................................Distribution............................... FDMA ...........5 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS ...62 3.................................................................................. Message Delay Distribution..................................................... 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS ......................................................................................................................................80 Throughput Analysis ....58 The Capture Phenomenon .......4.1........................................... Stabilizing the Aloha System...........67 3......3.............................CONTENTS PREFACE ......64 3...1.........................14 2................................................................................1 1.........................................2..........................1............................. DYNAMIC CONFLICTFREE PROTOCOLS ................................3..................................26 Expected Message Delay ..................................56 Expected Delay ............ SLOTTED ALOHA ........3....................................4....................47 3.........3...20 2...................38 EXERCISES ...1......83 Throughput Analysis ...........4.....2............................. 2.................... FREQUENCY DIVISION MULTIPLE ACCESS ...................2........................TDMA Comparison ............................. PROTOCOL CLASSIFICATION .......................................................................54 Throughput Analysis .............................. Expected Delay ..................1.........................1...............................FINITE NUMBER OF USERS ...............1................................ i CONTENTS ..............................................................................................................................2............. (IN)STABILITY OF ALOHA PROTOCOLS ....................................84 ............................................................................................................................. SLOTTED ALOHA ................................................................... Number of Packets at Allocated Slots ...............5.....80 4.20 Expected Number of Packets at Allocated Slots.............2..4...............................24 Message Delay Distribution.....................................................................................................................12 2............................................... iii CHAPTER 1: INTRODUCTION ..9 2........................1.......35 2.. 43 CHAPTER 3: ALOHA PROTOCOLS ............... 74 CHAPTER 4: CARRIER SENSING PROTOCOLS ................................................................................................................ NONPERSISTENT CARRIER SENSE MULTIPLE ACCESS ........79 4.2............................................................47 3...........9 2............................49 3.................. 2.......................4..........................12 2..................2. Analysis.29 2.......................... PURE ALOHA .....................................................................................................................3...........................................................................................................................................................................53 SteadyState Probabilities ..............................3...................... Delay Distribution......... RELATED ANALYSIS .............1.........71 EXERCISES .4........... 2..............15 2......................................................................... TIME DIVISION MULTIPLE ACCESS ....... THE SYSTEM MODEL ...........33 2...........................
.............................................................................................. 125 5................................................. 171 .............................................................. 129 5.............................................. Throughput Analysis .............................................1...................140 EXERCISES .107 5....2.............94 Throughput of Slotted Nonpersistent CSMA/CD .. The Modified BinaryTree Protocol............................. The Epoch Mechanism........................................ 96 Throughput of Slotted 1Persistent CSMA/CD ............................... 139 5.2..........................................................135 5................................ 90 Throughput of Slotted 1PCSMA .......................................................................3.............................................................. 143 APPENDIX A: Moments of Collision Resolution Interval Length .... 153 APPENDIX A: MATHEMATICAL FORMULAE AND BACKGROUND ......149 Multihop Networks ................. 99 4................ Moments of the Conditional Length of a CRI.................................................................................... 119 5............. LIMITED SENSING PROTOCOLS ...................................123 5..................................................................................89 Throughput of Slotted Nonpersistent CSMA ... ENHANCED PROTOCOLS ..3...............................................................................4...........2......................................... RELATED ANALYSIS ..................2......................101 EXERCISES ........... RELATED ANALYSIS ..... 92 4............................................ CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION ........ 87 4.............4................3.................................................................................................................................................................................1..1................ 114 5......... SLOTTED CARRIER SENSE MULTIPLE ACCESS .............................................................................................................................5...........................................................................................................................3............... 146 CHAPTER 6: ADDITIONAL TOPICS ......................................................................................... 167 GLOSSARY OF NOTATION ......................1....................... THE BINARYTREE PROTOCOL ....................... Bounds on Expected Packet Delay... 150 REFERENCES ........................ 150 Multichannel Systems .1..........................3........................ 104 CHAPTER 5: COLLISION RESOLUTION PROTOCOLS ........................................... 110 Bounds on the Moments ........................2................................................................................107 5........... 86 State Probabilities ...... Stability Analysis ..................................... 121 5...............................2........1.................................................................... 123 5........................................................................................................................................... 149 Multistation Networks .................................... The Clipped BinaryTree Protocol ..........4 : CONTENTS Transmission Period Lengths .........................1.............................................................
mostly determined at network design time. Subsequent topological changes are quite hard (and costly) to implement. The MAC layer is between the Data Link Control (DLC) layer and the Physical Layer. If. The channel then becomes the shared resource whose allocation is critical for proper operation of the network. not available. and some local area networks. To make a transmission successful interference must be avoided or at least controlled. This basic property has its advantages and disadvantages. In terms of known networking models. In other words. Pointtopoint channels. that is. these protocols reside mostly within a special layer called the Medium Access Control (MAC) layer. channels. To brieﬂy summarize the . however. However. Moreover. beyond being very economical. or when dynamic topologies are required broadcast channels can be used. where a resource is shared (and thus accessed) by a number of independent users. This book focuses on access schemes to such channels known in the literature as Multiple Access Protocols. and the channel is the physical medium over which signals. representing data. Broadcast channels appear naturally in radio. Traditional networks make use of pointtopoint channels. transmissions over a broadcast channel interfere. a storage facility or a server of any kind. and protocols. in the sense that one transmission coinciding in time with another may cause none of them to be received. When pointtopoint channels are not economical. The need for multiple access protocols arises not only in communications systems but also in many other systems such as a computer system. a message is destined to a large number of destinations then a broadcast channel is clearly superior. indeed. are advantageous due to their noninterference feature namely. These protocols are nothing but channel allocation schemes that posses desirable performance characteristics. The switches (or nodes) are the hardware entities that house the data communication functions. the success of a transmission between a pair of nodes is no longer independent of other transmissions. In this book we mainly address a shared communications channel. require the topology to be ﬁxed. the protocols are the sets of rules and agreements among the communicating parties that dictate the behavior of the switches. a broadcast channel is one in which more than a single receiver can potentially receive every transmitted message. Informally stated. travel from one switch to another. channels that are dedicated to an (ordered) pair of users. that transmission between a pair of nodes has no effect on the transmission between another pair of nodes even if these two pairs have a common node. These channels. satellite.CHAPTER 1 INTRODUCTION Three major components characterize computer communication networks: switches. in a typical case a message is destined to a single or a very small number of destinations and wasteful processing results in all those switches for whom the message is not intended. such as the OSI reference model.
First and foremost. the dynamic one allocates the channel based on demand so that a user who happens to be idle uses only little. There are various ways to classify multiple access protocols. all pollingtype access protocols The classiﬁcation of Figure 1. Hence. The MiniSlotted Alternat . leaving the majority of its share to the other. communication between one pair of nodes may inﬂuence the communication between other pairs. We therefore classify these protocols and take samples of the various classes to be analyzed in the text. In particular there is no single node coordinating the activities of the others (whose protocol. whenever made. 1.e. that is. differs from the rest). in general. At the highest level of the classiﬁcation we distinguish between conﬂictfree and contention protocols. Such an allocation can be done by various reservation schemes in which the users ﬁrst announce their intent to transmit and all those who have so announced will transmit before new users have a chance to announce their intent to transmit.1. we are interested in noncentralized multiple access protocols. To counter the static allocation. is a successful one.. the channel can be divided by giving the entire frequency range (bandwidth) to a single user for a fraction of the time as done in Time Division Multiple Access (TDMA). PROTOCOL CLASSIFICATION The multiple access protocols suggested and analyzed to date are too numerous to be all mentioned here. Another common scheme is referred to as token passing in which a single (logical or physical) token is passed among the users permitting only the token holder to transmit. for example. for this purpose. from a time. and (4) interference is inherent. These are protocols in which all nodes behave according to the same set of rules. This also excludes. The channel resources can be viewed. frequency. or giving a fraction of the frequency range to every user all of the time as done in Frequency Division Multiple Access (FDMA). (3) we conﬁne ourselves to the Medium Access Control layer freeing us from worrying about network wide functions such as routing and ﬂow control. more active users. (2) users typically hear one another. thereby guaranteeing noninterference. of the channel resources. by necessity. or mixed timefrequency standpoint. i.1 attempts to exhibit the underlying balance and symmetry behind existing multiple access protocols. that (1) sending a message to multiple users in a single transmission is an inherent capability. Our classiﬁcation is presented in Figure 1. will not be interfered by another transmission. if at all. Conﬂictfree transmission can be achieved by allocating the channel to the users either statically or dynamically. Examples of such classiﬁcations appear in [KSY84] and [Sac88]. More precise deﬁnitions appear later in the introduction and in the description of the various protocols. or providing every user a portion of the bandwidth for a fraction of the time as done in spreadspectrum based systems such as Code Division Multiple Access (CDMA).2 CHAPTER 1: INTRODUCTION environment in which we are interested we assume. In this description we consider the channel as the focal point and refer to the nodes transmitting through the channel as its users. Conﬂictfree protocols are those ensuring a transmission.1.
1.: PROTOCOL CLASSIFICATION Multiple Access Protocols Contention Conﬂict Free Dynamic Resolution Static Resolution Dynamic Allocation Static Allocation Time of Arrival Probabilistic ID Probabilistic Reservation Token Passing Time and Freq Freq.Section 1. Based Time Based FIGURE 1.1: Classiﬁcation of Multiple Access Protocols 3 .
such as might be the case with bursty users. Each example presents a system or a protocol whose nature differs from the others either in the system characteristics. A static resolution can be based. on user ID’s or any other ﬁxed priority assignment. here too. or in the purpose of the system. Dynamic resolution. Moreover. the smallest ID (this is done in some treeresolution protocols). idle users do consume a portion of the channel resources. say. Alternatively resolution can be probabilistic but such that the statistics change dynamically according to the extent of the interference. The protocol must prescribe a way to resolve conﬂicts once they occur so all messages are eventually transmitted successfully. For example. the necessity arises to resolve the conﬂicts. or in the method of analysis. . Static resolution means that the actual behavior is not inﬂuenced by the dynamics of the system. In contention schemes idle users do not transmit and thus do not consume any portion of the channel resources. taking the chance of having to resolve the interference compensates for the resources that have to be expanded to ensure freedom of conﬂicts. The main body of the text contains typical examples of multiple access protocols that are analyzed and discussed. If the probability of interference is small. As in the conﬂictfree case. resolution can be based on time of arrival giving highest (or lowest) priority to the oldest message in the system. The resolution process does consume resources and is one of the major differences among the various contention protocols. this portion becomes major when the number of potential users in the system is very large to the extent that conﬂictfree schemes are impractical. namely taking advantage and tracking system changes is also possible in contentionbased protocols. These examples were chosen judiciously to cover the above classes of protocols. meaning that the transmission schedule for the interfering users is chosen from a ﬁxed distribution that is independent of the actual number of interfering users. in most conﬂictfree protocols. as is done in Alohatype protocols and the various versions of Carrier Sense Multiple Access (CSMA) protocols. and the exponential backoff scheme of the Ethernet standard fall into this category. for example. Estimating the multiplicity of the interfering packets. Via this approach we cover all types of noncentralized multiple access protocols known to date and most of the analytical methods used in their analysis. When contentionbased multiple access protocols are used. whenever they occur. meaning that whenever a conﬂict arises the ﬁrst user to ﬁnally transmit a message will be the one with.4 CHAPTER 1: INTRODUCTION ing Priority (MSAP) and the Broadcast Recognition Access Method (BRAM) are examples of multiple access protocols that belong to this class. Contention schemes differ in principle from conﬂictfree schemes since a transmitting user is not guaranteed to be successful. A static resolution can also be probabilistic. both static and dynamic resolutions exist.
We take the throughput of the channel to mean the aggregate average amount of data that can be transported through the channel in a unit of time. THE SYSTEM MODEL The main body of the text contains typical examples of multiaccess protocols that are analyzed and discussed. In “fair”. meaning that either the source and destination can communicate directly or there exists a node that communicates directly with both the source and the destination. In those cases where only a single transmission can be successful at any time (as is typical in many singlehop systems) the throughput. Evaluation of those input rates for which the system remains stable is therefore essential. Two other criteria are also of interest: system stability and message storage requirement (buffer occupancy). Buffer occupancy is clearly an important performance measure since having to provide larger buffers generally translates into more costly and complex implementation. and hence no routing of messages is required. even smaller than the maximal transmission rate in the channel. In the analysis we are interested mainly in throughput and delay characteristics. and on the sensitivity of the receiver at the receiving node. that is. Dualhop topologies are those in which messages from a source to a destination do not have to pass more than two hops. dualhop and multihop topologies. Higher buffer occupancy usually also means longer message delays and vice versa. or homogeneous systems we expect these to be almost identical. In delay calculations we generally consider the time from the moment a message is generated until it makes it successfully across the channel. on the distance between the two nodes. In general. We postpone further deﬁnitions to the sections dealing directly with stability. Basically.: THE SYSTEM MODEL 5 1. This conﬁguration is peculiar in a broadcast channel context since the intermediary node can be affected by the behavior of two other nodes that do not hear one another.2. Here one must distinguish between the user and the system measures as it is possible that the average delay measured for the entire system does not necessarily reﬂect the average delay experienced by any of the users.2.Section 1. In this text we assume a symmetric connectivity pattern. The notion of stability arises in this context because the protocol characteristics may be such that some message generation rates. connectivity patterns can be classiﬁed into three categories known as singlehop. every node can successfully transmit to every node it can hear. To analyze multiple access protocols one must make assumptions regarding the environment in which they operate. The multihop topology is the most general one in which beyond (and in addition to) the problems encountered in the single and dualhop topologies one must address routing issues that become complex if the topology is allowed to vary dynamically. the ability of a node to hear the transmission of another node depends on the transmission power used. Hence. as thus deﬁned. In a singlehop topology all users hear one another. in each and every protocol we must address the following issues: • Connectivity. cannot be sustained by the system for a long time. equals the fraction of time in which the channel is engaged in the successful transmission of user data. .
by utilizing a different channel or by being able to determine the feedback locally. • Message size. in the case of a noisy channel). furthermore. Feedback is the information available to the users regarding activities on the channel at prior times. that due to its length.g. Common feedback information indicates whether a message was successfully transmitted. though. It is generally assumed that the feedback mechanism does not consume channel resources. if a single transmission heard at a node is always received correctly. Other operations. Various degrees of synchronism are required in the slotted protocols we consider. Another channel type is the capture channel in which one or more of the colliding transmissions “captures” the receiver and can be received correctly. or by explicit acknowledgment messages sent by the receiving node. a message cannot be transmitted in a single transmission and must therefore be broken into smaller units called packets each of which can be transmitted in a single channel access. for example. or providing uncertain feedback (e. • Feedback/Acknowledgment. A collision channel is one in which all the colliding transmissions are not received correctly and in most protocols have to be retransmitted. Packet size is measured by the time required to transmit the packet once access to the channel has been granted. or the channel was idle. Recently nofeedback protocols have also been proposed. • Message generation. In these systems transmissions of data start only at slot boundaries. For every protocol we assume that there exist some instants of time (typically slot boundaries or end of transmissions) in which feedback information is available. The channel is the medium through which data is transferred from its source to its destination. A message consists of an integral number of packets although the number of packets in a message can vary randomly. the channel may be such that errors between successive transmissions are not independent. This information can be obtained by listening to the channel. Other feedback variations include indication of the exact or the estimated number of colliding transmissions. Other possible channels include the noisy channel in which errors may occur even if only a single transmission is heard at a node and. Another important class of systems is that of. such as determining activities on the channel can be done at any time. Few analyses can be found in the literature accommodating nonPoisson generation processes. Typically. In this text we deal with an errorless collision channel. A channel is errorless. at the receiver. • Synchronism. or a collision took place. two or more transmissions overlap in time wholly or partially. All users are statistically identical and generate new messages according to a Poisson process.. . Collision is a situation in which. we assume all packets to be of equal size and variations include randomly varying packets. It is possible.slotted systems in which a global clock exists that marks equally long intervals of time called slots.6 CHAPTER 1: INTRODUCTION • Channel type. The basic unit of data generated by a user is a message. Yet another case is a channel in which coding is used so that even if transmissions collide the receiver can still decode some or all of the transmitted information. Variations include cases in which all users are not the same and in particular one heavy user and many identical small ones. Users are generally not assumed to be synchronized and are capable of accessing and transmitting their data on the channel at any time.
.Section 1. The expectation is denoted by x or just x. For such cases contention based protocols are the only possible solution.: THE SYSTEM MODEL 7 • User population. If x is a discrete random variable ˜ then X(z) denotes its generating function. • Buffering capability. A random variable is ˜ denoted by a letter with a tilde.2. Every user is a different. n ≥ 0 } . The number of users in the system can be ﬁnite or inﬁnite.g. generally independent entity. In a typical analysis it is assumed that every user has a buffer for a single message and that it does not generate new messages unless its buffer is empty. and by x k its kth moment. at each user. a discrete stochastic process is denoted { x n. by f x(x) its probability density function. Messages generated by the user are stored in a buffer. x . by F x ˜ ˜ ˜ ˜ Laplace transform of f x(x) . Other alternatives include more buffering. For this random variable we denote by F x(x) its ˜ *(s) the probability distribution function. One interesting observation is that most conﬂictfree protocols are useless if the user population increases beyond a certain point. A word about notation Throughout the text we adopt a consistent notation as follows. both inﬁnite and ﬁnite. In ˜ general. . e.
8 CHAPTER 1: INTRODUCTION .
The main advantage of FDMA is its simplicityit does not require any coordination or synchronization among the users since each can use its own frequency band without interference. 2. and the Time Division Multiple Access (TDMA) in which the entire bandwidth is used by each user for a fraction of the time. and a receiver for each band (which can be implemented as a single receiver for the entire range with a bank of bandpass ﬁlters for the individual bands). it is possible to divide the frequency range unevenly. The two most well known protocols in this class are the Frequency Division Multiple Access (FDMA) in which a fraction of the frequency bandwidth is allocated to every user all the time. parts of the channel might be idle even though some users have data to transmit. For both the FDMA and the TDMA protocols no overhead. in the form of control messages. adding a new user to the network requires equipment modiﬁcation (such as additional ﬁlters) in every other user.. . proportional to the demands. Dynamic channel allocation protocols attempt to overcome this drawback by changing the channel allocation based on the current demands of the users. FREQUENCY DIVISION MULTIPLE ACCESS With Frequency Division Multiple Access (FDMA) the entire available frequency band is divided into bands each of which serves a single user. is incurred. For more details the reader may consult any of the many texts treating FDMA that have been published. since when one user is idle his share of the bandwidth cannot be used by other users. The ﬁrst three sections are devoted to static channel allocation strategies in which channel allocation is predetermined (typically at network design time) and does not change during the operation of the system. predetermined. However.1. An important advantage of conﬂictfree access protocols is the ability to ensure fairness among users and the ability to control the packet delaya feature that may be essential in realtime applications..CHAPTER 2 CONFLICTFREE ACCESS PROTOCOLS Conﬂictfree protocols are designed to ensure that a transmission. whenever made. Every user is therefore equipped with a transmitter for a given.e. frequency band. i. is also the cause of waste especially when the load is momentarily uneven. e. however. It should be noted that if the users have uneven long term demands. This. This is achieved by allocating the channel to the users without any overlap between the portions of the channel allocated to different users. due to the static and ﬁxed assignment. FDMA is also not ﬂexible. by Stallings [Sta85] or the one by Martin [Mar78].g. is not interfered by any other transmission and is therefore successful.
then the service time ˜ ˜ afforded to every packet is the random variable T = M P ⁄ R . Each node can therefore be viewed as an M/G/1 queue.10 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS To evaluate the performance of the FDMA protocol let us assume that the entire channel can sustain a rate of R bits/sec which is equally divided among M usersR/M bits/sec for each. all parameters relating to FDMA can be obtained by applying known results of the corresponding queue discipline. . Consider a typical user that generates packets according to a Poisson process with rate λ packets/sec.= T + 2 ( 1 – λT ) 2 ( 1 – λx ) ˜ ˜ where T = E [ T ] and T 2 = E [ T 2 ] .. and his buffering capabilities are not limited. using the known system delay time formula for M/G/1 queueing systems we get that the expected delay of a packet is (see Appendix) λx 2 λT 2 D = x + . Since the individual bands are disjoint.e.1).1: FDMA System Model ˜ process for that user. there is no interference among users’ transmissions and the system can therefore be viewed as M independent queues (See Figure 2. To evaluate the throughput of the individual user we note that every bit transmitted is a “good” bit and thus the individual throughput is the fraction of time the individual server is busy (i. Each of these queues has an individual input process governing the packet generation ˜ T λ 1 ˜ T λ 2 ˜ T λ Μ FIGURE 2. In general. The total throughput is M times the individual throughput while the average packet delay can be obtained by applying Little’s result to the individual queue. nonempty system). If the packet length is a random variable P . The time required for the trans˜ mission of a packet is T . Thus.
Section 2.1 + R 2 ( 1 – λT ) 2(1 – ρ) 2(1 – ρ) where ρ ∆ λT .1) by P/R. For values of throughput beyond 0.= T 1 + .9 1 Throughput ( S ) FIGURE 2.1 0.6 0. Normalizing the expected delay given in (2.1.5 0.= .1) ˆ Expected Delay ( D ) M=1000 1000 M=100 100 M=10 10 M=5 1 0 0.M = M . and substitutˆ ing S for ρ we get the normalized expected delay. Graphs depicting the throughput delay for various population sizes is depicted in Figure 2.4 0.2 0. 100000 FDMA 10000 (2.= 1 + . 1 +  2(1 – S ) 2 P⁄R 2(1 – S ) 1 – S which is the desired throughputdelay characteristic for FDMA with constant packet size.3 0. therefore.2: ThroughputDelay Characteristic for FDMA . In our case.: FREQUENCY DIVISION MULTIPLE ACCESS 11 When all packets are of equal length consisting of P bits each then the transmission time of every packet is (deterministically) equal to T= MP/R seconds. ρ equals the individual user’s throughput. D . Note that for a wide range the delay is rather insensitive to the throughput.= . = For M/G/1 systems ρ = λx is the fraction of time the server is busy.8 0.8 the delay increases quickly to values which cannot be tolerated .7 0. 2–S M D S 1 ˆ D = . In this case the queueing model corresponds to an M/D/1 queue in which T 2 = T 2 and therefore MP ρ ρ λT 2 D = T + .2. the time required to transmit a packet in a channel with rate R.
TIME DIVISION MULTIPLE ACCESS In the time division multiple access (TDMA) scheme the time axis is divided into time slots. that is. the users must be synchronized so that each one knows exactly when and for how long he can transmit. Higher moments can be obtained by taking higher order derivatives. In the most basic TDMA scheme every user has exactly one slot in every frame (see Figure 2.3). each such period is called a cycle or a frame. Note that for proper operation of a TDMA scheme. More general TDMA schemes in which several slots are assigned to one user within a frame. The slot assignments follow a predetermined pattern that repeats itself periodically. Every user is allowed to transmit freely during the slot assigned to it. is given by (see Appendix) s(1 – ρ) D *(s) = X *(s) s – λ + λX *(s) where X *(s) is the Laplace transform of the packet transmission time. 2. D *(s) = E [ e –sD ] .2. if ˜ D is a random variable representing the packet delay. during the assigned slot the entire system resources are devoted to that user. Delay Distribution The delay distribution of FDMA is also taken directly from M/G/1 results. For the case of equally sized packets we have X *(s) = e –sT and therefore s(1 – ρ) D *(s) = λ + ( s – λ )e sT The expected delay can be obtained by taking the derivative of D *(s) with respect to s at s=0 and for equally sized packets the expression is given in (2. referred to as generalized TDMA. preassigned to the different users. are considered in the next section. ˜ X *(s) = E [ e –sT ] . i.3: TDMA Slot Allocation .. Frame Frame Slot for User 1 2 3 4 5 6 7 1 2 3 4 5 6 7 t FIGURE 2. Further details on TDMA schemes can be found in texts such as [Kuo81.Sta85].e. then the Laplace transform of the ˜ ˜ probability density function (pdf) of D . Speciﬁcally.1).12 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS 2.1.1.
in the following we concentrate on the characteristics of one user. If the total rate of transmission is R bits/sec then the packet transmission time is T=P/R which is taken to be the slot size. To compute the queueing time (once the end of the current frame is reached) we observe that the queue behaves exactly like one with deterministic service time of Tc. Consequently. (3) the packet transmission time itself. Since all frames are of equal length. assume that the user transmits a packet. at the ﬁrst slot of every frame. as in the FDMA case. in other words. Wq. Assuming that the packet arrival processes of new packets to the different users are independent.T c + W q + T = ..2. the throughput equals the fraction of time the server is busy.: TIME DIVISION MULTIPLE ACCESS 13 To analyze the performance of a TDMA scheme consider a system composed of M users each transmitting equally long packets of P bits each.5 Tc. if he has any. The reason is that a user transmits a packet. every bit transmitted should be counted in the throughput or. (2) the queueing time to allow all the packets already queued to be transmitted and. Thus. If we assume a Poisson arrival process of λ packets/sec for the user and that the number of packets that can be stored in a queue is not bounded. The delay suffered by this packet has three components: (1) the time between its generation and the end of the current frame. is given by (see Appendix) ρ ρ ρ W q = . if he has any. Consider a typical packet generated by the user. 2 2 2(1 – ρ) 2(1 – ρ) As in the FDMA case. and without loss of generality.T c = . the average time between the packet generation time and the end of the current frame is 0. it follows that the queueing behavior of the queue at one user is independent of the queueing behavior of all others. every Tc seconds.. 2(1 – S ) and the normalized expected packet delay is obtained by dividing D by the time required to transmit a packet. The duration of the entire cycle is therefore Tc = MT.x = .MT 2(1 – ρ) 2(1 – ρ) 2(1 – ρ) where ρ=λTc=λMP/R (note that this is the same value for ρ as in the FDMA case). The total expected packet delay is therefore 1 1 ρ M D = . Of these components the ﬁrst and the third are readily known. we have S = ρ leading to M D = T 1 + .Section 2. We thus have that the expected queueing time of a packet.MT + T = T 1 + . Packet transmission time is T. which for an M/ D/1 queue equals ρ. .MT + . then the queueing time is identical to that of the queueing time in an M/D/1 queueing system in which the deterministic service time x is Tc. independently of any event in any of the other queues of other users.
. They are. 100000 TDMA 10000 ˆ Expected Delay ( D ) M=1000 1000 M=100 100 M=10 10 M=5 1 0 0.4: ThroughputDelay Characteristic for TDMA 2. similar to those of the FDMA (Figure 2.2).e.3 0. This difference is somewhat offset by the fact that a packet arriving to an empty queue may have to wait until the proper slot when a TDMA scheme is employed.8 0.7 0. A family of graphs of the expected packet delay versus the throughput for various values of M is given in Figure 2.14 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS D M ˆ D = .4 0.2.5 0.1 0.= 1 + T 2(1 – S ) (2..2) which is the desired throughput delay characteristic of a TDMA scheme.TDMA Comparison Comparing the throughput delay characteristics of FDMA and TDMA we note that P M .6 0.1. whereas in FDMA transmission starts right away.D FDMA = D TDMA + . FDMA . not surprisingly. every meaningful case) the TDMA expected delay is always less than that of FDMA and the difference grows linearly with the number of users and independent of the load! The difference stems from the fact that the actual transmission of a packet in TDMA takes only a single slot while in FDMA it lasts the equivalent of an entire frame..9 1 Throughput ( S ) FIGURE 2.– 1 ≥ D TDMA .4. R 2 We thus conclude that for M ≥ 2 (i.2 0.
2.. the queueing behavior of one user is independent of the queueing behavior of other users and therefore we consider a typical user in an M user system in which the slot size T equals the duration of packet transmission. that is k ≥ 0 messages arrived prior to it in the k ˜ ˜ ˜ ˜ same frame ( k and w are random variables and when w is given. As a practical matter. L its mean and L 2 its second moment. compute the distribution of the delay in such a system.e. A mes˜ ˜ sage consists of a random bulk of L packets. Assume that our tagged message is the ˜ + 1 st message arriving in that frame. 2. that at high throughput the dominant factor in the expected delay is proportional to 1/(1S) in both TDMA and FDMA and therefore the ratio of the expected delays between the two schemes approaches unity when the load increases. i. In this section we derive the Laplace transform of the message delay distribution. If one takes L(z)=z.2. k has a Poisson distri˜ bution with parameter λ ( T c – w ) ).. then each message consists of a single packet and the general arrival process we consider here degenerates to that of the previous section.2. At each arrival epoch a new message arrives.Section 2. .: TIME DIVISION MULTIPLE ACCESS 15 It must be remembered. which is necessary to keep the TDMA users from transmitting in a slot which is not their own. while FDMA performs slightly worse than TDMA. ˜ Consider an arbitrary “tagged” message arriving T c – w seconds after the beginning of the ˜ (j+1)st frame (i. ˜ The message delay D is deﬁned as the time elapsing between the message arrival epoch until after the transmission of the last packet of that message is completed. however.5 shows the relation among these quantities. if he has any. it is somewhat easier to implement since it does not require any synchronization among users. In this section we generalize the arrival process and demonstrate how one can. Let L(z) be the generating function of L . L ( z) = ˜ ∑ Prob [ L = l ]z l l=1 ∞ L = L'(z) z=1 = ∑ l=1 ∞ ˜ l ⋅ Prob [ L = l ] L2 + L = L''(z) z=1 = ˜ ∑ l 2 ⋅ Prob [ L = l ] l=1 ∞ We assume that messages arrive to the user according to a Poisson process at a rate of λ messages/sec. As before. Message Delay Distribution In the previous section we derived the expected packet delay in a TDMA system. The buffering capabilities of the user are not limited. Figure 2. in the ﬁrst slot of every frame. with fairly straightforward queueing theory techniques. Notice that the arrival process considered here is somewhat more general than that considered in the previous section. The analysis presented is essentially due to Lam [Lam77].e. w seconds before its end). The user transmits a packet.
The expression in (2. The k ˜ other component T c ∑ L i corresponds to waiting for transmission of packets that i=1 arrived since the beginning of the (j+1)st frame until the arrival of the tagged message.3) represents the time required to transmit all the packets already queued in the user buffer upon the tagged message arrival. and the last one to the tagged message itself.3) can be rewritten as k ˜ ˜ ˜ D = max(q j – 1. Thus. This second term is ˜ composed of two components: the ﬁrst T c [ max ( q j – 1. except the last one that requires only T seconds. . then this is the time the message will wait from the end of the (j+1)st frame until its ﬁrst packet will be transmitted.3) represents the time required to transmit the tagged message ˜˜ itself: ( L k + 1 – 1 )T c seconds to transmit all packets of the tagged message. 0)T c + w + T c ∑ ˜ i=1 ˜ ˜˜ Li + [ Lk + 1 T c + T – T c ] (2.3) i=1 ˜ where the ﬁrst term w represents the waiting time of the tagged message until the end of the (j+1)st frame. The second term in (2. the second to the current frame up to the arrival of the tagged message.4) The three components in this expression are statistically independent of one another: the ﬁrst contains quantities relating to previous frames. Then the delay of the tagged message is given by ˜ ˜˜ L i T c + [ ( L k + 1 – 1 )T c + T ] (2. 0 ) ] corresponds to waiting for transmission of packets˜ present in the queue at the beginning of the (j+1)st frame. When messages are transmitted in a FIFO manner.5: TDMA Packet Delay Components ˜ Denote by q j the number of packets awaiting transmission at the beginning of the (j+1)st frame.16 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS jth Frame (j+1)st Frame_ (j+2)nd Frame Tc ˜ qj 1 ˜ qj + 1 1 t ˜ w Arrivals of User 1 k ˜ ˜ ˜ D = w + max(q j – 1. the third term in (2. 0) + ∑ ˜ 1 2 3 ˜ k ˜ ( k + 1 )st Arrival of User 1 FIGURE 2. Finally.
: TIME DIVISION MULTIPLE ACCESS 17 ˜ D *(s) . ˜ Q j + 1(z) = A(z) { z –1 Q j(z) + ( 1 – z –1 )P [ q j = 0 ] } (2. Let aj be the number of packets arriving during the jth frame. the Laplace transform of the probability density function of D is the product of the Laplace transforms of these three components. j = 1. Q j(z) = E [ z q j ] . ˜ ˜ Let Qj(z) be the generating function of q j . 2. ˜ ˜ Let A(z) be the generating function of a i.6)..8) = + (1 – ˜ z –1 )P [ q j Combining (2. ….5) we obtain.[ L(z) ] m = e λT c [ L(z) – 1 ] ∑m = 0 m! ∞ (2. The packets awaiting transmission at the beginning of the (j+1)st frame are those packets that were queued up at the beginning of the jth frame.8) we obtain.6) ˜ where we used the fact that the arrival process is independent of the queue size. hence a is ˜ independent of q j . In addition. less the packet (if there were any) that has been transmitted in the ﬁrst slot of the jth frame. ˜ ˜ E [ z q j – ∆( q j ) ] = ˜ ∑k = 0 z k – ∆(k) P [ q j = k ] z –1 Q j(z) ∞ 1 ˜ ˜ = P [ q j = 0 ] + . the existence of steadystate (invariant) proba . A(z) = E [ z a ] . ˜ ˜ ˜ ˜ ˜ ˜ ˜ Q j + 1(z) = E [ z q j + 1 ] = E [ z q j – ∆(q j) + a ] = E [ z a ]E [ z q j – ∆(q j) ] (2. i. and because all frames are equally long we conclude that aj is independent of j.. ˜ ˜ A(z) = E [ z a ] = E { E [ z a m messages arrive in frame ] } = E { [ L ( z) m ] } = e –λT c ( λT c ) m .e.e.6) and (2. The derivation of A(z) is simple.5) is simple.[ Q j(z) – P [ q j = 0 ] ] z = 0] (2. Assuming that this Markov chain is ergodic (see below). the packets that arrived during the jth frame are also queued up at the beginning of the (j+1)st frame.2. } is clearly a Markov chain (see Appendix). ˜ ˜ We now turn to compute the quantity E [ z q j – ∆(q j) ] that appears in (2.7) where we used the fact that the generating function of the number of packets in a message is L(z). The explanation of (2.9) ˜ The chain { q j.Section 2. From (2. The relation ˜ among the values of q in consecutive frames is therefore given by ˜ ˜ ˜ ˜ q j + 1 = q j – ∆(q j) + a (2. Because the number of messages arriving in a frame depends only on the arrival process and on the frame length.5) where ∆(i) equals 0 for i=0 and equals 1 elsewhere.
e. Then from (2.7) into the last equation and denoting L *(s) ∆ L(e –sTc) we get = ( 1 – ρ ) ( 1 – e –sT c ) ˜ ˜ E [ exp [ – sT c ( q – ∆(q) ) ] ] = exp { λT c [ L *(s) – 1 ] } – exp { – sT c } ˜ (2. Let Q(z) be the gener˜ ating function of this invariant distribution. This is done by computing the conditional transform and then relaxing the conditions one by one as follows: . we get (by substituting z = e –sT c ) ˜ ˜ ˜ ˜ ˜ ˜ E [ exp [ – sT c ( q – ∆(q) ) ] ] = E [ [ exp ( – sT c ) ] q – ∆(q) ] = E [ z q – ∆(q) ] z = exp ( –sT c ) Substituting (2. Q(z) = lim Q j(z) and let P [ q = 0 ] be j→∞ the probability that the queue will be empty at the beginning of a frame in steadystate. 0 ) = q j – ∆(q j) . We now turn to compute the Laplace transform of the probability density function of the three components in equation (2. using (2. we use the normalization condition that ˜ Q(z) z = 1 = 1 and obtain that P [ q = 0 ] = 1 – A'(z) z = 1 = 1 – λLT c ∆ 1 – ρ (we used = L’Hopital’s rule in this calculation).18 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS bilities for the queue size at the beginning of a frame is guaranteed. ˜ Q(z) = A(z) { z –1 Q(z) + ( 1 – z –1 )P [ q = 0 ] } or ˜ P[q = 0](z – 1) Q(z) = A(z) z – A(z) ˜ The computation of P [ q = 0 ] is now easy. From (2. which renders the q’s an ergodic Markov chain and an altogether stable system.13) k ˜ To handle the second component of equation (2.12) and (2.9) we obtain. ˜ ˜ ˜ E [ z q j – ∆(q j) ] = z –1 Q j(z) + ( 1 – z –1 )P [ q = 0 ] (2. Therefore.8) we have.4)..11) We are interested in the steadystate behavior.= e λT c [ L(z) – 1 ] z – A(z) z – e λT c [ L(z) – 1 ] The derivation of equation (2. i.12) where we used the expression for Q(z) from (2.10) (2.7) we obtain (1 – ρ)(z – 1) (1 – ρ)(z – 1) Q(z) = A(z) . (1 – ρ)(z – 1) ˜ ˜ ˜ E [ z q – ∆(q) ] = z –1 Q(z) + ( 1 – z –1 )P [ q = 0 ] = z – A(z) (2. Starting with the ﬁrst component we note that ˜ ˜ max ( q j – 1.11) to obtain. Consequently.10). hence we let j → ∞ in (2.4) we deﬁne y ∆ w + T c ∑ L i and =˜ i=1 ˜ thus we need to compute E [ e –sy ] .10) holds if and only if ρ <1.
the Laplace transform of the message delay.e –sw Tc ∑ k=0 ∞ [ λ( T c – w ) ]k e L *k(s) .⋅ e s ( T c – T ) L *(s) T c s – λ + λL *(s) which is our desired result.⋅ .16) .4) we get ˜ ˜˜ E [ exp [ – s ( L k + 1 T c + T – T c ) ] ] = e s ( T c – T ) E [ e –sL T c ] = e s ( T c – T ) L *(s) (2. and (2.e. k is distributed according to a Poisson distribution with mean ˜ λ(Tc w) we now remove the condition on k ˜ ˜ ˜ ˜ ˜ ˜ E [ e –sy w = w ] = E [ e –sw L *k(s) w = w ] = e –sw E [ L *k(s) w = w ] ˜ ˜ = e –sw E [ z k w = w ] z = L *(s) = e –sw e λ ( T c – w ) ( z – 1 ) z = L *(s) = exp [ λT c [ L *(s) – 1 ] ] exp [ – w [ s – λ + λL *(s) ] ] ˜ Finally.: TIME DIVISION MULTIPLE ACCESS 19 ˜ E [ e –sy ˜ ˜ ˜ k = k. relaxing the condition on w and recalling that the latter is uniformly distributed on [0.13).15) Putting together the results of equations (2.–λ ( T c – w ) dw k! Finally. i. treating the third component of equation (2. Tc] we get ˜ ˜ E [ e –sy ] = E [ exp [ λT c [ L *(s) – 1 ] ] exp [ – w [ s – λ + λL *(s) ] ] ] ˜ = exp [ λT c [ L *(s) – 1 ] ] E [ exp [ – w [ s – λ + λL *(s) ] ] ] exp [ λT c – 1 ] ] – exp [ – sT c ] = T c [ s – λ + λL *(s) ] This last result can be obviously computed directly by Tc ˜ E [ e –sy ] [ L *( s ) (2.Section 2. (2.2. w = w ] = E exp – sw – sT c ∑ L i i=1 k = e –sw E ˜ exp – sT c ∑ L i i=1 k ˜ = e –sw [ E [ exp ( – sT c L ) ] ] k = ( e –sw )L *k(s) ˜ Noting that for a given w. To ﬁnd the expected message delay we take (the negative) derivative of D *(s) with respect to s computed at s=0 to obtain (2.15) we ﬁnally get 1 – e –sT c 1–ρ D *(s) = .14).14) = ∫ w=0 1 ..
20
CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
2 λT c L 2 1 D = T c L –  +  + T , 2 2 ( 1 – λT c L )
and in a normalized form M L2 ρ 1 D ˆ D =  = M L –  +  ⋅  + 1 L 2(1 – ρ) T 2 which, for the case that each message consists of a single packet ( L = L 2 = 1 ), reduces to the previously obtained result given in equation (2.2).
2.3. GENERALIZED TDMA
The allocation of a single slot within a frame to each user is reasonable in a homogeneous system. However, when the communication requirements of the users in a system are unequal, the channel will be utilized more efﬁciently if users with greater requirements have more slots allocated within each frame than users with light trafﬁc. This is done in the generalized TDMA scheme in which a user might be allocated more than one slot within a frame, with arbitrary distances between successive allocated slots. An example is depicted in Figure 2.6 where the frame consists of 7 slots and 4 users share the channel. Slots 1,2 and 4 are allocated to one user, slots 3 and 5 to another user, and slots 6 and 7 one to each of the other users. Consider a user to whom K slots are allocated in every frame and let d(k)≥1 (1≤ k≤K) be the distance between the (k+1)modK allocated slot and the K kmodK allocated slot. Notice that ∑k = 1 d(k) = T c . In each allocated slot, the user transmits one packet, if it has any. Without loss of generality we assume that the ﬁrst slot in a frame belongs to our user. The analysis of the performance of a user in a generalized TDMA scheme is more complicated than the analysis presented in the previous section for regular TDMA scheme. This chapter is interesting mainly due to the mathematical tools and techniques used. 2.3.1. Number of Packets at Allocated Slots  Distribution
˜ Let q j(k) be the number of packets awaiting transmission at the beginning of the kth allocated slot (1≤ k ≤ K) in the (j+1)st frame. We start by determining the generating function ˜ of the steadystate distribution of q j(k) (1≤ k ≤K). Steadystate distribution exists when λLT c ρ ∆  < 1 = K since λLTc is the expected number of packets arriving at the user during a frame and the user can transmit at most K packets during a frame. In steadystate the user’s throughput is S = Kρ = λLT c .
Section 2.3.: GENERALIZED TDMA
21
Frame
Frame
Slot for User
1
1
2
1
2
3
4
1
1
2
1
2
3
4
t
d(1)
d(2)
d(3)
1
1
2
1
2
3
4
1
1
2
1
2
3
4
t
FIGURE 2.6: GeneralizedTDMA Slot Allocation
From the operation of generalized TDMA we have that the number of packets awaiting transmission at the beginning of the (k+1)st allocated slot (1 ≤ k ≤K1) in the (j+1)st frame equals the number of packets awaiting transmission at the beginning of the kth allocated slot, less the packet (if there were any) that has been transmitted in the kth allocated slot, plus the packets that arrived during d(k). In addition, the number of packets awaiting transmission at the beginning of the ﬁrst allocated slot in the (j+1)st frame equals the number of packets awaiting transmission at the beginning of the last (kth) allocated slot in the jth frame, less the packet (if there were any) that has been transmitted in the Kth allocated slot, plus the packets that arrived during d(k). Therefore, ˜ ˜ ˜ ˜ q j(k + 1) = q j(k) = ∆(q j(k)) + a(k) ˜ ˜ ˜ ˜ q j + 1(1) = q j(K ) = ∆(q j(K )) + a(K ) ˜ where a(k) (1≤ k ≤K) is the number of packets arriving to the user during d(k), and ˜ ˜ ˜ ˜ ˜ ∆(q) = 0 if q = 0 , ∆(q) = 1 if q > 0 . Notice that a(k) does not depend on k. ˜ Let Q k(z) be the steadystate generating function of q j(k) . Then from (2.17) we have Q k + 1(z) = A k(z) { Q k(0) + [ Q k(z) – Q k(0) ]z –1 } Q 1(z) = A K (z) { Q K (0) + [ Q K (z) – Q K (0) ]z –1 } 1≤k≤K–1
(2.18)
1≤k≤K–1
(2.17)
˜ ˜ where A k(z) = E [ z ak ] = e λd(k) [ L(z) – 1 ] 1≤ k ≤ K). In (2.18) we used the fact that a(k) is ˜ j(k) . The derivation of (2.18) is identical to that of equation (2.8) in the independent of q regular TDMA. From (2.18) we obtain
22
CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
k–1
Q k(z) = Q 1
(z)z –( k – 1 )
k–1
∏
A m(z) +
k–1
m=1
2≤k≤K
.
K
+ ( 1 – z –1 )
∑ Qv(0)z –( k – v – 1 ) ∏ Am(z) ∏
A m(z) + ( 1 –
m=v K z –1 )
(2.19)
Q 1(z) = Q 1 We therefore have
v=1 K (z)z –K
∑
Q v(0)z –( K – v )
∏ Am(z)
m=v
m=1
v=1
(z – 1)
∑ Qu
K
(0)z u – 1
∏
K
A m(z)
(2.20)
m=u u=1 Q 1(z) =  . K zK – ∏ A m(z) m=1
˜ Had the boundary probabilities Q k(0) = Prob [ q(k) = 0 ] (1 ≤ k ≤ K) been known, the generating functions Q k(z) (1 ≤ k ≤ K) would be completely determined. To complete the calculation we must therefore compute these unknown probabilities which we do, using a standard method (see for instance the book by Hayes [Hay84]). The method exploits the fact that Q 1(z) , being a generating function, must be analytic within the unit disk and thus any zero of the denominator within the unit disk must also be a zero of the numerator. The approach is to prove that there are exactly K zeroes of the denominator of equation (2.20) within the unit disk, all of them distinct. Let their values be denoted by zn. These values must also be zeroes of the numerator of equation (2.20) which results in K linear equations in the K unknowns Q k(0) . Consider the zeroes of the denominator of (2.20) within the unit disk. Any such zero, z n ≤ 1 ,satisﬁes the equation
K zn
=
∏
m=1
K
A m(z) = e λT c [ L(zn) – 1 ] .
(2.21)
We ﬁrst prove that each root of (2.21) is a simple root. If there were a multiple root zn, then the derivative of the denominator of (2.20) with respect to z computed at z = zn would also vanish, i.e.,
K K z n – 1 = λT c L'(z n)e λT c [ L(zn) – 1 ]
which when substituted into (2.21) yields K = λT c L'(z n)z n .
(2.22)
its numerator must vanish too.23) imply that K ≤ λTc L which contradicts the stability condition.Section 2.: GENERALIZED TDMA 23 We also have L'(z n) = ˜ ˜ l ∑l = 1 l ⋅ zn – 1 ⋅ Prob [ L = l ] ≤ ∑l = 1 l ⋅ Prob [ L = l ] ∞ ∞ = L. zK and e λT c [ L(z) – 1 ] are both analytic in R since they are analytic in z ≤1. To apply this theorem we identify f (z) = z K and g(z) = – e λT c [ L(z) – 1 ] . we can use a Taylor series expansion to obtain z k = ( 1 + δ' ) K ≈ 1 + K δ' e λT c [ L(z) – 1 ] ≈ 1 + λT c L(δ') . and hence the denominator of (2. But zK has K roots in the unit circle (actually a root of multiplicity K at the origin). To do so we apply Rouche’s theorem.e. Also. and therefore each root of (2. We have already indicated that whenever the denominator of (2. namely ∑ Qi(0) . z < 1 + δ ) for some δ>0.20) vanishes within the unit disk. that are all distinct.. If δ is small enough. From the stability condition λLTc <K we see that on the ring z = 1+δ’ we have z k < e λT c [ L(z) – 1 ] . We can thus substitute the values of zn (1≤n≤K1) into the numerator of (2.21) is a simple one. if on C we have f (z) ≠ 0 and f (z) > g(z) .22) and (2. Rouche’s Theorem: Given two functions f(z) and g(z) analytic in a region R. because δ is strictly positive we can ﬁnd some δ’ such that δ > δ’ > 0 so that z = 1+δ’ is an appropriate contour for Rouche’s theorem. implying that z K and z K – e λT c [ L(z) – 1 ] have the same number of roots within z= 1+δ’.20) and obtain the following K1 equations: ∑ v=1 K v Q v(0)z n – 1 ∏ A m( z n ) m=v K = 0 1 ≤ n ≤ K – 1.20) has K roots within the unit circle.3. …z K – 1 . (2. consider a closed contour C in R. Next we determine the number of roots of the denominator of (2. The other roots are denoted by z 1. The region R is the disk of radius 1+δ (i. If δ is small enough.25) . i=1 K (2.24) An additional equation comes from the normalization condition Q 1(z) (we use L’Hospital’s rule in (2.20)) K – λT c L = z=1 = 1 . (2. then f(z) and f(z)+g(z) have the same number of zeroes within C. One of these roots is zK =1.23) Equations (2. z 2.20) within the unit disk.
To summarize. The solution of equations (2. the actual solution procedure is ﬁnding the roots of equation (2. For a complete description of the computation of the underlying determinant the interested reader is referred to the book by Hayes [Hay84]. and converges almost quadratically in the vicinity of a root. take expectations of both sides of (2.25). The solutions are then substituted into equation(2.17) we have .24)(2. …K we obtain 2 ∑ q(k) [ 1 – a(k) ] k=1 K = ∑ k=1 K ˜ E [ a 2( k ) ] + ∑ [ 1 – Qk(0) ] [ 1 – 2a(k) ] k=1 K (2. We obtain ˜ ˜ ˜ ˜ ˜ ˜ E [ q 2(k + 1) ] = E [ q 2(k) ] + E [ ∆(q(k)) 2 ] + E [ a 2(k) ] + 2E [ q(k)a(k) ] ˜ ˜ ˜ ˜ – 2E [ ∆(q(k))a(k) ] – 2E [ q(k)∆(q(k)) ] ˜ ˜ ˜ ˜ ˜ ˜ E [ q 2(1) ] = E [ q 2(K ) ] + E [ ∆(q(K )) 2 ] + E [ a 2(K ) ] + 2E [ q(K )a(K ) ] ˜ ˜ ˜ ˜ – 2E [ q(K )∆(q(K )) ] – 2E [ ∆(q(K ))a(K ) ] 1≤k≤K–1 Let q(k) ∆ E [ q(k) ] and a(k) ∆ E [ a(k) ] for 1 ≤ k ≤ K.20). This method is particularly useful since it is iterative. Solving (2. To that end.19) and (2. the roots are computed in an increasing absolute value order and therefore the roots within the unit disk are computed ﬁrst.25) has a unique solution.3. E [ q(k)∆(q(k)) ] = E [ q(k) ] (which stem from ˜ ˜ ˜ the structure of the ∆(. One quite efﬁcient method to do it is due to Mueller [Mue56.17) directly. In addition.24) and (2.25) determines Q k(0) (1 ≤ k ≤ K).20)).24 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS It is not difﬁcult to verify that the set of K equations (2.27) Using (2. ) function) we obtain ˜ q 2(k + 1) = q 2(k) + E [ a 2(k) ] + [ 1 – Q k(0) ] [ 1 – 2a(k) ] 1≤k≤K–1 . – 2q(k) [ 1 – a(k) ] ˜ q 2(1) = q 2(K ) + E [ a 2(K ) ] + [ 1 – Q K (0) ] [ 1 – 2a(K ) ] – 2q(K ) [ 1 – a(K ) ] (2.21) is. 2. Expected Number of Packets at Allocated Slots The expected number of packets at the beginning of an allocated slot in steadystate can be computed by evaluating the derivative of Q k(z) with respect to z at z=1 (see (2. we square. 2. CoB80]. An alternative method (the one we employ here) is to use (2. obtains both real and complex roots even when these are not simple. Another alternative for computing the boundary probabilities is to use Neuts’ theory of matrix geometric computation [Neu81] as is described in [HoR87].24)(2.2. does not require the evaluation of derivatives. by all counts.26) for all k = 1. the toughest part of the procedure.26) Summing(2.21) within the unit disk and then solving the set of equations (2. With these notations and using the = ˜ = ˜ ˜ ˜ independence between q and a and the identities ˜ ˜ E [ ∆(q(k)) 2 ] = E [ ∆(q(k)) ] = 1 – Q k(0) .17) and let j → ∞.
Section 2. we obtain from (2. (2. Therefore.29) ˜ ( a(m) – E [ ∆(q(m)) ] ) k=1 m=1 –  ∑ a(k) k=1 K ˜ Because the arrival is Poisson we have a(k) = E [ a(k) ] = λLd(k) and K ˜ E [ a 2(k) ] = λL 2 d(k) + λ 2 L 2 d 2(k) .28) = q(1) [ 2 – a(1) – a(2) ] + K–1 ˜ ∑ ( a(k) – E [ ∆(q(k)) ] ) [ 1 – a(k + 1) ] = … = k–1 k=1 + ∑ q(k) [ 1 – a(k + 1) ] K K k=2 ˜ = q(1) K – ∑ a(k) + ∑ [ 1 – a(k) ] ∑ ( a(m) – E [ ∆(q(m)) ] ) k=1 k=1 m=1 where an empty sum vanishes. from (2. Substituting (2.27) we obtain ∑ K ˜ E [ a 2( k ) ] + ∑ [ 1 – Qk(0) ] [ 1 – 2a(k) ] ∑ a(k) k=1 K K k=1 q(1) = k = 1 2 K– 2 ∑ [ 1 – a(k) ] ∑ 2 K– K k–1 .29) .25) ∑k = 1 [ 1 – Q k(0) ] = λLT c .28) into (2. Also.3.: GENERALIZED TDMA 25 ∑ q(k) [ 1 – a(k) ] k=1 K = q(1) [ 1 – a(1) ] + = q(1) [ 1 – a(1) ] + K–1 ∑ q(k) [ 1 – a(k) ] k=2 K–1 K ∑ q(k) [ 1 – a(k + 1) ] k=1 + ˜ ∑ ( a(k) – E [ ∆(q(k)) ] ) [ 1 – a(k + 1) ] K–1 k=1 (2.
In other words we can write ˜(k) = ˜ (k) ⋅ K + J (k) ˜ l f ˜ 0 ≤ J (k) ≤ K – 1 ˜ f where ˜ designates the number of complete frames of delay and J designates the number ˜ f l of intervals left over.17) we get k–1 q(k) = q(1) + ∑ ˜ [ a(m) – E [ ∆(q(m)) ] ] . 2. As a matter of fact ˜ and J depend only on ˜(k) and not on k itself. if ˜(k) is a random variable representing the total number of packets that are to be transmitted before the tagged packet then the delay of the ˜ l tagged packet is w(k) plus the time needed to transmit the ˜(k) packets plus the time to ˜ l transmit the tagged packet itself (note that as in the regular TDMA ˜(k) depends on w(k) ).26 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS λLT c + λ ∑ d(k) ( L 2 + λL 2 d(k) – 2L [ 1 – Qk(0) ] ) k–1 K k=1 q(1) = 2 ( K – λLT c ) 2 ∑ [ 1 – λLd(k) ] ∑ K . This number can be (uniquely) decomposed into a number of complete frames and some leftover.30) m=1 k–1 = q(1) + ∑ [ λLd(m) – 1 + Q m(0) ] 2≤k≤K m=1 Note that once the expected number of packets at the beginning of allocated slots is determined. The quantity ˜(k) is actually the number of intervals that have to elapse before the interval in which the tagged packet is transmitted. . Thus. it is straight forward to compute the expected number of packets at the beginning of an arbitrary slot (see the Exercises section).3.e. it is the l delay of the tagged packet. l l Some insight into ˜(k) is appropriate. Message Delay Distribution ˜ Consider a tagged message arriving within d(k) (1 ≤ k ≤ K). w(k) seconds before the beginning of the (k+1)st allocated slot and tag its last packet.3. until its last packet is transmitted. Both J (k) and ˜ (k) are ˜ ˜ ˜ f f We use the notation f integervalued random variables and their distributions are derived in Appendix A at the end of this section.. i. The delay of the tagged message is the time elapsed from its arrival. ˜ (k) and J (k) (or ˜ and J as a shorthand. [ λLd(m) – 1 + Q m(0) ] k=1 m=1 – 2 ( K – λLT c ) and ﬁnally from (2. (2.
k) .32) ˜ ˜ Note that ˜(k) . l l K K (2.33) we get ˜ of l . Continuing from equation (2.3. given w(k) the total delay is ˜ w(k) + ˜ ˜ D(k w(k). ˜(k)) ] l = ˜ ˜ ˜ ˜ e –sT e –sw(k) e –sT c l(k) ⁄ K e sT c J (k) ⁄ K e –sV (J . k) . k ) ∆ = k + 1 + J (k) ∑ d(u) – d(k + 1 + J (k)) u = k+1 which then turns equation (2. ˜ the transmission time of the tagged packet is T. In all cases. Let l k(z. l the above equation turns into ˜ * ˜ D k (s w(k). w(k)) be the generating function l ˜ given w(k) . let us deﬁne V (J .: GENERALIZED TDMA 27 ˜ The delay of the tagged message after waiting the initial w(k) seconds is f (k)T c seconds ˜ (representing the number of complete frames) plus the time to transmit the J (k) left over ˜ k + J (k ) packets which requires ∑u = k + 1 d(u) seconds if there are any packets left.– J (k) . ˜(k)) = w(k) + T + ˜(k) .Section 2.˜ D(k w(k). Moving to the Laplace transform domain.33) ˜ We proceed by eliminating the condition on ˜ .˜ D(k w(k). ˜(k)) = E [ e –sD(k l ˜ w(k).+ V (J . (2.+ ∑ d(u) – d(k + 1 + J (k)) .– J (k) . The above can be rewrit˜ ten (along with the relation ˜ K = 1 – J ) as follows: f T c k + 1 + J (k) Tc ˜ ˜ ˜ ˜ . and hence J depend on w(k) . ˜(k)) = w(k) + T + ˜(k) .31) into Tc Tc ˜ ˜ ˜ ˜ . In summary. (2.31) l l K K u = k+1 ˜ For notational convenience. ˜(k)) = l w(k) + ˜ ˜ (k)T + T f c ˜ k + J (k ) ˜ J (k ) = 0 ˜ (k)T + f c ∑ d(u) + T ˜ 1 ≤ J (k) ≤ K – 1 u = k+1 where the summation wrapsaround from K to 1 when necessary.
(ii) Packets arriving from the beginning of the kth allocated slot until the ˜ arrival of the tagged message (generating function e λ [ d(k) – w(k) ] [ L(z) – 1 ] . k) ] ˜ ˜ = ˜ e –sT e –sw(k) ∑ e –sT l ⁄ K e sT J(k) ⁄ K e –sV (J. w(k)) ∑ e –sV ( j.34) 1 ˜ = . (iii) Packets of the tagged message. f =0j=0 = ˜ e –sT e –sw(k) ∑ e –sV ( j..8). see equation (2. k) Prob [˜l(k) = l ] c c ∞ ˜ ˜ l=0 = ˜ e –sT e –sw(k) ∑ ∑ K–1 ∞ K–1 e –sT c ( fK + j ) ⁄ K e sT c ( j ⁄ K ) e –sV ( j. the ˜ ˜ generating function of l ˜(k) . w(k)) ∆ E [ z l(k) w(k) ] = ˜ = [ ( 1 – z –1 )Q k(0) + z –1 Q k(z) ]e λ [ d(k) – w(k) ] [ L(z) – 1 ] L(z)z –1 (2. k) Prob [ ˜(k) = fK + j ] l .35) By deﬁning h k(z) ∆ [ ( 1 – z –1 )Q k(0) + z –1 Q k(z) ]e λd(k) [ L(z) – 1 ] L(z)z –1 = Q k + 1(z)L(z)z –1 = equation (2. k) ( zs βm ) – j m=0 j=0 is the unit root of order K. Making the substitution we get = K–1 * D k (s ˜ ˜ w(k)) = e –sT e –sw(k) ∑ K–1 e –sV ( j. l sum of three independent random variables: (i) the packets already waiting at the beginning of the kth allocated slot less one packet (if there were any) that is transmitted in the ˜ ˜ kth allocated slot.e –sT e –sw(k) K where β m = 2πm i e K K–1 ∑ K–1 ˜ l(z s β m. Therefore.28 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS ˜ ˜ * ˜ D k (s w(k)) = e –sT e –sw(k) E [ e –sT c l(k) ⁄ K e sT c J (k) ⁄ K e –sV (J . see equation (2.e.35) can be written as . q(k) – ∆(q(k)) generating function [ ( 1 – z –1 )Q k(0) + z –1 Q k(z) ] .( z s β m ) – j l(z s β m. i.7). ˜ ˜ ˜ l k(z. k) ∑ j=0 m=0 1 ˜ . not including the tagged packet itself (generating function L(z)z –1 ). but before doing so we must evaluate the ˜ since it depends on w(k) . k) ∑ ( e –sT f =0 ∞ c ⁄ K fK ) Prob [ ˜(k) = fK + j ] l j=0 We recognize the bracketed term as the basic relation derived in Appendix A at the end of this chapter. ˜ The next step is to remove the condition on w(k) . with z replaced by z s ∆ e –sT c ⁄ k . To do so we notice that given w(k) . w(k)) K (2. is the total number of packets that are to be transmitted before the tagged packet.
+ E [ V (J . k) ( zs βm ) – j m=0 j=0 ˜ Since w(k) is uniformly distributed between 0 and d(k) we have from the above equation * D k (s) ˜ E [ e –s dkD ] K–1 * E [ D k (s = = 1 ˜ w(k)) ] = d(k) d (k ) d (k ) ∫ 0 * D k (s w) dw 1 = . we take expectation on (2. An alternative method (the one we employ here) is to use (2.36) since d(k)/Tc is the probability that the tagged message will arrive within d(k).32) and obtain Tc Tc ˜ ˜ ˜ dD(k) = E [ w(k) ] + T + E [ ˜(k) ] . Expected Message Delay The expected delay of a message can be computed by evaluating the derivative of D*(s) with respect to s at s=0 (see (2.e –sT e –sw(k) K 1 = .d(k) { s – λ [ 1 – L(z s β m) ] } K–1 ∑ e –sV ( j. k) ( zs βm ) – j ˜ e –w(k) [ s – λ [ 1 – L(zs βm) ] ] m=0 j=0 ∑ h k ( z s β m) ∑ e –sV ( j. 1 ˜ E [ w(k) ] = . k) ] l K K We now compute each of the terms in (2. d(k) ∫ 0 e – w { s – λ [ 1 – L(z s β m) ] } dw K–1 ∑ e –sV ( j.Section 2.37) .37). (2.e –sT K 1 = .34) we get * D k (s 1 ˜ ˜ w(k)) = .d(k) 2 From (2. w(k)) = h k(z)e λw(k) [ 1 – L(z) ] and when substituted into (2.4.e –sT K K–1 K–1 ∑ h k(z s β m)e K–1 ˜ λw(k) [ 1 – L(z s β m) ] K–1 ∑ e –sV ( j.: GENERALIZED TDMA 29 ˜ ˜ l k(z.36)). k) ( zs βm ) – j j=0 Finally.30). k) ( zs βm ) – j j=0 m=0 1 1 – e – d (k ) { s – λ [ 1 – L ( z s β m ) ] } h k(z s β m) . the Laplace transform of the delay distribution is given by D *(s) = k=1 ∑ Dk*(s) Tc K d (k ) (2. To that end. 2.35) and (2.3. Clearly.– E [ J (k) ] .e –sT K ∑ ∑ m=0 K–1 1 h k(z s β m) .3.32) directly. 1≤k≤K.
w(k)) ∑ 1 1 – m=1 βm K–1 K–1 = . k) ] = E [ E [ V (J . k)β mJ h k(β m)E [ e cm w(k) ] . In a similar manner.– ∑ .– 2 h k(β m) 1 ∑ .– 2 h k ( β m ) e c m d (k ) – 1 . w(k)) n=0 ˜ – V (J . J=0 K–1K–1 ∑ ∑ J = 0m = 0 Considering that V(J.35).– 2 K–1 ˜ l k(β m.k)=0 for J=0. k) w(k) ] ] = E 1 = E K 1 = K K–1 K–1 ˜ ∑ V (J . the distribution of J (from Appendix A at the end of this chapter) and the deﬁnition of hk(z) we have K–1 ˜ ˜ E [ J (k) ] = E [ E [ J (k) w(k) ] ] = E . w(k)) ] K–1 ˜ ∑ .. k) + K K–1K–1 ∑ ∑ m = 1J = 1 e c m d (k ) – 1 – V (J . c m d(k) Combining all terms we obtain .30 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS 1 E [ ˜(k) ] = q(k) – [ 1 – Q k(0) ] + . and c0 =0. the above yields 1 ˜ E [ V (J .E [ e cmw(k) ] 1 1 2 1 – m=1 m = 1 1 – βm βm K–1 = .⋅ ∑ c m d (k ) 1 m = 1 1 – βm K–1 where c m ∆ λ [ 1 – L(β m) ] .– 2 K–1 ˜ h k (β m ) E [ l k(β m. k)β mJ h k(β m) . k) Prob [ J = J J=0 K ˜ w(k) ] – ˜ ∑ V (J .= . β0 = 1. k) ] = K K–1 ∑ J=1 1 V (J .λLd(k) + 2 ∑ m=1 k [ λLd(m) – 1 + Q m(0) ] ˜ Using (2.⋅ 1 d (k ) m = 1 1 – βm K–1 d(k) ∫ 0 e cm w K–1 dw = . = ˜ ˜ ˜ E [ V (J .λLd(k) + L – 1 l 2 1 = q(1) + L – 1 – . k) ∑ βmJ lk(βm.
2 1.⋅ .0 2.+ .d(k) + T + .7 and Figure 2.38) Figure 2. four of which are allocated to the user under consideration.λLd(k) + ∑ [ λLd(m) – 1 + Q m(0) ] K 2 2 m=1 T c(K – 1) 1 – . q(1) + L – 1 – .0 Throughput (S=Kρ) FIGURE 2.: GENERALIZED TDMA 31 T c 1 1 D(k) = .4 2. k )β mJ ∑ cm a 1 – .Section 2.+ ∑ V ( J .+ K d (k ) K–1 k ∑ V (J .3.K 2K Tc + .4 0.6 2.8 contain some numerical results for a frame with 24 slots.8 3.69 4.J = 1 m=1 βm K–1 Finally. Tc = 24 d(1) = d(2) = d(3) = d(4) = 6 200 Expected Delay (D) 150 100 50 0 0 0.7: Throughput Delay for Generalized TDMA L(z) = ( z + z 2 + z 3 + z 4 ) ⁄ 4 .2 3. the expected delay of a message is D = k=1 ∑ D(k) . Tc K d (k ) (2. e c m d (k ) – 1 1 – h k(β m) .8 1. k ) K–1 J=1 . We assume that an arriving message 250 K=4.
8 3. the evenly space allocation is better.6 2.0 2..8: Throughput Delay for Generalized TDMA L ( z) = ( z + z 2 + z 3 + z 4 ) ⁄ 4 A natural question to ask is how to allocate the K slots available to a user in a frame. Figure 2. numerical experimentation leads us to believe. It is interesting to mention that in this case the evenly spaced allocation gives the minimal expected delay and the contiguous allocation gives the maximal expected delay for all values of arrival rate. i.8 contains the evenly spaced case as well as the contiguous allocation in which the ﬁrst four slots are allocated to the user.32 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS can contain one. When the expected message delay is used as the performance measure. Hofri and Rosberg showed [HoR87] that the best allocation is the uniform one.e. In Figure 2. d(1)=d(2)=d(3)=1.e. or equivalently. namely. in order to improve the performance. We observe that for the arrival pattern under consideration. d(4)=21. When the expected number of packets in the user’s buffer is the performance measure. that the optimal allocation depends both on the arrival rate λ .7 the allocated slots are evenly spaced. but the differences between the two allocations is rather small. 50 45 40 Expected Delay (D) 35 30 25 20 15 10 d(1) = d(2) = d(3) = 1 d(4)=21 d(1) = d(2) = d(3) = d(4) = 6 K=4.4 0. three or four packets with equal probabilities.. so that d(1)=d(2)=d(3)=d(4)=6. when expected packet delay is the measure. L(z)=(z+z2+z3+z4)/4 . Furthermore. Thus any other allocation results in expected delay that is between the curves of Figure 2. this allocation remains optimal for all arrival rates. two. all the internal periods d(k) (1 ≤ k ≤ K) should be equal.4 2. Tc = 24 0 0.2 Throughput (S=Kρ) FIGURE 2. i.8.8 1.2 1.
Thus.. The more responsive and better usage of the channel achieved with dynamic protocols does not come for free: it requires control overhead that is unnecessary with static protocols and consumes a portion of the channel. Dynamic channel allocation protocols are designed to overcome this drawback. As an example of a protocol that belongs to the family of dynamic conﬂictfree protocols we take the protocol known as Mini Slotted Alternating Priority (MSAP) protocol. . the following captures some of our observations.Section 2. then the expected delay is completely independent of the interallocation distances and if γi= γKi+1 for i=2. its delay is affected by the number of packets ahead of it in the buffer.2. then for γ1 > γK the optimal allocation is the uniform (equidistant) one while for γ1 <γK the K slots should be contiguous in order to minimize Dλ→ 0. The static and ﬁxed assignment in these protocols. This number amounts to a number of whole frames plus some leftover. K1. To ensure freedom of transmission conﬂicts it is necessary to reach an agreement among the users on who transmits in a given slot. The presentation here follows that of Kleinrock and Scholl [KlS80]. With dynamic allocation strategies. Whereas complete characterization of the optimal allocation pattern is still an open question. 2.4. The allocation of slots within a frame can affect only this leftover. cause the channel (or part of it) to be idle even though some users have data to transmit. The MSAP protocol handles properly various such structures. the expected message delay might be sensitive to the allocation distances. the expected message delay is not very sensitive to changes in the interallocation distances since the major portion of the delay is due to the large number of whole frames a message must wait before its transmission starts.: DYNAMIC CONFLICTFREE PROTOCOLS 33 and on the speciﬁc distribution of the message length. i. DYNAMIC CONFLICTFREE PROTOCOLS Static conﬂictfree protocols such as the FDMA and TDMA protocols do not utilize the shared channel very efﬁciently. For light load (ρ → 0 or equivalently λ → 0)..4. It can be shown (we leave the details as an exercise) that if γi= γKi+1 for i=1. When a message arrives at the user’s buffer.. especially when the system is lightly loaded or when the loads of different users are asymmetric. This agreement entails collecting information as to who are the ready users. i. or in other words the probability that a message K length is imodK (clearly ∑i = 1 γ i = 1 ).. one in which the time axis is divided into slots of equal duration and where a user’s transmission is limited to within the slot (the TDMA system is also such a system). .e.. those who request channel allocation. Let γi (1 ≤ k ≤ K) be the probability that a message transmission requires i slots beyond the number of whole frames. for heavy load (ρ → 1). and an arbitration algorithm by which one of these users is selected for transmission.3. ... It is designed for a slotted system.. This latter mechanism is nothing but imposing a priority structure on the set of users each of which constitutes a separate priority class. K.e. the channel allocation changes with time and is based on current (and possibly changing) demands of the various users.
9). Denote by τ the maximum system propagation delay. The alternating priorities structure allows a user to whom the channel is allocated to transmit all the messages in its buffer before the channel is allocated to another user. Given then. The roundrobin priority structure implies a channel that is allocated to the users in a cyclic order and ensures that between any two transmissions of any user all other users have a chance to transmit at least once. 1. thus it treats preferentially the smaller numbered users. i+M1.34 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS Let the users be numbered sequentially 0..... The priority enforcement is based on the observation that if in the most recent slot the channel was allocated to user i then it must have been the one with the highest priority.. MSAP does this by means of reservations as follows.9: Msap Slot Structure Only those users wishing to transmit in a slot take any action.. that user i transmitted last we deﬁne the following priority structures: • Fixed Priorities... and is somewhat “less fair” (although desired in some cases). • Alternating Priorities. that is. Transmission order: i+1. i+2. i+M (user arithmetic is modulo M). in other respects this structure is a roundrobin one since the channel is allocated to the different users in a cyclic order. • RoundRobin. Transmission order: 0.. Given that every user wishing to transmit knows his own priority they behave as follows: . Once the priority structure is decided upon the only issue left is to identify the one user with the highest priority among those wishing to transmit. a user that does not wish to transmit in a given slot remains quiet for the entire slot duration... followed by a data transmission period of duration T. the longest time it takes for a signal emitted at one end of the network to reach the other end. M1. followed by another minislot (see Figure 2. M1. Deﬁning the priority structure is thus the determination of the transmission order after the transmission of some user. Transmission order: i. Slot Reservation Minislots Data Transmission M1 1 2 1 2 τ T t FIGURE 2.1... i+1.. The ﬁxed priority structure implies that user i has always higher priority than user i+1... Let every slot consist of initial M1 reservation “minislots” each of duration τ.
Section 2.4.: DYNAMIC CONFLICTFREE PROTOCOLS
35
• If the user of the highest priority wishes to transmit in this slot then he starts immediately. His transmission consists of an unmodulated carrier for a duration of M1 minislots followed by a message of duration T. • A user of the ith priority ( 1 ≤ i ≤ M – 1 ) wishing to transmit in this slot will do so only if the ﬁrst i minislots are idle. In this case he will transmit M1i minislots of unmodulated carrier followed by a message of duration T. The speciﬁc choice of the minislot duration ensures that when a given user transmits in a minislot all other users know it by the end of that minislot allowing them to react appropriately. The additional minislot at the end allows the data signals to reach every user of the network. This is needed to ensure that all start synchronized in the next slot, as required by the reservation scheme. The evaluation of throughput is fairly simple. Since transmission is conﬂictfree every nonempty slot conveys useful data. However, the ﬁrst M1 minislots as well as the one after data transmission are pure overhead and should not be counted in the throughput. Thus if all slots are used, that is in the highest possible load circumstances, we get a channel capacity (maximum throughput) of T 1 S max =  = T + Mτ 1 + Ma where a ∆ τ ⁄ T is a characteristic parameter of the system. It is evident that the capacity = increases both with the reduction of a and the number of users. For a typical value of a=0.01 only a few tens of users can be tolerated before the capacity reduces below an acceptable level. 2.4.1. Expected Delay
Consider the problem in a somewhat more general setting. Let the arrival processes of ˜ packets at user i be Poisson with rate λi ( 1 ≤ i ≤ M – 1 ). Let w be the waiting time of a ˜ packet at user i (with mean wi), and let x be the transmission (service) time of a single
packet (with mean xi). Denote ρ i ∆ λ i x i , and ρ = ∑ ρ i . In the following derivation we = assume a ﬁxed priority structure among the users; as we indicate later, some of the results apply also to other priority structures. In a ﬁxed priority structure λi can be interpreted as the arrival rate of packets of the ith priority. We also assume that the buffering capabilities of each user are not limited. ˜ Consider a random “tagged” packet that joins user k. Upon its arrival there are L i packets ˜ at the ith user already waiting. Let N i be the number of packets that arrive at user i duri
ing the waiting time of the tagged packet. The waiting time of that packet is composed of waiting for the currently transmitting user to complete his transmission (or if there is no transmitting user the packet will wait until the channel is available for transmission), the transmission time of all packets with equal or higher priority (anywhere in the system) that
36
CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS
are waiting upon his arrival, and the transmission time of all higher priority packets that t will arrive (anywhere in the system) before he starts transmission. Denote by ˜ the forced idletime, that is an idle time imposed by the server (the channel) when it empties. With the above deﬁnitions we have
M–1
˜ E [ wk ] = wk =
∑
k
i=0
˜ E [ x i2 ] E [˜2 ] t ρ i  + ( 1 – ρ ) ˜ 2E [ x i ] 2E [ ˜ ] t
k–1
(2.39)
+
˜ ˜ ∑ E [ xi ]E [ Li ]
i=0
+
˜ ˜ ∑ E [ xi ]E [ N i ]
i=0
The ﬁrst term in this equation is the average time before the next transmission starts. If the system is nonempty, then ρi is the probability that user i is transmitting, and the average ˜ ˜ residual transmission time is E [ x 2 ] ⁄ ( 2E [ x ] ) . Similarly, a user encountering an empty
i i
system (probability 1ρ) must wait the residual forced idletime. The second term denotes the total transmission time of all packets of equal or higher priority that are waiting for transmission. The third term is the total transmission time of all packets of higher priority that will arrive while the tagged packet is waiting. Note that within a priority class FCFS order prevails. ˜ ˜ Straightforward application of Little’s formula yields E [ L i ] = λ i w i and E [ N i ] = λ i w k which, when substituted into equation (2.39) yields
k–1 ˜ E [ x i2 ] E [˜ 2 ] t ∑ ρi ] + ( 1 – ρ ) ] + ∑ ρi wi ˜ 2E [ x i 2E [ ˜ t i=0 i=0 w k = (k) 1–ρ M–1 k where ρ ( k ) ∆ ∑i = 0 ρ i . The above equation can be recursively evaluated to yield =
˜ E [ x i2 ] E [˜2 ] t ρ i  + ( 1 – ρ ) ∑ 2E [ xi ] ˜ 2E [ ˜ ] t i=0 w k = ( 1 – ρ(k ) )( 1 – ρ(k – 1) )
M–1
The mean packet delay, averaged over all priority levels is ∑ ( λ i ⁄ λ )w i which in general cannot be evaluated in closed form. However, if the transmission times of all priority lev˜ els have the same distribution (that is x i is independent of i) then
M–1
˜ w = E[w] =
∑
i=0
˜ E [˜2 ] t E [ x2 ] ρ  + ( 1 – ρ ) λi ˜ 2E [ x ] 2E [ ˜ ] t  i = w λ 1–ρ
(2.40)
Section 2.4.: DYNAMIC CONFLICTFREE PROTOCOLS
37
˜ If, in addition ˜ is distributed in the same way as x we get t ˜ E [ x2 ] ˜ 2E [ x ] ˜ ] = E[w 1–ρ ˜ For the MSAP case, all the x i are indeed the same and equal the slot size namely Mτ + T = T ( 1 + Ma ) . The quantity ˜ represents the forced idletime incurred when the t system becomes empty (a transmission can start only at slot boundaries) and is also deterministically equal to the slot size. Thus the conditions of equation (2.41) hold and when substituted yields ( 1 + Ma )T w k = 2( 1 – ρ(k ) )( 1 – ρ(k – 1) ) with ρ ( k ) = ( 1 + Ma )T ∑i = 0 λ i . The mean packet delay is thus
k M–1
(2.41)
D =
∑
i=0
λi  i = D λ
M–1
∑
i=0
λi ˜  [ w i + ( 1 + Ma )T ] = E [ w ] + ( 1 + Ma )T λ
1 = ( 1 + Ma )T 1 + 2(1 – ρ) which ﬁnally yields D 1 ˆ D =  = ( 1 + Ma ) 1 + T 2(1 – ρ)
(2.42)
Note that since ρ is the fraction of periods in which transmission takes place and since Ma/ (1+Ma) of every slot is overhead, we conclude that S =ρ /(1+Ma), or ρ=(1+Ma)S. When substituted into the last equation the throughput delay characteristic of MSAP results. Although the preceding analysis was performed for ﬁxed priorities the ﬁnal results hold for other priority structures as well. The priority conservation law [Kle76] states that for any M/G/1 queue and any nonpreemptive workconserving queueing discipline (which includes all those of concern here) ∑ ρ i w i = Const .Thus, as surprising as it might appear, the average delay given by equations (2.40)and (2.42) holds for all priority structures mentioned in the beginning of this subsection. Higher moments are, as expected, different.
In MSAP there are M1 such bits for every packet transmitted. CDMA allows the coexistence of several systems in the same frequency bands. FDMA and TDMA A good treatment of TDMA and FDMA analysis can be found in [Hay84]. The conﬂictfree property of CDMA is achieved by using orthogonal signals in conjunction with matched ﬁlters in the corresponding receivers. The question of optimal allocation of slots to the users in the generalized TDMA scheme is addressed in [ItR84] where the throughput of the system is maximized (assuming single buffers for each user) and in [HoR87] where the expected packetdelay in the system is minimized.38 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS 2. such protocols are discussed in Section 3. This is the reason for the large volume of analyses of such protocols. The bitmap protocol requires synchronization among the users that is somewhat more sophisticated than the MSAP protocol. Code Division Multiple Access Both FDMA and TDMA does not allow any time overlap of transmissions. but the overhead paid per transmitted packet is less than the overhead in the MSAP protocol.code division multiple access (CDMA) [Pur87]. The use of multiple orthogonal signals increases the bandwidth required for transmission. Interconnecting all users in the system requires that matched ﬁlters corresponding to all signals to be available at all receivers. We mention here those reservation protocols which basically do not involve contention for the reservation itself. are the oldest and most popular protocols around.5. An improvement to the MSAP protocol is the bitmap protocol described in [Tan81]. Reservation Protocols The MSAP protocol presented in the text is a representative of an entire family of protocol that guarantee conﬂictfree transmission by way of reservation. as long as different signals are used in different systems. We suggest here several of those analyses along with some other conﬂictfree protocols. The idea behind this protocol is to use a single reservation preamble to schedule more than a single transmission. . This is done by utilizing the fact that all participating nodes are aware of the reservations made in the preamble. Yet. All these protocols have a sequence of bits precede serving to reserve or announce upcoming transmissions (this is known as the reservation preamble). RELATED ANALYSIS Conﬂictfree access protocols. A conﬂictfree scheme that does allow overlap of transmission both in the frequency and the time domain is the.5. A TDMA scheme in which the packets of each user are serviced according to a priority rule is analyzed in [MoR84]. A sample path comparison between FDMA and TDMA schemes is carried out in [Rub79] where it is shown that TDMA scheme is better than the FDMA scheme not only on the average. especially the static ones.
If there is a collision in the reservation minislot all users but the “owner” of that minislot will abstain from transmission. and if the reservation remains uncontended that reserving user will transmit. As in the MSAP protocol. There. Under heavy load BRAM reduces to regular TDMA.: RELATED ANALYSIS 39 Another variation of a reservation protocol has been described in [Rob75].Section 2. the reservation preamble does not necessarily contain all M1 minislots. This is essentially a combination between the bitmap and the MSAP protocols. One of the most efﬁcient reservation protocol is the Broadcast Recognition Access Method [CFL79]. a reservation preamble serves to reserve the channel for a single user but unlike the MSAP.5. The idea is that users start their transmission with a staggered delay not before they ensure that another transmission is not ongoing (the paper [KlS80] also refers to a similar scheme). this is a standard TDMA with idle slots made available to be grabbed by others. every user can make a reservation in every minislot of the reservation preamble. . Altogether.
sync bits. We are interested in the message delay distribution at the user. Problem 4. What will be the throughput delay characteristic of a TDMA channel? Problem 2. Problem 3. Assume a Poisson message arrival process with rate λ messages/second and the number of packets in a message distributed according to the generating function L(z).40 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS EXERCISES Problem 1. containing L 3. Deﬁne an appropriate set of random variables and write the equation for D the delay of an arbitrary (“tagged”) message. etc.g. Consider a TDMA system in which a user is assigned the ﬁrst two slots of every frame but a message transmission will start only at the ﬁrst slot in the frame. . Derive the Laplace transform of the message delay in FDMA in which every message contains a random number of packets. address. Problem 5. Assume that a portion y of every transmitted packet is overhead (e. ˜ 1. Find the generating function of ˜ the number of frames required to transmit a message ˜ packets. f 2. Derive the steadystate distribution and the ﬁrst two moments of the number of messages in a TDMA system where L(z) is the generating function of the number of packets in a message. What will be the throughput delay characteristic of an FDMA channel? 2.. 1. Compare the ﬁrst two moments of the distribution of the queueing time of FDMA with that of TDMA (Note: the queueing time does not include the actual transmission time). Compare the expected message delay with that of TDMA. Compute the Laplace transform of the message delay and derive from it the average message delay.).
. choosing d(k)) in the generalized TDMA protocol when the load is very light.3.∑ γ i 2 Tc i=1 K ∑ j=1 K d 2( j) + ∑ ∑ K j+i–1 d( j)d(i) j = 1l = j + 1 where γi (1 ≤ i ≤ K ) is the probability that a message transmission requires i slots beyond the number of whole frames..Section : EXERCISES 41 Problem 6.2.. Show that when the interallocation of slots in the generalized TDMA scheme is uniform.16) (see Sections 2. 3.. Problem 7.37) reduces to that in (2. Show that it is sufﬁcient to minimize F in order to minimize the expected delay under the light load circumstances 2.e. Let γi = γK^i+1 for i=2.e..∑ Tc 2 K K–1 ∑ d(l)d(l + k) ∑ i = k+1 K (γ i – γ K – i + 1) l = 1k = 1 Note from the above expression that when γi = γK^i+1 for i=1. K. 1..e. Compute also the expected number of packets at the beginning of an arbitrary slot. What is the optimal allocation when γ1>γK and when γ1<γK? Problem 9. For the generalized TDMA scheme derive the generating function of the number of packets at the beginning of an arbitrary slot.2. i. and 2.+ ( 1 – ρ ) ∑ 2E [ xi ] ˜ 2E [ ˜ ] t i=0 w k = ( 1 – ρ(k ) )( 1 – ρ(k – 1) ) M–1 . Let 1 1 F = 1 + . This problem addresses the optimal allocation of slots (i.3.. or in other words the probability that a message length is K imodK (clearly ∑i = 1 γ i = 1 ). Show that the expression for F reduces to 1 1 F = 1 + . Referring to the delay analysis of the MSAP protocol. Prove that ˜ E [ x i2 ] E [˜2 ] t ρ i . K1. then the Laplace transform of the delay distribution in (2. ρ → 0 or equivalently λ → 0 .. then the expected delay is completely independent of the interallocation distances. i. d(k) = Tc/K for 1 ≤ k ≤ K .T c + ...). Problem 8..
+ ( 1 – ρ ) λi ˜] 2E [ x 2E [ ˜ ] t .i = w λ ( 1 – ρ(k ) )( 1 – ρ(k – 1) ) .42 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS k ˜ (with ρ ( k ) = ∑i = 0 ρ i ) is a solution of the general equation. and that if all x i have the same distribution then M–1 ˜ E[w] = ∑ i=0 ˜ E [˜2 ] t E [ x2 ] ρ .
l would like to compute the distributions of J f Let βm be the unit roots of order K namely. From this J(z). We ˜ l In another form this can be written as J = ˜ modK and ˜ = ˜ ⁄ K l f ˜ and ˜ from that of ˜ . the generating function of J . l ˜ = ( ˜ – J ) ⁄ K. These roots obey β n = 1 – ∆(nmodK ) = 1 K divides n 0 Otherwise m=0 ∑ Our most basic relation is derived as follows: 1 K K–1 ∑ ( zβ m ) –n l(zβ m=0 1 m) = K K–1 ∑ ( zβ m ) –n ∑ Prob [˜l = l ] ( zβm ) l l=0 K–1 ∞ 1 l = .∑ Prob [ ˜ = l ]z l – n K l=0 ∞ m=0 ∞ ∑ ( βm ) l – n (2. can be computed as follows: . β m = e 1 K K–1 2πm i K .43) m=0 = ∑ Prob [˜l = l ]z l – n [ 1 – ∆(( l – n )modK ) ] l=0 ∞ = ∑ Prob [ ˜ = f ⋅ K + n ] l f =0 By setting z=1 in equation (2. The quantity ˜ can be uniquely l decomposed into ˜ = ˜ ⋅K+J ˜ l f ˜ 0≤J≤K–1.Section : APPENDIX A Distribution of the Mod Function 43 APPENDIX A Distribution of the Mod Function Let ˜ be a nonnegative integer valued random variable with a known distribution and a l generating function l(z).43) we get ˜ Prob [ J = n ] = ∑ f =0 ∞ 1 Prob [ ˜ = f ⋅ K + n ] = l K K–1 ∑ – β mn l(β m) 0≤n≤K–1 m=0 ˜ ˜ which gives us the distribution of J . and let K be a known positive integer constant.
l(β m) = . we thus have 1 – zK J (z) = K l (β m ) l(β m ) 1 – zK 1 – zK ∑ .∑ K z z m = 0 1 – m = 1 1 – βm βm K–1 K–1 Taking the derivative at z=1 yields the expectation K–1 ˜ E [ J ] = . K–1 K–1 β m l(β m) 1 1 – zK 1 – zK .l(β m) = . Overall.∑ z K z K z 1 – m = 0 1 – m = 0 1 – βm βm βm K where in the step before last we used the fact that β m = 1 .44 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS K–1 ∑ K–1 ˜ Prob [ J = n ]z n = ∑ n=0 K–1 n=0 1 K K–1 ∑ m=0 1 – β mn l(β m) z n = K K–1 K–1 m=0 n=0 ∑ ∑  β m z n l(β m) 1 = K ∑ m=0 z K 1 – . Clearly f K–1 f Prob [ ˜ = f ] = and thus ( z) = ∑ Prob [ ˜ = f ⋅ K + n ] l f ≥0 n=0 ∑ f =0 K–1 ∞ Prob [ ˜ = f ] z f = f ∞ ∑ ∑ ∞ K–1 Prob [ ˜ = f ⋅ K + n ] z f = l f =0 n=0 = ∑ ∑ Prob [ ˜ = f ⋅ K + n ] z f l n=0 f =0 We note that the bracketed term in the summation appears in equation (2. Hence .∑ .= K ( 1 – z ) + .+ 2 l(β m ) 1 m = 1 1 – βm K–1 ∑ We turn now to calculate the generating function of ˜ .43) when z1/K is substituted for z.
= .∑ .Section : APPENDIX A Distribution of the Mod Function 45 K–1 F (z) = ∑ n=0 1 K K–1 ∑ ( z 1 / K β m ) – n l ( z 1 / K β m) ( z1 / K β ) –n l( z 1 / K β 1 m) = K K–1 m=0 1 = K 1 = K K–1 K–1 ∑ ∑ ∑ m ∑ m=0 n=0 K–1 m=0 1 – ( z 1 / K β m ) –K .l(z 1 / K β m) = K 1 – ( z 1 / K β m ) –1 Calculating the expected value of ˜ can be done by taking the derivative of the above f equation at z=1 or using the direct approach.e.l E[ f .. K K 2K K 1 1 – m=1 βm . i.l(z 1 / K β m) 1 – ( z 1 / K β m ) –1 l ( z 1 / K β m) ∑ 1 – ( z 1 / K β m ) –1 m=0 K–1 m=0 1 – z –1 1 – z –1 .E [ ˜ ] – K – 1 + . K–1 ˜ l(β m) 1 1 E [˜ ] – E [ J ] l ˜ ] = ..
46 CHAPTER 2: CONFLICTFREE ACCESS PROTOCOLS .
successful transmissions are indicated by blank rectangles and collided packets are hatched. PURE ALOHA The pure Aloha protocol is the basic protocol in the family of the Aloha protocols. It considers a singlehop system with an inﬁnite population generating packets of equal length T according to a Poisson process with rate λ packets/sec. The reason is that whenever two or more users are transmitting on the shared channel simultaneously. Second. This chapter covers a few of these variations. Many local area networks of today implement some sophisticated variants of this family’s protocols. a collision occurs and the data cannot be received correctly. This randomness is required to ensure that the same set of packets does not continue to collide indeﬁnitely. The pure Aloha protocol is very simple [Abr70].1 where the arrows indicate arrival instants. Transmission scheduling is therefore the focal concern of contentiontype protocols. every colliding user. schedules its retransmission to a random time in the future. The users whose packets collide with one another are called the colliding users. The variations present different protocols for transmission and retransmission schedules as well as adaptation to different circumstances and channel features. 3. a collision is caused and none of the colliding packets is received correctly and they have to be retransmitted. Should the transmission be unsuccessful. It states that a newly generated packet is transmitted immediately hoping for no interference by others. The channel is errorfree without capture: whenever a transmission of a packet does not interfere with any other packet transmission. as it is the ﬁrst random access technique introduced.CHAPTER 3 ALOHA PROTOCOLS The Aloha family of protocols is probably the richest family of multiple access protocols. the transmitted packet is received correctly while if two or more packet transmissions overlap in time. With the conﬂictfree protocols that were discussed in Chapter 2. At the end of every transmission each user knows whether its transmission was successful or a collision took place. many of these protocols are so simple that their implementation is straightforward. Because of the great popularity of Aloha protocols. independently of the others. A simple example of the operation of the protocol is depicted in Figure 3. analyses have been carried out for a very large number of variations. every scheduled transmission is guaranteed to succeed. packets may have to be transmitted and retransmitted until eventually they are correctly received. Its popularity is due ﬁrst of all to seniority.1. This being the case. The Aloha family of protocols belongs to the contentiontype or random retransmission protocols in which the success of a transmission is not guaranteed in advance. .
48
CHAPTER 3: ALOHA PROTOCOLS
Vulnerable Period
T
tT
t
t+T Retransmissions
time
FIGURE 3.1: Pure aloha Packet Timing
Since the population is inﬁnite each packet can be considered as if it belongs to a different user. Hence, each newly arrived packet can be assigned to an idle user i.e., one that does not have a packet to retransmit. This allows us to interchange the roles of users and packets and consider only the points in time when packet transmission attempts are made. Observing the channel over time we deﬁne a point process consisting of scheduling points, i.e., the points in which packets are scheduled for transmission. The scheduling points include both the generation times of new packets and the retransmission times of previously collided packets. Let the rate of the scheduling points be g packets/sec. The parameter g is referred to as the offered load to the channel. Clearly, since not all packets are successful on their ﬁrst attempted transmission, g>λ. The exact characterization of the scheduling points process is extremely complicated. To overcome this complexity it is assumed that this process is a Poisson process (with rate g, of course). This assumption can, however, be a good approximation at best (as has indeed been shown by simulation). The reason is that a Poisson process implies independence between events in nonoverlapping intervals, which cannot be the case here because of the dependence between the interval containing the original transmission and the interval containing a retransmission of the same packet. It can be shown, however, that if the retransmission schedule is chosen uniformly from an arbitrarily large interval then the number of scheduling points in any interval approaches a Poisson distribution. The Poisson assumption is used because it makes the analysis of Alohatype systems tractable and predicts successfully their maximal throughput. Pure Aloha is a singlehop system. Hence, the throughput is the fraction of time the channel carries useful information, namely noncolliding packets. The channel capacity is the highest value of arrival rate λ for which the rate of departure (throughput) equals the total arrival rate (but see the discussion of stability in Section 3.4.).
Section 3.2.: SLOTTED ALOHA
49
Consider a packet (new or old) scheduled for transmission at some time t (see Figure 3.1). This packet will be successful if no other packet is scheduled for transmission in the interval (tT,t+T) (this period of 2T is called the vulnerable period). The probability of this happening, that is, the probability of success, is that no packet is scheduled in an interval of length 2T and since scheduling is Poisson we have P suc = e –2gT Now, packets are scheduled at a rate of g per second of which only a fraction Psuc are successful. Thus, the rate of successfully transmitted packets is gPsuc. When a packet is successful the channel carries useful information for a period of T seconds; in any other case it carries no useful information at all. Using the deﬁnition that the throughput is the fraction of time that useful information is carried on the channel we get S = gT e –2gT which gives the channel throughput as a function of the offered load. Deﬁning G ∆ gT to = be the normalized offered load to the channel, i.e., the rate (per packet transmission time) packets are transmitted on the channel, we have S = Ge –2G The relation between S and G is depicted in Figure 3.2, which is typical to many Aloha type protocols. At G=1/2, S takes on its maximal value of 1 ⁄ ( 2e ) ≈ 0.18 . This value is referred to as the capacity of the pure Aloha channel. We recall that for a system to be stable the long term rate of input must equal the long term rate of output meaning that stability requires S= λT. Larger values of λ clearly cannot result in stable operation. Note however, that even for smaller values of λ there are two values of G to which it correspondsone larger and one smaller than 1/2. The smaller one is (conditionally) stable while the other one is conditionally unstable, meaning that if the offered load increases beyond that point the system will continue to drift to higher load and lower throughput. Thus, without additional measures of control, the stable throughput of pure Aloha is 0. We return to the stability issue in Section 3.4. (It is appropriate to add that this theoretical instability is rarely a severe problem in real systems, where the long term load including, of course, the “off hours” load, is fairly small although temporary problems may occur).
3.2. SLOTTED ALOHA
The slotted Aloha variation of the Aloha protocol is simply that of pure Aloha with a slotted channel. The slot size equals Tthe duration of packet transmission. Users are restricted to start transmission of packets only at slot boundaries. Thus, the vulnerable period is reduced to a single slot. In other words, a slot will be successful if and only if exactly one packet was scheduled for transmission sometime during the previous slot. The
50
CHAPTER 3: ALOHA PROTOCOLS
throughput is therefore the fraction of slots (or probability) in which a single packet is scheduled for transmission. Because the process composed of newly generated and retransmitted packets is Poisson we conclude that S = gT e –gT or using the deﬁnition of the normalized offered load G = gT S = Ge –G This relation is very similar to that of pure Aloha, except of increased throughput (see Figure 3.2). Channel capacity is 1 ⁄ e ≈ 0.36 and is achieved at G=1.
0.4 0.35 0.3
Throughput (S)
0.25 0.2 0.15 0.1 0.05 0 0.001
Slotted Aloha
Pure Aloha
0.01
0.1
1
10
100
Offered Load (G) FIGURE 3.2: ThroughputLoad of Pure and Slotted Aloha
We would like to introduce an additional method of calculating throughput which will be useful later and which can be easily demonstrated in the slotted Aloha case. When observing the channel over time we notice a cyclic behavior of busy and idle periods (see Figure 3.3 in which uparrows point at arrival instants, blank rectangles refer to successful slots, and hatched rectangles denote colliding packets, i.e., unsuccessful slots). A busy period is a succession of slots in which transmission takes place (successful or not). The idle period is deﬁned as the interval between two busy period. The starting times
these are points of a regenerative process. In fact. since the system is memoryless in the sense that its behavior in a given cycle does not depend on the behavior in any previous cycle. ˜ P [ I = 1 ] = P [ Some packets scheduled in first slot ] = 1 – P [ No packets scheduled in first slot ] = 1 – e –gT The probability that the idle period lasts exactly two slots is the probability of the event that no packets were scheduled in the ﬁrst slot and some were scheduled in the second (to be transmitted in the third slot). the length of the idle period is seen to be geometrically distributed. 2. The random ˜ variable I must be strictly positive since there must be at least one empty slot in an idle period. namely ˜ P [ I = k ] = ( e –gT ) k – 1 ⋅ ( 1 – e –gT ) yielding an average length (measured in slots) of 1 I = 1 – e –gT k = 1. Thus. the expected fraction of time the system is in a given state equals the expected fraction of time during a single cycle that the system is in that state (see Appendix).: SLOTTED ALOHA 51 of every cycle (just before the start of the busy period) deﬁne renewal points. As such. ˜ P [ I = 2 ] = e –gT ⋅ ( 1 – e –gT ) In general. The probability that the idle period consists of a single slot is the probability that there were some packets scheduled during that slot that will be transmitted in the next slot.Section 3. Thus. … .3: SlottedAloha Packet Timing ˜ Let I be a random variable describing the number of slots in the idle period.2. Cycle Busy Period Idle Period FIGURE 3.
the throughput is the expected fraction of slots within a cycle in which successful ˜ transmission takes place. some arrivals do occur). The probability that a given slot in the busy period is successful is gT e –gT 1 – e –gT which is the probability of a single arrival in a slot given that we are in the busy period ˜ (i.+ e –gT 1 – e –gT – gT . packets must be scheduled for transmission in each and every one of the ﬁrst k1 slots and none scheduled in the kth. Thus. 1 – . This leads to ˜ P [ B = k ] = ( 1 – e –gT ) k – 1 e –gT yielding an expected value of 1 B = – gT e ˜ Since not all the slots in the busy period are successful. let U denote the number of useful. or successful slots in a cycle and let U be its expected value.52 CHAPTER 3: ALOHA PROTOCOLS ˜ ˜ Similarly.e.⋅ e –gT 1 – e –gT . An argument similar to that used in calculating the distribution of the idle period leads to the ˜ derivation of the distribution of B ..= . k 1 – e –gT 1 – e –gT k = 1.= . Clearly B > 0 . For the busy period to be k>0 slots long. If C denotes the number of slots in a cycle then gT e –gT B ˜ E[U ] U 1 – e –gT S = . given that the duration of the busy period is B slots we have ˜ ˜ B gT e –gT k gT e –gT B – k ˜ ˜ P [ U = k B ] = .= gT e –gT 1 1 . 2.= ˜ B+I B+I E[C ] 1 gT e –gT . let us deﬁne B as the number of slots in the busy period. … ˜ 0≤k≤B and hence – gT ˜ ˜ ˜ gT e E [ U B ] = B 1 – e –gT from which we get gT e ˜ ˜ ˜ U = E [ U ] = E [ E [ U B ] ] = B 1 – e –gT Now.
in the next slot. Let every user be in one of two statesthinking and backlogged. the user generates a packet in every slot with probability σ and does not generate a packet in a slot with probability 1σ. The number of backlogged users at the beginning of the k+1st slot depends on the number of backlogged users at the beginning of the kth slot and the number of users that moved from state to state within the slot. if all states communicate (as we subsequently indicate) this Markov chain is also ergodic. requiring T seconds for transmission. To gain insight into the relation between transmission of new packets and retransmission of old ones we build the following packetscheduling model (referred to as the linear feedback model).. In the thinking state the user does not have a packet in its buffer and does not participate in any scheduling activities. The preceding means that packet generation is an independent process distributed geometrically with mean 1/σ. since collision of any number of packets is possible. When the packet is ﬁnally successfully transmitted the user moves back to the thinking state.. SLOTTED ALOHA . Once a packet is generated its transmission is attempted immediately. While in the backlogged state the user does not generate any new packets. packet generation is independent of any other activity.FINITE NUMBER OF USERS To make the previous model more realistic we analyze here the case of an Aloha system with a ﬁnite number of users. The random variable N (k) is referred to as the state of the system. The technique of using regenerative processes as done above is an important tool in deriving the throughput of more complicated protocols. thus. All packets are of the same size. We consider a case in which slotted Aloha is used by a group of M users each with a single packet buffer (this less general case makes the comparison with the inﬁnite population cases more meaningful since there too every user had only a single buffer). When in this state. … ) } is a Markov chain.3. “Downward” transitions are possible only to the adjacent . and let N (k) denote the ˜ number of backlogged users at the beginning of the kth slot..Section 3. The following analysis is based on that of Kleinrock and Lam [KlL75].FINITE NUMBER OF USERS 53 which is the result obtained previously. If the transmission was successful the user remains in the thinking state and the packet generating process starts anew. Because the number of backlogged users cannot exceed M this chain is ﬁnite. ( k = 1.4.1. “Upward” transitions are possible between every state and all the highernumbered states.: SLOTTED ALOHA . in every slot the user will retransmit the packet with probability ν and will refrain from doing so with probability 1ν. In other words. ˜ Let the slots of the system be numbered sequentially k=0. Since statetransition of the users is independent of the activities in any previous slot the process ˜ { N (k). meaning that steadystate distribution exists. 3. The analysis of this model enables us to derive packet delays which we were unable to do in the previous model. 2. If packet transmission was unsuccessful the user moves to the backlogged state and schedules the retransmission of the packet according to an independent geometric distribution with parameter ν. The transition diagram for the system is shown in Figure 3. that is.3. which is also the slotduration.
Finally. Note also the missing transition from state 0 to state 1 which is clear since if all users were thinking and a single user generated and transmitted a packet he could not cause a collision and become backlogged. denote by P the matrix whose elements are pij and by π the row vector whose elements are πi. Let πi be the steadystate probability of the system being in state i.4: State Transitions of Finite Population Aloha SteadyState Probabilities For analysis purposes we introduce the following notation (see Appendix). p ij = lim k → ∞ Prob [ N (k) = j N (k – 1) = i ] . From the above argumentation it follows that the steadystate probability vector is the solution to the ﬁnite set of linear equations π = πP ∑i π i = 1 to which the existence of a unique solution is guaranteed. let pij be the steadystate transition probability.. The fact that all states communicate is evident from the diagram. We must therefore construct the matrix P and derive the desired solution. that is ˜ π i = lim k → ∞ Prob [ N (k) = i ] . . ˜ ˜ i. at which time the backlog is reduced by unity.54 CHAPTER 3: ALOHA PROTOCOLS state since only one packet can be successfully transmitted in a slot.e. 0 1 2 i M From 0 1 i1 To M i+2 i+1 i To i1 From i+1 FIGURE 3. Further.
Consider the transition from state i to state i1. In this case the thinking user succeeds and therefore remains in the thinking state leaving the system in the same state. a collision must have taken place. because no transmission of new packets was attempted the backlog did not change. Furthermore. namely j Prob [ i backlogged users in a slot  j in backlog ] = ν i ( 1 – ν ) j – i i In a similar manner. exactly one thinking user has attempted together with at least one attempt from the backlog. collided. we obtain for the thinking users M – j i Prob [ i thinking users transmit in a slot  j in backlog ] = σ ( 1 – σ ) M – j – i (3.Section 3. Mj users are thinking. The matrix P can be constructed by applying equations (3. Here the backlog increased by two or more meaning that ji thinking users generated packets and. The transition from state i to the very same state can come about from two distinct reasons. since the backlog increased by unity. The next transition to consider is from state i to state i+1. The transmitting users clearly collide and remain in the backlog. The last case is the transition from state i to a state j>i+1.2) as follows.2) i since when j users are backlogged. This indicates a reduction in the backlog which is possible only if a single backlogged packet was transmitted (and no new packet was generated. The second reason for this transition results from a situation in which none of the backlogged users attempt retransmission and a single thinking user transmits. The activity of the backlogged users is immaterial in this case since the collision is generated by the thinking users alone. A special case of this latter one is when the slot remains idleneither a transmission of a new packet is attempted nor is a retransmission attempted. Clearly a transition from state i to state j<i1 is impossible implying that pij =0 for those cases. (3. of course. The above can be summarized by the union of the two independent events: “No backlogged user succeeds and no thinking user attempts” and “No backlogged user attempts and a single thinking user attempts”. The ﬁrst results from the circumstance in which no new packet was generated (and transmitted) while several backlogged users attempted transmission.FINITE NUMBER OF USERS 55 Since the retransmission process of every user is an independent geometric process the probability that i out of the j backlogged users will schedule a retransmission in a given slot is binomially distributed. of course).1) .: SLOTTED ALOHA .3.1) and (3. Since the backlog increased.
Hence.3) It can be easily veriﬁed that ∑ j p ij = 1 as required. It can be shown that the rank of IP is M1 which means. the long term fraction of time the channel carries useful informationthe throughputequals the expected fraction of slots containing useful transmission. Having done this M times and calculated the values of all xi we then compute xi π i = ∑ xj j The value of π thus computed is colinear with x and therefore solves the original set of equations. Consider the homogeneous set of equations x(IP)=0 where I is the identity matrix. that nontrivial solutions exist and are all collinear. Proceeding similarly. Throughput Analysis To evaluate the throughput of the system consider the epochs at the beginning of every slot. To ﬁnd one of these solutions assume x0=1. Since the activity within a given slot is independent of the activity in any previous slot these epochs are renewal points. note that p01 turns out to be identically zero. However. the special structure of the P matrixhaving nonzero elements in the upper right triangle and in the ﬁrst subdiagonalallows fairly easy computation. among others. every step involves a simple computation and results in determining the value of an additional xi. The πi clearly sum up to 1. from which x2 can be calculated. From the second equation we have – x 0 p 01 + x 1 ( 1 – p 11 ) – x 2 p 21 = 0 . Also. If we denote by Psuc the probability of a successful slot then . this result is correct and expected since it takes at least two colliding packets to increase the backlog and because no users were backlogged before. This is therefore our desired solution.56 CHAPTER 3: ALOHA PROTOCOLS The above can be summarized in the following formulae (where the bracketed terms correspond to the events described in the preceding explanation): p ij = 0 [ iν ( 1 – ν ) i – 1 ] ( 1 – σ ) M – i [ 1 – iν ( 1 – ν )i – 1 ]( 1 – σ)M – i + [ ( M – i )σ ( 1 – σ)M – i – 1](1 – ν )i [ ( M – i )σ ( 1 – σ ) M – i – 1 ] [ 1 – ( 1 – ν ) i ] M – i σ j – i ( 1 – σ ) M – j j–i j<i–1 j = i–1 j=i j = i+1 j>i+1 (3. it is impossible to have a single backlogged user at the end of the slot. The ﬁrst equation. Solving the set of equations π = πP for the above matrix cannot be done in closed form. namely x 0 ( 1 – p 00 ) – x 1 p 10 = 0 yields x 1 = ( 1 – p 00 ) ⁄ p 10 . This is the same set of equations rewritten with x replacing π .
6) Let us continue a bit with this line of thought. consider a situation in which we do not distinguish between backlogged packets and new packets. we set ν=σ. Given that there are i backlogged users this can be stated as P suc(i) ∆ Prob [ successful slot  i users in backlog ] = = = ( 1 – ν ) i ( M – i )σ ( 1 – σ ) M – i – 1 + iν ( 1 – ν ) i – 1 ( 1 – σ ) M – i The total throughput is therefore S = P suc = E [ P suc(i) ] = ..3. because Psuc(i) is independent of i we obtain from equation (3. This result is. i. (3. i. M Under these circumstances. namely E [ P suc(i) ] = Mσ ( 1 – σ ) M – 1 (3.5) a closed form expression for the throughput. As a special case. that the inﬁnite population model is indeed in some sense the limit of the ﬁnite population model if backlogged users are not .FINITE NUMBER OF USERS 57 S = P suc For a slot to be successful only a single transmission must take place within it. in our case this equals Mσ.4) yields P suc(i) = Mσ ( 1 – σ ) M – 1 indicating that Psuc(i) is independent of i. a result identical to the one derived in Section 3. the individual throughput is given by the value of S from the last equation divided by M. of course. In previous sections we denoted by G the total.5) Note that since all users are statistically identical.e. Substituting this value into the throughput equation above yields G M–1 S = G 1 – . average number of transmissions per slot.: SLOTTED ALOHA . This means that either all backlogged users remain silent and a single new user transmits. not surprising since if we cease to distinguish between backlogged and thinking users we cannot expect the probability of success to depend on the number of backlogged users. letting M increase to inﬁnity we ﬁnd that in the limit S = Ge –G . not distinguishing the backlogged from the thinking users.Section 3.e. Moreover. therefore.3 for the inﬁnite population slotted Aloha scheme.. system wide. We might conclude.4) ∑ Psuc(i)πi i=0 M (3.1. or a single backlogged user transmits while no new packet is generated. Substituting this into equation (3.
Denote by b the average rate at which packets (actually. Thus. If the system is to be stable then this rate must equal the average rate of new packet generation. It should be noted that this representation is parametric since σ inﬂuences the value of S. Since b is the rate of packets joining the backlog and S (the throughput) is the rate of packets leaving the system then a fraction (Sb)/S of the packets are never backlogged. when the system is in state i there are Mi thinking users each generating packets in every slot with probability σ. Thus.6). The delay.+ σ S (3. This last equation is the desired throughput delay relation. Taking expectation yields S = E [ ( M – i )σ ] = ∑ ( M – i )σπi = ( M – N )σ (3. normalized to the packet transmission time) the average delay is S–b b N ˆ D = . the average rate of new packet generation when in state i is (Mi)σ. as intuitively expected. S S b Using the value of N from equation (3.⋅ .5.+ 1 . for which the throughput is given in equation (3.58 CHAPTER 3: ALOHA PROTOCOLS distinguished from the thinking ones and if the number of users is increased under the constraint that the total average arrival rate remains ﬁnite. increases monotonically with σ.7) ﬁnally yields 1 M ˆ D = 1 – .⋅ 1 + .8) With the value of S taken from equation (3. thereafter the throughput decreases with increasing load.e. Expected Delay The previous derivation considered the throughput from the departure standpoint since Psuc is the average rate of packet departure from the system. users with packets) join the backlog.. All the others (whose fraction is b/S) suffer the backlog delay mentioned above plus the one slot in which their transmission is successful. then according to Little’s formula. Substituting this into equation (3. Throughput delay characteristics for several parameter choices are depicted in Figure 3. Consider again the special case in which σ=ν. Not all packets going through the system are backloggedthe lucky ones make it the ﬁrst time.7) where N s the average number of backlogged users. Now.5). Each of the curves in the ﬁgure represents one value of ν with σ varying from 0 to 1 along the curve.8) yields . the throughput ﬁrst increases with σ until capacity is achieved (for that value of ν). These packets suffer a delay of 1 slot only. Measured in slots (i. the average amount of time spent in the backlog is the ratio of the average number of backlogged users to the average rate of joining or N ⁄ b .
3 ν=0.FINITE NUMBER OF USERS 59 100000 10000 Finite Population Slotted Aloha M=10 ˆ Expected Delay ( D ) 1000 100 10 ν=0.5: ThroughputDelay of Finite Population Slotted Aloha (a) 10 Users (b) 25 Users .2 1 0 0.2 0.3 1 0 0.Section 3.3.3 0.25 0.35 0.4 ν=0.1 0.25 0.15 0.4 Throughput (S) 100000 10000 Finite Population Slotted Aloha M=25 ˆ Expected Delay ( D ) 1000 100 10 ν=0.5 ν=0.6 ν=0.15 0.3 Throughput (S) FIGURE 3.2 0.: SLOTTED ALOHA .05 0.7 ν=0.1 0.4 ν=0.05 0.
a result that may look surprising at ﬁrst.. Solving this set of equations yields M π M – 1 = M + ( M – 1 )σ ( M – 1 )σ π M = . in fact such that Mν<<1. This is not a surprising fact and is due to the instability of the inﬁnitepopulation Aloha systems. These assumptions in fact mean that eventually most users will become backlogged since the rate of exit from the backlog is very small. M = 1 – Mν and all other values of pij vanish. and remain in this state for a very long time since the average waiting time for a backlogged packet is inversely proportional to σ. in the rare case of a collision the colliding users become of course backlogged. and in most cases therefore the delay will be a single slotthat of the transmission itself. We shall introduce this phenomenon through the equations themselves. hardly ever will a collision result. M = ( M – 1 )νσ p M . Consider a situation in which σ accepts nonnegligible values (although not necessarily very close to 1) and ν accepts very small values. yielding a combined average delay of M slots. That is. very few packets having extremely large delays. However. keeping the product Mσ constant and increasing M shows an ever increasing delay.ν)i as 1iν equations (3.60 CHAPTER 3: ALOHA PROTOCOLS 1 – (1 – σ)M – 1 ˆ D = 1 + σ(1 – σ)M – 1 Two interesting observations can be made regarding this last result. M + ( M – 1 )σ The corresponding values of the probability of success are P suc(M – 1) = σ + ( M – 1 )ν + ( M – 2 )νσ P suc(M ) = Mν . the model we have developed cannot be used to evaluate the delay for the inﬁnite population case. Taking the limit we ﬁnd ˆ that D(σ → 0) → M . M – 1 = Mν p M – 1.3) become p M – 1. Putting all these together we obtain the following expression for the throughput . One can therefore safely assume that most of the time either M or M1 backlogged users will be observed. The second interesting observation relates to the expected delay when σ tends to zero. First. When σ is very small. Using these assumptions.4. M – 1 = 1 – ( M – 1 )νσ p M . The Capture Phenomenon There is a curious phenomenon in ﬁnite population Aloha systems that appears in certain situations when the retransmission probability is small. Putting it all together we ﬁnd most packets having a delay of unity. and approximating (1. a subject we shall discuss in more detail in Section 3.
Furthermore.05 ν=0.FINITE NUMBER OF USERS 61 σ + ( M – 1 )ν + ( M – 3 )νσ Mσ S = M .≈ . demonstrating our previous observation that system performance under these conditions is almost independent of ν.: SLOTTED ALOHA . substituting these values into equation (3.8) yields 1+σ M–1 ˆ D = 1 + ( M – 1 ) .1 0.4 0.5 0 0. 100 90 80 70 Finite Population Slotted Aloha M=10 ˆ Expected Delay ( D ) 60 50 40 30 20 10 ν=0.45 0. recall that the retransmission rate is very small and that backlogged users remain in that state for a fairly long time.01 ν=0.35 0.6: ThroughputDelay of ﬁnite population Aloha under Capture To understand the nature of this phenomenon.15 0. At higher values of σ we note a decrease of the delay while the throughput increases.Section 3.1 0.6 where the throughputdelay curve for a system with M=10 is shown (this ﬁgure depicts the result of equation (3. which is rather strange and is indicative of a curious phenomenon.= M + σ σ indicating that the average delay is actually decreasing with increasing load! Furthermore.2 0.05 0. Since the rate of .. The phenomenon is demonstrated in Figure 3.3 Throughput (S) FIGURE 3. both the throughput and the average delay do not depend on ν hence these systems exhibit identical performance for various values of ν.001 ν=0.8)). M + ( M – 1 )σ M + ( M – 1 )σ This expression is interesting because it indicates that the throughput increases with the load σ. Also note that the graphs for the various values of ν coincide for larger values of σ.25 0. One can clearly observe that as long as σ is small delay increases with throughput as one might generally expect.3.
If equilibrium between arrival and departure rates prevails. (IN)STABILITY OF ALOHA PROTOCOLS An underlying assumption in the analysis of the Aloha protocol is that the total arrival process of new and retransmitted (due to collisions) packets is a Poisson process. In other words. There is no justiﬁcation to this assumption except that it simpliﬁes the analysis of the Aloha protocol.e. another user retransmits and usually succeeds and the previous scenario repeats itself. however. We ﬁrst give an intuitive reasoning why this assumption is false and then prove it rigorously. Assume that the arrival rate of new packets is λ packets per slot and assume that λ < e –1 ( e –1 is the maximum throughput predicted for the slotted Aloha protocol). then the rate of the total trafﬁc on the channel (new and retransmitted pack . Another assumption that is implicitly used in the analysis of the Aloha protocol is a stability assumption. This user will generate another packet after a relatively short time and.2 where relation (3. When this happens all users are backlogged again. With this situation in place consider a user that retransmits its packet and (usually) succeeds. When σ becomes sufﬁciently high. eventually all users will become backlogged.9) is depicted. having generated a new packet. since retransmission rate is small. or there are two thinking users that are likely to collide. g is the expected number of new and retransmitted packets per second and G=gT.4. the throughput approaches 0. We recall that under this assumption. namely. What actually happens is that a random user captures the channel for a while. This chain of events continues for quite a while until the “unfortunate” event that a retransmission of a backlogged user is scheduled at which time either a collision occurs immediately.5 because the capturing user transmits in one slot and generates a new packet in the next and so forth. it is assumed that packets are entering and leaving the system at the same rate. This capture effect appears in all ﬁnitepopulation Alohalike protocols such as all the variants of the carriersensing protocols. with and without collision detection (see. Performance clearly improves with increasing σ because the more packets generated in between two retransmissions the better the throughput. is very likely to succeed again since the probability that its transmission will be interfered with that of a backlogged user is slight. [ShH82]). for example. Consider Figure 3.. 3. that the number of backlogged users with packets awaiting to be retransmitted is not steadily growing. remaining so for quite a while. the throughput of the slotted Aloha protocol is predicted to be S = gT e –gT = Ge –G (3. transmits successfully a succession of packets. i.9) where T is the (ﬁxed) transmission time of a packet. Eventually.62 CHAPTER 3: ALOHA PROTOCOLS new packet generation σ is nonnegligible.
However. We ﬁrst describe the concrete model that is used in the analysis.2 0.15 0. 0. Consequently. .001 λ 0.4. This is of course an “average” rate.1 G1 1 G2 10 100 Offered Load (G) FIGURE 3. If the actual trafﬁc rate moves a little above G1. al. this result is established in a more precise and formal manner. Thus. ∞ ) .7: Demonstrating the Instability of Slotted Aloha In the following.G)=(λ. etc.7. which causes further increase in the actual trafﬁc rate. the actual rate will ﬂuctuate around this mean. it is stable under small variations in G1.G1). the point (S. over any ﬁxed interval of time.: (IN)STABILITY OF ALOHA PROTOCOLS 63 ets) G=gT will be G1=g1T. but rather drifts relentlessly toward the catastrophic.G1) is a conditionally stable point. and. which causes the actual trafﬁc rate to increase back to G1. a further decrease in actual throughput. If the actual trafﬁc rate moves a little below G1. point ( S. packets leave the system faster than they arrive. unconditionally stable. [FLB74]. Thus. the actual throughput increases a little above λ. which causes the actual trafﬁc rate to decrease back to G1. G ) = ( 0. The following is based on the analysis by Fayolle et. The system never returns to the point (λ. packets leave the system at a slower rate than they enter. the actual throughput decreases a little below λ.25 0.7 then the actual throughput decreases below λ.4 0.1 0.01 0. if a large variation (and this will happen with probability one) causes the actual trafﬁc to exceed G2 in Figure 3.3 Slotted Aloha Throughput (S) 0. Thus. We conclude that the maximum stable throughput of the Aloha protocol is zero. as is shown in Figure 3.05 0 0. packets leave the system slower than they arrive.35 0. namely.Section 3.
independently of any other event in the system. 2. Let A(k) be the number of new packets that are generated (arrive) during the kth slot. The number of backlogged users is unchanged. 2. and no new packet was transmitted (a0). Analysis Consider the slotted Aloha system with an inﬁnite population of users. … ) } is a Markov chain. The number of backlogged users at the beginning of the k+1st slot depends on the number of backlogged users at the beginning of the kth slot (one of them might transmit a packet successfully) and the number of new packets that arrived within the slot.) random variables with a common distribution ˜ Prob [ A(k) = i ] = Prob [ i new packets arrive at slot k ] = a i and mean λ packets/slot λ = is Poisson. There were n backlogged users at the beginning of slot k ( π n(k) ).e. The retransmission delay of a packet is assumed to be geometric. none of them transmitted ( b 0(n) ) and a single new packet was transmitted (a1). The single new packet is successfully transmitted and therefore the number of backlogged users is unchanged. at least two of them transmitted ( 1 – b 0(n) – b 1(n) ) and hence collided. the chain { N (k). ( k = 0. 1.64 CHAPTER 3: ALOHA PROTOCOLS 3. Let us list all possible transitions that will lead into the state in which there are n backlogged users at the beginning of slot k+1. Backlogged users have packets that collided.11) n ˜ = Prob [ i backlogged users transmit in slot k  N (k) = n ] = ν i ( 1 – ν ) n – i i As in the case of a ﬁnite number of users (see Section 3.. Since the process of new arrivals is independent of the ˜ activities in any previous slot. 2.) the number of backlogged ˜ users N (k) is referred to as the state of the system. … ) } is a sequence of independent and identically distributed (i. 1. ˜ Let π n(k) denote the probability that N (k) = n . There were n backlogged users at the beginning of slot k ( π n(k) ). .i. so it is not obvious whether or not the chain is ergodic.. i. ˜ Unlike the case of a ﬁnite number of users. We assume that ˜ N (0) = 0 .1.4.. These packets are transmitted in ˜ slot k+1. … ) } is not ﬁnite. a i = ( λ i e –i ) ⁄ i! . retransmits its packet during slot k with probability ν.3. so that each new ˜ packet that arrives to the system is associated with a new user.10) ∑i = 1 iai . Notice that if the arrival process of new packets ˜ Let N (k) be the number of backlogged users at the beginning of slot k (k=0. a user that is backlogged at the beginning of slot k.. From these deﬁnitions we have: b i(n) ∆ Prob [ i backlogged users transmit in slot k  n in backlog ] = (3. ( k = 0.d. 2. ( k = 0. ˜ N (k + 1) = n in parenthesis we indicate the probabilities of the corresponding events): 1. ∞ i≥0 (3. It is assumed that { A(k).1. 1. the process { N (k).2. and that have to be retransmitted. namely.).
In this case.4.12) is valid for all n ≥ 0 if one adopts the convention that π i(k) = 0 for i<0. This corresponds to an idle slot and the number of backlogged users is unchanged. and therefore the number of backlogged users decreases by one. k = 0. we obtain from (3. 2. 6. the transmission of the backlogged user was successful. There were nj 2 ≤ j ≤ n backlogged users at the beginning of slot k ( π n – 1(k) ) and at least two new packets were transmitted aj 2 ≤ j ≤ n . 4.13) for n=0. ˜ The Markov chain { N (k).. There were n backlogged users at the beginning of slot k ( π n(k) ).13) j = 0 n– j j ∑n = 0 πn . It is ergodic if an invariant probability distribution { π n. none of them transmitted ( b 0(n) ) and no new packet was transmitted (a0).. at least one of them transmitted ( 1 – b 0(n – 1) ) and a single new packet was transmitted (a1). Assuming the latter limit exists.12) Notice that (3. 1. There were n1 backlogged users at the beginning of slot k( π n – 1(k) ). N (3. N n π a (3. 5. … } exists satisfying (3. Summarizing the above.Section 3..12): π n = π n [ b 0(n)a 1 – b 1(n)a 0 ] + π n + 1 b 1(n + 1)a 0 – π n – 1 b 0(n – 1)a 1 + ∑ Deﬁne PN = and sum (3. In this case a collision occurs and the user with the new packet joins the backlogged users.14) . k = 0.: (IN)STABILITY OF ALOHA PROTOCOLS 65 3. There were n+1 backlogged users at the beginning of slot k ( π n + 1(k) ). … } is aperiodic and irreducible. exactly one of them transmitted ( b 1(n + 1) ) and no new packet was transmitted (a0). 1. the following balance equation can be written: π n(k + 1) = π n(k)b 0(n)a 1 + π n(k) [ 1 – b 0(n) – b 1(n) ]a 0 + π n(k)b 0(n)a 0 + π n + 1(k)b 1(n + 1)a 0 + π n – 1(k) [ 1 – b 0(n – 1) ]a 1 + ∑ n j=2 π n – j(k)a j (3. In this case a collision occurs and all users with the new packets join the backlogged users.1. ∑n = 0 π n = 1 and π n = lim k → ∞ π n(k) . 2..12) such that ∞ πn > 0 for all n.
the sum P ∞ ∆ lim N → ∞ P N exists only if π n = 0 for all ﬁnite = values of N. and the Aloha protocol is not stable. Thus.18) for any N ≥ 0 .16) a jPN – j N j=1 ≤ π N b 0(N )a 1 + π N + 1 b 1(N + 1)a 0 + P N – 1 ∑ aj (3.66 CHAPTER 3: ALOHA PROTOCOLS PN = ∑n = 0 πn N = ∑n = 0 πn [ b0(n)a1 – b1(n)a0 ] + ∑n = 0 πn + 1 b1(n + 1)a0 N N N N n=0 –∑ = π n – 1 b 0(n – 1)a 1 + ∑ π a n = 0 ∑j = 0 n – j j N n N ∑n = 0 πn [ b0(n)a1 – b1(n)a0 ] + ∑n = 0 πn + 1 b1(n + 1)a0 N n=0 (3. From (3. Therefore.19) implies that the ratio π N + 1 ⁄ π N increases without limit as N → ∞ . otherwise. P ∞ is divergent.17) we have: π N ( 1 – a 0 ) = ( P N – P N – 1 ) ( 1 – a 0 ) ≤ π N b 0(N )a 1 + π N + 1 b 1(N + 1)a 0 or (we use (3. the Markov chain { N (k). … } representing the number of backlogged users is not ergodic. we will see that the throughput of the system is zero. k = 0. The inequality in (3. given that N (k) = n . n ≥ 0 ˜ deﬁne a probability distribution.17) ≤ π N b 0(N )a 1 + π N + 1 b 1(N + 1)a 0 + P N – 1 ( 1 – a 0 ) where the ﬁrst inequality above is due to the fact that PN does not decrease as N increases. Let S n(k) be the conditional probability that one packet is successfully transmitted during ˜ the kth slot. 1. The throughput of the system is then .15) –∑ π n – 1 b 0(n – 1)a 1 + ∑ N j=0 a j∑ π n = j n– j N j=0 N = π N b 0(N )a 1 – π 0 b 1(0)a 0 + π N + 1 b 1(N + 1)a 0 + ∑ Using the fact that b1(0)=0 we have: P N = π N b 0(N )a 1 + π N + 1 b 1(N + 1)a 0 + ∑ or P N ( 1 – a 0 ) = π N b 0(N )a 1 + π N + 1 b 1(N + 1)a 0 + ∑ N j=1 N j=0 a jPN – j a jPN – j (3. 2. which cannot be the case when the πn.11)) π N + 1 1 – a o – b 0(N )a 1 1 – a0 – ( 1 – ν ) N a1 . N ∞ and the second inequality is due to the fact that ∑ j = 1 a j ≤ ∑ j = 1 a j = 1 – a 0 .= πN b 1(N + 1)a 0 ( N + 1 )ν ( 1 – ν ) N a 0 (3.≥ . Furthermore.19) (3.
it is clear that the Aloha system (with inﬁnitely many users) cannot be stable for a policy of retransmission of collided packets that does not take into account (somehow) the system state.20) A packet will be successfully transmitted only if exactly one backlogged user is transmitting and no new packet is transmitted or no backlogged user is transmitting and exactly one new packet is transmitted. All other backlogged users remain silent during that slot. no packets will be successfully transmitted. ˜ The facts that the Markov chain { N (k). meaning that the retransmission policy is independent of system state. when the number of backlogged users does not exceed θ. it . and the expected delay of a packet will be inﬁnite. the retransmission probabilities must somehow adapt in accordance with the state of the system. Assuming that some coordination among the backlogged users is possible prior to each slot. each of them with probability ν independently of the other users. rendering these schemes unstable. 1. For instance. k = 0. Stabilizing the Aloha System From the intuitive arguments given at the beginning of this chapter and from the analysis presented in the previous section.Section 3. The schemes presented thus far use ﬁxed retransmission probabilities. Hence. Therefore. 3.22) ∞ = lim ∑n = 0 [ b1(n)a0 + b0(n)a1 ]πn = 0 where we used our conclusion from (3. it is not difﬁcult to develop retransmission policies that stabilize the Aloha system.4.2. and even if it is known.21) ∑ [ b1(n)a0 + b0(n)a1 ]πn(k) k→∞ n = 0 N N→∞ ∞ = ∑n = 0 [ b1(n)a0 + b0(n)a1 ]πn (3.4. 2. indicate that the number of backlogged users will eventually grow to inﬁnity. It is not clear how the number of backlogged users would be known. consider the following threshold policy: At most θ (the threshold) of the backlogged users at the beginning of slot k. retransmit their packets during slot k. The implementation of the threshold policy is not speciﬁed but is clearly not simple.: (IN)STABILITY OF ALOHA PROTOCOLS 67 S = lim ∑ S n(k)πn(k) k→∞ n = 0 ∞ (3. S = lim (3. S n(k) = b 1(n)a 0 + b 0(n)a 1 ( S n(k) does not depend on k). In other words. … } is not ergodic and that the throughput of the system is zero. In order to stabilize the system.19) that πn=0 for any ﬁnite n. a subset of size θ is chosen from the backlogged users and each user in this subset retransmits its packet with probability ν. each of them retransmits its packet with probability ν. If the number exceeds θ.
k = 0.24) ∑ j = 2 ( i + j )a j – i – b1(i)a0 – b0(i)a1 Hence.23) we have: . 2. condition (a) holds if λ < ∞ . Assume that N (k) = i and i < θ. 1.23) we obtain ˜ ˜ ˜ ˜ ˜ ˜ ˜ E [ N (k + 1) – N (k) ( N (k) = i ) ] = E [ N (k + 1) ( N (k) = i ) ] – E [ N (k) ( N (k) = i ) ] = ( i – 1 )b 1(i)a 0 + i { [ 1 – b 1(i) ]a 0 + b 0(i)a 1 } + ( i + 1 ) [ 1 – b 0(i) ]a 1 + ∑ = ∞ ∞ j=2 ( i + j )a j – i = λ – b 1(i)a 0 – b 0(i)a 1 (3. Given that N (k) = i we have i–1 i ˜ N (k + 1) = i+1 with probability b 1(i)a 0 with probability [ 1 – b 1(i) ]a 0 + b 0(i)a 1 with probability [ 1 – b 0(i) ]a 1 (3. i→∞ ˜ It is clear that the Markov chain { N (k). Pakes’ Lemma: ˜ Let { Z k. … } be an irreducible. The following two conditions are sufﬁcient for the Markov chain to be ergodic: ˜ ˜ ˜ 1. since ˜ ˜ N (k) cannot increase to inﬁnity in this case. aperiodic and ˜ homogeneous for the threshold policy. ˜ Consider now the case that N (k) = i and i ≥ θ . As in (3.68 CHAPTER 3: ALOHA PROTOCOLS is not clear how the backlogged users coordinate to choose the subset of size θ.23) is similar to that of (3.23) i + j. 2. it is instructive to see why this policy stabilizes the Aloha system for a certain range of arrival rate.12). aperiodic homogeneous Markov chain whose state space is the set of nonnegative integers. j ≥ 2 with probability a j The explanation of (3. In this case we ˜ ˜ have only to show that condition (a) of Pakes’ Lemma ( N plays the role of Z ) holds. From (3. lim supE [ Z k + 1 – Z k ( Z k = i ) ] < 0 . Nevertheless. we use a lemma due to Pakes [Pak69] that is often useful in proving ergodicity of homogeneous Markov chains. k+1 k k ˜ ˜ ˜ 2. E [ Z – Z (Z = i)] < ∞ ∀i . To prove the stabilizing properties of the threshold policy. … } is irreducible. k = 0. 1.
Section 3. j ≥ 2 with probability a j The above stems from the fact that at most θ users are transmitting when the threshold policy is employed. Deﬁne (3.29) ∀i . equation (3. conditions (a) and (b) of Pakes’ Lemma will hold if λ < limsup n → ∞ [ b 1(n)a 0 + b 0(n)a 1 ] .24)): ˜ ˜ ˜ E [ N (k + 1) – N (k) ( N (k) = i ) ] = λ – b 1(i)a 0 – b 0(i)a 1 Consequently.4.28) (3. that guarantees stability. we saw that by limiting the number of backlogged users that contend for the channel. the system is stable. Moreover.27). In summary. (3. we derived a condition on the arrival rate.23) holds for all i and therefore (see (3.27). This policy is based on adaptively controlling the retransmission probabilities.26) For this retransmission policy.25) we have ˜ ˜ ˜ E [ N (k + 1) – N (k) ( N (k) = i ) ] = ( i – 1 )b 1(θ)a 0 + i { [ 1 – b 1(θ) ]a 0 + b 0(θ)a 1 } + ( i + 1 ) [ 1 – b 0(θ) ]a 1 + ∑ = λ – b 1(θ)a 0 – b 0(θ)a 1 Hence. Let us now quantify this value of λ.: (IN)STABILITY OF ALOHA PROTOCOLS 69 i–1 i ˜ N (k + 1) = i+1 with probability b 1(θ)a 0 with probability [ 1 – b 1(θ) ]a 0 + b 0(θ)a 1 with probability [ 1 – b 0(θ) ]a 1 (3. From (3.25) i + j.11)): n b i(n) = [ ν(n) ] i [ 1 – ν(n) ] n – i i (3. we allow the retransmission probability ν to be a function of that number. both conditions (a) and (b) hold if λ < b 1(θ)a 0 + b 0(θ)a 1 Therefore. if λ satisﬁes (3. assuming that the number of backlogged users is known. Denoting this function by ν(n) where n is the number of backlogged users we have (see (3. Speciﬁcally.27) ∞ j=2 ( i + j )a j – i (3. Let us now consider another stable policy for the Aloha system that requires only the knowledge of the number of backlogged users at the beginning of each slot. it is possible to stabilize the Aloha system.
namely retransmitting with probabilities that are (approximately) inversely proportional to the number of backlogged users. The general structure of these policies is: ν k + 1 = f (ν k. exactly as was predicted for constant retransmission probabilities. one should not forget that the Aloha system with constant retransmission probabilities is unstable for any arrival rate. The stabilizing policy presented above. To summarize. feedback of slot k). The protocols that stabilize the Aloha system are no longer as simple as the original protocol. all the functions f that are used) increase the retransmission probability when an idle slot occurs and decrease it when a collision occurs. However. according to what happened during that slot. a1/a0 = log a0 = λ and therefore the system would be stable if λ < e –1 . The policies of the form (3. the simple protocol yields an unstable system. In essence. (3. The reason for this low throughput is that in all the stabilizing policies dis .70 CHAPTER 3: ALOHA PROTOCOLS S n(ν) ∆ b 1(n)a 0 + b 0(n)a 1 = [ 1 – ν(n) ] n a 1 + nν(n) [ 1 – ν(n) ] n – i a 0 .30) By differentiating S n(ν) with respect to ν and setting the result equal to zero. one observes that S n(ν) is maximized for a0 – a1 ν *(n) = na 0 – a 1 The maximum value of S n(ν) is n–1 S *(ν *) = a 0 n – a1 ⁄ a0 n–1 Taking the limit as n → ∞ (see equation (3. all such policies (namely. However. These policies are based on updating the retransmission probabilities recursively in each slot.29)) we see that the system will be stable for λ<e a1 log a 0 + . the virtue of the Aloha protocol is its simplicity.– 1 a0 It is interesting to note that if the arrival process is Poisson. requires the knowledge of this number and it is not clear how the users would know this number. and yet. they only guarantee throughput of at most e –1 . There exist retransmission policies that do not require the knowledge of the exact number of backlogged users. Examples of such policies and their analysis can be found in the work by Hajek and Van Loon [HaL82]. = (3. the retransmission probability (of a backlogged user) in slot k +1 is some function f of the retransmission probability in the previous slot and of the event that occurred in slot k.31) namely.31) were proved to yield maximal stable throughput of at most e –1 .
Capture The assumption that whenever two or more packets overlap at the receiver.). all backlogged users are using the same retransmission probability.5. Hay84. Lee87] considered power capture. Thus. 3. higher stable throughput can be obtained. This phenomenon is known as capture and it can happen as a result of various characteristics of radio systems. Variablelength packets Abramson [Abr77] studied the performance of the inﬁnite population pure Aloha system with two different possible packet lengths. ﬁxed length packets and Poisson arrivals. we refer to some papers in which performance measures. namely the phenomenon whereby the strongest of several transmitted signals is correctly received at the receiver. Sha84. BeG87] cover part of these and other variations. Tas86. In the next chapter we shall see that if the decisions of users whether to transmit or not are based both upon their own history of retransmissions and the feedback history. Some other works (for instance. HaO86. Tann81. Arbitrary interarrival distribution Sant [San80] studied the performance of the pure Aloha system where packet interarrival times are statistically independent and identically distributed (i. Ferguson [Fer77b] and Bellini and Borgonovo [BeB80] considered a system with an arbitrary packet length distribution. RELATED ANALYSIS Numerous variations of the environment under which the Aloha protocol operates have been addressed in the literature. if a single high powered packet is transmitted then it is correctly received regardless of other transmissions and hence the utilization of the channel increases.i. In radio systems the receiver might correctly receive a packet despite the fact that it is timeoverlapping with other transmitted packets.: RELATED ANALYSIS 71 cussed in this chapter (excluding the threshold policy that is not practical). In addition. ﬁnite and inﬁnite population.d. all packets are lost. other than the throughput.Section 3. but not necessarily exponential. Most studies [Abr77. have been computed for the Aloha protocol.5. . slotted and nonslotted system. We considered a very small part of these variations. is overly pessimistic. Davis and Gronemeyer [DaG80]) studied the effect of delay capture (the receiver captures a packet since it arrived a short time before any other packet that is transmitted during the same slot) in slotted Aloha protocol (see Exercises). It is interesting to note that it was shown in [BeB80] that constant length packets yield the maximum throughput over all packet length distributions (see Exercises). In the following we list a few of the variations that we did not describe. Met76. Several books such as [Kl75.
The interdeparture process of the Aloha protocols has been studied in [TaK85b]. When a ﬁnite number of buffers are available at each user. the success probability of a transmission of a certain user depends on the activity of the other users that have packets to transmit.1. When the buffering capability at each user is not limited. to use a reservation scheme with contention. Approximate analysis of a reservation Aloha protocol can be found in [Lam80]. the analysis proceeds along the same lines. Speciﬁcally. The case of a single buffer per user has been considered in Section 3. that users contend during a reservation period an those who succeed in making reservations transmit without interference. Additional variations of reservation protocols and their analysis can be found in [ToK76. although the number of states increases dramatically. In the category of reservation schemes fall the works by Binder [Bin75] that requires knowledge of the number of users (or an upper bound thereof). and by Ephremides and Zhu [EpZ87] for a nonsymmetric system. Reservation and adaptive protocols Reservation schemes are designed to have the advantages of both the Aloha and the TDMA approaches. and the works by Crowther et.72 CHAPTER 3: ALOHA PROTOCOLS Buffered users In some practical systems the users can provide buffer space for queueing of exogeneous packets that arrive at the user. al. The operation of such schemes is discussed in Section 2. of course. [CRW73] and Roberts [Rob75] that do not require this knowledge.5. i.e.. where only conﬂictfree reservation schemes are discussed. Fer77b] and compared it to TDMA [Fer77a]. The urn scheme [KlY78] is an example of such protocol. Bounds for the expected queue lengths have been derived by Szpankowski [Szp86]. An immediate extension is. Sufﬁcient conditions for ergodicity of the system have been provided by Tsybakov and Mikhailov [TsM79] and Tsybakov and Bakirov [TsB88]. Delay and interdeparture times Ferguson presented an approximate analysis of the delay in the Aloha protocol [Fer75. one faces a very complicated queueing problem due to the strong interaction among the various queues of the system. Approximate analysis of an Muser system has been carried out by Saadawi and Ephremides [SaE81] and by Sidi and Segall [SiS83] for a symmetric system. Exact analysis of a twouser system has been carried by Sidi and Segall [SiS83]. .3. Exact packet delay and interdeparture time distribution for a ﬁnite population slotted Aloha system with a single buffer per each user has been derived in [Tob82b]. TsC86]. These schemes derive their efﬁciency from the fact that reservation periods are shorter than transmission periods by several orders of magnitude. TaI84. Another kind of protocols that are designed to operate as the Aloha protocol in light load and according to a TDMA scheme in heavy load are known as adaptive protocols.
The operation of the stable algorithms depends very strongly on the feedback information that is obtained in each slot (see (3. Kelly [Kel85] proposed an additive rule for determining νk+1 (as opposed to the multiplicative rule suggested in [HaL82]). Therefore. the tuning of the algorithms should be adjusted as is discussed in [MeK85].31) have been suggested in several studies. .Section 3. Another additive rule known as the pseudoBayesian rule has been suggested in [Riv87] and analyzed in [Tsi87].5.: RELATED ANALYSIS 73 Stability Instability issues of the Aloha protocol were ﬁrst identiﬁed in [CaH75] and [LaK75]. similar issues were identiﬁed for the CSMA family of protocols in [ToK77]. For instance. OnN85]. Stable protocols for the Aloha system of the form (3. MeL83.31)). if the feedback information is not reliable. Later. Other papers that discussed these issues are [Jen80.
Draw S as a function of G for this system and for slotted Aloha without acknowledgments on the same ﬁgure. 1. However. (Busy period of pure Aloha) Consider a pure Aloha system with Poisson offered load g packets/sec and packets of ˜ equal size 1 (T=1).2. the receiver of the station “locks on” the packet and can receive it correctly.. ˜ tion of F 4. The slot size is T sec and . Consequently. After each successful transmission the receiving station sends an acknowledgment packet indicating the transmission was successful. By conditioning on L ﬁnd F F (s) the Laplace transform of the probability density func˜ . Compute the expectation and variance of F . A collision between a data packet and an acknowledgment packet destroys both of them. Problem 3. (Slotted Aloha with time capture) Consider a slotted Aloha system with a large number of terminals transmitting to a central station. if the ﬁrst packet arrives at the station θ seconds before any other packet in that slot. (Slotted Aloha with acknowledgments) Consider a system that employs the slotted Aloha protocol.. Find ˜ L ( z) = E [ z L ] . Let A i (i=1. ˜ 1. Problem 2. In the following use the standard Poisson assumptions. the distribution of A i . ˜ * 3.74 CHAPTER 3: ALOHA PROTOCOLS EXERCISES Problem 1. when two or more packets are transmitted concurrently all are lost. Find the relation between the throughput S and the offered load G in this system. Denote by F the length of an unsuccessful transmission period. What is the maximal throughput of this system? 2. its pdf f A(t) .) be the ith interarrival time between packets arriving within a single ˜ unsuccessful transmission period. its Laplace transform F A ˜ 2. Typically.. and * (s ) . Acknowledgment packets are transmitted on the same channel used for data packets and their length is the same as that of a data packet. Using the result of part (3) ﬁnd the average length of a transmission period (successful or not) and the throughput of the system. Find F A(t) . The terminals are evenly distributed around the station so that in terms of propagation time the most remote terminal is τ seconds away from the station. Let L be the number of transmissions in an unsuccessful transmission period. the collided data packets and the data packet that has been acknowledged by the collided acknowledgment packet have to be retransmitted. This phenomenon is called time capture.
Section : EXERCISES 75 packets arrive for transmission on the channel according to a Poisson process with average g packets per second. Use the standard Poisson assumptions. Problem 4. . (Slotted Aloha with power capture [Met76]) This problem deals with a different model for the capture phenomenon. The model used here is typical to power capture. Draw your result. (Pure Aloha with variable length packets [BeB80]) Consider a network whose nodes transmit (arbitrarily distributed) variable length packets according to the pure Aloha protocol. In the following we would like to compute the probability of capture and the channel throughput. 2. then the transmitted packet of node i is captured and thus successfully received. Determine the maximal throughput for K=2. determine the possible values of the couple (S1. Show that regardless of θ.3. 1.. Determine the maximal throughput for any K. 1. Let S = ∑ j = 1 S i be the total throughput of the systems. a capture means receiving correctly a packet even when other packets are transmitted during the same time... Explain this result. What is the throughput S of the system? 5. All other nodes that transmitted during the slot have to retransmit their packets at some later time. As explained in Problem 3. Let Si denote the rate of generation of new packets (per slot) by nodes of class i (1≤ i ≤K) and let Gi denote the total rate of transmitted packets (per slot) by nodes of class i. Deﬁne t = t2t1. Find the pdf f t(w) conditioned on k). What is the probability of capture in a slot (conditioned on k)? 3. namely.2. S2). If a single node of class i (2≤ i ≤K) is transmitting simultaneously with any number of users of classes 1. i1. K 2. that some nodes transmit with higher power than others. For K=2. How should the Gi’s (1≤ i ≤K) be chosen in order to obtain the maximal throughput? What happens when K→ ∞? 3. Let t1 be the time of the ﬁrst arrival in the station in the given slot and let t2 be the time of the second arrival in the station in the given slot. (In questions (1) and (2) below we compute the quantities conditioned on the event that k packets are ready to transmit at the beginning of the slot). maximum throughput is achieved for gT>1. What is the (unconditional) probability of capture in a slot? 4. Assume that the population of users in the system is divided into K classes all using the slotted Aloha protocol.. Problem 5.
In a similar manner we ˜ deﬁne x c . Assume that new packets are generated according to Poisson process with mean λ0 and the total trafﬁc on the channel is also Poisson with mean λc. Let S = λ0 x0 and G = λc xc. Xc*(x) and xc for the packets transmitted on the channel (new packets and those that are retransmitted because of collisions). Xc(x). Let X0(x) be its distribution function. Therefore. 1.Constant packet length: x 0 = x 0 with probability 1). X c (s) where X 0 = ds 0 ds ∞ * 4. x0(x) its density function. and x0 its mean. Find the relation between S and G in the following three special cases: ˜ . Let λ0(x) = λ0 x0 (x) and Λ 0 = ∫0 – e –sx λ 0(x) dx . X0*(s) its Laplace transform. Prove that the probability P suc(x) for successful transmission of a packet whose length ˜ is x c = x (the packet is from the total trafﬁc on the channel) is given by: P suc(x) = exp { – λ c ( x c + x ) } ˜ Note that P suc(x) depends only on xc and not on the distribution of x c . xc(x).X c () = xc xc *(1) – G X 0 (. Prove the following relationships: * X 0 (s – λ c) * X c (s) = * X 0 (– λ c) * X c (s + λ c) * X 0 (s) = * X c (λ c) 3. ˜ Let x 0 be a random variable representing the length of packets (the time required to transmit a packet) generated by the nodes (this random variable represents also the length of packets that are successfully transmitted). the length distribution of packets that are successfully transmitted is different from the length distribution of an arbitrary packet transmitted in the channel. Prove that the throughput is given by: x 0 Ge –G Ge –G *(1) G S = . . 2. Prove that * * Λ o (s) = λ c X c (s + λ c)exp { – λ c x c } * dΛ o (s) S = – ds s=0 5.76 CHAPTER 3: ALOHA PROTOCOLS Observe that long packets are more likely to collide than short packets. .) xc d * d * *(1) X (s) .
(Aloha with K channels) Consider a system consisting of K separate slotted channels (with slot boundaries synchronized).Section : EXERCISES 77 .Each user selects randomly and uniformly one of the K channels and performs all its activity on that channel..Every user and for every packet transmission (or retransmission) selects randomly. all packets have the same constant length. 6. uniformly. There are packets of two types . . .The users are divided in advance into K equally sized groups each assigned one channel. 2. 1.Exponential packet length: x 0 = µe –µx . The probability that a packet is of the ﬁrst type is α. Find the throughput of such a system for the inﬁnite population case if the channels are selected uniformly at random. Prove that the throughput of a pure Aloha system is maximized when the packet length distribution is deterministic i. The system operates according to the slotted Aloha protocol with the speciﬁc channel chosen according to some rule. . . .Dual packet size: x 0 = αδ(x – x 1) + ( 1 – α )δ(x – x 2) .Every user upon generating a new packet selects randomly and uniformly one channel to be used for transmission and all retransmissions of that packet.e. Problem 6.those with constant length x 1 and those with constant length x 2 ( x 1 < x 2 ). and independent of the past the channel over which this speciﬁc packet is to be transmitted. For the ﬁnite population case ﬁnd (and compare) the throughput delay characteristics if .
78 CHAPTER 3: ALOHA PROTOCOLS .
Every user is equipped with a receiver anyway. disturbing another user will cause the currently transmitted packet to be retransmitted. his packet will clearly not be successful if transmitted and. The carrier sensing family of protocols is characterized by sensing the carrier and deciding according to it whether another transmission is ongoing. further. “Concurrently” here does not really mean at the very same time. The variations on the CSMA scheme are due to the behavior of users that wish to transmit and ﬁnd (by sensing) the channel busy. For more detail the reader is referred to any of the many books and surveys that deal with various aspects of CSMA protocols such as Tannenbaum’s [Tan81]. Suppose the channel has been idle for a while and two users concurrently generate a packet. Carrier sensing does not. All the carrier sense multiple access (CSMA) protocols share the same philosophy: when a user generates a new packet the channel is sensed and if found idle the packet is transmitted without further ado. however. ToK77]. exhibited fairly poor performance which can be attributed to the “impolite” behavior of the users namely. relieve us from collisions. Moreover. When a collision takes place every transmitting user reschedules a retransmission of the collided packet to some other time in the future (chosen with some randomization) at which time the same operation is repeated. Hence “concurrently” actually means within a time window of duration equal to signal propagation time. that is. . The process of listening to the channel is not that demanding. Consider a behavior that we generically characterize as “listen before talk”. or Clark Pogran and Reed’s [CPR78]. possibly disturbing yet another packet. if one user starts transmitting it takes some time for the signal to propagate and arrive at the other user. ToK75. Each will sense the channel. described in the previous section. Stallings’ [Sta85]. to detect another user’s transmission does not require receiving the information. It does not take much to realize that even little consideration can beneﬁt all. This latter quantity becomes therefore a crucial parameter in the performance of these protocols. Most of these variations were introduced and ﬁrst analyzed in a series of papers by Tobagi and Kleinrock [KlT75. discover it is idle.CHAPTER 4 CARRIER SENSING PROTOCOLS The Aloha schemes. whenever one has a packet to transmit he does so without consideration of others. every user before attempting any transmission listens whether somebody else is already using the channel. it sufﬁces to sense the carrier that is present when signals are transmitted. If this is the case the user will refrain from transmission to the beneﬁt of all. and transmit the packet to result in collision.
80
CHAPTER 4: CARRIER SENSING PROTOCOLS
4.1. NONPERSISTENT CARRIER SENSE MULTIPLE ACCESS
In the nonpersistent versions of CSMA (NPCSMA) a user that generated a packet and found the channel to be busy refrains from transmitting the packet and behaves exactly as if its packet collided, i.e, it schedules (randomly) the retransmission of the packet to some time in the future. The following analysis is based on Kleinrock and Tobagi [KlT75]. Throughput Analysis To evaluate the performance of NPCSMA let us adopt a model similar to that used in the evaluation of the performance of the Aloha protocol. We assume an inﬁnite population of users aggregately generating packets according to a Poisson process with parameter λ. All packets are of the same length and require T seconds of transmission. When observing the channel, packets (new and retransmitted) arrive according to a Poisson process with parameter g packets/sec. In addition to the assumptions of the model used to analyze the Aloha protocol, the model used for CSMA deals also with system conﬁguration which is manifested by a propagation delay among users. Denote by τ the maximum propagation delay in the system (measured in seconds) and deﬁne a ∆ τ ⁄ T to be the normalized propagation time. We assume = that all users are “τ seconds apart” that is, τ is the propagation delay between every pair of users. With this assumption the following analysis provides a lower bound to the actual performance. Consider an idle channel and a user scheduling a transmission at some time t (see Figure 4.1). This user senses the channel, starts transmitting at time t and does so for T seconds; once he is done it will take τ additional seconds before the packet arrives at the destination. This transmission therefore causes the channel to be busy for a period of T+ τ seconds. If, at time t’> t+τ another user scheduled a packet for transmission, that user would sense the channel busy and refrain from transmission. If, however, some other user scheduled a packet for transmission during the period [t,t+τ], that user would sense the channel idle, transmit its packet, and cause a collision. The initial period of the ﬁrst τ seconds of transmission is called the vulnerable period since only within this period is a packet vulnerable to interference. Figure 4.1(b) depicts a situation in which a packet transmission starting at time t is interfered by two other transmissions that start in the interval [t,t+τ]. In the case of a collision the channel will therefore be busy for some (random) duration between T+τ and T+2τ. This period in which transmission takes place is referred to as the transmission period (TP). In the case of NPCSMA the transmission period coincides with the busy period. Having completed a transmission period the channel will be idle for some time until the next packet transmission is scheduled. We therefore observe along the time axis a succession of cycles each consisting of a transmission period followed by an idle period (see Figure 4.1(a)). Because packet scheduling is memoryless, the times in which these cycles start are renewal points. As we did before,
Section 4.1.: NONPERSISTENT CARRIER SENSE MULTIPLE ACCESS
81
Cycle Busy Period ˜ B T
Cycle Idle Period ˜ I τ
Cycle
Cycle
˜ B
˜ I
time (a) Cycle Structure
˜ B
Vulnerable Period
˜ Y
t+τ t
t+T
t+T+τ ˜ t+T +τ+Y
(b) Unsuccessful Transmission Period FIGURE 4.1: Non Persistent CSMA Packet Timing ˜ ˜ we denote by B the duration of the busy (transmission) period, and by B its mean. Let U be the time duration within the transmission period in which a successful packet is being ˜ transmitted (mean U), and let I the duration of the idle period (with mean I). The cycle ˜ ˜ length is clearly B + I and the throughput is given by S = U / (B+I). We now derive these quantities. Consider ﬁrst the idle period. Its duration is the same as the duration between the end of packet transmission and the arrival of the next packet. Because packet scheduling is memoryless we get ˜ ˜ F I (x) = Prob [ I ≤ x ] = 1 – Prob [ I > x ] = 1 – P [ No packet scheduling during x ] = 1 – e –gx
82
CHAPTER 4: CARRIER SENSING PROTOCOLS
which means that I is exponentially distributed with mean 1 I = g The expected useful time U can also be easily computed. When a packet is successful the channel carries useful information for a duration of T secondsthe time it takes to transmit the packet; in the unsuccessful case no useful information is carried at all or, in other words U = T Successful Period 0 Unsuccessful Period If P suc denotes the probability that a transmitted packet is successful then ˜ U = E [ U ] = T ⋅ P suc + 0 ⋅ ( 1 – P suc ) = T P suc . The probability of a successful transmission, P suc is the probability that no packet is scheduled during the vulnerable period [t, t+ τ]. Hence, P suc = Prob [ No arrival in the period [ t, t + τ ] ] = e –gτ and thus U = T e –gτ . ˜ To compute B, the average transmission period duration, let Y be a random variable such ˜ that t + Y denotes the time at which the last interfering packet was scheduled within a ˜ transmission period that started at time t (see Figure 4.1(b)). Clearly, Y < τ and for a suc˜ = 0 . Using this notation the duration of the transmission cessful transmission period Y period is ˜ ˜ B = T +τ+Y ˜ The period Y is characterized by the fact that no other packet is scheduled for transmis˜ ˜ sion during the period [ t + Y , t + τ ] for otherwise the packet that is transmitted at t + Y ˜ would not have been the last packet to be transmitted in [ t + Y , t + τ ] . Thus, the probabil˜ ity distribution function of Y is ˜ F Y (y) = Prob [ Y ≤ y ] = Prob [ No packet arrival during τ – y ] = e –g ( τ – y ) The above relation holds for 0 ≤ y ≤ τ . For negative values of y the probability distribution function vanishes and for values greater than τ it equals unity. It is important to notice that F Y (y) has a discontinuity at y=0 which means that care must be taken when the probability density function is derived. Denoting by δ(t) the Dirac impulse function we get
With our previous deﬁnition of the normalized propagation time a we get Ge –aG S = . the lower a the better the performance. G ( 1 + 2a ) + e –aG A sketch of the throughput versus normalized offered load G for various values of the normalized propagation time a is shown in Figure 4.2. let G denote the average scheduling rate of packet measured in packets per packet transmission time.: 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS 83 f Y (y) = e –gτ δ(y) + ge –g ( τ – y ) from which 1 – e –gτ ˜ E [ Y ] = τ – g and ﬁnally 1 – e –gτ ˜ B = E [ T + τ + Y ] = T + 2τ – g Putting all these results together we get U T e –gτ gT e –gτ S = . persists to wait and transmits as soon as the channel becomes idle. The 1persistent CSMA (1PCSMA) is an alternative to nonpersistent CSMA that avoids such situations. the channel is always used if there is a user with a packet.+ g g which is the desired relation we were seeking. In fact. in other words G=gT. To that end. This is achieved by applying the following rule: A user that senses the channel and ﬁnds it busy. 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS With nonpersistent CSMA there are situations in which the channel is idle although one or more users have packets to transmit. Consequently.2.1) (4. These graphs have the same shape as those for the Aloha system except for the evidently improved throughput. We remark also that having the same characteristic shape as the Aloha protocol means that NPCSMA (as the other protocols in this family) suffer from the same instability problems from which Aloha suffers. .Section 4.= .. (4. the extreme case of a → 0 yields a throughput of G/(1+G) which does not decrease to zero with increasing load. As is expected.2.= B+I 1 – e –gτ 1 g ( T + 2τ ) + e –gτ T + 2τ – .2) As before we normalize the quantities with respect to the packet transmission time. 4.
When observing the channel over time.2 0. All users that sense the channel busy in some transmission period.1 1 10 100 1000 Offered Load (G) FIGURE 4. The type of a transmission period that follows another transmission period depends on the number of those persistent users waiting for the current transmis . If no packet is scheduled for transmission during some transmission period.01 0.84 CHAPTER 4: CARRIER SENSING PROTOCOLS 1 0.4 0. each consisting of an idle period (no packet is scheduled for transmission during this period).3 0.5 a=0. Notice that a transmission period starts either with the transmission of a single packet (call it type 1 transmission period).01 Throughput (S) 0.6 0. The analysis presented in the following is considerably simpler and is based on Shoraby et. then an idle period begins as soon as this transmission period ends. transmit their scheduled packets at the beginning of the successive transmission period. or with the transmission of at least two packets (call it type 2 transmission period). al.001 a=0. A transmission period that follows an idle period is always a type 1 transmission period.3). [SMV87].1 0 0. followed by a busy period that consists of several successive transmission periods (see Figure 4.9 0.7 a=0 Nonpersistent CSMA a=0.001 a=1.2: ThroughputLoad of Nonpersistent CSMA The performance of the 1persistent CSMA scheme was ﬁrst analyzed by Kleinrock and Tobagi [KlT75].1 0.8 0. Throughput Analysis We adopt the same model as the one used in the analysis of nonpersistent CSMA.0 0. one sees a sequence of cycles.
These states (0. This is depicted in Figure 4.3.Section 4. For consistency.3: 1Persistent CSMA Timing sion to end. The knowledge of the system state at the beginning of some transmission period (together with the scheduling points of packets during this transmission period) is sufﬁcient to determine the system state at the beginning of the successive transmission period. . Deﬁne the state of the system at the beginning of a transmission period to be the type of that transmission period.1 and 2) correspond to a threestate Markov chain embedded at the beginning of the transmission periods.2. The possible transitions among the three states of the embedded Markov chain are depicted in Figure 4.: 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS 85 Cycle Cycle Cycle ˜ I TP Type 1 T τ Busy Period ˜ B TP Type 2 TP Type 1 Idle Period ˜ I Busy Period ˜ B TP Type 1 time (a) Cycle Structure ˜ ˜ Y τ –Y ˜ T +Y ˜ t t +Y t+τ t+T t+T+τ ˜ t+T +τ+Y (b) Unsuccessful Transmission Period FIGURE 4. an idle period is also viewed as a transmission period that starts with no transmitted packets (call it type 0 transmission period).4.
hence T 0 = 1 ⁄ g . which we do next. i=0. .3) To compute the throughput we still have to compute π i and Ti. it is necessary that no packets arrive during its ﬁrst τ seconds that constitute its vulnerable period (the probability of the latter event is e –gτ .2 be the stationary probability of being in state i. and from the same renewal arguments we used before the throughput is given by T π 1 e –gτ S = 2 πi T i ∑ i=0 1 P12 (4. namely that the system is in ˜ a transmission period of type i.. The reason is that the length of a transmission period with either a single packet or with two or more packets.1. i. for a type 1 transmission period to be successful.2. Let π i i=0.86 CHAPTER 4: CARRIER SENSING PROTOCOLS P11 P01 0 P10 P21 P20 2 P22 FIGURE 4. Yet. Regarding the ˜ ˜ random variables T 1 and T 2 the important observation is that they have the same distribution.1. Consequently. Let T i i=0. Since the length of a packet is T seconds.1. is determined only by the time of arrival of the last packet (if any) within the vulnerable period (the ﬁrst τ seconds of the transmission period). and does not depend at all on the type of the transmission period.e.4: State Transitions of 1PCSMA During a type 0 transmission period no packets are transmitted and during type 2 transmission periods two or more packets are transmitted and collide. exponentially distributed with mean 1/g. only type 1 transmission periods may result in a successful transmission.2 be a random variable representing the ˜ length of type i transmission period and let E [ T i ] = T i . Transmission Period Lengths The nature of the idle period in 1persistent CSMA is identical to that of nonpersistent CSMA.
1(b)). if no packets arrive within the transmission period.t+τ]). the next transmission period will be of type 1.2.6) we obtain: 1 – e –gτ ˜ T 1 = T 2 = T + τ + E [ Y ] = T + 2τ – g State Probabilities From the state diagram in Figure 4. Let a transmission period start at time t and let Y be a random variable representing the time (after t) of the last packet that arrived during the vulnerable period [t.1) and (4. P1 j = P2 j Using (4. Speciﬁcally.10) (4. 2 .4) We already derived the probability distribution function and probability density function ˜ of Y and found (see equations (4. If a single packet arrives within the transmission period. P 10 π 0 = 1 + P 10 P 10 + P 11 π 1 = 1 + P 10 1 – P 10 – P 11 π 2 = 1 + P 10 (4. Finally. the next transmission period will be of type 2.4) and (4.: 1PERSISTENT CARRIER SENSE MULTIPLE ACCESS 87 The computation of T 1 and T 2 is identical to that of B in the nonpersistent CSMA (Fig˜ ure 4.9) . the next transmission period will be of type 0.t+τ] ˜ of a transmission period that started at time t ( Y = 0 if no packets arrive during [t.5) (4. Then ˜ ˜ ˜ T1 = T2 = T + τ + Y (4. 1. if at lease two packets arrive within the transmission period.4 we have. Therefore. at least τ seconds after it begins.9) we have. the type of the next transmission period is determined (only) by those packets scheduled for transmission after the transmission period begins.6) j = 0. at least τ seconds after it begins.8) (4.Section 4.7) 0≤y≤τ (4.2)) f Y (y) = e –gτ δ(y) + ge –g ( τ – y ) 1 – e –gτ ˜ E [ Y ] = τ – g Combining (4. π 0 = π 1 P 10 + π 2 P 20 π 1 P 12 = π 2 ( P 21 + P 20 ) π0 + π1 + π2 = 1 When a type 1 or a type 2 transmission period starts.8) and (4. (4.
03 .12) τ = ge –g ( T + τ ) T + gτ T +  2 Combining (4. Recall that the 1persistent CSMA was devised in an attempt to improve the performance of the nonpersistent CSMA by reducing the extent of the idle periods. however. While. In particular note that. (4. Conditioning on Y = y .7). For low load. evidently. generally.10). in the best possible case. we obtain τ τ P 10 = ∫ 0 e –g ( T + y ) f ˜ ( y) d y Y = ∫ e –g ( T + y ) ( e –gτ δ(y) + ge –g ( τ – y ) ) dy 0 (4. S → 0 when G → ∞ . .3). these graphs have the same form as those of the nonpersistent CSMA. the maximum for S in this case is obtained for G ≈ 1. (4. (4.5. an idle period) only if no packet is scheduled for transmission after time t+τ and before the end of the type 1 transmission period (namely. and (4. the next transmission period will be of type 0 (namely.11). This attempt is. The probability of this event is e –g ( T + y ) . Unconditioning.11) = ( 1 + gτ )e –g ( T + τ ) In a similar manner we obtain τ τ P 11 = ∫ g(T + 0 y )e –g ( T + y ) f ˜ ( y) d y Y = ∫ g ( T + y )e –g( T + y ) ( e –gτ δ(y) + ge –g( τ – y ) ) dy 0 (4. performance is less than expected.88 CHAPTER 4: CARRIER SENSING PROTOCOLS Assume that a type 1 transmission period starts at time t. time t+y+T+τ). not quite successful since for high load the nonpersistent CSMA outperforms the 1persistent CSMA. 1persistent CSMA shows a slightly better throughput and improved performance.12) we obtain the throughput gT e –g ( T + 2τ ) [ 1 + gT + gτ ( 1 + gT + gτ ⁄ 2 ) ] S = g ( T + 2τ ) – ( 1 – e –gτ ) + ( 1 + gτ )e –g ( T + τ ) or in a normalized form: Ge –G ( 1 + 2a ) [ 1 + G + aG ( 1 + G + aG ⁄ 2 ) ] S = G ( 1 + 2a ) – ( 1 – e –Ga ) + ( 1 + Ga )e –G ( 1 + a ) This relation is depicted in Figure 4. G(1 + G) S a → 0 = 1 + Ge G and thus.
001 0.Section 4. These slots are sometimes referred to as minislots since they are shorter than the time required to transmit a packet.: SLOTTED CARRIER SENSE MULTIPLE ACCESS 89 1 1Persistent CSMA 0.5: ThroughputLoad of 1Persistent CSMA 4. Users behave as follows. for Nonpersistent CSMA the packet is rescheduled to some randomly chosen time in the future.3.8 Throughput (S) 0.e. SLOTTED CARRIER SENSE MULTIPLE ACCESS Consider an environment similar to that described for the CSMA protocols except for a slotted time axis. All packets are of the same length and require T seconds for transmission. As in every slotted system users are restricted to start transmissions only at minislot boundaries. Let the slot size equal the maximum propagation delay τ which means that any transmission starting at the beginning of a slot reaches (and could be sensed by) each and every user by the end of that slot. If the channel is sensed busy. i.01 0. and for 1PCSMA the .6 a=0 0.1 0.3. We also assume that the packet size T is an integer multiple of the propagation delay τ and denote by a the ratio between τ and T (1/a is therefore an integer). 1/a minislots (the packet occupies the channel one more minislot before all other users have received it). namely. When a packet is scheduled for transmission at a given time the user waits to the beginning of the next minislot at which time it senses the channel and if idle transmits its packet for T seconds. then the corresponding CSMA protocol is applied.1 1 10 Offered Load (G) FIGURE 4.4 a=0. We assume that carrier sensing can be done in zero time (we may assume that τ includes the propagation delay as well as the carrier sensing time).0 0 0..2 a=1.
6). τ I = 1 – e –gτ (4. Successful Transmission Period Unsuccessful Transmission Period Idle Period T τ Idle Period FIGURE 4. Observing the channel we see that a busy period B consists of con˜ secutive transmission periods.6). By our deﬁnition.2).2. For it to be two minislots long means that there are no arrivals in its ﬁrst minislot and there is at least one arrival in its second minislot. In both cases collided packets are retransmitted at some random time in the future. Throughput of Slotted Nonpersistent CSMA We adopt a similar approach to that taken in the corresponding unslotted systems (see Sec˜ tions 3. The idle period I is the time elapsed between every two successive busy periods (see Figure 4. the length of an idle period is at least one minislot. A collision occurs if two or more packets arrive within the same minislot and are scheduled for transmission in the next minislot.2. Continuing this reasoning and considering the Poisson scheduling process we have.6: Slotted Nonpersistent CSMA Packet Timing For the idle period to be one minislot long means that there is at least one arrival in the ﬁrst minislot of the idle period. A busy period will contain k transmission periods if there is at least one arrival .90 CHAPTER 4: CARRIER SENSING PROTOCOLS user waits until the channel becomes idle and then starts transmission.13) k = 1. ˜ P [ I = kτ ] = ( e –gτ ) k – 1 ( 1 – e –gτ ) so. … An outcome of the deﬁnition of the model is the fact that both successful and unsuccessful transmission periods last T+τ seconds (see Figure 4. 2.1 and 3.
= . We have P suc = Prob [ Successful Transmission Period ] = Prob single arrival in last minislot some arrivals before the transmission period Prob [ Single arrival in last minislot ] gτe –gτ = . we have B ˜ E [ U ] = T P suc T +τ where P suc is the probability of a successful transmission period. … .: SLOTTED CARRIER SENSE MULTIPLE ACCESS 91 in the last minislot of each of the ﬁrst k1 transmission periods. When a transmission period is successful the channel carries useful information for T seconds. since a transmission period has been initiated.= Prob [ Some arrivals in last minislot ] 1 – e –g τ The division by the probability of “some arrivals” is noteworthy. It is necessary because we are computing the probability of a single arrival in the last minislot of the preceding transmission period knowing that there was at least one arrival. while it carries no useful information in unsuccessful transmission periods. Since the number of ˜ ˜ transmission periods during B is B ⁄ ( T + τ ) . ˜ Prob [ B = k ( T + τ ) ] = ( 1 – e –gτ ) k – 1 e –gτ so.Section 4. 2.= T +τ B+I τ T + τ – Te –gτ .+ e –gτ 1 – e –gτ Using a=τ/T and G=gT we have: aGe –aG S = 1 + a – e –aG k = 1.3. and no arrival in the last minislot of the kth transmission period. Thus. Putting all these together we get B T P suc Tgτe –gτ U T +τ S = . T +τ B = e –gτ Following a similar approach to that used in the slotted Aloha case we deﬁne a cycle as the ˜ period consisting of a busy period followed by an idle period and denote by U the amount of time within a cycle during which the channel carries useful information.
and no packet arrives in the kth transmission period.92 CHAPTER 4: CARRIER SENSING PROTOCOLS When a is very small we obtain: G S a → 0 = 1+G which is identical to the unslotted case when a → 0 . is different from the success probability in any other transmission period within the busy period. For the ﬁrst transmission period in a busy period to be successful we need the last minislot of the idle period to contain exactly one arrival (notice that we know there is at least one arrival there. P suc1 = Prob Successful transmission in first transmission period of a busy period gτe –gτ = Prob [ Single transmission in a minislot at leat one arrival ] = 1 – e –gτ For any transmission period in a busy period. Hence. So. T +τ B = –g ( T + τ ) e The probability of success in the ﬁrst transmission period in a busy period. The probability of success of the ﬁrst transmission period in a busy period is P suc1 and therefore T ⋅ P suc1 is the expected amount of time the channel carries useful information during these periods. P suc2 . The expected number of transmission periods (other than the ﬁrst) in a busy . i. P suc2 = Prob [ Successful transmission in nonfirst period in a busy period ] = Prob [ Single arrival in a transmission period at least one arrival ] g ( T + τ )e –g ( T + τ ) = 1 – e –g ( T + τ ) The channel carries useful information only during successful transmission periods. Throughput of Slotted 1PCSMA The analysis of the slotted 1PCSMA is similar to that of slotted Nonpersistent CSMA.13). other than the ﬁrst.e. The distribution of the busy period is ˜ Prob [ B = k ( T + τ ) ] = ( 1 – e –g ( T + τ ) ) k – 1 e –g ( T + τ ) since a busy period will contain k transmission periods if at least one packet arrives in each of the ﬁrst k1 transmission periods (as opposed to minislots in the nonpersistent case). P suc1 . The mean of the idle period is given by (4.. to be successful we must have exactly one arrival during the previous transmission period. since it is the last minislot of the idle period).
: SLOTTED CARRIER SENSE MULTIPLE ACCESS 93 period is B ⁄ ( T + τ ) – 1 .6 Nonpersistent 0.8 Slotted Nonpersistent Throughput (S) 0.4 0.01 0.7 1 CSMA a=0.2 Slotted 1Persistent 0 0. In summary.01 0.= B+I ( T + τ ) [ 1 – e –gτ ] + τe –g ( T + τ ) or in a normalized form Ge –( 1 + a )G [ 1 + a – e –aG ] S = ( 1 + a ) [ 1 – e –aG ] + ae –( 1 + a )G This relation is depicted in Figure 4.7: ThroughputLoad of Slotted 1Persistent and Nonpersistent CSMA For the case of a very small minislot size we have . since each transmission period lasts T+τ seconds. the expected amount of time within a cycle that the channel carries useful information is B – (T + τ) U = T P suc1 + . T +τ The throughput is therefore given by U gT e –g ( T + τ ) [ T + τ – T e –gτ ] S = .Section 4.3. The probability of success in each of these transmission periods is P suc2 and therefore [ B ⁄ ( T + τ ) – 1 ] ⋅ T ⋅ P suc2 is the expected amount of time the channel carries useful information during these periods.1 1 10 100 1000 Offered Load (G) FIGURE 4.P suc2 .001 0.
Together with carrier sensing this produces a variation of CSMA that is known as CSMA/CD (Carrier Sense Multiple Access with Collision Detection). Clearly. . 4. a new avenue must therefore be sought. namely that users can detect interference among several transmissions (including their own) while transmission is in progress and abort transmission of their collided packets. the very small gain achieved is probably not worth the cost of keeping the users synchronized. a cycle is composed of a transmission period followed by an idle period. The pure Aloha protocol suffers most. as no precautions to reduce collisions are taken. even a bit. CSMA reduces the level of interference caused by overlapping packets by allowing users to sense the carrier due to other users’ transmissions. all are lost and must be retransmitted. Beside the ability to sense carrier. CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION The Aloha family of protocols suffers from the inherent interference of concurrently transmitted packets. and inhibit transmission when the channel is in use and a collision is inevitable. This is advantageous since analysis of slotted systems is often much simpler than the unslotted ones. unfortunately. as expected. an observation that is the foundation of the protocols described in this section. Finding the way to shorten the busy period is therefore our only recourse. Hence. From a theoretical standpoint the close performance means that the slotted system can serve as an approximation of the unslotted one. To improve the throughput we must therefore reduce the cycle length. some local area networks (such as Ethernet) have an additional feature.4. whenever the transmission of two or more packets overlap in time. performance can be improved by shortening the duration of unsuccessful transmission periods. the duration of successful transmission periods should not be changed for this is the time the channel is used best. a slightly better performance of the slotted systems. as we now explain. our measure of performance. Shortening the idle period is possible by means of 1persistent protocols which. As we have seen. If this can be done sufﬁciently fast then the duration of an unsuccessful transmission would be shorter than that of a successful one which is the effect we were looking for.94 CHAPTER 4: CARRIER SENSING PROTOCOLS Ge –G [ 1 + G ] S a → 0 = G + e –G Comparing these graphs with those of the corresponding unslotted systems we note. is the ratio between the expected useful time spent in a cycle to the cycle duration itself. Practically speaking. Throughput. perform poorly under most loads. that is. CSMA protocols appear to be the best possible solution since their performance depends only on the endtoend propagation delaya quantity that is not alterable (except by a different topological design). To further improve performance.
In this ﬁgure Unsuccessful Transmission Period Distance t 0+τ + τ cd t 1 t 0+τ t 0+τ +τ cd+τ cr t 1+γ t 0+ T +τ B A t0 t 0+γ t 1+τ +τ cd+τ cr t 1+τ t 1+τ +τ cd Successful Transmission Period FIGURE 4. By similar calculation. at t0+τ+τcd +τcr user B completed the consensus reenforcement procedure which reaches user A at t0 +2τ+τcd +τcr. hence in the worst case. From user A’s standpoint this transmission period lasted γ = 2τ + τ cd + τ cr . the channel will be sensed busy. Suppose that user A starts transmission at time t0 when the channel is idle. user B completes this transmission period at time t1+γ. The channel is therefore busy for a period of t1+γ t0. the transmission is aborted and the packet is scheduled for transmission at some later time. i. In all CSMA protocols. the propagation between whom is τ. Beyond this time. except that if a collision is detected during transmission. Suppose. so that at time t0 +τ+τcd user B positively determines the collision.e. a transmission that is initiated when the channel is idle reaches all users after at most one endtoend propagation delay.8 captures this situation.. Thus. It takes τcd for a user to detect the collision. then its transmission reaches user B at t0 +τ. every user upon detection of a collision initiates a consensus reenforcement procedure.4. τ. at time t1=t0+τ.Section 4.8: Collision detection Timing t 0+ T time we consider two users A and B. In the worst case user B starts transmission just prior to the arrival of A’s packet. that B initiates a transmission at time t1 <t0 +τ (when B still senses an idle channel). in an . which is manifested by jamming the channel with a collision signal for a duration of τcr to ensure that all network users indeed determine that a collision took place. In many local area networks such as Ethernet. further. The spacetime diagram of Figure 4.: CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION 95 The operation of all CSMA/CD protocols is identical to the operation of the corresponding CSMA protocols.
. Throughput of Slotted Nonpersistent CSMA/CD With the nonpersistent CSMA/CD time alternates between busy periods (that contain both successful and unsuccessful transmission periods) and idle periods.9: Slotted Nonpersistent CSMA/CD Packet Timing ˜ ˜ We denote. Thus X takes on only values that are certain integer multiples of τ. senses the channel. Denoting ˜ by X the length of the transmission period we have ˜ X = T + τ Successful transmission period γ + τ Unsuccessful transmission period In the following we analyze the slotted versions of CSMA/CD.9). the user with that packet waits until the end of that minislot. In addition. the length of the busy period by B . The analysis is based on the work by Tobagi and Hunt [ToH80]. and follows the corresponding version of the CSMA/ CD protocol. as before.e. Thus. it is assumed that time is quantized into minislots of length τ seconds and that all users are synchronized so that transmissions can begin only at the start of a minislot. Busy Period Successful Transmission Period Unsuccessful Transmission Period Idle Period T Idle Period γ τ time FIGURE 4. the length of the idle period by I ˜ The distribution of the idle period is identical to that and the useful time in a cycle by U computed for slotted nonpersistent CSMA. i. namely..96 CHAPTER 4: CARRIER SENSING PROTOCOLS unsuccessful transmission period the channel remains busy for a duration of γ+τ. we assume that both γ and T (the transmission time of a packet) ˜ are integer multiples of τ. A cycle is a busy period followed by an idle period (see Figure 4. when a packet is scheduled for transmission during some minislot.
…. 2. A busy period will contain l transmission periods if there was at least one arrival in the last minislot of each of the ﬁrst l1 transmission periods.4. the probability that the busy period contains exactly l ( l ≥ 1 ) transmission periods is e –gτ ( 1 – e –gτ ) l – 1 and the average number of transmission periods within the busy period is 1 ⁄ e –gτ . l where l corresponds to the total number of transmission periods in the busy period and k corresponds to the successful transmission periods. 2.e. we have that the probability distribution of the length of the busy period is ˜ Prob [ B = k ( T + τ ) + ( l – k ) ( γ + τ ) ] l k = e –gτ ( 1 – e –gτ ) l – 1 P suc ( 1 – P suc ) l – k k l = 1.. and no arrival in the last minislot of the lth transmission period. Therefore. Every successful transmission period contributes T ˜ to U while unsuccessful transmission periods do not contribute anything.Section 4. k = 0.: CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION 97 ˜ Prob ( I = kτ ) = ( e –gτ ) k – 1 ( 1 – e –gτ ) so the expected length of the idle period is τ I = 1 – e –gτ k = 1.16) Each transmission period that contains a successful transmission is of length T+τ seconds while a transmission period with an unsuccessful transmission is of length γ+τ seconds.17) ˜ We now turn to compute U = E [ U ] . …. (4. In addition. Therefore. … (4.14) (4. Thus. ˜ Prob ( U = kT ) = Prob [ k successful transmission periods in a busy period ] = ˜ ∑ Prob [ B = k ( T + τ ) + ( l – k ) ( γ + τ ) ] l=k ∞ . i.15) The probability that a certain transmission in a busy period is successful is the probability that the transmission period contains exactly one packet (given that it contains at least one packet). B = ∑ ∑ [ k ( T + τ ) + ( l – k ) ( γ + τ ) ] Prob [ k ( T + τ ) + ( l – k ) ( γ + τ ) ] l = 1k = 0 ∞ l P suc ( T + τ ) + ( 1 – P suc ) ( γ + τ ) = e –gτ . the probability of a single arrival in a minislot (given that there was at least one arrival): gτe –gτ P suc = Prob [ Single tranmission at least one transmission ] = 1 – e –gτ (4. 1.
4 0.18) Combining (4.10 depicts the throughputload characteristics of the nonpersistent CSMA with collision detection.19) where γ’ is the ratio between γ and the transmission time of a packet (γ’ =γ/T).10: ThroughputLoad of Slotted Nonpersistent CSMA/CD .01 0.20) (4.98 CHAPTER 4: CARRIER SENSING PROTOCOLS from which U = ˜ ∑ kT Prob [ U = kT ] k=0 ∞ T = P suc .3 0.2 0. The improvement in performance is readily apparent. – gτ e (4.5 0. B+I gτT e –gτ + [ ( 1 – e –gτ ) – gτe –gτ ]γ + τ In a normalized form: aGe –aG S = aGe –aG + ( 1 – e –aG – aGe –aG )γ' + a (4. (4.18) we compute the throughput: U gτT e –gτ S = .1 1 10 100 1000 Offered Load (G) FIGURE 4.1 a=1.8 0. Figure 4. 1 0.1 0 0.= .20) is identical to slotted nonpersistent CSMA. Notice that when γ’=1 the result in (4.01 Throughput (S) 0. and (4..17).0 0.7 γ’=2a a=0 a=0.9 0.6 0.001 a=0.16).
and a cycle is a busy period followed by an idle period (see Figure 4. the length of the remainder of the busy period is a function of x.11).Section 4.[ γ + τ + [ 1 – a 0(γ + τ) ]B(γ + τ) ] 1 – a 0(x) (4. given that a transmission period is of length x.21) . The quantity B(x) is given by: a 1(x) B(x) = . then the duration of the i+1st transmission period depends only on x i .11: Slotted 1Persistent CSMA/CD Packet Timing a transmission period in the busy period depends (only) on the length of the preceding transmission period. given that a transmission period is of length x. Under the Poisson assumption a i(x) = ( ( gx ) i e –gx ) ⁄ i! . Similarly. and its average is denoted by B(x). This is so since the type of the ith transmission period (success or collision) is determined by the number of arrivals during the previous transmission period which. depends only on its duration. Hence. the average time the channel is carrying successful transmissions in the remainder of the busy period is denoted by U(x). except for the ﬁrst transmission period that depends on arrivals on the ˜ preceding minislot.4. Let ai(x) be the probability of i arrivals during a period of length x.: CARRIER SENSE MULTIPLE ACCESS WITH COLLISION DETECTION 99 Throughput of Slotted 1Persistent CSMA/CD With the 1persistent CSMA/CD the time also alternates between busy periods (containing successful and unsuccessful transmission periods) and idle periods. Notice that here a success or failure of Busy Period Successful Transmission Period Unsuccessful Transmission Period Idle Period T Idle Period γ τ time FIGURE 4.[ T + τ + [ 1 – a 0(T + τ) ]B(T + τ) ] 1 – a 0(x) a 1(x) + 1 – . in turn. Denoting by x i the duration of the ith transmission period in the busy ˜ period.
using (4. the remainder of the busy period is of length B(T+τ). the argument of B(.22) with x=T+τ and x=γ+τ one obtains two equations with two unknowns U(T+τ) and U(γ+τ). (4. computation is straight forward. in which case the remainder of the busy period will be of length γ+τ (the length of an unsuccessful transmission period). The average time during a cycle that the channel is carrying successful transmissions. U(x) is given by a 1(x) U (x) = . B(x) can be determined for any x.13 shows the characteristics of the slotted nonpersistent and 1persistent CSMA with and without collision detection. In addition. if there is at least one arrival within T+τ (probability 1a0(T+τ)). the “gap” in performance between the nonpersistent and the 1persistent CSMA when collision detection is used has narrowed down.12 shows the throughput load characteristic of the 1persistent CSMA with collision detection.= B+I τ B(τ) + 1 – e –gτ (4. a rather dramatic change is seen here.21) is not quite enough since values of B(. in which case the remainder of the busy period will be of length T+τ (the length of a successful transmission period). Because of its better performance in low load and .21) corresponds to a successful transmission of the single packet that arrives during x.23) While the above evaluation of the throughput does not result in a closed form.) can be interpreted as an arrival period for the next transmission period and clearly the arrival period for the entire busy period is the ﬁrst minislot before it started. As opposed to the nonpersistent case. Having determined these values.21).21) we notice that substituting τ for x in (4.22) is similar to that of (4.22) The explanation of (4.01. In a similar manner. The second term in (4. In addition. Moreover.[ 1 – a 0(γ + τ) U (γ + τ) ] 1 – a 0(x) . U. an additional length of B(γ+ τ) is the remainder of the busy period.) appear on the right handside as well.15). Observing (4. The average length of an idle period is τ ⁄ ( 1 – e –gτ ) (see (4. and the throughput is given by U U (τ) S = .21) and obtaining two equations with two unknowns B(T+τ) and B(γ+τ) which can be solved easily. expressed in terms of U(T+τ) and U(γ+τ). if there is at least one arrival within γ+τ. The expected duration of the entire busy period is B(τ) since x. For comparison Figure 4. Again. is given by U(τ).[ T + [ 1 – a 0(T + τ) ]U (T + τ) ] 1 – a 0(x) a 1( x ) + 1 – .100 CHAPTER 4: CARRIER SENSING PROTOCOLS The ﬁrst term in (4. to yield the expected length of a busy period. Superiority of the collision detection mechanism is evident. in particular x=τ. Figure 4.21) corresponds to an unsuccessful transmission due to arrivals during x. all for a=0. This is overcome by setting x=T+τ and x =γ+τ in (4.
0 0.5. .1 1 10 100 1000 10000 Offered Load (G) FIGURE 4.9 0.4 0.001 Throughput (S) 0. RELATED ANALYSIS Carrier sensing has become extremely popular in recent year for one major reasonlocal area networks (LANs). Variablelength packets The performance of CSMA and CSMA/CD protocols with two different possible packet lengths has been studied in [ToH80]. 4. the amount of published material is so large that it is impossible to cover it all or even classify it properly.2 0.8 0.3 0.5. In fact. In the next few paragraphs we point the reader at some relevant additional work on the subject. This is a result of the ease of implementing collision detection in broadcast LANs.1 0 0.6 a=0.12: ThroughputLoad of Slotted 1Persistent CSMA/CD because its delay characteristic is favorable the 1persistent CSMA with collision detection is so popular in local area networks.: RELATED ANALYSIS 101 1 0. The direct outcome is an enormous amount of research and analysis of these protocols under all types of circumstances.7 γ’=2a a=0.01 a=1.Section 4.01 0. Batch packet arrivals are considered in [Hey82] for the CSMA protocol and in [Hey86] the constant length packets are allowed to be grouped into messages of random size (with a geometric distribution) and the performance of CSMA/CD is studied.5 a=0.1 0. For a broader survey the reader is referred to [Tob80].
a system with ﬁnite number of buffers per user is studied in [ApP86] and approximate analysis of a system with inﬁnite buffers is presented in [TTH88].02 1Persistent/CD Nonpersistent Throughput (S) 0. .1 0 0.8 0.2 0.6 0.5 0.102 1 0.01 0.4 0. Packet delay was analyzed in [CoL85. BeC88] using the matrixgeometric approach [Neu81] and interdeparture time (both distributions and moments) was derived by Tobagi [Tob82b]. Delay and interdeparture times Numerous papers studied the throughput delay characteristics of CSMA and CSMA/CD protocols. For instance.001 Nonpersistent/CD 1Persistent 0.1 1 10 100 1000 Offered Load (G) FIGURE 4. Coyle and Liu [CoL83] treated a ﬁnite population as did Tobagi and Hunt in [ToH80]. A twouser system with inﬁnite buffers is analyzed in [TaK85a].01 γ’=.7 CHAPTER 4: CARRIER SENSING PROTOCOLS Slotted CSMA a=0.01 Nonpersistent without collision detection. 1persistent without collision detection Nonpersistent with collision detection 1persistent with collision detection Buffered users The performance of CSMA and CSMA/CD with ﬁnite number of users having ﬁnite or inﬁnite buffering capabilities has been considered in several works.13: Comparison of ThroughputLoad of Slotted systems with a=0.3 0.9 0.
As such. Other versions of virtual time CSMA have been considered in [ZhR87. by Limb and Flores [LiF82]. An idea of how to implement collision detection in radio systems is described in [Rom86] (see Exercises). It was shown by Rom [Rom84] that users can identify themselves their ordinal number on the network. Takagi and Yamada [TaY83] proposed to resolve collisions deterministically while Molle and Kleinrock [MoK85] proposed using virtual time to resolve collisions. Tob82a.: RELATED ANALYSIS 103 Ordered users Most CSMAtype local networks are implemented using a coax as the transmission means. KiK83]. CuM88]. Collision detection in radio systems Implementing collision detection in local area networks is relatively simple since the transmitted and received signals are of the same order of magnitude. if properly used. the attachment of the users to the cable introduces an inherent order. RoT81]. In radio systems the received signal is considerably weak compared with the transmitted signal and therefore collision detection cannot be implemented via a simple comparison. . which. Attempts to incorporate priority structures are presented in [RoT81. and by Tobagi and Fine [FiT84]. Performance improvement In an attempt to improve the performance of CSMA and CSMA/CD protocols Meditch and Lea [MeL83] derive some optimized versions (keeping stability in mind).Section 4. Such attempts were done by Tobagi and Rom [ToR80. can improve performance substantially.5.
Find E [ x ] and E [ x 2 ] . 4. Find the throughput of a slotted 1persistent CSMA for a→ 0 and compare with the result for an unslotted system. Increasing a uniformly decreases the throughput. Find the throughput of nonpersistent and 1persistent CSMA for a→ 0 and g→∞. There is a single load for which the channel attains its capacity. and S. The random variable x is the time from the moment the channel is ﬁrst sensed for that packet until its transmis˜ ˜ sion is complete. the partial throughputs S1 = λ1 T. Deﬁne. 1. S2=λ2 T. What is the average delay under the conditions of part (2)? . (CSMA with a heavy user [ScK79] Consider a network containing one central computer and a large (read: inﬁnite) number of terminals all operating as follows. a terminal packet transmission consists of a seconds of carrier followed by T seconds of information (and. When is the throughput maximal? What is the throughput in this case? ˜ ˜ 3. Find the throughputs S1. of course. a more seconds to ‘clean’ the channel). S2. From the results of part (3) compute the average delay of a computer packet. (For convenience deﬁne Λi=λi (T+2a)). as usual. at the beginning of the slot where transmission is attempted. Problem 2. For Nonpersistent CSMA show that 1. listens to the channel and will transmit a packet of duration T only if the channel is sensed idle.104 CHAPTER 4: CARRIER SENSING PROTOCOLS EXERCISES Problem 1. All packets are of equal length T. Let x denote the service time of a central computer’s packet. 2. The terminals each have a single packet buffer and communicate using the slotted ALOHA scheme. Let λ1 and λ2 be the Poisson arrival rate (in packets/second) of the computer and the terminals respectively and let g be the combined offered load of the terminals. Problem 3. The computer. 2. and the total throughput S = S1+S2. (Note that the computer has in fact a lower priority since it defers transmission to an ongoing terminal transmission). The computer has an inﬁnite buffer and uses a modiﬁed CSMA (see below). To increase the total throughput the slot size is set to T+2a and the terminal packets carry a preamble of length a (where a is the propagation delay in the system). Explain the difference. That is.
a user arriving in a noisy slot within an idle period thinks (erroneously) that the channel is busy and behaves accordingly. If the channel is sensed busy the user concludes that his packet collided and will not transmit the entire packet. To analyze the system we assume an inﬁnite number of users collectively forming a Poisson arrival process with average g packets per slot. All packets are of equal length T and occupy an integer number of slots. The analysis of such a system is based on a slotted nonpersistent CSMA. Let the slot size equal the endtoend propagation delay τ and let all the packets be of equal length T (assumed to occupy an integer number of slots). If the channel is busy. a coin is ﬂipped. In addition. When a slot is noisy while a packet is being transmitted. The slot size τ equals the endtoend propagation delay in the system. The . Each transmitting user pauses during transmission and senses the channel. (Collision detection in radio systems [Rom86]) This problem deals with a collision detection scheme usable also in radio network. What is the probability that a given slot in the idle period is not the last slot of that period. Each slot is noisy with probability pe which is independent of the system and of the noise in previous slots. The period from the start of transmission to the time of abortion is called the collision detection interval. 2. 1. 1. Compute the throughput of the system. If the channel is sensed idle transmission proceeds as usual. How should p be chosen to maximize the throughput? Problem 5. Let there be inﬁnitely many users collectively generating (on the channel) a Poisson distributed offered load with mean g packets/second. (Mixed mode CSMA) Consider the following version of slotted CSMA: Whenever a packet is scheduled for transmission. that packet is destroyed. the user will not abort transmission immediately but rather will continue transmitting for some period of time and then abort. the corresponding user senses the channel. Verify your answer for the case pe = 0. with probability p the packet is scheduled for transmission at some later time (nonpersistent) and with probability 1p the user waits until the channel becomes idle and then transmits the packet (1persistent). However. (CSMA with a noisy channel) Consider a slotted nonpersistent CSMA system with a noisy channel. Check your results for p=0 and p=1. Repeat part (a) for a ≠ 0 . If the channel is idle the packet is transmitted. For the analysis we deﬁne an embedded Markov chain in the beginning and end of each transmission period. Assume a=0 and use the standard Poisson assumption to determine the relation between S and G as a function of p. The scheme is essentially a nonpersistent CSMA with the following exception. 2. Problem 6.Section : EXERCISES 105 Problem 4.
. 3. What is the probability of a long transmission period? What is the probability of a short transmission period? 4. 6. Deﬁne an embedded Markov chain with which you intend to analyze the system. Why does such an r exist? What system parameters does it depend on? What is the value of r for low load. Let ropt be that r that maximizes the throughput. Is it necessary for R to be identical for all users? 2.106 CHAPTER 4: CARRIER SENSING PROTOCOLS collision detection interval is R=rτ (where r is an integer) and a transmitting user will pause for one slot randomly and uniformly chosen among the r slots of transmission starting with the second slot. 1. Watching the channel we observe long and short transmission periods. Derive an expression for the channel throughput. What is the probability of a successful transmission period? 5.
The users involved in the collision split randomly into two subsets. If we look into the philosophy behind the Aloha protocol. Moreover. If slot k+1 is either idle or contains a successful transmission. whether the slot was idle (no packet was transmitted). If two or more packets are transmitted in a slot. the attempts to resolve collisions are always deferred to the future. in most versions of these protocols. At the end of each slot the users of the system know what happened in the slot. so that collisions are resolved more efﬁciently and without chaotic events. This is known as the ternary feedback model. The latter is referred to as binary feedback model. THE BINARYTREE PROTOCOL The most basic collision resolution protocol is called the binarytree CRP (or binarytree protocol) and was proposed almost concurrently by Capetanakis [Cap79]. by (for instance) each ﬂipping a coin. a collision occurs and the packets involved in the collision have to be retransmitted. we notice that there is no sincere attempt to resolve collisions among packets as soon as they occur. all users that are not involved in the collision wait until the collision is resolved. Instead. those that ﬂipped 0. New packets arrive to the system according to a Poisson process with rate λ packets/slot. The underlying model and the assumptions used here are identical to those assumed for the slotted Aloha protocol. in slot k say. 5.CHAPTER 5 COLLISION RESOLUTION PROTOCOLS We have seen that the original Aloha protocol is inherently unstable in the absence of some external control. namely. with the hope that things will then work out. In these protocols. then the system is stable. The channel is slotted and the users can transmit packets (whose length is one slot) only at beginning of slots. the efforts are concentrated on resolving collisions as soon as they occur. retransmit in slot k+1 while those that ﬂipped 1 wait until all those that ﬂipped 0 transmit successfully their packets. or contained a successful transmission (exactly one packet was transmitted) or there was a collision (at least two packets were transmitted). This ensures that if the rate of arrival of new packets to the system is smaller than the rate at which collisions can be resolved (the maximal rate of departing packets). but they never do. For some protocols it sufﬁces for the users to know whether the slot contained a collision or not. called Collision Resolution Protocols (CRP). Tsybakov and Mikhailov[TsM78]. According to this protocol when a collision occurs. somehow. The users in the ﬁrst subset. new packets that arrive to the system are inhibited from being transmitted while the resolution of collisions is in progress.1. The basic idea behind these protocols is to exploit in a more sophisticated manner the feedback information that is available to the users in order to control the retransmission process. and Hayes[Hay78]. the users . In this chapter we introduce and analyze multi access protocols with a different philosophy.
i. corresponding to the splitting of the subset into two new subsets. The root of the tree corresponds to the slot of the original collision.1.108 CHAPTER 5: COLLISION RESOLUTION of the second subset (those that ﬂipped 1) retransmit in slot k+2. the users whose packets collided in slot k+1 (the “colliding users”) ﬂip a coin again and operate according to the outcome of the coin ﬂipping.1) in which every vertex corresponds to a time slot. This recursion will be manifested later when these protocols are analyzed. the users that ﬂipped 1 in that slot transmit (in slot 7). then the procedure is repeated. One user ﬂips 0 and transmits (successfully) in slot 5 causing all users that ﬂipped 1 in slot 4 to transmit in slot 6.e. Another . Vertices whose subsets contain at least two users (labeled “≥ 2”) indicate collisions and have two outgoing branches. respectively. 0 0 1 ≥2 0 ≥2 1 1 2 3 4 5 6 0 1 ≥2 1 1 0 1 ≥2 7 8 9 10 0 1 ≥2 1 1 0 11 12 Collision resolution Interval (CRI) FIGURE 5. But even at the description level the recursive operation of the protocol is best described by a binarytree (see Figure 5. so that only users that collided in slot 1 and ﬂipped 1 transmit in slot 2. If slot k+1 contains another collision. Each vertex of the tree also designates a subset (perhaps empty) of backlogged users. and so on. Vertices corresponding to empty subsets (labeled “0”). At this point it is neither known how many users nor who are the users that collided in this slot.1: Example of the BinaryTree Protocol Operation To further understand the operation of the protocol we consider in detail the example depicted in Figure 5. We refer to a user having a packet that collided (at least once) as a backlogged user. A collision occurs in slot 1. By the rules of the protocol no newly arrived packet is transmitted while the resolution of a collision is in progress. In this example there is one such user and therefore slot 6 is a successful one. The users that ﬂipped 1 in slot 2 transmit again in slot 4. Another collision occurs in slot 2 and the users involved in that collision ﬂip a coin again. resulting in another collision and forcing the users involved in it to ﬂip a coin once more. The above explanation shows that the protocol is speciﬁed by a recursion: a group of colliding packets is split into two subgroups each of which is subjected to the same procedure as the original group.. In this example. Each of the colliding users ﬂip a coin and those that ﬂipped 0 transmit in slot 2. or subsets containing one user (labeled “1”) are leaves of the tree and indicate an idle and a successful slot. all the colliding users of slot 2 ﬂipped 1 and therefore slot 3 is idle. Now that the collision among all users that ﬂipped 0 in slot 1 has been resolved.
and all the packets that were at the top of the stack are transmitted. There is only one such user. To complete the description of the protocol. and thus can determine when to transmit his packet. The operation of the binarytree protocol can also be described in terms of a stack (see Figure 5. meaning that at least two users ﬂipped 0 in slot 7. Observing again Figure 5.1.: THE BINARYTREE PROTOCOL 109 collision occurs. each backlogged user can keep track of his own position on that tree (while a collision is being resolved). Users that are not involved in the collision. hence slot 11 is idle. The time interval starting with the original collision (if any) and ending when this collision is resolved is called collision resolution interval (CRI). In the same manner. The users that collided in slot 8 ﬂip a coin and. The users that ﬂip 0 remain therefore at the top of the stack to be popped and transmitted in the next slot. For the correct operation of the binarytree protocol. Another collision is observed in slot 8. wait until the collision is resolved.. can also follow the binarytree and thus know exactly when the collision is resolved. It is clear from this example that each user can construct the binarytree shown in Figure 5. in slot 10. is that new packets are inhibited from being transmitted while a resolution of a collision is in progress. packets that arrive to the system while a resolution of a collision is in progress. and his transmission is. In the above example the length of the CRI is 11 slots. and the users involved in it ﬂip a coin. Then. transmit the users that ﬂipped 1 in slot 8. the same results.Section 5. In this example there is no such user. in fact. of course. This is. That is. successful. as it happens. Performance analysis of binarytree CRP can be done either with the tree or stack representations reaching. at the same time. In the example of Figure 5. the one in the ﬁrst slot. a standard description of tree traversal by a stack representation. users do not have to distinguish idle slots from successful ones. in other words. the CRI is over and newly arrived packets (if any) are pushed onto the top of the stack and operation proceeds as before. i. to specify the ﬁrsttime transmission rule. Finally. completing the resolution of the collision that occurred in slot 7 and. the binary feedback sufﬁces.1 all new packets arriving to the system during slots 1 through 11 are transmitted for the ﬁrst time in slot 12.2). We say that a collision is resolved when the users of the system know that all packets involved in the collision have been transmitted successfully. that which we assumed all along (known as the obviousaccess scheme). In each slot the stack is popped.1 we see that the order of transmission corresponds exactly to the traversal of the tree. at which time they are transmitted.1 by following the feedback signals corresponding to each slot. In this book we conﬁne ourselves to the . One alternative. In case of a successful transmission or an idle slot no further operations are done on the stack. The binarytree protocol dictates how to resolve collisions once they occur. of course. there is a single user that ﬂipped 0 and it transmits (successfully) in slot 9.e. the users that ﬂipped 1 in slot 7 must transmit in slot 11. In case of a collision. when the stack empties the collision is resolved. we need to specify when newly generated packets are transmitted for the ﬁrst time or. the stack is pushed with the users that ﬂip 1 and then pushed again with those that ﬂip 0. Clearly then.
1. When no packet. Clearly. then the system is stable.2: BinaryTree Protocol Stack Representation tree approach. These results are then used to derive the stability condition for this system. is transmitted in the ﬁrst slot of a CRI then the CRI lasts exactly one slot. One would expect that if the arrival rate of new packets is smaller than the “effective service rate” even when the system is highly loaded (n is large).110 CHAPTER 5: COLLISION RESOLUTION Flip 0 Flip 1 Transmission Level Level 0 Level 1 Level 2 Level 3 Idle or Success Collision FIGURE 5. al. Loosely speaking. The quantity Bn plays a crucial role in the analysis of the binarytree CRP. or a single packet. For analysis via the stack representation the interested reader is referred to the work by Fayolle et. To resolve the collision each participating user ﬂips a coin and proceeds correspondingly. 5.1. [FFH85].2. The analysis in this section is based on Massey [Mas81]. This statement is made rigorous in Section 5. the expected delay of a packet is bounded. In the following we ﬁrst compute the moments of the time required to resolve a collision among n packets and obtain tight bounds for these moments. Finally. the number of slots required to ˜ resolve such a collision is a random variable.1. But ﬁrst we compute the ˜ moments of B n .1) . Denote therefore by B n the length (in slots) of a CRI given that it starts (in its ﬁrst slot) with a collision among n packets. since n packets are transmitted successfully during Bn slots. and let ˜ B n = E [ B n ] . Moments of the Conditional Length of a CRI Assume that at some given slot n packets collide. the ratio n/Bn represents the “effective service rate” of packets in a CRI that starts with the transmission of n packets.. we show that when the system is stable. hence ˜ ˜ B0 = B1 = 1 (5.
from equation (5.Section 5. solving for Bn we obtain the recursion n–1 1+ ∑ Qi(n)Bi + ∑ Qi(n)Bn – i n–1 n i=1 B n = 1 – Q 0(n) – Q n(n) i=0 (5. Then it takes B i ˜ slots to resolve the collision among the i users that ﬂipped 0.4) provides the initial values. First.: THE BINARYTREE PROTOCOL 111 When n ≥ 2.3) Bn leading to Bn = 1 + i (5. from equation (5.1) we have B0 = B1 = 1 Then.2) ˜ ˜ = 1 + Bi + Bn – i n≥2 (5. the probability that exactly i of the n colliding users ﬂip 0 (and hence transmit in the next slot) is n Q i(n) = p i ( 1 – p ) n – i i Given that i users ﬂipped 0.4) ˜ = E [ Bn i ] = 1 + Bi + Bn – i n≥2 ∑ Qi(n) ( Bi + Bn – i ) i=0 n n≥2 (5. .5) In this last equation Bn appears on both sides of the equation.1. there is a collision in the ﬁrst slot of the CRI and the random coin ﬂipping takes place. Let p be the probability that a user ﬂips 0 whenever it ﬂips the binary coin. the length of the CRI is ˜ Bn i 0≤i≤n (5.3) ˜ The 1 corresponds to the slot of the initial collision among the n users. From the above rela˜ tion the ﬁrst moment of B n can be recursively derived as follows. Finally.6) 1+ ∑ [ Qi(n) + Qn – i(n) ]Bi i=0 = 1 – Q 0(n) – Q n(n) n≥2 where equation (5. it takes B n – i i additional slots to resolve the collision among the ni users that ﬂipped 1. Then.
However. obviously.e.7) Gn(z) appears on both sides so. Taking the ﬁrst derivative at z=1 leads.8) allowing recursive computation of Gn(z). to the result of equation (5.7) twice with respect to z and obtain ˙ ˙ ˙ ˙ ˙˙ G n(z) = 2 ∑ Q i(n)G i(z)G n – i(z) + 2 ∑ Q i(n)G i(z)G n – i(z) + 2 ∑ Q i(n)G i(z)G n – i(z) + ˙˙ ˙˙ ∑ Qi(n)Gi(z)Gn – i(z) + ∑ Qi(n)Gi(z)Gn – i(z) i=0 i=0 i=0 n i=0 n i=0 n n n (5.112 CHAPTER 5: COLLISION RESOLUTION Computation of a similar nature can be done for any moment desired. ˜ ˜2 Let Vn be the second moment of B n i. From (5.7) In equation (5..1) we immediately have V0 = V1 = 1 To compute the second moment for higher values of n we differentiate equation (5.6). To compute this gen= ˜ ˜ erating function notice that the random variables B i and B n – i are independent and thus G n(z) = or G n(z) = z ∑ Q i(n)G i(z)G n – i(z) i=0 n ˜ E [ z Bn ] = ∑ i=0 n ˜ ˜ Q i(n)E [ z 1 + Bi + Bn – i ] = ∑ Qi(n)zE [ z B ]E [ z B ˜i i=0 n ˜n–i ] n≥2 n≥2 (5. V n = E [ B n ] .9) ˙˙ ˙ Substituting z=1 in (5.1) translate into G 0(z) = G 1(z) = z (5. a more ˜ general approach is that of the generating function G n(z) ∆ E [ z Bn ] . We now proceed to calculate the second moment.9) and using the facts that G n(1) = B n and G n(1) = V n – B n we obtain . From this generating function the moments of ˜ B n can be computed recursively by taking derivatives at z=1. as we did before. solving for Gn(z) we obtain: z ∑ Q i(n)G i(z)G n – i(z) i=1 G n(z) = 1 – z 2 [ Q 0(n) + Q n(n) ] n–1 n≥2 and the initial conditions of equation (5.
Differentiating (5.12) Vn = 1 + ∑ k=2 n 2n! .10) n≥2 where we used (5. Table 1 contains some of the values of Bn and Vn when p = 1 ⁄ 2 along with values of the “effective service rate” n/Bn. Indeed. it is possible to obtain closed˜ form expressions for the moments of B n .1.: THE BINARYTREE PROTOCOL 113 V n – B n = 2 ∑ Q i(n) ( B i + B n – i + B i B n – i ) + i=0 n ∑ Qi(n) ( V i – Bi + V n – i – Bn – i ) i=0 n = B n – 1 + 2 ∑ Q i(n)B i B n – i + i=0 n ∑ Qi(n) ( V i + V n – i ) i=0 n (5.1 – Q 0(n) – Q n(n) n≥2 (5.5) and (5.5).Section 5. p = 1 ⁄ 2 is the only real value for which the ﬁrst derivative vanishes.* k – ( 1 – p )k ] ( n – k )! [ 1 – p n 4 ( k – 1 ) ( –1 ) k ∑ Bi* Bk* – i i=0 + where * B0 k=2 ∑ k ] [ 1 – pk – ( 1 – p )k n≥2 = 1 * B1 = 0 * Bk = k=2 ∑ k ] k! [ 1 – p k – ( 1 – p ) k n n 2 ( k – 1 ) ( –1 ) k k≥2 It is interesting to investigate the behavior of Bn as a function of p. that Bn is minimized for p = 1 ⁄ 2 for all n. Solving for Vn we obtain the following recursion: 2B n – 1 + 2 ∑ Q i(n)B i B n – i + n n–1 ∑ [ Qi(n) + Qn – i(n) ]V i i=0 i=0 V n = . The derivations of these expressions are quite lengthy and for the interested reader are given in Appendix A at the end of this chapter.12) twice with respect to p we note that independent of n. The resulting expressions are: Bn = 1 + n ∑ k=2 n 2 ( k – 1 ) ( –1 ) k n  k [ 1 – p k – ( 1 – p ) k ] n≥2 k (5. In fact. Judging by the values of n/Bn it is interesting to note that the protocol resolves collisions among a small number of packets .11) The recursive nature of equations (5. We conclude therefore.B k + . at p = 1 ⁄ 2 the ﬁrst derivative vanishes while the second derivative is positive.11) are sometimes inconvenient to work with and a direct expression might be preferred.
5097 0.28 9 24.0000 0.8532 0.4191 0. We return to this topic in Section 5. we assume that (5. We are seeking an upper bound for Bn of the form Bn ≤ αm n – 1 n≥m (5. ˜ Table 1: The ﬁrst and second moments of B n for p = 1 ⁄ 2 .114 CHAPTER 5: COLLISION RESOLUTION more efﬁciently than collisions among a large number of packets.3131 0.000 7 19.82 13 36.7382 0.≥ . This also motivates looking for the smallest αm for which the bound holds.0000 1. The motivation for a bound of this form is that it guarantees a strictly positive “effective service rate” (n/Bn) for large n. The approach to determine αm is as follows.1 3 7.2.3 5 13.3955 0.3579 982.3548 1838. We ﬁx m (m ≥2) and choose some αm so that Bn ≤αmn 1 for n= m (any αm≥ (B m +1)/m is feasible).3561 1376.2.555 8 22. If indeed a bound of that form can be found then one would be able to write n 1 1 .3554 1599.89 14 39.6238 0.9691 0.5238 0.3726 197.3569 1171.13) for some arbitrary m and with some αm>0.3678 286.0000 6 16.36 12 33.4 Bounds on the Moments In the previous section the ﬁrst two moments of the conditional length of a CRI are computed. In this section we derive upper bounds for these moments.3622 514.2813 0.6667 0. Next.2010 0.+ Bn αm αm Bn n≥m and hence the effective service rate is guaranteed to be larger than 1/αm.42 11 30.13) holds .3913 68.9 4 10.3604 653.4000 33.3801 124.0000 1.05 2 5.3646 392.0854 0. n Bn n/Bn Vn n Bn n/Bn Vn n Bn n/Bn Vn 1 1.3590 809.63 15 42.00 10 27.
∑i = 0 iQ i( j) = jp and j ∑i = 0 iQ j – i( j) = j ( 1 – p ) . i.: THE BINARYTREE PROTOCOL 115 up to j1. .. i.Section 5. i. j1 and by induction establish the validity of (5.. The point of departure is equation (5.. such that m–1 ∑ [ Qi( j) + Q j – i( j) ] ( Bi – αm i + 1 ) ≤ 0 j>m i=0 or.1.13) for j.6). Bn ≤ αmn 1 for n= m.14) It thus follows that (5..13) holds if we choose αm so that the summation in (5.e. m–1 i=0 B j ≤ ( α m j – 1 ) + 1 – Q 0( j) – Q j( j) ∑ [ Qi( j) + Q j – i( j) ] ( Bi – αm i + 1 ) (5..e.m+1.e...14) is nonpositive for all j>m. j–1 B j [ 1 – Q 0( j) – Q j( j) ] = 1 + ∑ [ Qi( j) + Q j – i( j) ]Bi ∑ [ Qi( j) + Q j – i( j) ]Bi + ∑ [ Qi( j) + Q j – i( j) ]Bi i=m j–1 i=0 m–1 = 1+ i=0 Applying the induction hypothesis we obtain m–1 B j [ 1 – Q 0( j) – Q j( j) ] ≤ 1 + m–1 ∑ [ Qi( j) + Q j – i( j) ]Bi + ∑ [ Qi( j) + Q j – i( j) ] ( αm i – 1 ) i=m j–1 i=0 = 1+ ∑ [ Qi( j) + Q j – i( j) ] ( Bi – αm i + 1 ) i=0 j + ∑ [ Qi( j) + Q j – i( j) ] ( αm i – 1 ) – [ Q0( j) + Q j( j) ] ( αm j – 1 ) ∑ [ Qi( j) + Q j – i( j) ] ( Bi – αm i + 1 ) + αm j – 2 – [ Q0( j) + Q j( j) ] ( αm j – 1 ) i=0 m–1 = 1+ i=0 m–1 = ∑ [ Qi( j) + Q j – i( j) ] ( Bi – αm i + 1 ) + [ 1 – Q0( j) – Q j( j) ] ( αm j – 1 ) j j i=0 where we used the facts that ∑i = 0 Q i( j) = 1 . Therefore.
885 2.182 2.896 2. one can compute the supremum in (5.250 2.886 2.6) or (5. choosing say.500 2. Hence. after computing Bi for i <m (using (5.16) To summarize.886 9 2. for a given m.15) for p = 1 ⁄ 2 .333 2.400 2.886 2.926 2.914 2.886 2.886 2.898 2.667 4 2.886 2. m–1 j > m ∑ i [ Qi( j) + Q j – i( j) ] i=0 the induction step follows and hence (5. m 2 3 4 5 6 7 8 9 j 3 2.892 2. Table 2: Values of right side of (5. m=6.16) and hence determine αm as the maximum between that supremum and (Bm+1)/m.886 7 2.886 10 2.887 2.885 6 2.875 5 2.12)).886 m–1 (5.886 11 2. we see from Table 2 that the supremum is 2.886 2.286 2.13) holds.116 CHAPTER 5: COLLISION RESOLUTION m–1 i=0 α m ≥ m–1 ∑ [ Qi( j) + Q j – i( j) ] ( Bi + 1 ) ∑ i [ Qi( j) + Q j – i( j) ] i=0 j>m (5.889 2. We recall that Bn is minimized for p = 1 ⁄ 2 .886 2.222 2.895 2.920 2.885 2. Recall that we started with αm that satisﬁes B m ≤ α m m – 1 . the best bound of the form (5.889 2.886 2.15) for p = 1 ⁄ 2 . Table 2 depicts some of the values of the expression on the right side of (5.880 2. thus .200 2.885 2.887 2.886 8 2.886 2. if we choose αm as the maximum between (Bm+1)/m and the following supremum: ∑ [ Q i( j) + Q j – i( j) ] ( B i + 1 ) i = 0 sup .15) can also be computed for any desired value of m.886 when p = 1 ⁄ 2 .15) Having computed previously the values of Bi. the right hand side of equation (5.896 2.907 2.894 2.13) is obtained when p = 1 ⁄ 2 . Therefore. For example.
17) Although inequality (5.886n – 1 n≥6 (5.17) does not hold for n<6.886 we see from Table 1 that (5.346.18) We conclude that the “effective service rate”.. Next. we assume that (5.11) that can be rewritten for j≥2 (using (5.19) holds. we shall shortly present improved versions of the binarytree protocol that yield much better performance.19) for n=m.e. This effective rate is smaller than e1 the maximum throughput of the slotted Aloha protocol. m =6.. Using the above methodology.m+1.19) for j. The point of departure is equation (5. In subsequent computations of system parameters and performance we shall take the value αm = 2.346 and thus the system is expected to be stable for arrival rates smaller than 0. and αm=2. V n = α m n + 1 for n=m.886.5) as 1 + 2 ∑ Q i( j) ( B i + B j – i + B i B j – i ) + j j–1 ∑ [ Qi( j) + Q j – i( j) ]V i (5.19) where αm is the same one used to bound Bn and is determined by the procedure described above.13) we have . Vn.19) holds for all values of n up to 2 j1 i. it is possible to develop a bound on the second moment of the conditional length of a CRI.1.. We ﬁrst check the validity of (5. j1 and by induction establish the validity of (5.17) and Table 1) by B n ≤ 2.. for large n is 1/2. This bound will be required in developing an upper bound on the expected delay of a packet.20) i=0 i=0 V j = 1 – Q 0( j) – Q j( j) Using (5. n/ Bn.886n + 1 n≥0 (5.886 ≅ 0. For p = 1 ⁄ 2 .Section 5.: THE BINARYTREE PROTOCOL 117 B n ≤ 2. of the form 2 V n ≤ αm n 2 + 1 n≥m (5.. it is very easy to bound Bn for all n ≥0 (by using (5.
for the induction hypothesis to hold we require that .22) in (5.21) and (5.21) i=0 m–1 = 2 ∑ Q i ( j) [ B i + B j – i + B i B j – i – i=0 m–1 = 2 2 2 ∑ Qi( j) [ Bi + B j – i + Bi B j – i – αm i ( j – i ) + 1 ] + 2αm [ j 2 p – jp ( 1 – p ) – j 2 p 2 ] – 2 i=0 2 Similarly.20) we obtain: m–1 2 i=0 2 V j ≤ α m j 2 + 1 + 1 – Q 0( j) – Q j( j) m–1 2 ∑ [ Qi( j) + Q j – i( j) ] ( V i – αm i 2 – 1 ) 2 ∑ Q i ( j) [ B i + B j – i + B i B j – i – α m i ( j – i ) + 1 ] i=0 + 1 – Q 0( j) – Q j( j) Therefore. assuming that V i ≤ α m i 2 + 1 up to i ≤ j – 1 we have j–1 2 ∑ [ Qi( j) – Q j – i( j) ]V i ≤ ∑ [ Qi( j) – Q j – i( j) ]V i + ∑ [ Qi( j) – Q j – i( j) ] ( αm i 2 + 1 ) i=0 m–1 i=m 2 ∑ [ Qi( j) – Q j – i( j) ] ( V i – αm i 2 – 1 ) 2 2 ∑ [ Qi( j) + Q j – i( j) ] ( αm i 2 + 1 ) – [ Q0( j) – Q j( j) ] ( αm j 2 + 1 ) m–1 j–1 i=0 = i=0 j (5.118 CHAPTER 5: COLLISION RESOLUTION 2 ∑ Q i( j) ( B i + B j – i + B i B j – i ) ≤ 2 i=0 j j m–1 ∑ Qi( j) ( Bi + B j – i + Bi B j – i ) i=0 +2 ∑ Q i ( j) [ α m i – 1 + α m ( j – i ) – 1 + ( α m i – 1 ) ( α m ( j – i ) – 1 ) ] Q i ( j) ( B i + B j – i + B i B j – i ) + 2 2 ∑ Qi( j) [ αm i ( j – i ) – 1 ] i=m 2 αm i ( j 2 – i ) + 1 ] + 2 ∑ Q i( j) [ α m i ( j – i ) – 1 ] i=0 j j i=m m–1 = 2 ∑ (5.22) + i=0 m–1 = 2 ∑ [ Qi( j) – Q j – i( j) ] ( V i – αm i 2 – 1 ) i=0 2 2 + α m [ 2 jp ( 1 – p ) + j 2 p 2 + j 2 ( 1 – p ) 2 ] + 2 – [ Q 0( j) + Q j( j) ] ( α m j 2 + 1 ) Substituting (5.
The system is said to be stable if the Markov chain { B(k). For instance. i→∞ We start by computing the following conditional expectation: . We are now ready to prove this claim. ( k = 0. limsup E [ ( B(k + 1) – B(k) ) ( B(k) = i ) ] < 0 .881 ≅ 0. m=6 and αm=2.23) for p = 1 ⁄ 2 . We ﬁrst notice that the chain is irreducible. … ) } is ergodic. Let B(k) be an integervalued random variable that represents the length (in slots) of the kth CRI. 2. 5.347. it can be shown that (see [Mas81]) B n ≥ 2.1. the expected delay of a packet is ﬁnite). (We shall see in the next section that when the system is stable.1. ˜ To obtain sufﬁcient conditions for which the Markov chain { B(k). 1. 2. 1. 2. When the binarytree CRP is executed then along the time axis we observe a sequence of collision resolu˜ tion intervals. … ) } forms a Markov chain because the length of the k+1st CRI is determined by the number of packets transmitted in its ﬁrst slot. aperiodic and homogeneous. ( k = 0.Section 5. To be ergodic. Stability Analysis One of the most important properties of the binarytree CRP is its stable behavior to which we alluded in previous sections. When the obvious access scheme is employed.).886 can be checked directly. … ) } is ergodic. E [ ( B(k + 1) – B(k) ) ( B(k) = i ) ] < ∞ ˜ ˜ ˜ 2. This number equals the number of packets arriving during the kth CRI and depends only on the length of the kth ˜ CRI. 1. we use again Pakes’ Lemma (see Section 3.24) and therefore one can show that the system is unstable for arrival rate that is larger than 1/ 2. it is possible to establish lower bounds on these moments. ( k = 0.881n – 1 n≥6 (5. the chain ˜ { B(k).23) + 2 ∑ [ Qi( j) + Q j – i( j) ] ( V i – αm i 2 – 1 ) ≤ 0 j>m i=0 The correctness of (5. 1. it is sufﬁcient that the chain fulﬁll the following two conditions: ˜ ˜ ˜ ∀i .2.: THE BINARYTREE PROTOCOL 119 m–1 2 2 ∑ Qi( j) [ Bi + B j – i + Bi B j – i – αm i ( j – i ) + 1 ] m–1 i=0 (5. By analogous arguments to those we used to establish upper bounds of the ﬁrst two ˜ moments of B n .4.
Substituting this bound in (5.26) where α = 2. Consequently. . be less than 1/αm.25) we used the fact that the distribution of the length of a CRI. number of arrivals in kth CRI = n ] n! ˜ ˜ ˜ ∑ E [ B(k + 1) B(k) = i. α It follows then.27) n=0 ∑ ( αn + 1 ) n! ∞ ( λi ) n e –λi = αλi + 1 From (5. and that the arrival process is Poisson.346 (packets per slot). that a sufﬁcient condition for stability of the system is that λ.120 CHAPTER 5: COLLISION RESOLUTION ˜ ˜ E [ B(k + 1) B(k) = i ] = ˜ ˜ ∑ E [ B(k + 1) B(k) = i. the arrival rate of new packets. does not depend on the length of the previous CRI.25) n! ( λi ) n e –λi ∞ ( λi ) n e –λi n=0 ∞ = n=0 ∞ = ∑ ∑ n=0 ∞ ( λi ) n e –λi ˜ ˜ E [ B(k + 1) n packets transmitted at start of B(k + 1) ] n! ( λi ) n e –λi B n n! = n=0 where Bn is the expected length (in slots) of a CRI given that it started with a collision among n packets and λ is the expected number of packets that arrive to the system in a slot. In (5. In the previous section we have shown that Bn is ﬁnite for n ≥0 and. n packets transmitted at start of B(k + 1) ] (5. ˜ B(k) = i ] – i ≤ ( αλ – 1 )i + 1 (5. given that it starts with the transmission of n packets.25) we have ˜ ˜ E [ B(k + 1) B(k) = i ] ≤ Therefore ˜ ˜ ˜ ˜ E [ B(k + 1) – B(k) B(k) = i ] = E [ B(k + 1) ˜ = E [ B(k + 1) ˜ ˜ ˜ B(k) = i ] – E [ B(k) B(k) = i ] . moreover.. the system is stable for arrival rates that are smaller than 0. is bounded by B n ≤ αn + 1 n≥0 (5.27) we see that conditions (a) and (b) of Pakes’ Lemma hold if 1 λ < .886.
( k = 0. Bounds on Expected Packet Delay ˜ Let D be the delay of a randomly chosen packet (a tagged packet). … ) } is also ergodic when λ < α1. In addition we have that ˜ ˜ E [ ( A(k + 1) ) 2 B(k) = l ] = λl + ( λl ) 2 which implies A 2 = λB + λ 2 B 2 = A + λ 2 B 2 (5. namely. the difference between its arrival time to the system and the time it is transmitted successfully.1. Then ˜ ˜ D ≤ E [ B(a) ] + E [ B(d ) ] (5. 2.28) ˜ Let B ( a ) denote the length of the CRI in progress when the tagged packet arrives at the ˜ system and let B ( d ) be the length of the subsequent CRI during which the tagged packet departs from the system. ( k = 0.29) (5.: THE BINARYTREE PROTOCOL 121 5. namely. steadystate distribution of B(k) and A(k) exist. 1.Section 5. respectively. 2. B = E [ B ] and B 2 = E [ B 2 ] . Let B 2 be the ﬁrst and second denote the length of a CRI in steadystate and let B and B ˜ ˜ ˜ ˜ moments of B . From (5. 1. Since the arrival process of new packets is Poisson we have ˜ ˜ E [ A(k + 1) A(k) = i ] = λB i and therefore. .30) since at the earliest the tagged packet arrives at the beginning of the interval whose length ˜ is B ( a ) and at the latest it will be transmitted successfully at the end of the next interval ˜ whose length is B ( d ) .1. 2. 1. ˜ ˜ ˜ E [ A(k + 1) – A(k) ( A(k) = i ) ] = λB i – i ≤ ( λα – 1 )i – λ ˜ Testing conditions (a) and (b) of Pakes’ Lemma for the chain { A(k).346. The pur˜ pose of this section is to show that the expected delay D = E [ D ] of a randomly chosen 1 where α1 = 0. … ) } is ergodic when λ<α1. We now ˜ show that the Markov chain { A(k).3. ˜ Let A(k) be the number of packets transmitted at the beginning of the kth CRI. ( k = 0. ˜ ˜ ˜ Being ergodic Markov chains. namely. ˜ Let A ( d ) be the number of packets transmitted in steadystate at the beginning of the CRI in which the packet leaves the system. Similarly. packet is ﬁnite when λ < α ˜ We already proved that the Markov chain { B(k). let A denote the number of packets transmitted at the beginning of a CRI in steadystate and let ˜ ˜ ˜ A and A 2 be the ﬁrst and second moments of A . A = E [ A ] and A 2 = E [ A 2 ] . By deﬁnition. … ) } we conclude that it is ergodic when λ < α1.28) we see that A=λ B.
18).30).29) implies that B2 ≤ α2( A + λ2 B2 ) + 1 and since λ < α1 we have α2 A + 1 B 2 ≤ 1 – λ2α2 .34) = ˜ ˜ ∑ V n Prob [ A = n ] ≤ ∑ ( α 2 n 2 + 1 )Prob [ A = n ] n=0 n=0 = α2 A2 + 1 The above along with (5.33) ˜ ˜ ˜ ∑ E [ ( B ) 2 ( A = n ) ]Prob [ A = n ] n=0 ∞ ∞ (5. We have that B2 ˜ = E [ ( B )2 ] = ∞ (5. the residual life theorem states that (see Appendix) B2 ˜ E [ B ( a ) ] = B and since B≥1 we have D ≤ ( 1 + λα )B 2 + 1 and therefore we only have to bound B 2 in order to bound D.886. (5.32) (5.31) and (5.28) we have ˜ ˜ E [ A ( d ) ] = λE [ B ( a ) ] Combining (5. ˜ ˜ E [ B ( d ) A ( d ) = n ] ≤ αn + 1 ˜ where α= 2.32) we have ˜ D ≤ ( 1 + λα )E [ B ( a ) ] + 1 Lastly. Unconditioning with respect to A ( d ) .31) (5. we obtain ˜ ˜ E [ B ( d ) ] ≤ αE [ A ( d ) ] + 1 Similarly to (5.122 CHAPTER 5: COLLISION RESOLUTION ˜ ˜ E [ B ( d ) A ( d ) = n ] = Bn and therefore with the help of (5.
2.+ 1 1 – λα 1 – λα 1 – λ2α2 where we used the fact that A=λ B. The ﬁrst is to speed up the collision resolution process by avoiding certain.: ENHANCED PROTOCOLS 123 Substituting the above result in (5.36) in (5.36) (5. Similarly to (5.34) and using (5. The modiﬁed binarytree protocol is due to Massey [Mas81] and eliminates such avoidable collisions by . collisions.1. ENHANCED PROTOCOLS The performance of the binarytree protocol can be improved in two ways.2. substituting (5.1.+ 1 = . the performance of the protocol is expected to improve. (5.+ 1 ( 1 – λα ) 2 Therefore. the expected delay of a packet is ﬁnite and an explicit upper bound for this quantity is given above.35) we obtain α 2 λ + 1 – λα D ≤ . Therefore.18) we have ˜ B = E[B] = ≤ or 1 B ≤ 1 – λα Thus.1) that collisions among a small number of packets are resolved more efﬁciently than collisions among a large number of packets (compare the ratio n/Bn for small n and for large n). This implies that in slot 2 all users (and there were at least two of them) ﬂipped 1. The second is based on the observation (see Table 4. The binarytree protocol dictates that these users must transmit in slot 4. 5.33) we obtain α2 A + 1 α2 A + 1 α 2 λB + 1 D ≤ ( 1 + λα ) . These ideas are the basis of the protocols presented in this section. we showed that when the system is stable. In slots 2 and 3 a collision is followed by an idle slot.2. although it is obvious that this will generate a collision that can be avoided. if most CRIs start with a small number of packets.Section 5. avoidable. The Modiﬁed BinaryTree Protocol Consider again the example depicted in Figure 5.35) ∑ n=0 ∞ n=0 ∞ ˜ ˜ ˜ E [ B A = n ]Prob [ A = n ] = ˜ ∑ Bn Prob [ A = n ] n=0 ∞ ˜ ∑ ( αn + 1 )Prob [ A = n ] = αA + 1 = αλB + 1 5.+ 1 = .
. The equation analogous to (5. ﬂip coins before transmitting is slot 4.124 CHAPTER 5: COLLISION RESOLUTION letting the users that ﬂipped 1 in slot 2 in the example above.3.. the slot in which a avoidable collision would occur is skipped and the evolution of the protocol for the same example of Section 5.e. Consequently. Except for eliminating these avoidable collisions the modiﬁed binarytree protocol evolves exactly as the binarytree protocol. the conditional length of the CRI is given by ˜ ˜ 1 + Bi + Bn – i 1 ≤ i ≤ n = ˜ 1 + Bn i=0 ˜ Bn i which accommodates for the saving of one slot when no users ﬂip 0 (i=0).6) is. i. Note that the correct operation of the modiﬁed binarytree protocol requires ternary feedback. The procedure of determining the expected length of a CRI given that it starts with a collision among n users is entirely analogous to the derivation in the previous section. n–1 1 – Q 0(n) + ∑ [ Qi(n) + Qn – i(n) ]Bi i=0 B n = 1 – Q 0(n) – Q n(n) n≥2 with the initial values B0=B1=1. the users have to be able to distinguish between idle and successful slots. 0 1 0 0 1 1 ≥2 ≥2 0 ≥2 1 1 2 3 4 5 1 1 0 1 ≥2 6 7 8 9 0 1 ≥2 1 0 10 11 12 Collision resolution Interval FIGURE 5.1. We have ˜ ˜ B0 = B1 = 1 and given that a CRI starts with a collision of n (n≥2) packets and that i users ﬂip 0. is depicted in Figure 5.3: Example of the Modiﬁed BinaryTree Protocol Operation The analysis of the modiﬁed binarytree protocol is essentially the same as that of the binarytree protocol.
When obvious access is employed.17) and (5. it is very likely that a CRI will start with a collision among a large number of packets when the previous CRI was long.2. When the system operates near its maximal throughput most CRIs are long.623 ≅ 0. each having an expected number of packets on the order of one and handling each group separately. then the system is stable for arrival rates up to 1/ 2.(i+1)∆ ).24) imply that n/ Bn ≅ 0.37) k=2 In this case. The Epoch Mechanism From Table 1 we see that1/B1=1.4. The operation of the epoch mechanism is illustrated in Figure 5. There are several ways to achieve this goal by determining the ﬁrsttime transmission rule. then the system is stable for arrival rates up to 1/2. if it were possible to start each CRI with the transmission of exactly one packet. when packets are transmitted for the ﬁrst time. the throughput of the system would have been 1.Section 5. On the channel we observe a sequence of collision resolution intervals each corresponding to arrivals during some time interval on the arrival axis. Since this is not possible.4175.664. there is no single value of p that minimizes all the Bn. and is described next. 2/B2=0. suggested by Capetanakis [Cap79].623. Consider the arrivals of packets to the system and divide the time axis into consecutive epochs (called the arrival epochs). Ideally. Packets that arrive during the ith arrival epoch are transmitted for the ﬁrst time in the ﬁrst slot after the collision among packets that arrived during the (i1)st arrival epoch is resolved.375 while if biased coins are used. The parameter ∆ is chosen to optimize the performance of the system. 3/B3=0.2. i. known as the epoch mechanism has been suggested by Gallager [Gal78] and by Tsybakov and Mikhailov [TsM80]. collisions among a large number of packets have to be resolved frequently. The ith arrival epoch is the time interval ( i∆.4.346. This implies that when the modiﬁed binarytree protocol is employed with fair coins.3913 and when n is large (5. If we number these collision resolution intervals sequentially then in the ith CRI all packets (if any) that . however. the Bn are not minimized for p=1/2.: ENHANCED PROTOCOLS 125 The equation analogous to (5.. If we choose p=1/2. yielding non efﬁcient operation. we can establish an upper bound on Bn of the form (5. we can establish an upper bound on Bn with αm=2. each of length ∆ slots (∆ is not necessarily an integer). Bn = 1 + ∑ k  [ 1 – pk – ( 1 – p )k ] n n ( –1 ) k [ k ( 1 + p ) – 1 – pk ] n≥2 (5.381 which is higher than e –1 the maximal throughput for the slotted Aloha protocol. Another way.664 ≅ 0. One way.e. while if we use p=0. 5. The conclusion is that the binarytree protocol resolves collisions among a small number of packets more efﬁciently than among a large number of packets.13) with αm=2.12) is.2. one should try to design the system so that in most cases a CRI starts with the transmission of about one packet. is to have an estimate on the number of packets that arrived in the previous CRI and divide them into smaller groups. hence. Moreover.
It turns out that throughput in both cases is the same although the average delay in the latter method is slightly higher (see Huang and Berger [HuB85]. The example of Section 5. i. Waiting Period } CRI0 CRI1 CRI2 CRI3 CRI4 Channel Axis 4∆ Arrival 5∆ Axis 0 ∆ 2∆ 3∆ FIGURE 5. ∆) are transmitted in the ﬁrst slot of CRI0. it is possible to describe the binarytree protocol via interval splitting.126 CHAPTER 5: COLLISION RESOLUTION arrive during the ith epoch are successfully transmitted. in the period (0.b)none in our example.d) behave as if they ﬂipped 0 while all those that arrived during (d. requiring further splitting of the interval (a. There are two options at this point (corresponding to two different protocols): one could shorten the third arrival epoch to match the end of CRI2 or. those arriving in the period (∆. The nodes that arrived in the left part correspond to users that would have ﬂipped 0 in the binarytree protocol and the nodes that arrived in the right part correspond to users that would have ﬂipped 1. and so on. Since a collision occurred one needs to split the group but instead of ﬂipping a coin we split the interval in two halves: all those users that arrived during (a. This causes packets that arrived during (b. thus all packets that arrived during (a. Note that the probability p that a user ﬂips 0 corresponds here to the interval splitting .4: Example of the Epoch Mechanism When the epoch mechanism is employed as the ﬁrsttime transmission rule.e. All the packets arriving in the 0th epoch.d) to be transmitted in slot 4. are transmitted. they collide.5 to demonstrate the interval splitting procedure. since the CRI2 ended before the third epoch.. as is shown in the ﬁgure (and analyzed in this section).1.g) in the ﬁgure. An interesting phenomenon occurs at the end of CRI2 in our example. newly arrived packets wait and when CRI0 ends all packets belonging to the ﬁrst epoch.g) are transmitted in slot 1. the interval is split in two. is depicted again in Figure 5.e. Slot number 1 is the ﬁrst slot of a CRI whose corresponding arrival epoch is the period (a.4). Whenever the transmission of packets that arrive during some interval results in a collision.. This results in another collision in slot 2. and a resolution process starts (see Figure 5. i. and so on. 2∆).g) behave as if they ﬂipped 1.d) so that in slot 3 transmit those users that arrived during (a. In the meantime. enter a waiting period lasting from the end of CRI2 to the end of the third epoch. it is guaranteed that packets are transmitted in the order they arrive (FCFS). In this way.
Section 5.b) (b. and when p=0. Thus. ˜ ˜ ˜ then A(0) . This means that the values Bn and Vn are identical to those of the binarytree CRP.: ENHANCED PROTOCOLS 127 ratio..k) 1 2 3 4 5 6 7 8 9 10 11 12 CRI Channel Axis Collision resolution Interval (CRI) a bcdef g h i j k ∆ ∆ FIGURE 5.3 the left part of the split interval is 0.d) (d.f) (f. Since the arrival process is Poisson. there is no sta˜ tistical dependencies among the corresponding collision resolution intervals.3 times the length of the original interval. Since the length of the kth CRI is completely determined by the number ˜ ˜ ˜ of packets transmitted in its ﬁrst slot. when p=1/2 the interval is halved. A(2) .d.) random variables.i. we conclude that the sequence B(0) . Thus.d) (a. ˜ is also a sequence of independent and identically distributed random variables. 0 ≥2 ≥2 1 1 ≥2 1 1 ≥2 ≥2 0 ≥2 (a..f) (d.e) (e. If A(k) denotes the number of new packets that are transmitted at the beginning of the kth CRI. .g) (g.2..5: IntervalSplitting Procedure for the Epoch Mechanism We now turn to evaluate the performance of this protocol. is a sequence of independent and identically distributed (i. The main difference lies in the region of stable throughput.g) (a. Let A and ˜ ˜ ˜ B denote an arbitrary pair A(k) and B(k) . . Since arrival process is Poisson the arrival points in every given interval are uniformly distributed. When the epoch mechanism is employed as the ﬁrsttime transmission rule.g) (d. splitting an interval is completely equivalent to ﬂipping a coin.c) (c.d) (b. B(2) . we have: ( λ∆ ) n e –λ∆ ˜ Prob [ A = n ] = n! Also. B(1) . A(1) .. .
The fact that . namely the expected change in the difference between the current time and the time of the last epoch whose packets were transmitted successfully. For p=1/2 the function f(z) is maximized for z*=1.38) can be written as: n=0 ∑ Bn . the system is stable.6 for various values of p (recall = that p is the probability that a user will ﬂip 0 and it is equivalent to the splitting ratio of the interval when a collision is observed). The total service time of these arrivals has a ﬁrst and a second moment B and B 2 .38) holds. respectively. to successfully transmit all packets that arrive in a period of duration ∆ must be less than ∆. ˜ B = E[B] = ˜ ˜ ∑ Prob [ Bn = i ]Prob [ A = n ] n=0 ∑ n=0 ∞ ˜ B n Prob [ A = n ] B2 ˜ = E [ B2 ] = ˜ ∑ V n Prob [ A = n ] n=0 ∞ where Bn and Vn are given by (5. The quantity B∆ is the expected change in the time backlog of the system. the time it takes.∑ Bn n! n! n=0 n=0 (5. When condition (5. The system can be viewed as a (discretetime) queueing system in which packets that arrive during the interval (i ∆.128 CHAPTER 5: COLLISION RESOLUTION ˜ Prob [ B = i ] = ˜ ∑ Prob [ Bn = i n=0 ∞ ∞ ˜ ˜ A = n ]Prob [ A = n ] i≥1 = from which we obtain.6) and (5. the expected delay of a packet is ﬁnite. respectively.11).15 and the maximum value is 0.429. Condition (5.∆ f (z) = ∞ ∞ ( λ∆ ) n e –λ∆ z n e –z ∑ Bn . the system is stable for arrival rates λ<0.429. Hence. The function f(z) is depicted in Figure 5.39) where z ∆ λ∆ .38) in other words. (i+1)∆) are served in the ith CRI.< ∆ n! ∞ ( λ∆ ) n e –λ∆ and rewritten as: z λ∆ λ < . on the average. The maximal value of f(z) is smaller for other values of p. In order for the “server” not to fall behind the arrivals we need that B<∆ (5.= . and if B 2 is ﬁnite.
25 p=0. that idea was not fully exploited by the above enhancements.7.35 0. The conclusion is that slightly longer epochs (slightly larger ∆) will not cause the maximal achievable throughput to deteriorate by a large amount.4 0.05 0 0 1 2 3 4 5 z=λ∆ FIGURE 5.39) one should use the values of Bn that correspond to this protocol (equation (5.15 is not surprising.2 0.1 0. 5.15 0.3 0. The results are as follows: When p=1/2 the system is stable for λ<0. 0. When the epoch mechanism is used with the modiﬁed binarytree protocol resolution method the analysis is the same.Section 5.45 0.37). It conforms with the intuition that most of the CRIs should start with the transmission of a single packet. The fact that z* is slightly higher than 1 is due to the waste incurred by idle slots. The length ∆* that should be chosen to obtain the maximal throughput is ∆*=1.429=2. assume that the epoch mechanism is employed and consider what happens when a collision is followed by another collision (see slots 1 and 2 .4 P=0.5 p=0.462 and when p=0. The Clipped BinaryTree Protocol In the previous analysis we made the point that to improve performance a CRI should start with the transmission of about one packet. especially for values of z larger than z*.2.: ENHANCED PROTOCOLS 129 z*=λ∆*=1.3.468.3 f( z ) 0. 0.6 we observe that the function f(z) is not very sensitive to small changes in z.6: Determining Permissible Arrival Rates for the Epoch Mechanism The description and the analysis above corresponds to use of the epoch mechanism and resolving collisions as done in the binarytree protocol. 0.68. However.6.4175 the system is stable for input rates up to λ<0. To see why.2.15/0. From Figure 5. except that in (5.
To remedy this. since after receiving the feedback indicating that a collision took place..e.1. The packets that arrived in the left part are then transmitted in slot 2 and collide again. it would be better to choose a new interval of length ∆ and let packets that arrived in this interval transmit. At this point the packets that arrived during the right part of the original interval are transmitted. that is. In the modiﬁed clipped binarytree protocol the collision in slot 3 will be avoided (skipped) as was the case in the example depicted in Figure 5.e. let x ˜ l and x r be the number of ˜ be the number of packets involved in the ﬁrst collision and let x packets in the left part and the right part of the interval. yet the expected number of arrivals during that period is different from the desired quantity. The users whose packets collided are divided into two groups by splitting the interval into two parts (for explanation purposes we assume p=1/2 so the length of either of the two parts equals ∆/2). rather than enabling the interval (d.1. if the distribution of the number of packets in that part is as that of new packets i.5.e. This assertion is not as trivial as it ﬁrst appears. Protocols based on this observation have been suggested by Gallager [Gal80] and Tsybakov and Mikhailov [TsM80]. slightly larger than one. Indeed. since part of the binarytree is clipped and not enabled (see Figure 5. To incorporate this strategy into the protocol we adopt the rule that whenever a collision is followed by two successive successful transmissions. a new epoch of length ∆ from the arrival axis is enabled. The argument that leads to the enhancement described above is based on the assertion that the distribution of the number of packets in the right part of the interval is the same as that of the newly arrived ones i. The protocol that results from such an operation is called the clipped binarytree protocol.7 when the clipped binarytree protocol is employed. Note that in slot 7 a new CRI is started. One might add that the clipping idea can be used with the original binarytree protocol (Section 5. By repeated interval splitting all packets in the left part are eventually successfully transmitted (slot 6 in our example).7). We compute .f) as would be the case in the regular epoch mechanism.) as well as the modiﬁed one (Section 5. then the transmission of packets in that part is identical to starting a CRI with half the optimal expected number of packets. it is not clear that the distribution of the ˜ right part remains Poisson with the same parameter as before. The collision in the ﬁrst slot is among the packets that arrived during an interval of length ∆. respectively. We need to show that ˜ ˜ ˜ ˜ ˜ if it is known that x = x l + x r ≥ 2 and x l ≥ 2 then the distribution of x r is Poisson. It is at this point that we have not done as best we can.i). The underlying observation is that there is no information regarding the number of packets in the part we attempt to resolve.130 CHAPTER 5: COLLISION RESOLUTION in Figure 5.3.). Poisson with parameter λ. the packets that arrived in that interval are transmitted. corresponding to the arrival epoch (d.2. The example of Section 5... namely. those that arrived in the part of the arrival axis that was never explored. i. the one that avoids transmission of packets that are guaranteed to collide (the latter is called the modiﬁed clipped binary tree protocol).1. is depicted in Figure 5. To show this property.
˜ as before. x l ≥ 2 ] = Prob [ x r = i x r ≥ 0.1. x l ≥ 2 ] ( λ∆ ⁄ 2 ) i e –λ∆ ⁄ 2 ˜ ˜ = Prob [ x r = i x l ≥ 2 ] = i! ˜ ˜ where we used the fact that x r and x l are independent. i of which arrived in its left part.2. Generally. a result stemming from the known property of the Poisson process that the number of events in nonoverlapping intervals are independent. but.c) (c. For n=0 and n=1 we have ˜ ˜ B0 = B1 = 1 .7: Example of the ClippedBinaryTree Protocol Operation ˜ ˜ ˜ ˜ ˜ ˜ ˜ Prob [ x r = i x l + x r ≥ 2. Denote.: ENHANCED PROTOCOLS 131 1 0 ≥2 ≥2 1 1 ≥2 ≥2 1 ≥2 ≥2 0 (a. and by Bn and B n i their respective expected values..f) (f. We present the analysis for the modiﬁed clipped binarytree (see Section 5. those that can be predicted beforehand. The analysis of the clipped binarytree protocol is similar to that of the binarytree protocol.2.b) (b. i.j) 1 2 3 4 5 6 7 8 9 10 11 12 CRI Arrival Axis Channel Axis CRI a bcdef ∆ g h i j CRI k ∆ ∆ FIGURE 5.g) (a.g) (d.) in which deﬁnite collisions.Section 5. results in a nonPoisson distribution of the number of colliding packets.d) (b.d) (a.e) (e.f) (d. are avoided. by B n the length of the collision resolution interval that starts with a collision of ˜ n packets and by B n i the length of the collision resolution interval that starts with a colli˜ sion of n packets. receiving the additional feedback of a collision in the left part of the interval means that the number of packets that arrived in the right part is greater or equal to zeroinformation we had to start withmeaning that we have the same Poisson distribution.i) (d.e.d) (d. receiving the collision feedback.
41) The quantity Bn can be computed recursively from (5. as before. .132 CHAPTER 5: COLLISION RESOLUTION Consider a CRI that starts with a collision of n (n≥2) packets. The conditional length of a CRI for the clipped binarytree protocol can therefore be summarized by: ˜ 1 + Bi 2≤i≤n ˜ = 2 + Bn – 1 i = 1 ˜ 1 + Bn i=0 ˜ Bn i and correspondingly the expected values are 1 + Bi 2≤i≤n = 2 + Bn – 1 i = 1 i=0 1 + Bn Bn i (5. If i=1 then the CRI includes the original collision slot.40). Finally. and then the number of slots needed to resolve the remaining n1 packets. when 2≤i≤n the CRI includes the original collision slot and then the number of slots needed to resolve the i packets that arrived in the left part. Note that in the latter case (that is unique to the clipped tree protocol) the resolution of the remaining ni packets is not a part of the current CRI. by Qi(n) the probability that in an interval containing n arrivals i occurred in its left part. i of which arrived in the left part and ni in the right part (0≤i≤n). the successful slot for the left part. based on equation (5.40) Denoting. an expression for Bn: ˜ Bn = E [ Bn i ] = ∑ Bn i Qi(n) i=0 n = 1 + Q 0(n)B n + Q 1(n) ( 1 + B n – 1 ) + or n–1 ∑ Qi(n)Bi i=2 n n≥2 1 + Q 1(n) ( 1 + B n – 1 ) + ∑ Qi(n)Bi i=2 B n = 1 – Q 0(n) – Q n(n) n≥2 (5.41) with the initial values B0=B1=1. we can write. If i=0 then the length of the CRI will include the original collision slot followed by the need to still resolve all n packets.
7 during the CRI that starts with a collision among four packets. For n=0 and n=1 we have U0 = 0 U1 = 1 and for n≥2. one must compute the rate of ˜ successful transmissions during a CRI. similarly to equation (5.43) and Un can be computed recursively from (5. It is interesting to note the slow growth of Bn with n compared to the linear growth of Bn with n when the binarytree protocol is employed (see Table 1). To evaluate the performance of the protocol.42) Leading to U n = Q 0(n)U n + Q 1(n) ( 1 + U n – 1 ) + or n–1 ∑ Qi(n)U i i=2 n n≥2 Q 1(n) ( 1 + U n – 1 ) + ∑ Qi(n)U i i=2 U n = 1 – Q 0(n) – Q n(n) n≥2 (5. only two packets are transmitted successfully (the rest will be part of the next CRI). we have Ui 2≤i≤n = 1 + Un – 1 i = 1 i=0 Un Un i (5.: ENHANCED PROTOCOLS 133 With the clipped binarytree protocol not all packets that collide in the ﬁrst slot of a CRI are successfully transmitted during that CRI (since part of the tree is clipped). and the expected number of packets that are successfully transmitted in a CRI.43) with the initial values U0=0 and U1=1. Values of Bn and Un are given in Table 3 for p=1/2. For Instance.40). and by Un and U n i their respective expected values. observe that the expected number of packets transmitted successfully during a CRI is almost a constant for n≥3. In addition. To that end let U n be a random variable representing the number of packets that are successfully transmitted during a CRI given that it ˜ started with the transmission of n packets and U n i be the same variable conditioned on ˜ having i packets in the left part.2.Section 5. are given by . U. in the example depicted in Figure 5. B. The expected length of a CRI.
44) leads to ( λ∆ ) n e –λ∆ z n e –z U n U n ∑ ∑ n! n! .0980 2. on the average.44) n=0 ∑ U n n! ( λ∆ ) n e –λ∆ The expected number of packets transmitted in the ﬁrst slot of a CRI is λ∆ and therefore the fraction of packets successfully transmitted during a CRI is U/λ∆.8363 2.0000 6 6.0286 2. it must be able to resolve collisions at least at the rate in which time progresses or.5052 14 8.6741 2.5073 ∞ 3 5.7937 2. Thus.45) where we have substituted z ∆ λ∆ . Hence.∑ Bn n! n! n=0 n=0 ∞ ∞ (5.4977 11 7.9027 2. The right handside is a function of z (with parameter = p).5064 4 6.5238 10 7.= n = 0 λ<n=0 ∞ ∞ ( λ∆ ) n e –λ∆ z n e –z ∑ Bn .5714 9 7.5406 2. U B < λ which.5079 2 4.000 7 7. For a given p this function can be plotted (the curves obtained are similar to those in .0000 2.5055 5 6. the portion of the resolved interval.5008 13 7. if a fraction U/λ∆ of the packets are successfully transmitted it means that U/λ∆ is also the fraction of the interval resolved.134 CHAPTER 5: COLLISION RESOLUTION ˜ ˜ Table 3: The ﬁrst moment of B n and U n for p=1/2.5049 B = E [ Bn ] = U = E[Un] = n=0 ∞ ∑ Bn n! ( λ∆ ) n e –λ∆ (5.3894 2.5000 8 7. the arrival points are uniformly distributed in the interval.0000 1. The Poisson process has an interesting property that given a number of arrivals in an interval. On the other hand.2180 2. upon substitution of equation (5.0035 2. for the system to remain stable.) n Bn Un n Bn Un n Bn Un 1 1.5075 15 8.8333 2.6698 2. (U/ λ∆)∆=U/ λ is. on the average.4762 2. Consequently.4958 12 7. it takes B slots to resolve a collision.
is the freeaccess protocol analyzed by Mathys and Flajolet [MaF83] and by Fayolle et.2. i. new packets are transmitted as soon as possible. This is necessary because a newly generated packet can be transmitted for the ﬁrst time only at the end of the current CRI.).60 slots. at the beginning of the slot subsequent to their arrival time.e. the system is stable for arrival rates λ<0. a crucial feature required for the correct operation of the full sensing protocols is the ability of all users to determine the end of a CRI. and the resulting maximal throughput is 0. even if that user has no packet to transmit.487 and the length ∆ that should be chosen is ∆=1.487 packets/sec. The maximal throughput of this limited sensing protocol when the modiﬁed binarytree protocol is employed is 0. For p=1/2 the right side of (5. [FFH85]. i.. The simplest. This kind of feedback monitoring is known as fullsensing. avoidable collisions are eliminated.3. a user that due to some fault did not receive properly all signals may actually disturb the others and decrease the efﬁciency of the protocol. The analysis presented above was carried out for the modiﬁed clipped binarytree protocol.4877 as was demonstrated by Mosley and Humblet [MoH85]. This is necessary so that users with new packets know exactly when they may transmit for the ﬁrst time in a manner that would not interfere with an ongoing resolution of a collision. When the splitting probability p is optimized it is possible to slightly increase the throughput to 0. As mentioned earlier. Thus.26 and the maximum value is 0. albeit not the most efﬁcient one. In this protocol.360. LIMITED SENSING PROTOCOLS The protocols described so far require every user to monitor the channel feedback at all times.449. furthermore. preferably after having generated a packet for transmission and until the packet is transmitted successfully.) or the modiﬁed binarytree protocol (see Section 5. The major change involved in the adaptation of the clipped binarytree protocol . When the modiﬁcation is not used.6) and its maximal value found.. it is desirable that users monitor the feedback signals only during limited periods. and protocols with such monitoring are referred to as protocols operating in a limited sensing environment. the most efﬁcient protocol for a limitedsensing environment is the one introduced by Humblet [Humb86] and Georgiadis and PapantoniKazakos [GeK87] which is essentially an adaptation of the (fullsensing) modiﬁed clipped binarytree protocol described in Section 5. This protocol works in conjunction with both the basic binarytree protocol (see Section 5. a user that transmitted a new packet monitors the feedback signals and continues operating as if he were an “old” user.3. thereafter.26/0. the only change in the analysis is the addition of Q0 (n) to the numerator of equation (5.2. requiring every user to positively determine the end of the CRI.Section 5.41).45) is maximized for z=1.e.1.: LIMITED SENSING PROTOCOLS 135 Figure 5. Several collision resolution protocols have been devised for the limitedsensing environment.487=2.3.1. To date. 5. Thus.. al. This kind of monitoring is known as limitedsensing. This mode of operation is quite impractical because a user that crashed can never again join the system and.
To complete the speciﬁcation of the protocol. but the delay incurred by a new packet upon arrival to an idle system will be small. and then to detect the end of a CRI. we recall that the end of a CRI that started with a collision is characterized by two consecutive successful slots. A large value for R may cause a packet arriving to an idle system to wait quite long before it can determine that the system is idle. that is. retransmit in the next slot to cause a collision. Thus. A user generating a new packet at this state will wait forever for the endofCRI marker. newly generated packets are not considered for transmission until the ongoing collision resolution is done (and all waiting users receive sufﬁcient feedback to detect the end of a CRI). while when R is very large. the difﬁculty is to realize whether or not a collision resolution is in progress when a series of idle and successful slots is observed. if we have a successful slot followed by either an idle slot or another successful slot. there are CRIs that consist of a single slot. only to announce an ongoing CRI. There is a clear performance tradeoff in the choice of R.. Thus. where R is a globally known protocol parameter. reveals a sequence of idle slots. Observing the channel once the system becomes idle with no user having a new packet to transmit. A collision feedback obviously indicates that a CRI is in progress. we avoid deﬁnite collisions at most R1 times in succession. a user that observes R+1 consecutive idle slots concludes that a CRI ended. Consequently. In addition. One of the rules has already been deﬁned. the protocol is similar to the clipped binarytree protocol.2. To overcome this potential deadlock. and continue regularly thereafter.2. Going back to the clipped binarytree protocol described in Section 5. a successful slot followed immediately by an idle slot or another successful slot denotes the end of the CRI. This indication of the end of a CRI is correct only if R+1 consecutive idle slots never occur during an actual resolution of a collision. he will be able to determine its end. those that are guaranteed to occur. either an idle one or a successful one. not enough. but it will eliminate more deﬁnite collisions. and requires a participating user to retransmit his packet after having sensed R consecutive idle slots.136 CHAPTER 5: COLLISION RESOLUTION to the limited sensing environment is therefore controlling the extent of CRIs so that their end remains uniquely and easily detectable. we are assured that a CRI just ended. the endofCRI marker is augmented to include an event consisting of R+1 idle slots. In summary.1.3.e. As before. The other rules deﬁne the exact transmission schedules and are similar to the rules of the clipped binarytree protocol. a user that generates a new packet at some slot will be able to decide whether or not a CRI is in progress within at most R+1 slots and if he ﬁnds out that a CRI is in progress. Observe that the case R=1 corresponds to elimination of the modiﬁcation introduced in Section 5. Yet. in the limited sensing environment. in the modiﬁed clipped binarytree protocol. We take this event to be the endofCRI marker. A small value for R requires fairly often to retransmit a packet unnecessarily. we must deﬁne the behavior of the users involved in the resolution process. Users that participate in a collision resolution and observe R consecutive idle slots. however. Such marking is. namely. a new epoch is .. upon detection of the end of a CRI. such event is possible since the protocol dictates to avoid definite collisions.. i.
Users that previously belonged to class A have now sensed a collision and therefore belong to class B.8(a) depicts the initial situation and shows the arrival times of the various classes. The uparrows indicate arrival instants of packets. Figure 5. Each snapshot is taken at the time marked CT (Current Time) and depicts the time axis starting at some time t0. and so forth.8(d). Note that the two idle slots. At this time. `1’=success. Class B contains those users that deﬁnitely know a CRI is in progress (they heard a collision feedback) but do not know when it started. Since it is desirable to select an initial epoch whose length is optimal we select it so that its end coincides with the latest class C slot.: LIMITED SENSING PROTOCOLS 137 chosen and packets that arrived during this epoch are transmitted. This is depicted in snapshot (b). and the feedback for the slot starting at CT is shown above the axis (`0’=idle. those in classes A and B cannot. the numbers are packet numbers used for explanation. At this point in time the rightmost half of the previous TI is enabled. In this sense the protocol is a lastcome ﬁrstserved protocol. However. causing a (deﬁnite) collision as is illustrated in Figure 5. ` ≥2’=collision). Figure 5. both the newly arrived and the old. whether some activity is going on.Section 5. depicted in snapshots (b) and (c) do not increase the extent of class B users since two idle slots leave all users arriving during these slots in uncertainty . The beginning of the CRI is not known to the users since they start monitoring the channel only upon a packet arrival meaning that the only common time reference is the end of the CRI (which is known to all class C users). Class A contains those users that cannot decide thus far whether or not some CRI is in progress. taken at seven consecutive slots. as illustrated in Figure 5. All but those users that generated packets in the last slot before CT belong to class C while those that arrived in the last slot belong to class A since they have not sensed a collision and cannot decide. untransmitted. packets that arrive during this transmission interval (marked TI in the ﬁgure) are transmitted. a moment for which all previously arrived packets have already left. having sensed two idle slots in a system with R=2 the entire left part of the previous TI is enabled. ones can be divided into three classes. of a system with R=2. To describe how new epochs are chosen and how epochs are split we note that at each point in time all users having packets awaiting transmission.8 presents seven snapshots.3. and since it contains no packets an idle slot occurs and therefore it is concluded that the left part contains at least two packets and its rightmost half is enabled. The same applies to packet 4 that arrived during the most recent slot. In other words. the protocol in the limited sensing environment evolves in such a way that the epochs for transmission are chosen so that the most recent arrivals belonging to class C attempt transmission ﬁrst. While all users in class C can simultaneously decide which epoch will be chosen for transmission. Class C contains those users that know a CRI is in progress as well as the time it started. a portion of the rightmost part of the class C time is enabled namely. Another idle slot occurs and it is concluded that the corresponding left part contains at least two packets.8(c). and a collision between packets 2 and 3 occurs. Such a diagram is often called the arrival timeaxis diagram. upon a collision the epoch is split in two. based on a single slot.
8: Enabled Intervals in Limited Sensing .138 CHAPTER 5: COLLISION RESOLUTION C TI (a) t0 1 2 3 C TI (b) t0 1 2 C (c) 3 TI A CT ≥2 4 B CT 0 4 B A CT 0 t0 1 C TI 2 3 B 4 A CT (d) ≥2 3 TI 4 B CT 1 t0 1 C 2 (e) t0 1 C 2 TI 3 4 B A CT 1 (f) t0 1 C TI 2 4 C TI A CT (g) t0 1 4 FIGURE 5.
Given that a CRI starts with a collision of n (n≥2) packets. It starts by enabling a transmission interval that contains a portion of the time covered by class C users which consists of a small interval after t0 and the entire period during which the most recent CRI took place. The number of transmitting users in this slot can be any number between 1 and n.2. At this point two consecutive successful slots took place and an endofCRI marker is detected. followed by exactly l consecutive idle slots.3. 0 ≤ l ≤ R – 1 ˜ 1 + R + Bn l=R ˜ Bn i (5.3. Throughput Analysis The analysis of the above protocol is similar to the analysis presented in Section 5. All users but those arriving in the most recent slot join class C while those arriving in the most recent slot join class B (having sensed a single success they know the system is not idle). except that one must accommodate the rule that R+1 slots ˜ will not appear within a CRI.: LIMITED SENSING PROTOCOLS 139 they have not sensed a collision and cannot determine whether or not the system is idle. intervals corresponding to different classes do not overlap. Having sensed a collision (snapshot (d)) all class A users become class B users. and whenever an interval is resolved.46) . 5.1. A CRI that starts with n≥2 packets. where already transmitted successfully. The previously enabled interval is halved and packet 3 is transmitted successfully (snapshot (e)) after which packet 2 is transmitted successfully (snapshot (f)). class B is empty. Note that in general. depending on the speciﬁc arrival times.3. Note that the extent of time covered by class C users is not contiguousit is separated by a period for which all arrivals.Section 5. We ﬁrst derive the evolution of B n the length of a CRI that starts with n users. but these will eventually be resolved if the system is to be stable. if any. for the clipped binarytree protocol. After the collision of snapshot (d). then the length of a CRI for the above protocol is ˜ 1 + l + Bi 2 ≤ i ≤ n. and having i packets in the interval that is enabled after the l idle slots. For the cases n=0 and n=1 there is no change and hence ˜ ˜ B0 = B1 = 1 . operation continues as before. starts with a collision slot followed by 0 or more idle slots (belonging to the rightmost subintervals without arrivals) after which a nonidle slot must occur. Snapshot (g) illustrates the start of the next CRI. As we have seen the subset of the arrival axis that contains packets from class C may consist of disjoint subintervals. 0 ≤ l ≤ R – 1 ˜ = 2 + l + B n – 1 i = 1.
and when l < R. We considered but a very small number of these variations: slotted time.3. When R=1 the maximal throughput is 0. RELATED ANALYSIS Numerous variations of the environment under which collision resolution protocols operate have been addressed in the literature and excellent surveys on the subject have been written by Gallager [Gal85] and Tsybakov [Tsy85].42) and (5. Recalling that Qin is the probability of i arrivals occurring during the right portion of the an interval containing n arrivals (see equation (5.46) in the same manner as before B n = 1 + [ Q 0(n) ] R ( R + B n ) R–1 = ∑ [ Q0 (n) ] l Q 1(n) ( 1 + l + B n – i ) + ∑ Qi(n) ( l + Bi ) i=2 n n≥2 l=0 or after some algebra we have for n≥2 that n–1 1 – Q 0(n) + [ 1 – [ Q 0 (n) ] R ] Q 0(n) + Q 1(n) ( 1 + B n – i ) + ∑ Qi(n)Bi B n = [ 1 – [ Q 0(n) ] R ] [ 1 – Q 0(n) – Q n(n) ] and Bn is computed recursively from the above equation with the initial values B0 = B1=1. the maximal throughput is 0. The maximal throughput of the protocol is given by equation (5.2.4. i users belonged to the right part of the interval.44). i=2 5. The derivation of the expected number of packets transmitted successfully during a CRI is identical to that presented for the clipped binarytree (see Section 5. Chapter 4 in Bertsekas and Gallager’s book [BeG87] is also an excellent source on collision resolution protocols. the same as the throughput for the fullsensing environment.140 CHAPTER 5: COLLISION RESOLUTION ˜ where B n i is the length of a CRI that started with the transmission of n packets given that there were l consecutive idle slots at the beginning of a CRI.449the same throughput obtained for the clipped binarytree protocol (without the modiﬁcation).2)) we obtain from (5. equations (5. as is indicated by the third component of equation (5. These equations are similar to those of the modiﬁed clipped binarytree protocol with the exception that after R idle slots the original n packets are retransmitted and collide.487.46). When R is large (for that matter R=5 is already large enough). Poisson arrivals and reliable ternary feedback. In the following we list a few of the other variations.41)). Bounds on throughput Considerable effort has been spent on ﬁnding upper bounds to the maximum throughput that can be achieved in an inﬁnite population model with Poisson arrivals and ternary feed . inﬁnite population.
Mol82. a branch of applied statistics. The most common are the noise errors that are intrinsic in any physical radio channel. erasures and captures have been studied in [VvT83. Reasons for erasures in practical systems are mobile users that may occasionally be hidden (for example. addresses the problem of classifying items of some population as either defective or nondefective. These errors correspond to situations in which one or several nodes are transmitting. There are two other binary feedback types that can be considered: (i) Something/Nothing feedback that informs the users whether or not a slot was idle. TML83. BeZ88]. SiC85. When a packet is transmitted successfully it is possible to use the contents of the packet in the feedback in order to improve the performance of the CRP. Noise errors. The latter information can be obtained by using energy or power detectors and this kind of feedback is termed known multiplicity feedback [Tsy80] and [GeP82]. CrH82. but the feedback detected by the users indicates that the slot was idle. MiT81. or by indicating the exact number of users involved in a collision. successful and collision slots. Such errors cause the feedback to indicate a collision although no user or a single user transmitted. Group testing Group testing. Feedback types In the text we considered mainly the ternary feedback distinguishing among idle. CKS88]. (ii) Success/Failure feedback that informs the users whether or not a slot contained exactly one packet.: RELATED ANALYSIS 141 back [Pip81. The best upper bound known todate is 0. Another type of errors are erasures. because of physical obstacles) or because of fading problems. The capture phenomenon can also be considered as erroneous operation of the system. erasures and captures Practical multipleaccess communication systems are prone to various types of errors. and the binary feedback that informs the users only whether or not there was a conﬂict in the slot. In some cases it might be possible to increase the amount of feedback detail by using an extra control channel for reservation [HuB85. Collision resolution protocols that operate in presence of noise errors. corresponding to the ability to receive a packet successfully although more than one packet is transmitted at the same time. MaM85].4. CiS85. Collision resolution protocols for these binary feedback channels have been studied in [MeB84]. Multiple access protocols without feedback have been considered in [TsL83. It has been discovered that group . ToV87]. A protocol that uses an extra bit that can be read only when a packet is transmitted successfully is presented in [KeS88]. CiS87.Section 5.568 and is due to Tsybakov and Likhanov [TsL88].
Collision resolution protocols based on group testing ideas have been developed in [BMT84] for a homogeneous population of users and in [KuS88] for a nonhomogeneous population of users. Collision resolution protocols yielding high throughputs for general arrival processes (even if their statistics are unknown) were developed in [GFL87] and [CiS88]. Bounds on the expected packet delay of the clipped binarytree with the epoch mechanism have been obtained in [TsM80. It is not difﬁcult to realize. Delay analysis Several delay analyses of collision resolution protocols appear in the literature. some of the parameters of the protocols were carefully tuned.142 CHAPTER 5: COLLISION RESOLUTION testing ideas and algorithms can be applied in the design of protocols. PMV87]. for random access communication. that the performance of some of the algorithms is not sensitive to the speciﬁc arrival process. though. the maximal throughput of the basic binarytree protocol with the obvious access scheme is 0. to yield the best performance (the epoch length ∆. General arrival processes In the analysis of the collision resolution protocols it has been assumed that the arrival process of new packets to the system is a stationary Poisson process. For instance. The expected packet delay of the binarytree protocol has been derived in [TsM78. GMP87]. FFH85]. based on the Poisson assumption. for instance). Other variations of delay analysis appear in [HuB85. MaF85. similar to the collision resolution protocols. The basic idea is that a defective item in the group testing problem corresponds to an active user in the communication problem and a nondefective item corresponds to an idle user. independently of the arrival process.346. Furthermore. .
B(0) = B 0 dB(z) dz = B1 – B0 z=0 . What are the conditions on the error probabilities to insure that these quantities are ﬁnite? 2. ∑ k=2 n ( 2k – 1 ) ( – 1 ) k n  k 1 – p k – ( 1 – p ) k n≥2 k=2 ∑ k  2k – 1 – 1 ∞ k=1 n n ( 2k – 1 ) ( – 1 ) k n≥2 n≥3 Bn = Bn – 1 + 2 ( n – 1 ) ∑ 2 –k ( 1 – 2 –k ) n – 2 Problem 2.Section : EXERCISES 143 EXERCISES Problem 1. All error events are independent of each other and of the system state. 1. so that an idle slot is interpreted as a collision with probability π0. B1 and write recursive relations for computing Bn.C = π1. 3.B(z) – 2B ( z ⁄ 2 ) = 1 – ( 1 + z ) ( 1 + B 0 )e – z . Bn = 3 + 2. n ≥ 2 for this system.The optimal z that maximizes the throughput (z*) does not depend on π1. Assume that the channel is noisy. . For π0. Problem 3. Compute B0.C.C (in the latter event the user that transmitted has to retransmit his packet again according to the protocol).12)) Bn = 1 + For p =1/2 prove: 1. For the binarytree algorithm we found that (see (5. (Noise errors) Assume that the binarytree protocol with the epoch mechanism is employed.C and a slot that contains a single transmission is interpreted as a collision with probability π1. Write an expression for computing the throughput.C = 1/4 determine z* and the maximal throughput.37) and determine the corresponding equation for Vn in this case. How would you explain this phenomenon? 4. Let z n e –z ∞ B(z) ∆ ∑n = 0 B n = n! Prove that . Prove equation (5.
the probability that Y k = l in steadystate. How would you compute Py(l)? Problem 5. Assume that the epoch mechanism is used and devise a collision resolution protocol for this system. 1. ˜ Let A k be the number of new packets transmitted at the beginning of the kth CRI and let ˜ Y k be the number of packets that due to captures are not transmitted successfully during the kth CRI and hence are transmitted at the beginning of the (k+1)st CRI. The latter implies that the users transmitting during a slot that contained a capture. namely. When n ≥ 2 users are transmitting in a slot. n ≥ 0 for the wait and the persist schemes. Write recursive equations for Pn(l) for the wait and the persist schemes. GeP82]) Consider a system in which at the end of each slot the users are informed of the exact number of users that transmitted during the slot. (Power capture [CKS88]) This problem deals with the capture phenomenon employing a model similar to that of Problem 3. Write an expression for the throughput of the protocol as a function of Bn. The question then is when will the users whose packets failed due to a capture retransmit their packets. As explained there a capture means receiving correctly a packet even when other packets are transmitted during the same time. Compute the maximal throughput of your protocol.4. . Assume that the users are executing the binarytree protocol with the epoch mechanism. Let ˜ ˜ ˜ ˜ ˜ X k = A k + Y k – 1 and for n ≥ 0 . Consider the following alternatives: (i) They transmit immediately in the subsequent slot after the capture occurs (persist scheme). the users are informed which packet was received correctly. We therefore assume that in case of success all users (including those that transmitted and failed in that slot) behave as if the slot was successful. Let Bn be the expected length of a CRI that starts with the transmission of n packets. namely. that some nodes transmit with higher power than others. a capture occurs with probability πn. (ii) They wait until the current CRI ends and retransmit in the ﬁrst slot of the subsequent CRI (wait scheme). Indicate the order in which Pn(l) should be computed. 0 ≤ l ≤ n let P n(l) = Prob [ Y k = l X k = n ] .144 CHAPTER 5: COLLISION RESOLUTION Problem 4. 4. Let p(n 2 n 1) = Prob [ ( Y k = n 2 ) Y k – 1 = n 1 ] . (Known Multiplicity [Tsy80. n ≥ 0 and ˜ Py(l). Also assume that when a transmission succeeds. know of that event (but other users are not aware that a capture occurred). Write recursive equations for computing Bn. How would you compute p(n 2 n 1) from Pn(l)? 3. one of the n transmitted packets is successfully received and the other n1 packets have to be retransmitted. ˜ ˜ 2. The model used is typical to power capture.1.
Section : EXERCISES 145 Problem 6. lost packets) Consider a system that uses the binarytree CRP with the epoch mechanism. Write expressions for the expected number of packets that are lost during a CRI that starts with the transmission of n packets. Assume that lost packets are never retransmitted. 5. πi=0 i>1.e.. How would you compute the optimal epoch length in this case. Derive the throughput of this system when the arrival rate is λ and the epoch length is ∆. Repeat (1)(5) when the modiﬁed binarytree CRP is used. i. 4. Write expressions for the expected number of packets that are transmitted successfully during a CRI that starts with the transmission of n packets. Denote by πn the probability of this event.5. there is a possibility that all these packets will be erased. . they will be lost and the feedback signal will indicate all users that the slot was empty. (Erasures. 3. 1. Assume that whenever n ( n ≥ 1 ) packets are transmitted. 2. Write expressions for the expected length of a CRI that starts with the transmission of n packets. 6. What is the rate (packets/slot) in which packets are lost in this case. How should λ ∆ be chosen in order to maximize the throughput. Let π1=0.
47) Multiplying equation (5.146 CHAPTER 5: COLLISION RESOLUTION APPENDIX A Moments of Collision Resolution Interval Length ˜ In this appendix we derive a closed form expression for the moments of B n .1)) B(z) = ez + ∑ n=0 zn .5) by zn/n! and summing both sides for n ≥ 2 we obtain ∑ n=2 ∞ zn B n .∑ Q i(n) ( B i + B n – i ) – [ 2 + 2z ( 1 – p + p ) ] n! i=0 ∞ n = ez – 2(1 + z) + n = 0i = 0 ∞ ∞ ∑ ∑ i! ( n – i )! p i ( 1 – p ) n – i ( Bi + Bn – i ) n zn = ez – 2( 1 + z ) + i = 0n = i ∞ ∞ ∑ ∑ . We ﬁrst demonstrate the method to obtain a closed form expression for Bn. n! (5.47) we have B(z) – z – 1 = ez –z–1+ ∑ n=0 zn . n! i=0 n Using (5..= n! ∑ n=2 ∞ zn .∑ Q i(n) ( B i + B n – i ) .∑ Q i(n) ( B i + B n – i ) n! i=0 n ∑ n=0 ∞ zn . The reader must bear with the lengthy (yet straight forward) algebraic manipulation involved in this derivation.( Bi + Bn – i ) ( n – i )! i! [ z ( 1 – p ) ] n – i [ zp ] i ..+ n! ∑ n=2 ∞ zn .p i ( 1 – p ) n – i ( Bi + Bn – i ) i! ( n – i )! i = 0n = i zn = ez – 2( 1 + z ) + ∑ ∑ . Deﬁne the exponential generating function of Bn by B(z) ∆ = ∑ n=0 ∞ ∞ zn B n .∑ Q i(n) ( B i + B n – i ) n! i=0 1 n – or (using (5.
– 2 .= k! ( k – 1 )! k! or 2 ( k – 1 ) ( –1 ) k * B k = k! [ 1 – p k – ( 1 – p ) k ] From the deﬁnitions (5.49) (5.B i ∑ i! ( n – i )! n=i ∞ + i=0 ∑ .50) ∑k = 0 Bk* z k ( –1 ) k – 1 ( –1 ) k 2 ( k – 1 ) ( –1 ) k * * * B k – p k B k – ( 1 – p ) k B k = – 2 .Section : APPENDIX A Moments of Collision Resolution Interval Length 147 = ez – 2(1 + z) + ∞ ∑ i=0 ∞ ∞ [ zp ] i [ z( 1 – p ) ]n – i .50) we get for k ≥ 2 . e – z B(z) = 1 – 2e – z ( 1 + z ) + B(zp)e – zp + B(z ( 1 – p ))e – z ( 1 – p ) .51) we obtain .= n! ∑ k=0 n * Bk . and using (5. by equating corresponding coefﬁcients of zn..48) to B *(z) – B *(zp) – B *(z ( 1 – p )) = 1 – 2e – z ( 1 + z ) Expanding both sides of (5.= B(z) = e z B *(z) = n! ∑ ∑ j=0 k=0 ∞ zj j! ∞ * Bk z k = ∑ ∑ k =0j=0 ∞ ∞ * Bk .48) (5. ∞ (5.49) we obtain k ≥ 2. since B 0 = 1 and B 1 = 0 .50) to a Taylor series around z=0. (5. letting B *(z) = and equating the coefﬁcients of zk on both sides of (5.z k + j j! from which. we obtain: Bn . It is convenient to deﬁne B *(z) ∆ e – z B(z) = which transforms (5.47) and (5.Bi ∑ .51) ∑ n=0 ∞ zn B n . ( n – k )! * * Finally.Bn – i i! ( n – i )! n=i [ zp ] i [ z( 1 – p ) ]n – i = e z – 2 ( 1 + z ) + B(zp)e z ( 1 – p ) + B(z ( 1 – p ))e zp therefore.
a closed form expression for Vn can be obtained.B k + . n≥2 . In almost the same manner.52) which is the closed form expression for Bn we were seeking.10) we obtain V (z) = 2B(z) – e z – 4 ( 1 + z ) + 2B(zp)B(z ( 1 – p )) + V (zp)e z ( 1 – p ) + V (z ( 1 – p ))e zp and with a deﬁnition V *(z) = e – z V (z) we get V *(z) – V *(zp) – V *(z ( 1 – p )) = 2B *(z) – 1 – 4e – z ( 1 + z ) + 2B *(zp)B *(z ( 1 – p )) from which Vn = 1 + ∑ V n n! ∞ zn n=0 ∑ k=2 n 2n! .* ( n – k )! [ 1 – p k – ( 1 – p ) k ] n 4(k – 1 ) ( –1 ) k ∑ Bi* Bk* – i i=0 k + k=2 ∑ k  [ [ 1 – pk – ( 1 – p )k ] ] n . Let V ( z) ∆ = Then from (5.148 CHAPTER 5: COLLISION RESOLUTION Bn = 1 + ∑ k=2 n 2 ( k – 1 ) ( –1 ) k n  k [ 1 – p k – ( 1 – p ) k ] n≥2 (5.
Carrier sensing can also be used. PYS87].e.i “receiving” node. . some of the basic protocols such as the pure and the slotted Aloha are still applicable in multihop systems. Multihop Networks The basic topology assumed throughout this book is the singlehop topology in which all users hear one another. The fact that each receiver hears a subset of transmitters rather than all of them renders the analysis of multihop systems far more complex than that of singlehop systems.CHAPTER 6 ADDITIONAL TOPICS The ﬁeld of multipleaccess systems is much too broad to be contained in a single book. This gives rise to routing issues i. A transmission is successful only if it is the only transmission currently being heard by the. Although we treated in depth many fundamental protocols and systems. An extensive survey of recent developments in the analysis of multihop systems appears in [Tob87].. we were able to uncover just the tip of the iceberg. either because they are beyond the scope we planned for the book or because they are still in a formative and fragmentary stage of research. The term “multihop” alludes to the need of packets to hop over several intermediate users in order to arrive at their destinations. since not all pairs of users communicate directly.i “multihop” network. To improve the effectiveness of carrier sensing. SiS81. known as a. BrT85. provided each receiver receives only a single transmission at a time. Yet. TaK85b. KlS87. The application of collision resolution or controlled Aloha protocols require substantial revisions of the protocol since it is difﬁcult to obtain the correct feedback information at the end of each slot in a multihop environment. The multihop topology allows for concurrent successful transmissions. CiR86]). BKM87. Many important and interesting subjects in this ﬁeld were not discussed in the book. A detailed description of multihop packet radio technology appears in [KGB78]. is characterized by the feature that each user is in reception range of only a subset of the users and similarly the transmission of a user is heard by a subset of all users. which of the receiving users should forward a received packet towards its destination. SiK83. this property is utilized to increase the capacity of the total network by an approach called spatial reuse Multihop networks appear naturally in radio networks with low powered transmitters or in interconnected local area networks. although it will not be as effective as in singlehop systems since a transmission cannot be sensed by all users. This section is devoted to short descriptions of several of these subjects. ShK87. or there is a common receiver that can hear all transmissions in the network. KBC87. the idea of busytone can be used (see ToK75. An important topology. Other relevant papers are [BoK80.
is forwarded to a station via the shared channel by employing some multiaccess protocol. The second is characteristic to carrier sensing systems. The network operates as follows: a packet that is generated at some user. In the multistation model the users of the network are originators of the data that are transmitted through a shared channel to the stations. the users should be equipped with several receivers. There are two ways in which interference is reduced compared to the single channel environment. Alohatype protocols in this environment are studied in [SiC88. Many studies that consider multichannel networks. The advantages of the multistation conﬁguration over the singlehop one are that the former allows for lower power transmitters at the users. it simpliﬁes nodal protocols. In addition. each frequency band in the multichannel system is narrower and the transmission time of a packet is longer. These networks are characterized by the ability of the users to communicate via several different. some are not mobile. The ﬁrst is obviousless users use each frequency band. CiR86]. To keep the same level of connectivity as in the single channel environment. since the total available bandwidth is ﬁxed. and is also adequate for a large number of naturally structured hierarchical networks. the population of users is not homogeneous: some users are more powerful than others. noninterfering.150 CHAPTER 6: ADDITIONAL TOPICS Multistation Networks The multihop topology assumes that every user in the network can and should serve as a repeater for packets it receives from its neighbors and that needs to be forwarded to their destinations. by forwarding them to their respective destinations (other stations or users). It is also advantageous over the multihop conﬁguration because it simpliﬁes both the design and the analysis of the network. some inherent hierarchy may exist among the users of the network. etc. In such systems it is natural to build a backbone network of stations that is responsible for the routing and other network functions. The station then forwards it to some other station through the backbone network of stations to be ﬁnally transmitted on the stationtouser channel to its destination (cellular phone systems use this approach). The propagation delay is constant and hence the ratio between the propa . Multistation networks in which a TDMA scheme is used are considered in [RoS89]. and collision resolution protocols in [BaS88]. Multichannel Systems The networks considered in this book contain a single shared channel used by the users for communication. results in better utilization of the common radio channel due to spatial reuse and allows distribution of control among several stations. communication channels at different bands. The stations may be either the ﬁnal destinations for some packets sent by the users or can act as relays for other packets. In many applications of packetradio networks. The main advantage of using multiple channels is the reduced interference level among the users.
ShK87. . Kim85. yielding better performance. Multichannel systems for various Alohatype protocols are discussed in [MaR83. Kim87. Tod85. MaB87. ChG88].Section : ADDITIONAL TOPICS 151 gation delay and the packet transmission time becomes smaller.
152 CHAPTER 6: ADDITIONAL TOPICS .
San Francisco. Protonotarios. Brussels (December 1987). T. D. 117128 (January 1977).” pp. Beuerman and E. Prentice Hall.” IEEE Trans. D. A. Towsley. Boorstyn.Another Alternative for Computer Communications.” IEEE Trans.REFERENCES Abr70. “Random MultipleAccess Communication and Group Testing. Abramson.K. Borgonovo. Coyle. COM32(7) pp. Kershenbaum. Wolf. 267274 (March 1987). Abr77.. “On the throughput of an ALOHA channel with variable length packets. COM35(3) pp. of the Fall Joint Computer Conference. 898905 (September 1986). COM34(9) pp. Abramson. N. “Collision Resolution Algorithms in MultiStation PacketRadio Networks. Data Networks. S. of ICC’75. “The ALOHA System . 281285 in Proc.R.” IEEE Trans. Bellini and P. 385400 in PERFORMANCE’87. 17 pp. Communications.” IEEE Trans. BaS88. on Communications. . on Communications. “An Improved Upper Bound for the Capacity of a Channel with Multiple Random Access.N. S. Sahin. Apostolopoulos and E. “The Delay Characteristics of CSMA/CD Networks. on Communications. and V. BeT88. Inc.” Problemy Peredachi Informatsii.” pp. N. Sidi. N. A. on Communications. BeC88. BarDavid and M. Tszan.” pp. R.” Problemy Peredachi Informatsii. Binder.L.Z. 19321935 (November 1980). BeZ88. COM36(5) pp. NewJersey (1987).S. “Upper Bound for the Capacity of a Random Multiple Access System.” IEEE Trans. Berger and T. 41. B. “Queueing Analysis of Buffered CSMA/CD Protocols. COM25(1) pp. “Throughput Analysis in Multihop CSMA Packet Radio Networks. 553563 (May 1988). COM28(11) pp. “A Dynamic Packet Switching System for Satellite Broadcast Channels. Berger. 21(4) pp. ApP86. Berger and R. T. R. 769779 (July 1984).J.” IEEE Trans. Gallager. Zhu. T. and J.141. Mehravari. BMT84. Bertsekas and R. “The Throughput of Packet Broadcasting Channels. BKM87. 8387 (January 1985). Bin75.5 in Proc. (1970). BeG87. 9095 (January 1988). on Communications. Maglaris. BeB80. California (1975). T.
154 : REFERENCES BoK80. Chlamtac and A. “Throughput Analysis of Multihop Packet Radio. 317324 (April 1985). Hellman. IT25(5) pp. Sidi.” IEEE Trans. Cidon and M. “BRAM: The Broadcast Recognizing Access Mode. BrT85. 263271 (March 1988). “The Effect of Capture on CollisionResolution Algorithms. of ICC’80. “Erasure. “The Multiple Access Broadcast Channel: Protocol and Capacity Considerations. S. Coyle and B.” IEEE Trans. Cidon and R. on Information Theory. on Information Theory. CaH75. Cap77. “Erasures and Noise in Multiple Access Algorithms. “Channel Allocation Protocols in FrequencyTime Controlled High Speed Networks. 12(1) pp. “Bistable Behavior of ALOHAtype Systems.” IEEE Trans.6. R. Franta. 401410 (April 1975). CoB80.C. I.I. Ganz. IT34(1) pp. Seattle.” pp. Communications. W. 110 (August 1986). Washington (1980). Levin. D. Capetanakis. “Tree Algorithm for Packet Broadcast Channels.D. COM27(8) pp. Liu.” IEEE Trans. ChG88.I. H. “Throughput Analysis of Spread Spectrum Multihop Packet Radio Networks. Carleial and M.D.” IEEE Trans. on Communications. COM33(4) pp.R.113. CiR86. de Boor. E. Capture and Random Power Level Selection in MultipleAccess Systems. 256265 in Proceeding of IEEE INFOCOM’85. J. (March 1985).). IT33(1) pp. Washington. Cidon. Rom. COM23(4) pp. MIT (August 1977). 13. and K. 505515 (September 1979). I. on Communications.R.” in Ph. J.” IEEE Trans. and M.J. CFL79. CoL83. COM31(11) pp. .6 in Proc. 101110 (January 1988).” pp. on Communications. CiS87. 12471251 (November 1983).A. “Carrier Sense Access in a Two Interfering Channels Environment. Sidi.” IEEE Trans.E. I. Chlamtac. on Communications. COM36(3) pp. Cidon and M. I. Cap79. I. Tobagi. Brazio and F. Kodesh. 132143 (January 1987). Boorstyn and A. J. on Information Theory.D. Sidi. Kershenbaum. Dissertation. I. “Finite Population CSMACD Networks. “Conﬂict Multiplicity Estimation and Batch Resolution Algorithms. CiS88.B. CKS88. I. 430440 (April 1988). Department of Electrical Engineering. Conte and C. Sidi. Communications. 11831189 (August 1979). COM36(4) pp.6. Elementary Numerical Analysis: An Algorithmic Approach (3rd ed. CiS85. McGraw Hill (1980).” IEEE Trans. A.M. Cidon and M. Capetanakis.” Computer Networks and ISDN Systems.” IEEE Trans.
644654 (July 1977). COM25(1) pp. FFH85. on Communications.D.Z. Clark.” IEEE Trans.” IEEE Trans. D. W.J.” IEEE Trans. R. and D. K.” pp. . Jacquet. “Delay Analysis of Interacting Queues with an Approximate Model. D. C33(12) pp.” IEEE Trans. IT31(2) pp. “An Approximate Analysis of Delay for Fixed and Variable Length Packets in an Unslotted ALOHA Channel. Fayolle. on Communications.A. A. IT28 pp. 194201 (February 1987). 136139 (January 1977). 685691 (June 1988). Reed. Flajolet. Computers. “Analysis of a Stack Algorithm for Random MultipleAccess Communication. 402405 (May 1982). COM23(11) pp. “A System for Broadcast Communication: ReservationALOHA. Walden. M.S. P. Information Theory. G. on Communications. Ferguson.” IEEE Trans. Honolulu. Stability. “Demand Assignment Multiple Access Schemes in Broadcast Bus Local Area Networks. “Performance of Slotted ALOHA Random Access with Delay Capture and Randomized Time of Arrival. Hawaii (1973). Zhu. Fine and F. 703710 (May 1980). M. G. 13061311 (November 1975). Crowther.” IEEE Trans.” IEEE Trans. Davis and S. Rettberg. Ephremides and R. Meditch.J. and Virtual Time CSMA. 14971517 (November 1978). CrH82. “A New Upper Bound to the Throughput of a MultiAccess Broadcast Channel.Section : REFERENCES 155 CoL85.J. R. Fer77a. M. Liu. CuM88. DaG80. “A Matrix Representation of CSMA/CD Networks. COM36(6) pp.P. on Communications.A. COM28(5) pp.T. COM35(2) pp.” Proc. Ornstein. “On the Control. of the 6th Hawaii International Conference on Systems Sciences. on Communications. 244254 (March 1985). 66(11) pp. and F. 113059 (December 1984). “Distributed Retransmission Controls for Slotted. Ferguson. D.” IEEE Trans. CPR78.” IEEE Trans. E.A. on Communications. Cunningham and J. on Information Theory. Fer77b. S.J. Hofri. and Waiting Time in a Slotted ALOHA RandomAccess System. 5364 (January 1985). “A Bound and Approximation of Delay Distribution for FixedLength Packets in an Unslotted ALOHA Channel and a Comparison with Time Division Multiplexing (TDM). CRW73. Pogran. Tobagi.” IEEE Trans.H. Fer75. COM25(7) pp. Gronemeyer. M. M. Ferguson. on Communications. 371374 in Proc. Cruz and B. of the IEEE. Coyle and B. EpZ87. Heart. FiT84. COM33(1) pp. “Introduction to Local Area Networks. Nonpersistent. and P. Hajek.
R.P. 289325 (April 1987). G. Hayes. D. Plenum Press. Gallager. New York (1984).” Bell System Technical Journal.” IEEE Trans.P. L.” IEEE Trans.G. IT33(2) pp. 10511062 (July 1987).F.487 Throughput Limited sensing Algorithm. J. and E. GMP87. and P. 4(1) pp. D. “An Adaptive Technique for Local Distribution.” Journal of the ACM. PapantoniKazakos. on Information Theory.J. on Communications. O’Reilly. Hey82. COM27(6) pp. and R. P.” IEEE Trans. 4953 (1974). on Communications. Hammond and P. Van Loon. Provincetown (September 1978).” IEEE Trans.” IEEE Trans. 233237 (March 1987). 20232051 (October 1982). COM26(8) pp. “The Stability Problem of Broadcast Packet Switching Computer networks. B. “A Perspective on Multiaccess Channels. Performance Analysis of Local Computer Networks. Labetoulle.” IEEE Trans. IT31(2) pp. A. 7476 in Proc. L. Modeling and Analysis of Computer Communications Networks. 34(2) pp. Hayes.” pp.P. Gek87. 61 pp.G. Gelenbe.” IEEE Trans. Gal78. 11781186 (August 1978). Flajolet. Fayolle. Heyman. 124142 (March 1985). AddisonWesley Publishing Company (1986). Greenberg. Ladner.G. on Automatic Control. D. L. AC27 pp. Schwartz. HaO86. HaS79. “Conﬂict Resolution in Random Access Broadcast Networks. “A Collision Resolution Protocol for Random Access Channels with Energy Detectors. on Information Theory. Hey86. Hay78. AFOSR Workshop Communication Theory and Applications. COM34(6) pp. Gal85.E. “An Analysis of the CarrierSense MultipleAccess Protocol. PapantoniKazakos. “A 0. Hajek and T.F.” IEEE Journal on Selected Areas in Communications. “The Effects of Random Message Sizes on the Performance of the CSMA/CD Protocol.L. Georgiadis and P. COM30(11) pp. Hansen and M. “Estimating the Multiplicities of Conﬂicts to Speed Their Resolution in Multiple Access Channels. GGFL87. HaL82. 846856 (June 1979). “A Method for the Delay Analysis of Random MultipleAccess Algorithms whose Delay Process is Regenerative. PapantonyKazakos. Gallager. Hay84. GeP82. Georgiadis and P. 559569 (June 1982). J. SAC5(6) pp.” Acta Informatica. 24132420 (November 1982). Georgiadis. “Decentralized Dynamic Control of a Multiaccess Broadcast Channel. “An AssignedSlot ListenBeforeTransmission Protocol for a Multiaccess Data Channel.W.F. R. L. L. on Communications. 547553 (June 1986). Heyman. .156 : REFERENCES FLB74. Bastin. on Communications. J. J. Merakos.
Section : REFERENCES 157 HoR87. 7885 (May 1987). KiK83. 531536 (May 1988). M. Kiesel and P.C. F. 345347 (April 1986). and R.” IEEE Trans.C. IT33(3) pp.” Proceedings of the IEEE. Kimura. Communications. Berger.” IEEE Trans. ItR84. on Information Theory. Sidi. “An Algorithm for Evaluation of Throughput in Multihop Packet Radio Networks with Complex Topologies. Electronics and Communication Engineering. Communications.” Journal of the Royal Statistical Society. Itai and Z. on Information Theory. 341349 (1987). 14681496 (November 1978).C. Kim85. Chen. Electronics and Communication Engineering. Automatic Control.” IEEE Trans. IT34(3) pp.J. Berger. AC29(8) pp. “Delay Analysis of 0. “Packet Delay Under the Golden Ration Weighted TDM Policy in a Multiple Access Channel. A. A. Burchﬁel. “Advances in packet radio technology. 869876 (November 1983). and M.E. Huang and T. COM34(4) pp. KBC87. KGB78. Y. IT31(2) pp. J. “An Analysis of the Multichannel CSMA/CD Protocol by Nonslotted Model. J68B (Part B)(12) pp. 264273 (March 1985). HuB86. “Stochastic Models of Computer Communication Systems.A. “On the Throughput of Channel Access Algorithms with Limited Sensing. Jen80. “Delay Analysis of IntervalSearching Contention Resolution Algorithms.” IEEE Journal on Selected Areas in Communications. Rosberg. Kuehn. A. “A Golden Ratio Control Policy for a MultipleAccess Channel.” Trans.” IEEE Journal on Selected Areas in Communications. KeS88. Jenq. P. 70(5) pp. of the Japanese Inst. 712718 (August 1984).S. G. HuB85. of the Japanese Inst. J. G. J.M. 47(1) (1985). I.” IEEE Trans.P Kelly. COM28(11) pp. “Mixing Collision Resolution Algorithms Exploiting Information of Successful Messages. Kunzelman. on Information Theory. SAC1(5) pp. Boorstyn. 13411348 (December 1985). “An Analysis of the MultiChannel CSMA/CD Protocol by Nonslotted Model. Humblet.” IEEE Trans.A. Hum86. Kershenbaum. Kimura. .C. Gronemeyer. R.487 Contention Resolution Algorithms. Kel85. W. on Communications. Hofri and Z. SAC5(6) pp. R. “On the Stability of Slotted ALOHA Systems. Kessler and M. 19361939 (November 1980). 66(11) pp. Kim87. Huang and T.” IEEE Trans. Rosberg. COM34(9) pp.” IEEE Trans.” Trans. Kahn. “A new CSMACD protocol for local area networks with dynamic priorities and low collision probability. 10031012 (July 1987). 916926 (September 1986).
410423 (April 1975). M. Kleinrock and M. “Packet Switching in a Multiaccess Broadcast Channel: Dynamic Control Procedures. Protocols and Techniques for Data Communication Networks. Kleinrock. 10151029 (July 1980). on Communications. “Spatial Reuse in Multihop Packet Radio Networks. Kurtz and M. C29(7) pp. of ICC’78. F.S. S. 1976). S.” Proceedings of the IEEE. 596603 (July 1980). Kleinrock. Communications. 13161323 (December 1988). J.F.” IEEE Trans. “Multiple Access Algorithms via Group Testing for Heterogeneous Population of Users. on Computers. on Communications.” in Proc.S. “Controlling window protocols for timeconstrained communication in multiple access networks. Lee87. Tobagi.S. “Packet Broadcast Networksa Performance Analysis of The RALOHA Protocol. COM25(12) pp. COM36(12) pp.” IEEE Trans.” IEEE Trans. Prentice Hall. “Error Probabilities for Binary DirectSequence SpreadSpectrum Communications with Random KlS80. KSY88. . L. Wiley (1975. “Random Signal Levels for Channel Access in Packet Broadcast networks. 14891494 (December 1977).C. SAC5(6) pp. L. Lam. L. KlS87. 10261034 (July 1987). Lam80. Schwartz. Kurose. Canada (1978). Kleinrock and J. COM28(7) pp. 156167 (January 1987).” IEEE Trans. L. Kleinrock and F. COM23(12) pp.” IEEE Trans. Toronto. and Y. J. LeP87. Lam. Lam and L. “Delay Analysis of a Time Division Multiple Access (TDMA) Channel. Kuo81. Kleinrock and Y. Kleinrock and S. KlY78. on Communications. Sidi. D. on Communications.B.F. Queueing Systems (Vols. Scholl.S.” IEEE Trans.Carrier Sense MultipleAccess Modes and Their Throughput Delay Characteristics. on Communications.S. L.A. 891904 (September 1975). COM23(4) pp. on Communications. “Packet Switching in a Multiaccess Broadcast Channel: Performance Evaluation.” IEEE Trans. I. “Packet Switching in Radio Channels: New ConﬂictFree Multiple Access Schemes. 75(1) pp. Lee. KlT75. COM36(1) pp. L. Lehnert and M. “An optimal adaptive scheme for multiple access broadcast communication. Yemini.158 : REFERENCES Kle76. 14001416 (December 1975). COM23(9) pp. Lam. New Jersey (1981). Kuo. 4149 (January 1988). C. Pursley. S. KuS88. Kle75. Yemini.” IEEE Journal on Selected Areas in Communications.” IEEE Trans. J. Silvester. Lam77. “Packet Switching in Radio Channels: Part I . II). LaK75.
“Poisson MultipleAccess Contention with Binary Feedback. “On Retransmission Control Policies in MultipleAccess Communication Networks. COM24(4) pp. Bruscagin. IT31(2) pp. 217243 (March 1985).” IEEE Trans.” pp.L. on Communications. MeK85. COM31(6) pp. J. MaR83. Martin. Information Theory. MaB87. 109117 (February 1985).” IEEE Trans. MaF83.L.” IEEE Trans. 745751 (September 1984). 268275 in IEEE INFOCOM’87. J.Section : REFERENCES 159 Signature Sequence.” pp. San Francisco.J. Lea. Marsan and D. 61 (Part 1)(7) (September 1982). PrenticeHall. 763774 (June 1983). M. Meditch and C. “Multichannel ALOHA Networks with Reduced Connections. Massey. Honolulu. New York (1981). A R. “Collision Resolution Algorithms and RandomAccess Communications. "SteadyState Analysis of a Slotted and Controlled ALOHA System with Blocking. April 1980) N. “The Collision Channel Without Feedback. New Jersey (1978). Metcalfe. Mathys and P. Communication Satellite Systems. CA (April 1987). MaM85. “Qary Collision Resolution Algorithms in Randomaccess Systems with Free or Blocked ChannelAccess. 8798 (January 1987). SpringerVerlag. COM35(1) pp. AC30(2) pp. Mas81. M. Mathys. P. Rofﬁnella. L. Information Theory.S. . Hawaii (January 1973). J.A. Flajolet. Merakos and D. MeB84. Met76. on Information Theory.A. 6th Hawaii International Conference on System Sciences. “Multichannel Local Area Network Protocols. Automatic Control. “On Improving Utilization in ALOHA Networks. a Unidirectional Local Area Communication Network. Met73. Longo. Mehravari and T. Flores.” IEEE Trans.O." pp.” IEEE Trans. Metzner. 447448 (April 1976). Limb and C. on Communications. IT31(2) pp. LiF82. Marsan and M. (Also in UCLA Technical Report UCLAENG8016. J.T.” IEEE Trans. SAC1(5) pp. 192204 (March 1985). Mar78.” IEEE Journal on Selected Areas in Communications. “Description of Fasnet. J. on Communications. 885897 (November 1983).A. ed.” Bell System Technical Journal. Berger. 375380 in Proc. 73137 in MultiUser Communications Systems (CISM Courses and Lectures Series). J. “Stability and Optimization of the CSMA and CSMA/CD Channels. MeL83. Massey and P. G. Kazakos.” IEEE Trans. IT30(5) pp.
MoH85. De Moraes and I. Pip81. Information Theory. 9095 (January 1981). COM32(5) pp. Onozato and S. V.A. M. 145151 (February 1985). and A. The John Hopkins Press (1981). on Communications. 396401 (May 1982). Molle. 11634 (January 1987). Venetsanopoulos. PYS87. “Some Conditions of Ergodicity and Recurrence of Markov Chains. (1) pp. “Virtual Time CSMA: Why Two Clocks are Better Than One. IT27(2) pp.F. “The canonical approximation in the performance analysis of packet radio networks. A. J.B.160 : REFERENCES MiT81.” IEEE Trans. 11711182 (November 1985).C.” IEEE Trans. Noguchi. Rubin. COM35(11) pp. 795799 (August 1977).G.L. M. “A Class of Efﬁcient contention Resolution Algorithms for Multiple Access Channels. on Communications. “On the Thrashing Cusp in Slotted ALOHA Systems.” Problemy Peredachi Informatsii. Neuts. Pakes. Molle and L. Y. M.E. PMV87. D. COM33(11) pp. Polyzos.A. Molle. on Communications. N. L. M. 17 (1969). Mueller. OnN85.” Proceedings of IEEE. Mosley and P.L. Yemini. “Bounds on the Performance Of Protocols for a MultipleAccess Broadcast Channel. 17(1) pp. Pursley. “On the Capacity of Inﬁnite Population Multiple Access Protocols.” pp. MoR84.” IEEE Trans.S. “A Method of Solving Algebraic Equations Using an Automatic computer.” IEEE Trans. M. Neu81.L.” Mathematical Tables and Other Aids to Computation (MTAC). 145151 (March 1981). MoK85. “Performance Evaluation for PhaseCoded SpreadSpectrum MultipleAccess Communication (part I: System analysis). 208215 (1956). Pur77. “Message Delays for a TDMA Scheme under a Nonpreemptive Priority Discipline. Information Theory. 11241138 (November 1987). G. COM33(2) pp. MatrixGeometric Solutions in Stochastic Models: An Algorithmic Approach. M. on Communications. Sidi. Pinsky. 5838 (May 1984). Tsybakov. COM33(6) pp.” IEEE Trans. Y. “Upper Bound for the Capacity of a Random Multiple Access System.M. E. 919933 (June 1985). Mikhailov and B.” IEEE Trans. 10 pp. Mol82. 140162 in Pur87. and M.” IEEE Trans. Pursley. on Communications. Humblet.B. Mue56.” IEEE Trans. COM25(8) pp. “Performance Analysis of Finite Nonhomogeneous Population Tree Conﬂict Resolution Algorithms using Constant Size Window Access. Pak69.F.” Operations Research. IT28(3) pp. on Communications. . “The Role of Spread Spectrum in Packet Radio Networks. Kleinrock. Pippenger.N.
Rosberg and M.M. Sant.N. on Communications. 680689 (June 1981). RoS89. “TDM Policies in Multistation PacketRadio Networks. on Communications. Rob72. R. Netherlands (1975). I. Rom86. R. D. “Ordering Subscribers on Cable Networks. 906911 (June 1979). COM27(5) pp. S. R.R." IEEE Trans.” IEEE Trans. Introduction to Probability Models.” in Computer communication networks. Stability and Optimization of Slotted ALOHA with a Finite Number of Buffered Users. New York (1972). Yemini. Rubin. Rom and F.” IEEE Trans. T. “Analysis. Kuo. “Throughput of unslotted ALOHA channels with arbitrary packet interarrival time distributions.F. RoT81.” IEEE Trans. Computer Science Press. AC26(3) pp. Rom. 2(4) pp. R. Kleinrock." IEEE Communications Magazine (26)(3) pp. COM37(1) pp. Ephremides. Sidi. 235249 in Local Area and Multiple Access Networks. 3138 (January 1989). Rubin. Grimsdale and F. Y. Rub78. "Group RandomAccess Disciplines for MultiAccess Broadcast Channels. IT33(3) pp. “MessageBased Priority Functions in Local Multiaccess Communication Systems. Tobagi. on Automatic Control. Noordhoff Internat Publishing. Riv87. "Alternative Local Area Network Access Protocols. Z.L. Sac88. 322334 (November 1984). on Communications. SaE81. ScK79. Academic Press. 273286 (July 1981). (1986). M. Scholl and L. “Network Control by Bayesian Broadcast. COM27(6) pp. “Collision Detection in Radio Channels.” ACM Transaction on Computer Systems. . Computer Science Press.” IEEE Trans.” pp. ed. Saadawi and A. Ed. Sachs. “On a Mixed Mode Multiple Access Scheme for PacketSwitched Radio Channels. 578592 (September 1978). R. Ros72. 769777 (May 1979). Rub79. 14221425 (August 1980).L. Communications. 323328 (May 1987). S.” IEEE Trans. on Information Theory IT24(5) pp.G. COM28(8 (part 2)) pp. I. “Message Delays in FDMA and TDMA Communication Channels. (1987).Section : REFERENCES 161 Current advances in distributed computing and communications. 5(4) pp.” IEEE Trans. L. San80.A. Ross.” Computer Networks. on Information Theory. Roberts. “Dynamic Allocation of Satellite Capacity Through Packet Reservation. Rivest. 2545 (March 1988). Groningen. Rom84. Rom.
” IEEE Trans.L. ShH82. 295301 (March 1985). Segall. on Communications. “A BusyTone MultipleAccess Type Scheme for PacketRadio Networks. Shacham. 4(3) pp. Molle. 8(1) pp. 10131025 (20 July 1987). COM35(2) pp. “Splitting Protocols in Presence of Capture. COM31(1) pp.” Performance Evaluation. on Communications. Data and Computer Communications. Kleinrock. IFIP TC6. Hunt. Paris. and A. Shacham and P. Ishibashi. “A Reservation Protocol for Satellite Packet communicationa Performance Analysis and Stability Considerations. Macmillan Inc. Silvester and L. M.” Performance Evaluation. M.” IEEE Trans. Szp86. M. 123129 (January 1983). W. J. Communications. N. W. Cidon.N. COM31(8) pp. TaI84. Venetsanopoulos.” IEEE Trans. on Communications. 401414 in Proc. SiC88.J. “ThroughputDelay Performance of PacketSwitching MultipleAccess Channel with Power Capture. of the International Symposium on Local Computer Networks. 974982 (August 1983). Sidi and A.” IEEE Trans. COM32(8) pp. Sidi and I. 9207 (August 1984). Sohraby. S.B. 110 in The International Conference on Performance of Data Communication Systems and their Applications. COM34(11) pp. M. Segall. ShK87. Tasaka and Y. K. France (1416 September 1981). “On the Capacity of Multihop Slotted ALOHA Networks with Regular Structure.” IEEE Trans. SiS83. Sidi and I. SiC85. 6572 (February 1988). Sidi and A. SMV87.. Cidon. Szpankowski.” pp. “Two Interfering Queues in PacketRadio Networks. “Bounds for queue lengths in a contention packet broadcast system. IT31 pp. “Comments on ‘Throughput Analysis for Persistent CSMA Systems’. M. Stallings.” IEEE Trans. 153170 (August 1984). SiK83.162 : REFERENCES Sha84.A. . Shacham and V. Sta85. SiS81. “Architectures and Performance of Multichannel Multihop Packet Radio Networks. New York (1985). 240243 (February 1987). Information Theory.” pp. on Communications.B. Florence. SAC5(6) pp. N. King. “Performance Evaluation of the CSMACD 1Persistent Channel Access Protocol in Common Channel Local Networks. N. “A MultiStation PacketRadio Network. Italy (April 1982). 11321140 (November 1986).” IEEE Journal Selected Areas Communications.
F. 11911199 (November 1985). 75(1) pp. (1986). Tob82b. 245259 (October/November 1980). New Orleans.) pp. F. 185200 (January 1982). . Kleinrock. Kleinrock.” IEEE Trans. Tob82a. Tobagi. “Distributions of Packet Delay and Interdeparture Time in Slotted ALOHA and Carrier Sense Multiple Access. A. Performance Analysis of Multiple Access Protocols.A. (Corrected February 1987) H. 276280 in GLOBECOM’85.” IEEE Trans. (April 1980).” IEEE Trans. B. COM33(10) pp. 627638 (July 1985). Hunt. “Modeling and Performance Analysis of Multihop Packet Radio Networks. Tob87. TML83. “Throughput Analysis for Persistent CSMA Systems. 243245 (February 1987). 11361139 (October 1985). Takagi and L. “Carrier Sense Multiple Access With MessageBased Priority Functions. H. H.S.A. Tobagi. Tobagi and V. 19(1) pp. Mass. TaK857.S. F. on Communications. “Performance Analysis of Carrier Sense Multiple Access with Collision Detection. Cambridge. Kleinrock.A. Tasaka.A. on Communications. “Correction to ‘Throughput Analysis for Persistent CSMA Systems’. COM35(2) pp. Information Transmission. MIT Press. 29(4) pp. New Jersey (1981). T. Takagi and L. Tas86. V. SAC1(5) pp. LA (December 1985).” IEEE Journal on Selected Areas in Communications. Yamada. on Communications.” Probl. 907927 (October 1982).A. H.” Journal of the ACM. Mikhailov. Inc.” IEEE Trans. Takagi and S. Kleinrock. Tannenbaum. 877884 (November 1983).D.Section : REFERENCES 163 TaK85a. of the IEEE. 468488.” Proc. 4(5) pp. F. S. COM33(11) pp. Todd.B. F.B. Computer Networks. “Mean Packet Queueing Delay in a Buffered TwoUser CSMA/CD system.” Computer Networks. Prentice Hall. COM33(7) pp. on Communications. Tobagi.” IEEE Trans. “Output Processes in Contention Packet Broadcasting Systems. Tobagi. Takagi and L. Takagi and L. “Bounds for Packet Transmission Rate in a RandomMultipleAccess System.A.” IEEE Trans. Likhanov. on Communications. 135155 (January 1987). Tod85. and N. “CSMA/CD with Deterministic Contention Resolution.. ToH80. on Communications. “Throughput in Slotted Multichannel CSMA/CD Systems. TaY86. 30(1 pt 2. TaK85b. “Multiaccess Protocols in Packet Communication Systems.” pp. COM28(4) pp. Tob80. 5068 (JanuaryMarch 1983). A. Tsybakov. Tan81.
on Communications.A. COM25(10) pp. Bakirov. Likhanov.A. 832845 (August 1976). Zurich (August 1980). Tsybakov and N. COM35(5) pp. 19 pp.A.S. ToK76.” IEEE Trans.Polling and (Dynamic) SplitChannel Reservation MultipleAccess. Tsybakov and N. Tsai and J. F. on Communications. Rom.B.” IEEE Trans. Kleinrock. Towsley and G. Vales. on Communications. Tobagi and L. “Announced arrival random access protocols. Tsybakov and V. TsM78.S.L. B. of the IFIPWG 6.4 Local Area Networks Workshop. “Efﬁcient Round Robin and Priority Schemes in Unidirectional Broadcast Systems. ToK77. “Stability Analysis of a Packet Switching Network and Its Application to Asynchronous Aloha Radio Networks. “Packet Switching in Radio Channels: Part II .O. 24(2) pp.” IEEE Trans. “Free Synchronous Packet Access in a Broadcast Channel with Feedback. Tobagi and L.” Probl. 224236 (January 1988). Towsley and P. D. ToR80. “Packet Switching in Radio Channels: Part IV .” Problemy Peredachi Informatsii. Information Transmission. “Window Random Access Protocols for Local Computer Networks. COM23(12) pp. Information Transmission.S. J.” IEEE Trans. 139151 (October 1988). Tsi87.” IEEE Trans. on Communications. Tsybakov and V. Likhanov. “Packet Switching in a Channel Without Feedback. “Analysis of a Multiaccess Control Scheme. TsL83. ToV87. Chang. 14171433 (December 1975). COM24(8) pp. D. B. 259280 (OctoberDecember 1978). COM34(7) pp. on Computers.The Hidden Terminal Problem in Carrier Sense MultipleAccess and the Busy Tone Solution.164 : REFERENCES ToK75. B.” in Proc. Mikhailov. Kleinrock.S. 513521 (May 1987). Tsitsiklis.” IEEE Trans.A. on Communications. Venkatesh. ToV82. TsB88. 10171020 (November 1987). 11031119 (October 1977). 23(3) pp. .” IEEE Trans.” Probl. F.A. 6984 (AprilJune 1983). B. F. 715722 (August 1982). Kleinrock. TsL88. C31(8) pp.N.” Problemy Peredachi Informatsii. “Performance Study of an Adaptive Reservation Multiple Access Technique for Data Transmissions. Tobagi and L. Tobagi and R.Stability Considerations and Dynamic Control in Carrier Sense MultipleAccess. 14(4) pp.B.F. “Packet Switching in Radio Channels: Part III . on Automatic Control. AC32(11) pp. D. TsC86. “Upper Bound on the Capacity of a Random Multiple Access System. 725727 (July 1986). F.
Tsy85.S. SE13(8) pp. B. Takahashi.” IEEE Trans. 131147 (AprilJune 1983). . 305317 (OctoberDecember 1980). Tsybakov and V.” IEEE Trans. Y. on Communications. Mikhailov.” Probl. IT31(2) pp. Tsybakov.S. 301312 (OctoberDecember 1979). Zhao and K. and T. VvT83. Tsybakov. 15(4) pp. 19(2) pp. TTH88. on Information Theory.S.A. Information Transmission. 16(4) pp. Tsy80. B. Tsm80.A. 134144 (AprilJune 1980). B.S. ZhR87. COM36(8) pp. W.” IEEE Trans. N. Ramamritham.” Prob. Takin. “Random Multiple Packet Access: PartandTry Algorithm. “Survey of USSR Contributions to MultipleAccess Communications. “Resolution of a Conﬂict of Known Multiplicity. Mikhailov. Tsybakov and V.” Probl. Vvedenskaya and B. Hasegawa. “Virtual Time CSMA Protocols for Hard RealTime Communication.S. 938952 (August 1987). Software Engineering. Information Transmission. “An Approximate Analysis of a Buffered CSMA/CD. “Random Multiple Access of Packets to a Channel With Errors.” Prob. “Ergodicity of a slotted ALOHA system.D. T. B. 932941 (August 1985). 143165 (March 1985). Information Transmission. 16(2) pp.Section : REFERENCES 165 TsM79. Tsybakov. Information Transmission.
166 : REFERENCES .
Since .APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND This appendix summarizes some of the important properties and results regarding queueing and Markov processes that are used in the text. the probability that at time n the process is in state j depends only on its state at n time n1 and not on prior history. Markov Chains ˜ Consider a ﬁnite or countable set E = { E 0. p ij = p ij for all n. The expectation is denoted by ˜ x or just x. A Markov chain having but one class of states is called irreducible. Two states of a Markov chain are said to communicate if and only if there is a positive probability that the process ever be in state j after having been in state i.. i. There is a variety of other ways to characterize states in Markov chains. In general. E 1. In this appendix as well as throughout the text we adopt a consistent notation as follows. aperiodic and homogeneous chains. For conciseness we consider the states as being the set of integers. In fact all communicating states form a class of states.e. e. the reader is expected to be acquainted with the items on the list (and with stochastic processes in general) to the extent that he/she understands them and knows how to make use of them. and vice versa. by f x(x) its probability density func˜ ˜ * ˜ ˜ tion. The material here is based on textbooks by Ross [Ros72] and Kleinrock [Kle76]. A ˜ random variable is denoted by a letter with a tilde. n ≥ 0 } . and by x k its kth moment. Such a stochastic process is a Markov chain if ˜ ˜ ˜ ˜ ˜ ˜ Prob [ x n = j x n – 1 = i..e. This is only a list. x 0 = i 0 ] = Prob [ x n = j x n – 1 = i ] ∆ p ij = n that is. When the transition probabilities are time indepenn dent. We say that the process is in state j at ˜ n = E j . …. i. The onestep probabilities can be arranged in a matrix P called the transition matrix. time n if x E j = j . in this textbook we are interested only in irreducible. The mstep transition probability of a homogeneous Markov chain is deﬁned as ( ˜ ˜ p ijm ) = Prob [ x m + n = m + n x n = i ] and is the probability of transitioning from state i to state j in exactly m steps. The quantities p ij are called the onestep transition probabilities of the process at time n.g. by F x (s) the Laplace transform of f x(x) . x . If x is a discrete ˜ ˜ random variable then X(z) denotes its generating function. x n – 2 = i n – 2. the chain is called homogeneous.. n ≥ 0 } ˜ in which x n ∈ E designates the state of the process. notably periodicity and ergodicity (whose deﬁnition we leave out). For this random variable we denote by F x(x) its probability distribution function. … } and a stochastic process { x n. a discrete stochastic process is denoted { x n.
the set of equations can be written in matrix form as π = πP where π is the row vector of the values πi. steadystate probabilities. The limiting probability πi is the steady state probability that the process is in state i. The result most frequently used in the text stems from the following theorem: For an irreducible ergodic Markov chain the limit ( π j ∆ lim p ijn ) = n→∞ exists and the values πj are the unique nonnegative solutions of the set of equations πj = 1 = ∑i πi pij ∑i π i Several remarks and corollaries result from the above theorem. i.. it is also the proportion of time that the process stays in state i (the latter remains true for periodic chains). First. . Recurrent Markov chains are members of another family of stochastic processes known as regenerative processes. The probabilities πi are (interchangeably) referred to in the literature as limiting probabilities.e. In general the term “steadystate” refers to the operation of the process after a long time. or invariant probabilities.. This notation is especially useful when the number of states is ﬁnite as the tools of linear algebra can be put to work. of the behavior of the process after time ti. It can also be shown that if the set of equations has a solution such that ∑ π i < ∞ then the chain is ergodic. This special family of processes possesses the property that there exist times t0.. such that the behavior of the process after time ti+1 is a repetition. Referring to the time between two regeneration points as a cycle. for large values of n. t1. stationary probabilities. we have that Exected time in state j in a cycle Proportion of time in state j = Expected cycle length This relation is used extensively in the textbook when a (regenerative) system is modeled as having two statesuseful and uselessthe ratio of which is a good measure of efﬁciency.. in a probabilistic sense.168 : APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND ergodicity plays an important role in the analyses the proof of ergodicity is included in the text in the appropriate places.
the time from the beginning of the interval to the sampled ˜ point has the same distribution as y .)l The following holds for an M/G/1 queueing system: .) according to a common distribution F x(x) (or density f x(x) and ˜ ˜ expected value E [ x ] = x .Laplace transform of b(t). The M/G/1 Queue Consider a queueing system in which arrivals occur according to a Poisson process with ˜ parameter λ and in which x the service rendered to the customersis distributed according to a distribution B(t).Queueing time (time spent in queue) W .e.Average queueing time. With this in mind we make the following notation: b(t) .d. it is sampled and we are interested in the distribution and moments of the residual time i. In such a queueing system the number of customers in the system as seen by an outside observer equals that seen by an arriving customer which equals that seen by a departing customer. B*(s) ..Load factor ˜ q .Generating function of q ˜ .e. i.Steady state number of customers in queue ˜ Q(z) . the time until the next marked point.i. ˜ If y denotes the residual time then: 1 – F x ( y) ˜ f y(y) = ˜ x * 1 – F x (s ) ˜ * F Y (s) = ˜ sx ˜ x2 ˜ E [ y ] = 2x ˜ x3 ˜ E [ y ] 2 = 3x Where F*( . denoted x n n ≥ 0 . ) is the Laplace transform of the corresponding probability density function.Average delay time ˜ W . while the process is ongoing.Probability density function of the service time.Section : APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND 169 Residual Life Consider a stochastic (renewal) process that marks time instants on the time axis in a way ˜ that the length of the marked intervals. At some random time t.Time spent in the system (delay time) D D . The age of the process. are independent and identically distributed (i.. ρ = λx .
170 : APPENDIX A MATHEMATICAL FORMULAE AND BACKGROUND (1 – ρ)(1 – z) Q(z) = B *(λ – λz) B *(λ – λz) – z ˜ λ2 x2 ˜ E [ q ] = ρ + 2 ( 91 – ρ ) ( 1 – ρ )s W *(s) = s – λ + λB *(s) ˜ λx 2 W = 2(1 – ρ) ( 1 – ρ )s D *(s) = B *(s) s – λ + λB *(s) ˜ λx 2 D = x + W = x + 2(1 – ρ) .
D*(s). N. D . )) General function (f( . Bni) Backlog departure rate (b. Sn. q j . )) Normalized channel (offered) load Generating function (G(z). A(z). Gn(z)) Channel (offered) load ˜ Idle period ( I . Pn) probability (p. D. a ) Normalized endtoend delay ˜ Number of arrivals ( A . pij) Transition matrix Generating function of q (Q(z). Qk(z)) Probabilities (Qi(n)) ˜ ˜ Number of packets in queue ( q . cn) ˜ ˆ Delay ( D .GLOSSARY OF NOTATION Symbol a A B b C c D d E F f G g I L M N P p P Q q R S s T Meaning (Form of Usage) ˜ Arrival per slot (ai. bi(n)) ˜ Cycle length ( C ) Constant (c. L*(s)) Number of users in the system ˜ ˜ Population size ( N . L. U U n . pi. D(k)) Distance between assignments (Generalized TDMA) (d(k)) Expectation (E[ . Un) t U . length of CRI ( B . Tc) t General time ( ˜ . P ) Probabilities (Psuc. Sn(k)) Laplace variable Slot size Packet length (time) ˜ ˜ Transmission period ( T . B. ]) General function (usually distribution) (F( . N k ) ˜ Packet size (P. L(z). Ak(z)) ˜ Busy period. L 2 . I) ˜ ˜ Number of packets in a message ( L . q(k)) Channel transmission rate Throughput (S. t) ˜ ˜ Useful (successful) time in a cycle ( U . T. T i. T i . Bn.
)) General (bounding) number Step function (∆ ( . V(z)) Waiting time General variable (x. xl . endtoend propagation delay . )) CRI epoch length Collision detection time (CSMA/CD) Arrival rate General probability Invariant probabilities (πi). x ) Generating function variable CRI length bound (αm) Root of unity (βm) Impulse function (δ( .172 : GLOSSARY OF NOTATION Symbol V W x z α β δ ∆ γ λ ν π ρ σ τ Meaning (Form of Usage) Second moment of CRI length (Vn. xi. Probability vector load factor General probability Minislot duration. xr) ˜ General service time ( x .
This action might not be possible to undo. Are you sure you want to continue?