You are on page 1of 9

A DISTRIBUTED QUEUEING RANDOM ACCESS PROTOCOL FOR A BROADCAST CHANNEL

WenxinXu
Reuters Information Technology Inc.

Graham Campbell
D artment of Corn ttter Science
%noisInstituteo ?Te&nology

Abstract -- For decades there has been a search for a multiple area networks. In fact with zero propagation delay, collisions in
access protocol for a broadcast channel that would provide a per- slotted CSMA can be completely avoided and the performance of
formance that approached that of the ideal M/D/l queue. This ideal CSMA then corresponds to that of a perfect scheduling protocol,
performance would provide imrnedhte access at light loads end i.e. the M/D/l queue [5]. However, the CSMA protocols, like
then seamlessly move to a reservation system at high offered loads. ALOHA protocols, are not stable when traffic is heavy and whale
DQRAP (distributed queueing random access protocol) provides a dynamic control mechanisms can improve performance, the unsta-
performance which approaches this ideal. Furthermore it is accom- ble nature cannot be changed [5,15-16].
plished using as few as three control rninklots which suggests tha~
aside from establishing new theoretical bounds, DQRAP will be of The first CRA was proposed by Capetanakis in 1977 [7]. Capet-
great practical value. enakis showed that his tree algorithm [8] achieves a maximum
throughput of 0.43, and is stable for all input rates of less than
DQRAP requires that channel time be divided into slots each of 0.43. This stable characteristic of the tree algorithm has attracted
which consists of one data slot end m control minislots, and that much attention in both the communications and information theory
each station maintain two common distributed queues. One queue areas. Massey improved the tree algorithm and increased maxi-
is called the data transmission queue, or simply TQ, used to orga- mum throughput to 0.462 [9]. The next improvement, the 0.487
nize the order of data transmission, the other queue is the collision window protocol (algorithm) is due to Gallager [10], Tsybakov and
resolution queue, or simply RQ, which is used to resolve the colli- Mikhailov [13]. The tree end window protocols are based on effi-
sions end to prevent collisions by new arrivals. The protocol cient use of channel feedback to resolve collisions and require
includes data transmission rules, request transmission rules end transmitter coordination. Tsybakov and Likhanov showed that the
queuing discipline rules. Modelling and simulation indicate that upper bound of throughput of all algorithms based on ternary chan-
DQRAP, using es few as 3 minislots, achieves a performance level nel feedback is 0.568 [17], the tightest upper bound to date. It is
which approaches that of a hypothetical perfect scheduling proto- widely believed that the best achievable throughput is in the neigh-
col, i.e., the M/D/l system, with respect to throughput and delay. borhood of 0.5. Pippenger shows that if the amount of channel
DQRAP could prove useful in packet radio, satellite, broadband feedback is increased to indicate the number of packets involved in
cable, cellular voice, WAN, and passive optical networks. each collisiom then throughput up to one may be achieved [18].
However, the known algorithms in this context achieve only 0.533
throughput [19-21]. Some protocols achieve higher throughput
L INTRODUCTION than 0.5 by using control minislots (CMS) to obtain extra feedback
[22-26]. Among them the announced arrival random access proto-
cols [AARA] achieve the best performance with respect through-
Investigation of multiple, random access control methods has been
put and delay characteristics [26]. With three minislots the AARA
an active research area since the introduction of the ALOHA sys-
protocol achieve a throughput of 0.853. However to achieve
tem by Abremson in 1970 [1]. Roberts achieved twice the through-
throughput approaching one the AARA protocol must use an infi-
put of ALOHA with the development of Slotted ALOHA [2]. The
nite number of minislots. Obviously the AARA protocols do not
well known CSMA protocols were then developed and these were
achieve or approach the bound of performance in this context. We
then followed with multiple access methods which utilized various
observed that all existing tree protocols used data slots to resolve
forms of feedback to improve performance by reducing or avoid-
collisions. In this process channel capacity is lost either to empty
ing the occurrence of collisions. These include CSMA/CD,
slots or collisions. All suggested improvements to tree protocols
CSMA/CA [3-6], end collision resolution schemes, now called
improved the channel throughput by reducing empty slots and col-
tree-and-window collision resolution algorithms (CRA) [7-14].
lided slots, but none etiiinated this type of loss,
The CSMA protocols achieve high throughput with minimal delay
with low offered loads and have gained wide application in local
We introduce a new stable random multiple access protocol, the
Permission to copy without fee all or part of this material is distributed queueing random access protocol (DQRAP), for use in
granted provided that the copies are not mada or distributed for
a broadcast channel shared by an infinite number of bursty sta-
direct commercial advantage, the ACM ccpyright nctica and the
tions. DQRAP was originally inspired by DQDB (distributed
title of the publication and its date appaar, and notice is given
queueing dual bus, now the IEEE 802.6 standad for Metropolitan
that copying is by permission of the Association for Ccmputing
Machinery. To copy otherwise, cr to republish, requires a fae Area Networks) and its concept of network intelligence in the sta-
and/or specific permission. tions and distributed queues but is based on tree protocols with
SIGCOMM’93 - Ithaoa, N.Y,, USA /9/93
e 1993 ACM 0-89791 -619 -0193 [000910270 . ..$1 .50

270
minislots. These tree protocols use minislots to provide extra feed- from arrival time, first introduced by Gallager in [10]. We rephrase
back in order to reduce the number of empty and collided slots. it as follows. Let ;-1 represent the instant in the transmission axis
DQRAP, however, uses the rnirtk]ots for collision resolution and at which all messages arrived before instant of N. 1 in the arrival
reserves the data slots for data transmission. Implicitly, even axis have successfully resolved their confiicts (Fig. 2). The inter-
though counters, etc., are often used existing tree algorithms use a val (Al, ;-l) is called the waiting intend. The intend (Xi.l, xi) is
single queue which performs as a collision resolution queue. We called the enable transmission interval (ETI), which is determined
achieve the desired performance by introducing an additional from the following fornmk
queue, the data transmission queue, to schedule data transmission
parallel to contention
tion in the data slots. DQRAP,
resolution tlhus ahnost eliminating
using as few as three rninislots,
conten- x.
1
= x.
~_~+min{wo)ti_~–xi_~}
achieves a performance level which approaches that of a hypothet-
ical perfect scheduling protocol, i.e., the M/D/l system, with where W. is called the default window size which will be opti-

respect to throughput artd delay. mized by performance requirements. Obviously if the length of a
waiting interval is greater than the default window size, the ETI is
prnt of the waiting interval (see Fig. 2 (a)), otherwise the ETI is
The remain&r of the paper is organized as follows: Section II pro-
vides a complete channel model and description of the protocol, equal to the waiting window (see Fig. 2 (b)).

section III models the system and provides an evaluation of the


performance, section IV gives some numerical results and com-
x i-1 =i
pares DQRAF with the M/D/l syskm and the AARA protocols.
I I b
Wo
J Auival Axis
Il. CHANNEL MODEL, AND PROTOCOL

We consider a communication system serving an infinite number ~


of bursty stations which communicate over a multiaccess and Tmmnissinn Axis
noiseless broadcast channel. The stations generate single messages
of fixed length. The channel is divided into slots of fixed length. (a) WO <= ‘i.l - ‘i-l
Each slot consists of m control tninislots (CMS) followed by a sin-
gle data slot (DS) (Fig. 1). The size of a data slot is assumed to be
x x
of length 1, equal to the length of messages generated by each sta- i-1 1

tion. Each CMS is assumed to be of length of 6. The size of 5 is I I ➤


implementation dependent but 5 is assumed to be much smaller J
Wo Arrival Am
than the data slot, i.e, 6<<1. We take (1 + mi5) as the charmel time
unit (CU). Assume that the generation times of the messages form
a Poisson point process with intensity of k messages per unit time.
~
k is also called input rate. A station may trartsmit a message in the
data slot and/or a request in one of the control minislots. All sta- Tramnissinn Axis

tions can synchronize on both minislot and data slot boundwies


(b) Wo > Ii-1 - Xi.1
and all stations can detect ternary feedback information for each
trtinislot and data slot from the channel irrunedately after trans- Figure 2. ETI and CRI
mission. The assumption of immediate feedback is unrealistic,
however, the collision resolution algorithms can be moditied to
In DQRAF wllision resolution is based on the ETI. Only after all
accommodate delayed feedback [9].
messages in the current ETI have successfully resolved their con-
flicts, can the next ETf be initiated. If at instant ~ all messages in
the ETI (%.l, xi) have successfully resolved their conflicts, the
interval (~.l, ;) is called the contention resolution interval (CRI)

E=JP
corresponding to ETI (xi.l, xi). In DQRAF two distribute queues
are maintained by each statiorc the data transmission queue, or
simply TQ, and the collision resolution queue, or simply RQ.
TQ(t)l and lRQ(t)l represent the queue lengths of TQ and RQ at
instant t respectively. The term “transmit a request” means that a
station rolls an m-sided die and transmits a request signal in the
Data Slot Minislots
minislot so selected.

Figure 1. Slot Format


Let Fj, j = 1,2,...m, denote feedback horn the j-th minislot. Fj
belongs to the set of [E,S,C ), where E denotes an empty minislot,
S denotes the presence of a single request signal in a minislo~ and
The basic principle of the tree collision resolution algorithm is to
C denotes the presence of two or more request signals transmitted
resolve one initial collision before trying artother one. One trick in
in a single minislot.
the window-based tree protocol is to “decouple” transmission time

L/l
xl<
G

4
m

UI zr--
Om
t-co

II
1-

’88

II
W
.
X’a

II
In

‘a

Em
*

mw
&@

.“
ii
L(

Em
cc

“d ‘G

UI
c e-l

N
0, ‘a
:

c
.

Em UI
.

o 0

272
The protocol consists of three parts: data transmission rules multibits, ternary feedback and multiple minislots are used,
(DTR), request transmission rules (RTR), and queueing discipline
rules (QDR). FCFS (first come first scheduled) scheduling disci- An example (refer to Fig. 3) is now presented to describe the oper-
pline is used for both the TQ and the RQ but other scheduling di- ation of DQRAP. In this example the default window size is infi-
sciplines could be utilized. Basically the DTR, the RTR and the nite (W. = 00), i.e., the ETI is equal to the waiting interval. The
QDR answer the questions: (1) who can transmit data and when? time axis is divided into equal slots with length of one channel
(2) who can transmit requests and when? and (3) how does the unit. Above the time axis the contents of the CMS (control minii-
channel feedback affect the queues? lots) and DS (data slots) are shown. Below the time axis the con-
tents of the TQ and the RQ at each station me shown. The symbol
Data Transmission Rules (DTR) “*” denotes the arrival time of messages pl, p2, ... p10. In this
(1) If lTQ(t)l = O && lRQ(t)l = O) then the stations with messages example two minislots are used. Assume at t = O that both the TQ
which have arrived in the current ETI transmit messages in the and the RQ are empty. At t = 1, pl and p2 each transmit both
data slot at (t). requests and messages. At t = 2 the feedback shows that the pl and
(2) If (lTQ(t)l > O) then the station which owns the first entry in the p2 data messages have collided but not their requests. p2 and pl
TQ transmits its message in the data slot at (t). go into the TQ and p2 data is transmitted at t = 2. Meanwhile p3,
arriving in interval [1, 2), transmits a request, but no data since
Request lkansmiaaion Rules (RTR) ~Q(2)l >0. p3 enters the TQ as p2 leaves. While pl and p3 are
(1) If (lRQ(t)l = O) then the stations with messages which have waiting their turn to transmit data, p4, p5, aod p6 transmit requests
arrived in the current ETI transmit requests at (t). at t = 3. p6’s request is ok and p6 enters TQ but p4 and p5 collide
(2) If lRQ(t)l >0 then the stations which “own” the first entry in and thus enter the RQ. p4 and p5 collide at t = 4 on their first try to
the RQ transmit requests at (t). resolve the collision but on the next attempt at t = 5 they succeed
and enter the TQ, their order determined by their relative positions
Queueing Discipline Rules (QDR). in the minklots. p6 transmits at t = 5 since the TQ operates inde-
At time (t), using data slot or rninislot feedback pendently of the RQ. The RQ is empty at t = 6 thus p7, p8 and p9,
(1) Each station increments TQ(t)l fc,r each (Fj (j = 1,...m) =S). which had rurived in the interval [3, 5) and could not transmit
(2) Each station decrements lTQ(t)l by one at (t) for a successful requests or data join p10 at t = 6 in making their first try to trans-
message transmission commencing at (t-1). mit. p8 and p9 collide in the first minislot while p7 and p10 collide
(3) If lRQ(t)l = O each station increments lRQ(t)l by n where n is in the second minislot. Thii determines their order in RQ. The pro-
the number of collisions C in Fj, j=l,...m. cess continues.
(4) If lRQ(t)l >0 each station modifies lRQ(t)l by (n-1) where n is
the number of collisions, C, in Fj, j=l,...,m. Ill. MODELING AND EVALUATION
(5) The stations which transmit successful requests or collided
requests know their positions in the TQ or the RQ and adjust
A. Modeling of DQRAP
their pointers to the TQ or the RQ accordingly.

DQRAP can be modeled as a queueing system that consists of two


Using the rules presented above DQRAP can be described by the
subsystems: (1) a queueing contention resolution subsystem
following algorithm
(QCR), and (2) a data transmission (DT) subsystem [Fig. 4]. This
model is used to evaluate the throughput of DQRAP. DTR(l) is
Set (t) = O, lTQ(t)l = O, and lRQ(t)l = Q
not considered, because it does not affect system throughput, as
While (TRUE)
pointed out previously. The servers of the QCR subsystem are
( minislo~. The QCR subsystem can be evaluated using Markov
(l)t=t+l
chain theory. The DT subsystem can be modeled as a G/D/l
(2) transmit data obeying the DTR;
queue, the server is the data slot, the service time is one slot per
(3) transmit(s) request(s) obeying the RTR;
message.
(4) All stations modKy their counters of the TQ and the RQ
and their pointers to the TQ or the RQ following the
QDR. I 4
I 1“
) L

DTR (1) is important since it preserves the immediate access fea-


<
ture of random multiple access communications and distinguishes
DQRAP from reservation protocols. [t is emphasized that DTR(l) ETI RQ TQ
may permit a collision to occur in the data S1OL but without
DTR(I) the data slot would otherwise be empty. DTR(I) improves
the delay characteristics of the protocol, especially when the input Figure 4. Modeling of the DQRAP Protocol
rate is low.

The algorithm to resolve queueing contention in IQRAF is simi- B. CRI Length


lar to the mukibit feedback algorithm (MFA) reported by Husng
and Berger [23]. However, instead of using binary feedback and Analysis of the subsystems first requires the calculation of the

273
expected length of the CRI, LtL defined as the period commencing Table 2: Maximum Input Rate and the Corresponding Window
with the time slot containing the initial queueing contention (if Size as a Function of the Minislot Number
any) and ending with the slot in which the initial queueing conten-
tion is resolved. n represents the number of stations involved in the
initial queueing contention and is crdled the multiplicity of CRI in
the contention resolution algorithm literature. For consistence a

successful transmission is detined as a conflict of multiplicity 1


whale an empty ETI is defined as a confllct of multiplicity O. We
have the following formula for calculating LXX
I 3
I 1.2400
I 2.794
I 11
I 2.6063
I
2.483
I
Let Ln be the expeeted length of CRI with multiplicity n. 4 1.5156 2.835 12 2.7171 2.442
Then,
5 1.7353 2.799 13 2.8226 2.425
f-o=%=] (1)
6 1.9207 2.726 14 2.9234 2.409

7 2.0834 2.670 15 3.0201 2.376

L2=~ (2) 8 2.2299 2.611 16 3.1133 2.363


m –1
9 2.3642 2.552
and

Here m is the number of minislots which is chosen by performance


n –1
requirements. The proof of formula (l), (2) and (3) can be found in
n (m-l) n-kLk the Appendm. Table 1 contains values of Ln as obtained from for-
~ (k)
mulae (1), (2), and (3) with different values of m. From Table 1 we
can see that when m=3, Ln e n for D 1. Tids means a collision of
1+ ’=2
n –1 multiplicity of n can be resolved in less than n slots, which is the
m
Ln = , XD2 time to transmit n messages. Thus the speed of contention resolu-
1
l–— tion is faster than the speed of data transmission, a very important
n -1 result.
m
(3)
C. Stability Conditions of DQRAP.

DQRAP is stable if and only if both the QCR subsystem and the
Table 1: CRI Length - ~ DT subsystem are stable. Stability conditions of the QCR sub-
system can be determined by using Markov chain theory. There is
a standwd result for the window-based tree algorithm in the tree
algorithm literature. That is, the maximum stable input rate, or
throughput, can be determined by the following formuIae[14, 23]:

P
c = sup ~ (4)
1 I 1.000 ] 1.000 I 1.000 I 1.000 I 1.000
n
2 I 2.000 I 1.500 I 1.333 I 1.142 I 1.066 Ln$ e–P
x n.
3 I 3.333 I 2.250 I 1.866 I 1.396 I 1.192 o

4 4.761 3.115 2.514 1.736 1.369 where p.LWo.

5 6.209 4.026 3.222 2.139 1.591


The maximum input rates and the corresponding window sizes as

6 I 7.656 I 4.951 I 3.958 I 2.590 I 1.853 a function of minislot number are listed in Table 2. Table 2 shows
that if m >2 the QCR subsystem is stable even when the input rate
7 I 9.100 I 5.874 I 4.703 I 3.074 I 2.149 is greater ~~ 1. Next consider fie DT subsystem. The DT sub-
system can generally be modeled as a G/D/l queue. Though the
QCR subsystem can be stable with the input rate greater than 1, G/
D/l is stable only when the input rate is less than 1. Thus DQRAP
is stable when the traffic intensity is less than 1, The QCR sub-
system can resolve collisions faster than the speed of data trans-
mission thus guaranteeing that the QCR subsystem will not block
input traffic to the whole system.

274
D. Performance of DQRAP

The performance of DQRAP is detemdned by the QCR subsystem


— m.s
and the DT subsystem. The QCR subsystem does not affect data
........ m=g
transmission, and is stable even when the traffic intensity is greater
than 1 if three or more minislots are utilized, Thtts since the QCR — m.lfj
-
subsystem does not block trrdlic to the whole system, the system
throughput
DQRAP
is entirely
can achieve a maximum
determined by
theoretical
the DT subsystem,
throughput approach-
i.e.
‘-<;”’”’’%.... ““’%
...,,...
..
ing one if three or more tninislots are utilized. When the mittislot “%
‘-’%
overhead is includ~ the actual throughput (or utilization) that can --%.%
be achieved is

min{k, l}
u = —— m
I+mb

Figure 5 shows the throughput of DQRAP as a function of the


input rate and the number of minislots with the overhead equal to t-
0.01. Figure 6 shows the relatiortahilp of the throughput and the 0.05 0.1
number of tninislots. Obviously high actual throughput requires One Minislot Overhead
that the number of minislotis selected should be as small as possi-
ble. Fortunately evaluation shows that with as few as three mirtis- Figure 6.. The Maximum Actual Throughput As a Function
lots DQRAP achieves a maximum theoretical throughput of one. of One Minislot Overhead and Minislot Number
The analytical solution of delay characteristics for DQRAP is ds-
cussed in [30]. Here art accurate simulation has been used to
obtain the delay performance of DQRAP and this performance
may be evaluated by comparing it to a perfect scheduling protocol.
The simulation results are pres&tted til-Section IV. -

— M/D/l

I
o 0.5 1.0

— M/D/l — DQRAP

o 0.5 -1.0 Figure 5. The Percentage of First Access Throughput of


the DQRAP Protocol As a Fuction of Input Rate
Input Rae

Figure 5. The Actual Throughput of the DQRAP Protocol


As a Function of the Input Rate and Minislot Number The performance bound of all random access protocols for a slot-
ted broadcast channel shared by an infilte number of Poisson
sources is that of a hypothetical perfect scheduling protocol, i.e.
IV. SOME NUMERICAL RESULTS AND COMPARISONS the M/D/l system. Thus the performance of DQRAP is best dem-
onstrated by comparison with that of the M/D/l system. Figure 7
The simulations, based upon the algorithm described in Section II shows the ratio of the first access throughpu~ which is defined as
rather than the above model have been carried out. The simula- the ratio of messages successfully transmitted in the first slot after
tions show that DQRAP demonstrates good system stabitity, in their anival to system throughput of DQRAP, as a function of the
particular all messages are guaranteed to be transmitted with a lim- input rate, using three mirtislota as compared to the M/D/l system.
ited delay for all input rates less than or equal to 0.99. This is con- Table 3 contains average delay and corresponding deviation of
sistent with the system stability analysis presented in the previous DQRAP (three miniilots are used) as compared with the M/D/l
section. system. Table 3 shows that the average delay of DQRAP is very

275
close to the average delay of the M/D/l system, and the maximum Table 4: Comparisons of Average Delay and Deviation of
difference of average delays between the M/D/l system and DQRAP with Varying Minislot Number
DQRAP is less than three slots when the input rate is less than
0.95. Table 4 lists simulated average delay and deviation of
DQRAP
DQRAP with a varying number of minislots. Figure 8 plots simu-
Nominal
lation results showing the average delay of DQRAP along with m=3 m=4 m.8 m=16
Input
that of an M/D/l system. Table 4 shows that increasing the number
of mirtislots
and even though
minislots
does not impact the maximum
the average delay is affected
it appears that for most practical
theoretical throughput
by the number
purposes the number of
of
AVGJ
STD I AVGJ
STD

1.6761 1.6666
minislots need not be greater than four. Finally we compare
0.6298 0.5990
DQRAP with the best tree protocols with minislots, namely, the
announced arrival random access protocols (AARA). To achieve a
theoretical
protocols
throughput
require
approaching
an infinite number
one the announced
of tninislots,
arrival tree
but DQRAP
0.20
I 1.9661
1.0459 I 1.9218
0.9433
1.8747
0.9938
-t-
1.8567
0.7957

requires
announced
DQRAP
as few
arrival
provides
as three
tree protocols
minislots.

better performance
Using
achieve
three
a throughput
minislots
of 0.853.
than the best tree protocols
the 0.30
I 2.2533
1.2672 I 2.1786
1.1326 I
2.0989
0.9938 I
2.0715
09490 I
to date. 0.40 2.5867 2.4699 2.3604 2.3186
1.4663 1.3012 1.1457 1.0926

Table 3: Comparisons
Between DQRAP
of the Average
and an M/D/l
Delay and Deviation
System (the number of
0.50
I 2.9838
1.6732 I 2.8097
1.4639 I 2.6564
1.2899 I 2.6052
1.2364 I
0.60 3.4895 3.2445 3.0400 2.9795
mhtislota = 3)
1.9453 1.6836 1.4961 1.4434

Protocol 0.70 4.1923 3.8413 3.5808 3.5015


Nominal 2.3431 2.0288 1.8361 1.7911
Input M/D/l M/D/l DQRAP
Rate Analyzed Simulated Simulated 0.80 5.3407 4.8690 4.5353 4.4367
3.0835 2.7636 2.5939 2.5491
AVG AVG/STD AVG/STD
0.90 8.2555 7.5451 7.1088 7.0018
0.10 1.5556 1.5549 1.7152
5.3156 5.0126 4.8637 4.8503
0.3518 0.7617
0,95 13.5251 12.5975 12.1022 11.9715
0.20 1.6250 1.6239 1.9661
9.9712 9.6747 9.5977 9.5731
0.4273 1.0459

0.30 1.7143 1.7137 2.2533


0.5229 1.2672 V. CONCLUSION

0.40
I 1.8333
I 1.8365
0.6518 I 2.5867
1.4663 I DQRAP
performance
is a medium
with
access control
respect to
method
throughput
which
and
can provide
delay which
0.50 2.0000 2.0002 2.9838 approaches that of a perfect scheduling protocol. DQRAP is stable

0.8184 1.6732 at all input rates of less than 1 when three or more minislots are
utilized. The current version of DQRAP requires that stations
0.60 2.2500 2.2543 3.4895 monitor channel feedback continuously, even when idle. This
1.0805 1.9453 requirement is removed in the free access DQRAP ~otocol [28].
Other issues which must still be addressed include (1) the robust-
0.70
I 2.6667
I 2.6746
1.5057 I 4.1923
2.3431 I ness of DQRAP, (2) the applicability to different channel models.
The analytical solution of delay characteristics for DQRAP is dis-
cussed in [30].
0.80 3.5000 3.5113 5.3407
2.3585 3.0836
DQRAP can be implemented by overcoming the usual problems
0.90 6.0000 5.9918 8.2555 attendant with any medium access control method. The major
4.7571 5.3156 challenge is obtaining ternary feedback but it appears that this is
feasible in broadband signaling over copper, fiber, and air and
0.95 11.0000 10.7504 13.5251 with baseband signaling on copper and fiber, This paper assumes
9.4069 9.9712 instantaneous feedback but there are techniques for compensating
for propagation delay [9][29]. DQRAP was developed m part of a
research project at IIT which has as its major objective the utiliza-

276
tion of CAN (cable TV) facilities for the transmission of digital Proof. It is trivial that
data. However, follow-on research indicates that DQRAP can be
used in metropolitan area packet rdo, digital cellular, passive LO= L1=l

opticrd, wide area, and satellite networks as well as the originally


Next, consider the case n= 2. If two signals transmit onto the same
targeted broadband cable environment.
slot after an initial collision, they collide again, otherwise the coHi-
sion is resolved. We have,

L2 n 1+ (~)mL2 (A4)
m2

Solving for L2, we have (A2).

Now consider the case for n >2. After the initial collisio~ the
QDR subdNides the conflicting set with multiplicity of N = n into
m subsets, each of which has multiplicity of Nj, j = 1,2,..,m. If Nj
>= 2, the j-th subset will be subdivided further in the same way.

o k<=l

“k = ( (A5)
Lk k>l

Then Ln can be expressed as

0 ~~ mn
0.0 0.:5 1.0
Ln=l+ Pr{Nj=klN=n) L’k (A6)
xx
Input Rate J“=lk=O

Nj is set of identically d~tributed random variables, but not inde-


Figure 8. Delay Characteristics of the DQRAP Protccol
pendent. (A6) can be written as

APPENDIX L el+m Pr {Nj=klN=n) L’ k (A7)


n z
Derivation of Ln kO=

Theorem 1. Let Ln be the expected length of CRI with multiplicity n


n. Them
=l+m Pr {Nj=klN=n] Lk (A8)
x
k2 =
Lo=%=l (Al)

Note that

and
L2 = ~’:
m .-
(M)
Pr(Nj=klN=n) a
(k) n (~)
k
(1 – #
1 ‘–k (A9)

n -1

~ 0
n (m-l) n-kLk
.
0 n (m-l)

mn
n-k
(A1O)

~+k=2
mrt -1
Ln = . IE-2
1
1 . ——
n -1 (A3)
m

277
We have, [14] R. G. Gallager, “A Perspective on multiaccess channels,”
IEEE Trans. on Information Theory, vol. IT-31, no. 2, pp.

[)
n 124-142, March 1985.
n (m-’) *-k ~ ~A1l) [15] F. A. Tobagi and L. Kleirtrock, “Packet switching in radio
L= l+m
n x n k channels: Part IV -- Stability considerations and dynamic
m
control in carrier sense multiple access,” IEEE Trans. on
k=2
Communications, COM-25, no. 10, pp. 1103-1120, Oct.,
1977.
Solving (All) yields (A3).
[16] G. Fayolle, E. Gelembe, and J. Labetoulee, “ Stability and
optimal control of the packet-switching broadcast channels,”
J. Ass. Comput. Mach., vol. 24, pp. 357-386, July 1977.
REFERENCES
[17] B. S. Tsyb&ov, and N. B. Ltianov, “ Upper bowd on tie
capacity of a random multiple access system, “ Problems of
[1] N. Abramson, “The ALOHA system -- Another alternative for
Information Transmissio~ vol. 23, no. 3, pp. 224-236, July -
computer communication,” in AFIPS Conf. Proc. Fall Joint
Sept. 1987. Translated horn Prolemy Peredachi Inforrnatsii,
Comput. Conf. 1970, pp. 281-285.
vol. 23, no. 3, pp. 64-78, July - Sept, 1987.
[2] L. G. Roberts, “ALOHA packet system with and without slots
[18] N. Pippenger, “Bounds on performance of protocols for a
and capture,” ARPA Network Inform. Center, Stanford Inst.,
multiple-access broadcast channel,” IEEE Trans. on Informa-
Menlo Park Calif., ASS Note 8 (NIC 11290), June 1972. Also
tion Theory, vol. lT-27, no. 2, pp. 145-151, March 1981.
see Computer Communications Review, vol. 5, pp. 28-42,
[19] L. Georgiadis and P. Papantoni-KazaJcos, “A collision resolu-
Apr. 1975.
tion protocol for random access channel with energy detec-
[3] L. Kleinrock and F. A. Tobagi, “Packet switching in radio
tors, “IEEE Trans. Cornrnun., vol. COM-30, no.11, pp. 2413-
channel: Part I -- Carrier sense multiple-access modes and
2420, NOV. 1982.
their throughput-delay characteristics,” IEEE Trans. on Com-
[20] B. S. Tsybakov, ‘CResolution of conflict with known multi-
munications, vol. COM-23, no. 12, pp. 1400-1416, Dec.
plicity, “Problem of Information Transmission, vol. 16, no. 2,
1975.
pp. Apr.-June, 1980. Translated from Prolemy Peredachi
[4] F. A. Tobagi and V. B. Hunt, “Performance Analysis of carrier
Inforrnatsii, vol. 16, no. 2, pp. 69-82, Apr. -June., 1980, Mos-
sense multiple access with collision detection,” in Proc. Local
cow, USSR.
area comrnun. network symp. , Boston MA. May 1979.
[21] E. Gulko, “ Tree based mukiaccess where collision multiplic-
[5] F. A. Tobagi, “Multiaccess protocols in packet communication
ities are knowrL” IEEE Trans. Commun. vol. COM-33, no. 9,
systems,” IEEE Trarts. on Cornmttn., COM-28, no. 4, pp.
pp. 999-1001, Sept., 1985.
468-488, April, 1980.
[22] J. Huang and B. Xu, “Pretested random ~ee protocol for
[6] M. K. Molloy and L. Kleinroc~ “Viitual time CSMA: Why
packet conllict resolutio~” IEEE Trans. on Information The-
two clocks are better than one,” IEEE Trans. Comrnun. vol.
ory, vol. lT-29, no. 6, pp. 910-913, June 1983.
COM-33, no. 9, pp. 919-933, Sept. 1985.
[23] J. C. Huaog and Berger, “Delay analysis of interval searching
[7] J. I. Capetanakis, ” The mukiple access broadcast channe~ pro-
contention resolution algorithm,” IEEE Trans. on Information
tocol and capacity considerations,” Ph.D. Dissertatio~ Mass.
Theory vol. IT-31, no. 2, pp. 264-273, March 1985.
Inst. Tech. Cambridge, MA, Aug. 1977,
[24] L. Merakos and D. Kazakos, “Multiaccess of a slotted chan-
[8] J. I. Capetankis, ” Tree algorithm for packet broadcasting chan-
nel using a control mini-slo~” in Proc. Int. Conf. Commun.,
nel,” IEEE Trans. Inform. Theory, vol. lT-25, pp. 505-515,
1983. pp. C5.3.1 - C5.3.6.
Sept. 1979.
[25] D. Raychaudhuri, “Amotmced re@msmission random access
[9] J. L. Massey, “Collision resolution algorithm and random
protocols,” IEEE Trans. on Communications, COM-33, no.
access comrnrnunications,” in Multiuser Communication
11, pp. 1183 -1190, Nov. 1985.
Systems, G. Longo Ed. New York: Sprirtger-Verlag, 1981. pp.
[26] T. Towsley and P. O. Vales, “Announced arrival raudom
73-137.
access protocols,” IEEE Trans. Cornmun. vol. COM-35, no.
[10] R. G. Gallager, “ Conilict resolution in random access broad-
5, pp. 513-521, May 1987.
cast networks,” in Proc. AFOSR Workshop Commun. Theory
[27] L. Kleinrock, Queueing Systems, volume 1: Theory. New
Appl. Provincetown, MA. Sept. 1978.
York Wiley, 1975.
[11] J. Mosely and P. Httmble~ “A class of efficient contention res-
[28] W. Xu, “Distributed Queueing Random Access Protocols for
olution ~gorithtns for multiple access channels,” IEEE Trans.
a Broadcast Channel,” Ph. D. Dissertatio~ Illinois Institute of
on Cornnmn. vol. COM-33, no. 2, pp. 145-151, Feb. 1985.
Technology, 1990.
[12] TSYB 78 B. S. Tsybakov and V. A. Mikhailov, “Free syn-
[29] S. Sirazi, “TrlInet - A Demand-Adaptive Protocol Mediurn-
chronous packet access in a broadcast channel with feed-
Access Protocol for Metropolitan Area Networks.” Ph.D.
back,” Problems of Information Transmission, vol. 14, no. 4,
Dissertation, Illinois Institute of Technology, 1986.
pp. 259-280 Oct. - Dec. 1978. Translated from Prolemy Pere-
[30] Xianyu Zhang and Graham Campbell, “Performance Analysis
dachi Inforrnatsii, vol. 14, no. 4, pp. 32-59, Ott,-Dec. 1978.
of Distributed Queueing Random Access Protocol --
[13] B. S. Tsybakov and V. A. Mikhailov, “ Random multiple
DQRAP”, Internal Report, Computer Science Dept., Illinois
packet access: part-and-try algorithm;’ Problem of Irtforma-
Institute of Technology, 1993.
tion Transmission, vol. 16, no. 4, pp. 305-317 Oct.-Dee.
1980. Translated from Prolemy Peredachi Informatsii, vol.
16, no. 4, pp. 65-79, Oct.-Dee. 1980.

278

You might also like