Professional Documents
Culture Documents
Thesis Report Vijayalashmi
Thesis Report Vijayalashmi
Certified that this thesis titled “Bandwidth Estimation and Analysis of Multi-hop Ad
hoc Networks” is a bonafide work of Ms. K.Vijayalakshmi who carried out the
research under my supervision. Certified further, that to the best of my knowledge, the
work reported herein does not form part of any other thesis or dissertation on the basis
of which a degree or award was conferred on an earlier occasion on this or any other
candidate.
Dr.S.Srikanth,
Member Research Staff,
AU-KBC Research Centre,
Date : 4/07/2005 MIT Campus, Anna University,
Place : Chennai. Chromepet, Chennai,
TN – 600 044 – India
1
ABSTRACT
Medium access control protocols and its effect on capacity are important
aspects in communication system design for any shared medium like wireless. In
multi-hop ad hoc networks that use a distributed and contention based channel access
mechanism such as those specified in the IEEE 802.11 standard, the capacity of
individual links are not known. Existing methods in literature attempt to solve this
using measurement based approaches. This thesis proposes graph theoretical and
real-time approaches to estimate the capacities of individual links in a multi-hop ad
hoc network and uses analytical modeling to derive node throughputs and successful
transmission probabilities of individual nodes in multi-hop ad hoc networks.
In a contention based system based on the IEEE 802.11 standard, the capacity of
links depends on the time available for the links to be active which in turn depends on
the probability of nodes’ transmission. While the link capacity can be expressed by the
number of times it gets activated, it can be more accurately expressed in probabilistic
terms. This is due to the randomness introduced in the channel access procedure in the
form of virtual carrier sensing and binary exponential backoff. This thesis proposes
centralized and distributed methods to estimate the active time of links. The
centralized approaches have the knowledge of the entire network topology and use
graph theoretic approaches to derive individual link active times. The distributed
approaches on the other hand only have the partial knowledge of network topology to
derive the link active times. This thesis proposes another approach to the link capacity
problem using the analytical modeling of multi-hop network for string and grid
topologies. The analytical modeling uses Markov models to derive node throughputs
and successful transmission probabilities of individual nodes in multi-hop ad hoc
2
networks. The advantages of the estimation methods are topology independent nature
of the solution, a wide spectrum of applications using the estimates and real-time
applications using the distributed approach. It also provides insight into the
performance evaluation of multi-hop networks. The accuracy of the estimations is
validated through simulations and it is observed that there is a difference between the
estimated and simulated values only of the order of 10e-02. The values obtained
through analytical model too matches the values obtained in simulations to a good
degree.
3
ACKNOWLEDGEMENTS
for giving me a free hand in research along with continued guidance and support
supervisor Dr.V.Vaidehi for the encouragement and valuable guidance that motivated
I’m greatly indebted to Dr.C.N.Krishnan, the Director of AU- KBC Research centre
for being responsible for the inspirational department . His dedication to work and
idealogy are sure fires for young minds. My immense thanks to Dr.S.V.Ramanan for
the short but most effective periods of interaction that gave the much needed impetus
support and useful suggestions throughout the program. Special thanks to KRK for the
support and valuable suggestions during the last leg of the programme.
Immense gratitude to all my wonderful friends and specially to friend and guide
Rajesh for the incredible support throughout. The thesis would not have seen the light
of the day had it not been for the constant care and blessings of my parents and loving
support of my sister. Finally my salutations to God for being with me all through.
4
TABLE OF CONTENTS
CHAPTER 1.......................................................................................................9
1.1 OVERVIEW......................................................................................................9
1.4 ASSUMPTIONS..............................................................................................12
CHAPTER 2.....................................................................................................15
2.1 OVERVIEW....................................................................................................15
5
CHAPTER 3.....................................................................................................33
3.1 MOTIVATION................................................................................................33
CHAPTER 4.....................................................................................................45
6
LIST OF TABLES
7
LIST OF FIGURES
8
CHAPTER 1
INTRODUCTION
1.1 OVERVIEW
A multi-hop ad hoc network is a self-organized system that can comprise many mobile
nodes connected without a pre-determined topology or central control. It provides
quick and easy networking in circumstances that require un-tethered connectivity or
temporary network services. It finds widespread applications in scenarios such as
hospitals, search and rescue operation, battle fields and disaster sites. Hass et al
(1992) outline some of the applications. With the growing popularity of applications, it
has become necessary to provide solutions for efficient data or multimedia
communication over the network. Link bandwidth estimation is an essential
component of rate control and Quality of Service (QoS) support schemes like
admission control, resource reservation and QoS routing. The IEEE 802.11 standard
based Medium access control (MAC) detailed in the IEEE standard (1999) is the
commonly used channel access scheme for ad hoc networks.
9
problem, which is an important aspect of multi-hop networks. Chhaya and Gupta
(1997) analyzed the effect of capture and hidden nodes. Li and Blake (2001) studied
the capacities of multi-hop networks for standard topologies and random traffic
pattern. The analysis however is for the entire network and not on a per link basis.
Liaw et al (2004) proposed a method to estimate the throughput available to a node
based on local measurements and neighbor information and channel occupancy of the
node. However the approach is traffic dependent.
The efficiency of the IEEE 802.11 protocol directly affects the utilization of the
channel capacity and system performance. Performance evaluation of single-hop ad
hoc networks using IEEE 802.11 MAC is done in Bianchi (2000), Carvalho and
Aceves (2003), Li et al (2003), and Liwa et al (2004) In the case of multi-hop ad hoc
networks using DCF protocol, performance analysis needs consideration of many
factors. Much work has gone into studying the interaction between higher layers and
IEEE 802.11 MAC in multi-hop networks. Xu and Saadawi (2003) brought out the
problems in fairness and throughput variations when TCP is used with 802.11 MAC.
Fairness issues and enhancement of MAC are studied by Tang and Gerla (1999) and
Bensaou et al (2000). Eladly and Chen (2003) extended the saturation throughput
model of Bianchi (2000) to the case of multiple overlapping BSSs. However a
10
comprehensive analysis of performance of the IEEE 802.11 based multi-hop networks
is still an ongoing research work. Wang and Aceves (2004) analyzed the performance
of CSMA/CA based multi-hop networks using different Markov models for channel
and node. Yawen and Biaz (2005) combined the Bianchi (2000) model to get a three
dimensional model to analyze the performance of a multi-hop network under different
traffic loads and network densities.
In this work, we address the problem of estimating the active time of links in a multi-
hop network and give a general framework for such estimations. A general solution
that works for any topology is proposed. The analytical model for IEEE 802.11
provides a much needed insight into performance modeling of multi-hop networks.
Multi-hop ad hoc networks are gaining attention due its vast application potential. For
multimedia applications, it is necessary that Quality of Service (QoS) schemes are in
place. An essential component of QoS and traffic management schemes is the
knowledge of link capacity information. Most work in the literature use measurement
based approaches to get the values of link capacities. This thesis addresses the problem
of estimating link capacities using a number of approaches. Another way of addressing
is the issue is by modeling the network. This helps in performance evaluation and
getting probabilistic estimates of capacities. This thesis addresses it by the modeling
the IEEE 802.11 based multi-hop networks for string and grid topologies.
11
1.4 ASSUMPTIONS
The following assumptions have been made regarding the problem in hand
while arriving at the solution. These assumptions define the scope or the boundaries of
the solution formulated.
Each node can transmits to other nodes in its transmission range ‘d’. We
assume that all nodes employ a common range ‘d’ for their transmission. Let X i,
1 �i �n denote the location of node i. When node i transmits to node j directly, over
the channel, this transmission is received successfully by j if,
X i - X j �d
X m - X j �(1 + k )d
where . is the Euclidean norm. The quantity k>d models the interference range of the
nodes.
12
Traffic at every node is assumed to be saturated and hence a node always has a packet
to transmit. And the destination is assumed chosen randomly from one of its
neighbors.
Distributed approaches with very low complexity algorithms to estimate the link
capacities are outlined. They have the advantage of being used in real-time
applications.
Analytical model for multi-hop networks are obtained that study individual node
performance and provides insight into the modeling of multi-hop networks.
Chapter 2 details centralized and distributed approaches used for estimating the link
capacities. It discusses the approaches and algorithms involved and their complexity.
Chapter 3 deals with analytical modeling of IEEE 802.11 based multi-hop networks in
string and grid topologies.
13
Chapter 4 discusses the results of the estimation schemes, compares the schemes with
simulation. It also discusses the analytical results and validates them with simulation.
Chapter 5 concludes the thesis discussing the benefits of the solution schemes,
improvements possible in them and future scope of the work.
14
CHAPTER 2
2.1 OVERVIEW
The IEEE 802.11 standard (1999) specifies MAC protocols for operation in ad hoc
mode and centrally coordinated mode. The distributed coordination function (DCF) is
used for operation in the ad hoc mode and of interest to our work.
The DCF is based on carrier sense multiple access with collision avoidance
(CSMA/CA). The standard defines the common duration for carrier sensing by the
Clear Channel Assessment (CCA) function or DCF Inter Frame Space (DIFS) at every
15
node and also the contention window ranges (CWmin, CWmax). The time gap
between two packets in the handshake sequence or an atomic operation, defined as
Short Inter Frame Space (SIFS), is less than DIFS. This prevents other nodes from
capturing the channel when one transmission (in the same area) is already going on.
All nodes that can hear each other and agree to join and form an Independent Basic
Service Set (IBSS). These nodes remain in synchrony with a special management
frame called the ‘Beacon frame’. This frame is periodically generated by one of the
nodes in the IBSS. Once a node captures the channel, it can send data packet of at
most one MAC Service Data Unit (MSDU) which might be broken into fragments or
MAC Protocol Data Units (MPDUs) as specified by the variable Fragmentation
Threshold. A four-way handshake is optionally used.
A CK 1
ACK 2
M PD U
M PD U
BU SY
FRE E
RTS
CTS
(1)
(2)
t ime
S1
t ime
S2
t ime
S3
-------NAV------ ----NAV---
D IFS
SIFS
SIFS
SIFS
SIFS
SIFS
16
A node willing to transmit performs carrier sensing
If it finds the channel free for a DIFS duration, it may transmit.
To avoid collision situation, it backs off in random multiples of a constant time
(Slot-time) chosen from the contention window (CWmin and CWmax).
After the backoff countdown, a request to send (RTS) frame is sent by the
source node.
If the destination node ( S2 in Fig 1 ) is ready to receive then it initiates a clear
to send (CTS) frame within a SIFS duration.
The source node initiates the sending of the Data frame (MPDU or fragment)
within a SIFS duration.
If the destination node properly receives the MPDU it responds with an
acknowledgement (ACK) packet.
In case of failure the source node backs-off again with a random number chosen from
a doubled contention window. Absence of ACK is treated as collision in which case,
the node has to carrier sense for an extended IFS (EIFS).
The duration for which the transmission would go on is continuously updated through
the network allocation vector (NAV) field in the handshake frames. The NAV
information is used by other nodes (like S3 in Fig 1 ) in the network to stay away from
transmitting. When one transmission is going on the same area, another transmission
will lead to collision. This combined with the handshake helps overcome the hidden
node problem to a good extent but creates more number of exposed nodes. MSDUs in
MAC queue for duration longer than a (configurable) constant 'maxMSDULifetime'
are dropped from the queue. In addition to this a retransmit limit can also be fixed.
For handling data packets a simple First Come First Serve (FCFS) queue is used by all
the nodes. This protocol allows fair-channel access on the long-run, given the traffic
17
generation characteristics and channel condition due to environmental changes are
even across all nodes in the network. This doesn't differentiate nodes or traffic in
anyway and hence it is suitable for best-effort type of traffic only.
The network is represented as a graph G =(N,L) with vertices and edges in our model.
The nodes in the network correspond to the vertices and the links correspond to the
edges. A wireless link can be defined as follows. A link L is assumed to exist between
two nodes u and v if they are within the transmission range of each other. This is the
same as the first condition for successful transmission given in chapter 1. This link
assumption is also used by Kodialam and Nandagopal (2003). Some of the work in
literature consider a link to exist between two nodes only when there is a flow through
the link. In this work, a wirless link is assumed to exist between between two nodes if
the nodes are within the transmission range of each other. This is appropriate for
deriving bounds on link capacity as this doesn’t make any assumption about existence
or non-existence of flows which can change with time while our approach gives a
general solution based on network connectivity.
The link activation constraints can be represented using graphs. There are two ways of
representing the link activation constraints (contention between links) in multi-hop
networks. One is based on the set of links that can be activated simultaneously, Link
Independent sets and the other, based on the set of links that cannot be active
simultaneously due to mutual interference, the Interference graph. In our centralized
18
approach, both the models are used to obtain the active time of links. A sample
network with string topology is used to illustrate these representations by a graph.
Consider the graph of string network. Each node is separated from its neighbor by a
distance of 250m. We consider a transmission range of 250m and an interference range
of 500m.
Consider the given graph G = (N,L) of the network. To represent the set of links that
cannot be active simultaneously, a link graph F = (Nf, Lf) is constructed from the given
graph G. Here Nf denotes the set of all links in graph G and L f respresents the edge
set of F where an edge between any two links i, j �Nf exists if i and j interfere with
each other’s transmission. For the sake of simplicity and ease of analysis, the
condition for interference can be denoted in terms of link distance. In an edge graph,
link distance can be specified in terms of no of hops between links, which are edges in
the graph.
19
The above figure represents the interference graph constructed from the string
network. It is an edge graph of G where there is an edge between links if they were
either adjacent or have a link distance of m. In this case, m is considered as 2.
In this section, we introduce a construct that defines the set of links that cannot be
active simultaneously. An independent set of a graph G is a subset of vertices such that
no two vertices in the subset represent an edge in G. A Maximal Independent Set
(MIS) is an independent set such that no new vertex can be added to the set without
affecting the independence property
An independent set obtained from the interference graph F, gives the set of possible
links that can be simultaneously active in a network. An enumeration of all MIS for
the interference graph gives all the possible configuration of active links in the
network.
For example, the set of all MIS in the above graph F is (1,4,7), (3,7), (2,7), (1,5),
(2,5), (1,6), (3,6), and (2,7).
Unlike in a wired network, the notion of link capacity is not well defined in a
contention based wireless network. Capacity of a link may be described as the number
of bits that can be carried per second through the link. While link capacity in wired
networks is determined by fixed bandwidth cables laid between nodes, in some
wireless systems deploying TDMA and FDMA, the capacity can be calculated due to
existence of a central entity that allocates resources to each node.
20
In a contention based wireless system, it is the need for every node all over the domain
to share the portion of the channel it is utilizing with the nodes in its local
neighborhood. The IEEE 802.11 standard based MAC is one example of a distributed
contention based medium access system in which a common channel is shared among
the participating nodes. It is of interest to us to find the time available for each link to
access the common channel
In a single hop network, the probability for each link to be active is same for all links
given the assumption of equal traffic at all nodes whereas in a multi-hop network, a
link’s activity depends on all other contending neighboring links’ activity. With the
assumption of saturated traffic at nodes and its equal distribution to links, there is
competition among links at all times.
The capacity of a link is directly related to the time a link gets to be active. If this time
is obtained, it can be converted to capacity by multiplying it with channel capacity. In
a shared contention based system such as the IEEE 802.11 standard based system, the
amount of time for which a link becomes active per unit time cannot be determined
exactly as in a TDMA system. Neither can be rate available for a user or node be
determined exactly using the existing methods. This is due to the contention based
channel access where nodes get to transmit only when its neighbors relinquish the
channel. Also the Binary Exponential Back-off (BEB) introduces further probabilistic
component in channel access. In this context we would like to estimate the time for
which a link can be active and give bounds on the link throughput.
In this section, different approaches to estimate the capacity of wireless links between
nodes are presented . Also various schemes to derive upper bounds for link active
times and hence link capacities are outlined. The schemes can be classified into
centralized and distributed. Two approaches are proposed in each of the category.
21
2.5.1 ACTIVE TIME PROPORTION of LINKS
A notion of time available for each link relative to other links in its neighborhood is
defined. Active Time Proportion or ATP of links is defined as a fraction of time for
which a link gets to be active. This is obtained for each method of estimation and
compared across different approaches and with simulation.
In this approach, it is assumed that the topology of the entire network is known to a
central entity which makes use of it to derive the active times of links.
In this approach, a central entity with its knowledge of global topology constructs an
interference graph F from the network graph as illustrated for the string network
above. This corresponds to a graph where the neighboring vertices, which are links in
the edge graph are the ones that cannot be active simultaneously. From the graph F, all
the Maximal Independent Sets (MIS) are enumerated. This gives the set of all the
possible configurations of active links in the network. Each MIS corresponds to a
state of the network.
In this approach, called EW-MIS or Equally weighted MIS approach, the probability
of occurrence of MIS is uniformly distributed. Hence if the total number of MISs in
the network is M, then each MIS is assigned a uniform probability of 1/M. Active
Time Proportion or ATPs of links are obtained by summing the probabilities of
occurrence of MISs in which a link occurs. It is given by
22
M
ATPij = � Ri ---- (1)
i =1,ij�M i
Where
ATPij is the Active Time Proportion of link ij.
Ri = Probability of occurrence of ith MIS.
M – Total number of MIS in the network.
Let
Ti : Probability that node i starts transmitting in a given time slot.
Pij : Probability that node i’s transmission is intended to node j.
WMi : probabilistic weight given to the ith MIS.
Lij : Initial link active probability.
Cij : No. of MIS in which link ij occurs.
M : set of all MISs.
Mi : refers to the ith MIS.
23
to begin or initiate transmission. Based on this each node has equal probability of
initiating data packet transmission. Hence Ti is 1/N, where N refers to total number of
nodes in the network. Probability of a node’s transmission to its neighbor can be
calculated based on the traffic pattern. If traffic is going to be equally distributed to a
nodes neighbor, it can be taken as 1/no. of immediate neighbors. It is a generic method
which can take values depending on the assumed traffic model.
Now we calculate the available Time proportion of a link by summing the probabilistic
weights given to MIS in which the link occurs.
While the initial link probabilities Lij are assigned by considering local node
distribution, they don’t take into account the parallel transmission in the network due
to spatial reuse. By assigning weights to MIS based on the initial L ij and then re-
computing the active time of links we are able to capture the parallel activity of links
well.
24
2.7 DISTRIBUTED SCHEMES
The centralized methods for computing link active time proportion require global
knowledge of network topology. Hence only a central entity that is aware of the entire
network topology can perform this. Also the method of enumerating all the Maximal
Independent Sets for a given graph is known to be a NP-hard problem. Due to its
complexity and the requirement for global knowledge, the solution suggested is
limited in its application. In this section, we outline two distributed approaches to
estimate the active proportion of links using local topology knowledge. The
distributed nature of solution finds application in wide ranging applications from QoS
routing and admission control to rate based flow control.
In the distributed approach, the objective is to enable a node itself to calculate the
ATPs of links incident on it. To achieve this, a node needs to have knowledge of all
other nodes in its transmission and interfering range since these are the nodes it
competes with. We extend the notion of interference from nodes to links, by
introducing one-hop and two-hop interfering links. If we construct from a network
graph G, an edge graph E , where links are the vertices, immediate neighbors of the
link node are termed one-hop interfering links and the corresponding two-hop
neighbors are called as two-hop interfering links.
We consider that interferers for a link includes only its one-hop and least two-other
links between them alone can be active simultaneously or that one in m=3 consecutive
links can be active simultaneously. It is fair to assume that a link experiences
interference only till two-hop interfering links as illustrated in Guerien and Chaulet
(2003) since considering hops greater than or lesser than 2 results in over estimation or
under estimating the effect of interference. Evaluations on random graphs of the
25
number of missed or over-detected neighbors showed that 2-hops is the best number
that gives the least missed or over-detected neighbors Guerin and Chaudet (2002) two-
hop interfering links.
Looking from the protocol constraints too, the consideration of 2-hop neighbors would
suffice. In multi-hop networks, to address the hidden terminal problem, RTS/CTS
handshake is used before data transmission. When this mode is used, only links that
are separated at least by two hops can be active simultaneously. This can be explained
with the following figure. At any time in a saturated 802.11 network, nodes contend
for channel. This can be viewed as a set of links contending for activity. Each link has
its own set of interfering links that compete for channel. We call it the contention set
of the link. We illustrate that for any link, the contention set includes the one-hop and
two-hop interfering links.
g i
a b c d e f
h k
Consider the link cd in the above figure. When the link cd is active, its one-hop
neighboring links bc, gc, ch & de are made inactive due to the reception of either
RTS/CTS packets containing Network Allocation Vector (NAV) information. Apart
from no transmission, they don’t either respond to any request for transmission from
the other nodes in the form of RTS/DATA. This makes any activity impossible in
either the one-hop or two-hop interfering links. Due to symmetry, any activity in these
26
links makes the link cd too inactive. In the figure, links ab, ei and ef are the two-hop
interfering links.
In an IEEE 802.11 network, nodes decode packets both intended and not intended for
them. We describe a simple approach to update a node’s one-hop and two-hop
neighbors. Using our approach a node has to maintain one-hop and two-hop neighbor
list. The RTS and DATA packets carry two fields namely Transmitter Address (TA)
and Receiver Address (RA) in their header. For each RTS/DATA packet that a node
decodes, a node updates its neighbor list as follows.
Since TA field refers to the address of the sender, it is updated as the immediate
neighbor or 1-hop neighbor of a node. RA field refers to the destination address of the
packet which means that the sender of the packet which is on-hop to this node is
sending a packet to the node referred to by RA. This implies that the node referred to
by RA must be reachable from this node by two-hops. However its doesn’t mean that
only way of reaching RA is by two-hops. A node which is a two-hop neighbor could
be a one-hop neighbor too. This only reflects the presence of cycles in the network.
However its is possible that the topology of the network changes in due course of
time. To accommodate that in our topology discovery method, we maintain a timer
with each entry. The value of the timer will depend on the rate at which topology
changes.
27
This method of update is robust in that it discovers all cycles in the network. In a
saturated network, nodes can quickly discover its local topology this way. The stale
entries in the table are cleaned with associated timers. This method also
accommodates for low mobility as any change in node positions can be taken care of
by new updates and update period.
With the topology discovery method, a node is aware of two-hop node topology. But
we need to know the two-hop link neighbors. While both the end nodes of the link
together get a complete information about the one-hop and two-hop neighboring
links, it is possible that in some configurations, one of the end-nodes misses some
node at the other end. This is illustrated by the following figure.
In the figure, consider link 3--4, while node 3 discovers nodes 2 ,1, 4, 5and hence the
links 2-3 and 1-2, 3-4 and 4-5 it doesn’t know about the link 5-6 since it doesn’t know
its third hop neighbor 6. Similarly, node 4 is aware of the links 3-4, 4-5, 5-6, and 2-3
but not 1-2 since it has no knowledge of 1. To facilitate this, neighboring nodes can
exchange the information about missing nodes to get the complete information on
two-hop interfering links.
In this section we give a distributed approach for estimating the link active
probabilities. This computes the link active time based on the node transmission
probabilities. In this approach the probability that a link is active is the sum of product
of probability of node transmission and probability that the traffic from the node is
meant to the neighbor at the other end for either end nodes of a link.
28
Let
Ti : Transmission probability of node i.
CNi : number of contending neighbor nodes of node i.
Pij : Probability that node i’s transmission is intended to node j.
29
We have a method to express a links chance to transmit based on its link
neighborhood. The proportion of time each link gets to be active is dependent on the
number of interfering link neighbors it has. However the effect of each link on this
link’s activity is different and depends on its own link neighborhood. It will be a
recursive effect. We don’t account for this in our work and just give initial estimates
as ATPs.
The above estimate of the time available for each link can be converted into available
rate of the link in bps.
Available Bandwidth = ATP*Channel bandwidth. – (6)
We adopt the parameters for as 0.75 and 1 second as update period in our
simulation. The value of chosen helps in removing the short term fluctuation and
30
reflects long-term trend. Since this estimate is made per link, losses per link are
appropriately captured. Now the achievable bandwidth is given by
The results for centralized and distributed estimations are discussed in chapter 4.
31
CHAPTER 3
ANALYTICAL MODELING OF MULTI-HOP
NETWORKS
3.1 MOTIVATION
In the IEEE 802.11 standard based MAC protocol, channel as a common resource is
shared among contending nodes using a distributed and random scheme. Each node
competes with all other nodes in its neighborhood. In the previous section, the link
capacity problem was solved from the node perspective since links become active only
when one of the nodes it is attached to becomes active. In this section, a node’s
transmission probability is calculated using the analytical model developed for
channel and node. Once this is obtained, a link’s probability of being active can be
easily found out. This is a more accurate way of characterizing a link’s capacity in a
system where the channel access mechanism uses a random procedure.
In the IEEE 802.11 standard based multi-hop network, neighborhood of each node is
different. Hence the probablity of each node accessing the channel is affected
differently depending on the node distribution around it. Each neighboring node would
affect a node’s channel access probability to a different degree. We can represent this
with a circle around a node that includes the nodes that compete with this each
affecting the node’s time for channel access to a different degree.
32
Apart from the neighborhood contention of nodes, a node’s channel access is affected
by constraints imposed by the medium access protocol, the IEEE 802.11 DCF in this
case. We use the model developed by Wang and Acves (2004) as a basis for multi-hop
networks using CSMA/CA to derive the node transmission probabilities for IEEE
802.11 networks. While the work assumes a two-dimensional poisson node
distribution, we assume node distributions of string and grid topologies.
We make the following assumptions for the sake of simplicity and without loss of
generality
Each node has the same transmission and receiving range ‘R’.
Heavy traffic assumption : A node always has a packet in its buffer to be sent
and that traffic is distributed equally to its neighbors.
To simplify our analysis, we assume that nodes operate in time slotted mode,
where length of a time slot ‘τ’ equals one propagation delay, overhead due to
the transmit-to –receive turn around time, carrier sensing delay and processing
delay.
We assume that a node is ready to transmit with probability ‘p’ and not ready
with probability ‘1-p’. p is a slot-independent, protocol specific parameter.
However at the level of individual nodes, the probability of being ready to
transmit varies for each time slot depending on the current state of the channel
and the node. Hence we are interested in calculating the probability that a
node transmits in a time slot denoted by p’.
33
= p.Пl. ----- (9)
where Пl denotes the limiting probability that the channel is in idle state.
The DCF access procedure of the IEEE 802.11 standard requires a node to do most or
all of the following before transmitting a packet onto the common ‘Channel’. They
include carrier sensing, wait for inter frame space time, backoff. Hence even when a
node is ready to transmit, it may or may not transmit in a slot depending on the state of
a channel, the node’s neighborhood and the constraints imposed by the MAC protocol.
To model this, we use the Markov model to capture the effect of channel on a node.
The channel around each node is modeled as a circular region with some more nodes.
It is assumed that nodes within the region can communicate with each other while they
have weak interactions with nodes outside the region. The model is simplified by
considering weak interactions where the decision of inner nodes to transmit, defer and
backoff is almost not affected by that of outer nodes and vice-versa.
With the above assumptions, the channel is modeled as a four-state Markov chain as in
34
Fig 6 Channel moled for IEEE 802.11 MAC
Idle : is the state when the channel around the node x is sense idle, its duration is ‘τ’.
This denotes the state where there is no transmission.
Long is the state when a successful four-way handshake is initiated in the channel by
some other node.
Tlong = lrts + τ + lcts + τ + ldata + τ + lack
= lrts + lcts + ldata + lack + 3τ.
where lrts, lcts, ldata and lack represent the length of rts, cts, data and ack packets.
Short 1 is the state when multiple nodes around the channel transmit RTS packets
during the same time slot and their transmissions collide. The busy time of the channel
in this state is Tshort1.
Tshort1 = lrts + τ.
Short 2 is the state when one node around the channel initiates a failed handshake
with a node outside the region. Even though CTS packet may not be sent due to
collision at or deferral of the receiving node, those nodes that overhear the RTS as
well as the sending nodes do not know if the handshake is successfully continued,
until the time required for receiving a CTS packet elapses. Since the channel is in
35
effect unusable for that period, for all the nodes sharing the channel, the duration of
this state is
Tshort2 = lrts + τ + lcts + τ
= lrts + lcts + 2 τ.
Using the channel model, steady state probabilities of channel being idle can be
derived for string and grid topologies. The derivation applied to the centre nodes of
string and grid networks since for a string and grid network with large number of
nodes, the edge nodes can be neglected without loss of generality. Also the Markov
chain can be solved easily when the number of neighbors are fixed. Expressions for
edge nodes can be derived similarly.
First the transition probabilities of the channel Markov chain for string network is
calculated. The transition probabilities from any other state to idle is one since before
the start of any other transmission, the channel stays in idle state at least for τ seconds.
Pii is the transition probability from idle to idle state of channel. A node ‘x’ in the
middle of the string network, has two immediate neighbors. Since the node x senses
the channel to be busy even when the two immediate neighbors receive packets from
their neighbors, transmissions from their neighbors to this node is also considered
Since all these nodes are assumed to be in the centre of the network, p’ is same for all
these nodes.
Pii= (1- p´)4 ---- (10)
Pil is the transition prob. from idle to long state. For the node x to be observe a
successful packet transmission in the channel, only one neighboring node of x should
be involved in a successful transmission.
36
where ps denotes the probability that a node begins a successful four-way handshake.
It is not known yet and will be derived
The transition probability, Pis1 from idle to short1 state is the probability that more than
one node transmit RTS packets in the same slots. In our case, its is just the probability
that both the neighbors of node x transmit simultaneously, given by,
Now with these probabilities known, we can calculate the transition probability from
p i Pii + p i Pl + p i Ps1 + p s 2 = p i
p i Pii + 1 - p i = p i
1 1
pi = = ----- (14)
2 - Pii 1 - (1- p´) 4
The limiting probability p i or long run probability that the channel around the node x
p iTidle
pl =
p iTidle + p lTlong + p short1Tshort1 + p short 2Tshort 2
p i , p l , p s1 , p s 2
37
p iTidle
pl = .
p iTidle + p i Pil Tlong + p i Pis1Tshort1 + p i Pis 2Tshort 2
Tidle
pl = . ---- (15)
Tidle + Pil Tlong + Pis1Tshort1 + Pis 2Tshort 2
pTidle
p’ = .
Tidle + Pil Tlong + Pis1Tshort1 + Pis 2Tshort 2
The derivation for centre nodes of a grid network can be similarly derived and the
p´ = p. p l .
= p . [ + 4* ps *(1 - p´)3Tlong + p´ 2(3 p´2 – 8 p´ +6)Tshort1 + 4(1- p´)3(p´- ps)
Tshort2 ]. ------ (22)
38
In the above equation the values of ps is yet to be determined which will be obtained
From the above model the steady state probability of channel being idle is obtained. In
order to obtain the throughput performance of node, the states of a node needs to be
modeled. The node is modeled as a 3 state Markov chain again as in Wang and Aceves
(2004). Our intention is to use this model which was used for CSMA/CA based
networks to the IEEE 802.11 standard based networks.
The wait state is a state when the node defers for other nodes or backs off, succeed is
the state when the node can complete a successful transmission and fail is the state
when the node initiates an unsuccessful handshake. We define the length of succeed
Tsucceed = Tlong.
39
= lrts + lcts + ldata + lack + 3τ.
Tfail = Tshort2.
The duration of the node in wait state is τ, in this model. This is a simple approach and
the work evaluates how good can it be modeled using a single state with a fixed time.
While there are many transition probabilities involved, it is sufficient for our purpose
if the transition probability Pws, from wait state to succeed state is determined. express
Pws can be expressed as follows,
The transition probability from wait state to succeed state, Pws is found out considering
the transmission from this node to every other node. Since in a string network, a node
sends to any of its two neighbors, we take twice the transition probability for each of
its neighbors.
Similarly the probability that a node x continues to stay in the same state, P ww is just
the probability that no node including itself transmits in the same slot.
Pww = (1-p’)3. ------ (25)
40
Let p s , p w , p f denote the steady state probability of state succeed, wait, and fail
respectively. The transition probabilities from any state to wait state, P sw, Pfw is 1 since
before attempting to send a next packet, each node has to wait for DIFS time. From
the markov chain we have,
p w Pww + p s + p f = p w .
p w Pww +1 - p w = p w .
1
pw =
2 - Pww
1
= . ------ (26)
2 - (1 - p ')3
From the node markov model we can also obtain the throughput of a node as follows,
p s ldata
Th = .
p wTwait + p sTsucc + p f T fail
= ldata Pws ( + (lrts + lcts + 2 )(- Pws ) + (lrts + lcts + ldata + lack ) Pws )
-1
------ (28)
We can use similar methods to derive the throughput for grid topologies. We give the
transition probabilities for the channel model as follows.
41
Pii = (1 - p´)4. ------ (29)
1
pi = .
2 - Pii
1
= .
2 - (1 - p ') 4
p´ = p. p l .
= p . [ + 4* ps *(1 - p´)3Tlong + p´ 2(3 p´2 – 8 p´ +6)Tshort1 + 4(1- p´)3(p´- ps) Tshort2 ].
------ (33)
And using the Markov model for node as in fig, the successful transmission
probabilities and node throughput is derived for grid network.
1
= . ------ (36)
2 - (1 - p ')5
42
P
ps = p w . Pws. = 2 - (1 - p ')5 = ps. ------- (37)
ws
The expressions obtained for string and grid topologies are evaluated in chapter 4.
43
CHAPTER 4
RESULTS AND DISCUSSION
The simulation model considers an ad hoc network with no mobility, operating using
DCF. The simulation setup consists of identical nodes with half-duplex radios. The
transmission range of each node is 250 m with the carrier sensing set to 550 m. The
simulation parameters are taken from the IEEE 802.11b physical layer using DSSS as
given in the table.
RTS 20 byte
CTS 14 byte
Data 2048 byte
ACK 14 byte
DIFS 50 μs
SIFS 10 μs
Slot Time 20 μs
Contention Window 31 – 1023
Data rate 2 Mbps
Basic rate 1 Mbps
44
is equally distributed to all its neighbors. Shortest path routing is used and is sufficient
since we simulate an ad hoc network with no mobility.
The centralized approaches, EW-MIS and NP-MIS are compared with the simulation
results for string, grid and random topologies. Fig 8 and Fig 9 show the centralized
estimations for string topology of 10 and 18 nodes respectively. It can be observed
that the edge links of both the networks get higher active times as expected. It can
also be seen that the links next to the edge links get a lesser time than the edge links
but higher share compared to the other centre links. It can be explained by the fact
that since they have lesser number of interferers, their times to transmit is higher than
the centre links. In both the cases, the estimates match the simulation results closely
with the average deviation for EW-MIS being 2.51Ee-02 for 10 nodes and 3.14e-02
for 18 nodes and a deviation of 1.53e-02 for 10 nodes and 2.07e-02 for 18 nodes case
for NP-MIS.
45
10 Nodes String
0.7
Simulation
MIS-equalWeights
MIS-NodeBased
0.6
0.5
Active time of Links
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9
Links
0.5
Simulation
MIS-equalWeights
0.45 MIS-NodeBased
0.4
0.35
Active time of Links
0.3
0.25
0.2
0.15
0.1
0.05
0
0 2 4 6 8 10 12 14 16 18
Links
46
0.5
Simulation
MIS-equalWeights
0.45 MIS-NodeBased
0.4
0.35
Active time of Links
0.3
0.25
0.2
0.15
0.1
0.05
0
0 2 4 6 8 10 12 14 16 18
Links
0.4
Simulation
MIS-equalWeights
MIS-NodeBased
0.35
0.3
A ctive tim e of Links
0.25
0.2
0.15
0.1
0.05
0
0 5 10 15 20 25
Links
47
0.35
Simulation
MIS-equalWeights
MIS-NodeBased
0.3
0.25
Active time of Links
0.2
0.15
0.1
0.05
0
0 5 10 15 20 25
Links
The centralized estimations for grid topology is given in Fig 8 and Fig 9 for 9 and 16
nodes respectively. With the average deviation of 3.70e-02 and 3.88e-02 for EW-MIS
approach and a deviation of 2.01e-2 and 2.26e-2 for NP-MIS method, the centralized
schemes estimate the grid network well. We see that for both the string and grid
networks, NP-MIS scheme estimates the active times much better than EW-MIS
method. This has been true for string network of any length from our experiments.
This can been explained by the fact that in the NP-MIS method we give weights to
MISs based on the end node’s probability to transmit. Since transmission is scheduled
in the IEEE 802.11 based MAC only for nodes, this method of weighting the MISs
reflects a link’s actvity more accurately.
48
4.2.2 DISTRIBUTED ESTIMATIONS
The distributed shemes DLN and DNP are validated for different topologies as in the
centralized validation methods. The estimations for string topology is plotted in Fig
13, Fig 14 for 8 and 14 nodes case. The deviation of DLN happens to be 7.73e-02 and
6.86e-02 for 8 and 14 nodes case, while it is 4.88e-02 and 5.32e-02 for DNP. The
edge links and centre links behaviour with the edge links getting higher share of time
is evident in the plots for string topology .
In the case of grid topology experiments with 9 and 16 nodes, the deviations for DLN
estimates is 5.43e-02 and 5.30e-02 and that for DNP are 4.28e-02 and 4.48e-02
respectively. Similar to the Centralized estimations, we find that the node based
estimations, DNP approximate the link active times closer than the DLN.
0.5
Simulation
Dist-Link Neigh
0.45 Dist-Node Prob
0.4
0.35
Active time of Links
0.3
0.25
0.2
0.15
0.1
0.05
0
1 2 3 4 5 6 7
Links
49
0.7
Simulation
Dist-Link Neigh
Dist-Node Prob
0.6
0.5
Active time of Links
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9 10 11
Links
0.25
Simulation
Dist-Link Neigh
Dist-Node Prob
0.2
Active time of Links
0.15
0.1
0.05
0
1 2 3 4 5 6 7 8 9 10 11 12
Links
50
0.4
Simulation
Dist-Link Neigh
Dist-Node Prob
0.35
0.3
Active time of Links
0.25
0.2
0.15
0.1
0.05
0
0 2 4 6 8 10 12 14 16 18
Links
0.45
Simulation
Dist-Link Neigh
0.4 Dist-Node Prob
0.35
0.3
Active time of Links
0.25
0.2
0.15
0.1
0.05
0
1 2 3 4 5 6 7 8 9 10 11 12 13
Links
51
The distributed methods estimate some of the sample random networks more closely
than the centralized methods. This can be seen in Fig 16 and Fig 17. We have studied
random networks of size 8 and 10 nodes. The DLN estimates for 8 and 10 nodes have
a deviation of 4.2e-02 and 2.57e-02 and the DNP estimates have a deviation of 2.68e-
02 and 1.84e-02 for 8 and 10 nodes with random topology. In the case of random
approximating better than the DLN for string and grid repeats for the sample random
networks too. Some small number of experiments conducted using random topology
gave high deviations. Hence it is difficult to say whether the distributed method is
good for networks in random topologies with a small set of experiments and a more
In this section the plots obtained for node throughputs computed analytically for string
and grid topologies are discussed. In the model, the time a node spends in wait state,
Twait is taken as τ. However in the IEEE 802.11 standard based multi-hop networks,
each node stays in the back off stage for some duration. The time for which a node
stays in the back off stage is modeled by taking the average time for which it stays in
wait state, Twait. In this section, the plots obtained for node throughputs of string and
grid topologies for a range of p’ and Twait values are discussed.
52
Fig 18 : String topology – Analytical results
Fig 18 shows the results obtained for normalized throughput of nodes in the centre of
a string network calculated under saturated conditions. The node throughput increases
to reach its peak around p’ = 0.05 and approaches 0 as p’ tends to 0.3. Since p’ denotes
53
the probability that a node is ready to transmit and since p’ is same for all nodes, more
nodes in the neighborhood compete for the channel, the individual node throughput
decreases owing to competition for channel, resulting in collisions and back off.
The fall from maximum throughput is gradual in the case of string topology relative to
the grid topology shown in Fig 19. This is due to the higher density of nodes in a grid.
throughput even for small values of p’. Also in a string network, a node reaches its
maximum throughput value for p’ around 0.05. This is much less for grid. This too is
due to the fact that higher number of nodes to compete with a node in a grid.
In this section, the normalized node throughputs of centre nodes obtained from
analysis are validated through the values from simulation. Fig 20 shows the
normalized node throughputs of centre nodes in a string network with varying string
lengths. The analytical expression derived for string and grid topology doesn’t have
the number of nodes as a variable in its expression while it has the packet transmission
probability p’ as a varying parameter. Since only the Channel idle probability, П l is
variable, it is used as a common variable term for comparing simulation and analysis.
The normalized node throughputs of centre nodes as well as the corresponding П l s for
a string network of particular size are obtained.
54
Fig 20 Simulation results of string topology
55
Fig 22 : Simulation results of 16 Nodes grid toplogy
Fig 21 and Fig 22 compares the analytically computed node throughput values for
grid topology with that of the simulated values for a 9 and 16 node cases. Though the
values are derived for only centre nodes, the values for all the nodes are plotted by
fixing the p’ as a common parameter for both simulation and analysis. There is a better
approximation of calculated values for centre nodes and also the average deviation is
about 0.1 much less than that for string network. Also the centre nodes show a
deviation of only 0.05.
The simulation results in general shows reasonlably good estimates. However these
values can be improved by considering a more accurate model of backoff and other
factors.
56
CHAPTER 5
CONCLUSION AND FUTURE WORK
The IEEE 802.11 standard based Medium Access Control protocol and its effect
on capacity in a multi-hop ad hoc networks have been studied. Generalized
frameworks for estimating individual link capacities have been proposed along with
approaches that can be used for real-time applications. Analytical model for
performance evaluation of individual nodes under specific topologies has been
studied.
The estimation methods proposed have provided a ground work for all future
enhancements in this area. The centralized approaches give close estimates for string
and grid topologies and sample newtorks with random topologies too. The Distributed
methods give good enough estimates but with lesser accuracy compared to the former.
While the centralized approaches use an NP-hard algorithm for their estimations, the
distributed schemes have very less or no complexity. Given the advantage of real-time
estimations, the lesser accuracy of distributed schemes is acceptable. Its worth noting
that the deviation of both the schemes was only of the order of 10e-02. The analytical
modeling of string and grid topologies provides a much needed insight into the multi-
hop modeling problem. Considering that the model approximates the backoff process,
the accuracy of the results are reasonably good.
57
4.6 FUTURE RESEARCH
The ditributed estimates can be improved by taking into account the recursive
effect of interference between nodes or links in the network.
Regarding the analytical model, a more throrough model of backoff process can
be done to capture its randomness. Also the study of multi-hop networks will be more
complete only by taking into account the effect of higher layers like routing and
traffic.
58
REFERENCES
1. Bensaou B., Wang Y. and Ko C.C. (2000) 'Fair Media Access in 802.11 based
wireless ad-hoc Networks', Proc. Mobihoc., Boston, pp.99-106.
3. Cali F., Conti M. and Gregori. (1998), 'IEEE 802.11 Wireless Lan : capacity
analysis and protocol enhancement', in INFOCOM ’98, 17th Annual Conf of
the IEEE Computer and Communications societies.
7. Gupta P. and Kumar P.R. (2000), ‘The Capacity of Wireless Networks’, IEEE
Transactions on Information Theory, vol. IT-46, no. 2, pp. 388-404.
59
9. Hass Z.H., Gerla M., Johnson D.B., Perkins.C.E., Pursley.M.B. and
Steenstrup.M. (1992) 'Guest-Editorial : Wireless ad hoc networks', IEEE
Journal on Selected Areas in Communications 17(8), pp.1329–1332.
10. IEEE standard (1999) for Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specification, P802.11.
11. Jain K., Padhye J., Padmanabhan V. and Qiu L. (2003), 'Impact of Interference
On Multi-hop Wireless Network Performance', ACM Mobicom, San Diego,
CA, pp.66-80.
12. Kazantzidis M., Gerla M. and Sung-Ju Lee. (2001), 'Permissible throughput
Network Feedback for Adaptive Multimedia in AODV MANETs', IEEE
International Conference on communications, pp.1352-1356.
13. http://citi.insa-lyon.fr/~cchaudet/docs/Guerin-Chaudet-EW2002.pdf
15. Li J. and Blake C. (2001), 'Capacity of ad hoc wireless networks', Proc. 7th
ACM Mobicom , pp.61-69.
16. Li M., Prabhakaran B. and Sathish S. (2003) 'On Flow Reservation and
Admission Control for Distributed Scheduling Strategies in IEEE 802.11
Wireless LAN',Proc. ACM/MSWiM, pp.108-115.
60
18. Samrath H., Chen K. and Nahrstedt. (2005), 'Dynamic Bandwidth
Management for Single-hop Ad Hoc wireless Networks',ACM/Kluwer Mobile
Networks and Applications (MONET) Journal, Special Issue on Algorithmic
Solutions for Wireless, Mobile, Ad Hoc and Sensor Networks,vol. 10, No. 1,
pp. 199-217.
19. Tang K. and Gerla M. (1999), 'Fair Sharing of MAC under TCP in Wireless Ad
Hoc Networks', Proceedings of MMT'99, Venice, Italy.
21. Xu S. and Saadawi T. (2003), 'Revealing the problems with 802.11 medium
access control protocol in multi-hop wireless ad hoc networks', Computer
Networks: J.Computer and Telecommunications Networking.
22. Yaling Yang., Jun Wang., Robin Kravets. (2003), 'Achievable Bandwidth
prediction in Multihop Wireless Networks', Technical report, University of
Illinois Urbana-Champaign, Urbana, IL.
23. Yawen D. and Barowski Saad Biaz. (2005), 'Towards the Performance
Analysis of IEEE 802.11 in Multi-hop Ad-Hoc Networks', IEEE Wireless
Communications and Networking Conference.
61
List of Publications :
3. Rajesh S., Vijayalakshmi K., Srikanth S. and Vaidehi V. (2003), 'Capacity and
QoS enhancement of ad hoc networks with intermittent smart directional
nodes', Proc. 9th NCC, India, pp. 35-39.
62