Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

Modelling using Markov Chains

S.R.Rimitha, Vedasamhitha Abburu, Saumya Gupta


November 11, 2016

Abstract
1 Learning Objective
In this document, we aim to study and ex-
plain Markov Chain modelling as a useful Through our study, we hope to fully under-
tool in analysing computer networks for re- stand the working of Markov Chain Models,
liability, availability, and various other per- as well as to learn the various practical appli-
formance metrics. Modelling of computer cations of Markov Chain Modelling in com-
systems and networks is becoming very nec- puter networks. We aim at understanding
essary, as a way to quantitatively analyse how the various computer network applica-
the system, without relying only on intu- tions can be modelled using Markov mod-
itive arguments. Markov chains offer fairly els, as well as how these models can be used
simple yet effective and practical modelling to analyze and evaluate the network perfor-
solutions. The aim of this document is to mance. Markov models have been widely
study these topics in detail and also learn used in real time applications and we hope
about their practical applications. These ap- to have a fair understanding of the working
plications are very varied, and include using and implementation of the Markov models in
Markov chains in the web for link prediction all such real time applications, as well as the
in adaptive websites, page ranking, which is advantages of using such models over other
a popular algorithm used by search engines analytical methods and simulations.
to list relevant web pages, modelling of queu-
ing networks, as well as performance analysis
of the MAC layer of WLANs.
Keywords: Markov chains, IEEE 802.11, 2 Literature Survey
IEEE 802.15.4, Backoff algorithm, perfor-
mance analysis, throughput, access delay, This section discusses the work done by var-
CSMA/CA, link prediction, queueing net- ious researchers in regard to Markov chain
works, PageRank modelling in wireless computer networks.

1
2.1 IEEE 802.11 condition and an error-free channel. Under
the assumption of saturation condition, ev-
In IEEE 802.11, Distributed Coordination ery station always has a packet to trans-
Function (DCF) is defined as the fundamen- mit. That is, after a successful transmis-
tal MAC layer standard for WLAN. DCF sion, the station again has a packet ready
implements the CSMA/CA collision avoid- to be sent. This is not the case in the real
ance method of network access, and also uses world, but this assumption is useful as it
an exponential backoff (usually binary) algo- gives the network more traffic than the usual;
rithm as the standard. thereby setting an upper bound for this pa-
The method in which DCF works can be de- rameter. This paper includes derivations for
scribed in the following way: when a sta- both variations of the CSMA/CA protocol
tion wants to start transmitting, it senses – one which uses ACK, and another which
the medium, and checks whether or not it is uses RTS/CTS handshaking.
busy. If the link is sensed to be busy, the sta- Papers [21] and [9] fall back extensively on
tion waits, and executes a backoff procedure. Bianchi’s work. The environment in both the
Once the procedure has been completed, the papers operates under saturation conditions,
station can again try to transmit. This back- and assumes a noiseless channel. These pa-
off procedure has been specified in order to pers have investigated the medium access de-
reduce the chances of packets colliding in the lay as well.
medium when multiple stations attempt to Paper [9] has deployed the use of a one-
transmit at the same time. dimensional Markov model to capture the
The standard backoff algorithm specified in algorithm, and does so only for the ACK
the IEEE 802.11 standard is the binary ex- mechanism. The analysis has gone beyond
ponential backoff (BEB) algorithm. BEB to consider the general case of Expo-
Giuseppe Bianchi was among the first few re- nential Backoff (EB) with a backoff factor r.
searchers to begin analysing the performance BEB algorithm is but a special case with r
of the BEB algorithm. His work ([2] and = 2. The authors have further compared the
[3]) has paved the way for not only devel- performance parameters in the presence and
opment in new models for analysing the ex- absence of a maximum retry limit M. This
isting algorithms, but also proposal of new limit forces the dropping of a packet after it
algorithms to improve the current ones. He has been involved in M collisions. This as-
is truly a pioneer in this field. sumption is significant as real-world imple-
mentations usually set M as 16. Paper [5]
In his papers [2], [3], Bianchi proposed to also takes into account a retry limit; how-
derive ‘saturation throughput’ using a two- ever, it deploys a two-dimensional model to
dimensional Markov model. His work stems
do so.
slightly away from real-world implementa-
tion because of his assumption of saturation In paper [17], [1], two new algorithms that

2
can execute the backoff procedure have been transfer rate as 250 kbit/second.
suggested. These are the Exponential In- The focus of this architecture is to establish
crease Exponential Decrease (EIED), and a simple wireless network conceptually. It is
the Exponential Increase Linear Decrease noteworthy to mention here that this stan-
(EILD) algorithms, and have been modelled dard has focused a lot on IEEE 802.11 while
using a suitable Markov model. In [20], a developing certain implementations. The
new analytical Markov model of the BEB, network layers were mostly based on the OSI
as well as the EIED and the EILD has been model. There are various kinds of topolo-
proposed. Both have been analysed to find gies possible using this standard to create
the average delay. a PAN. However the most user is a peer to
In paper [6], the authors propose a dynam- peer topology or star topology. The most
ically “tuned” backoff algorithm, which es- important aspect with respect to this IEEE
timates the network parameters to dynam- standard is its reliability which is promised
ically calculate and set the optimal backoff through the CSMA/CA protocol which uses
window size. A suitable Markov model has random exponential backoff algorithm.
been used to analyse the performance of this This report focuses on the role of Markov
algorithm. models that has been used while propos-
The above papers have all considered only ing various analytical models for CSMA/CA
saturated conditions. New Markov models (Carrier Sense multiple access/Collision
have been proposed in [12] and [18] that aim Avoidance) protocol for the above mentioned
to simulate a more realistic network load, IEEE 802.15.4 standard by various research-
i.e., non-saturated network conditions. ing teams. Though the CSMA/CA proto-
col used in this standard is extremely similar
to the one used in wireless LANs i.e. IEEE
2.2 IEEE 802.15.4 802.11, there is one major difference. This is
the backoff algorithm used in 802.11 where
In this section we are going to look at a new when the channel is idle, the count is reduced
IEEE standard i.e. IEEE 802.15.4 which is by 1 whereas in 802.15.4 the count is lessened
used to specify protocols for Wireless Per- by one regardless of the channel state.
sonal Area Networks especially low rate ones.
It is used to specify the PL (Physical Layer) This technical difference has been addressed
and MAC(Media Access Control) for the by a number of papers. [12] has addressed
above specified network. It was a relatively this issue by proposing an analytical Markov
new standard being released in 2003. Its ma- model for CSMA/CA in saturated condi-
jor attention is to lower the cost and present tions. This was followed by [18] where they
a low speed ubiquitous communication be- have proposed a Markov model which modi-
tween nearby devices. It mostly is used for fies the Double Carrier Sensing protocol used
embedded devices with a 10 meter range and in the MAC Layer to yield better perfor-

3
mance results. Paper [14] have also proposed 3.1 Existing Backoff Algo-
a model for the same aspect with different rithms
aspects into consideration and various dif-
ferent assumptions. It must be noted here The work carried out by Bianchi in his pa-
that papers [12],[18],[14],[10] and [15] sug- pers [2] and [3] has been used as the standard
gest analytical models for CSMA/CA with by a lot of other researchers. Paper [3] analy-
different constants taken into consideration. ses the binary backoff algorithm and derives
For instance, the papers mentioned above an expression for throughput using a two-
mostly went along with the assumption that dimensional Markov model. The paper at-
the Contention Access Period(CAP) is infi- tempts to cap the value of throughput by as-
nite which is not the condition always. [15] suming constant traffic in the network. The
has its own assumptions while proceeding following are the assumptions in this paper:
with the model which might not hold true
at every situation. Paper [7], which we have • Saturation conditions: Under this as-
focused upon , has tried to specify a Markov sumption, every station always has a
model for the same solution trying to over- packet to transmit. That is, after a suc-
come all the drawbacks of its predecessors. cessful transmission, the station again
has a packet ready to be sent. This is
not the case in the real world, but this
3 Outcome of Literature assumption is useful as it gives the net-
Survey work more traffic than the usual.

From the literature survey, we deduce that • Finite number of stations: This assump-
extensive work using Markov modelling has tion sticks correctly with current imple-
been carried out in 3 major areas: mentations.

• Analyzing performance metrics • Noise-free channel: The paper assumes


(throughput and delay) in the ex- an ideal link with no errors. As we
isting backoff algorithms of IEEE know, no media is completely free from
802.11. errors.

• Improving the existing algorithms, and • Infinite attempts: There is no limit on


analyzing their metrics as well. the maximum number of times a sta-
tion can resend a packet. Hence, a
• Analyzing performance metrics in the packet will never be dropped. This
algorithms of IEEE 802.15.4. Conse- stems away from current implementa-
quently, we present a gap analysis of the tions which usually set a maximum retry
work done in each of these areas. limit of 16.

4
Through mathematical manipulation, transmitted out onto the network.
Bianchi derived the following dependencies: The paper derives formulae for the two per-
• For the protocol making use of the ACK formance parameters under the following as-
mechanism: The backoff window size sumptions:
is usually the propagation delay. In • Saturation condition is in place.
this mechanism, throughput increases
with increase in the minimum window • There are no errors in the channel.
size. In simple terms, the probability • Number of nodes in the network is a fi-
for number of collisions decreases when nite constant.
the waiting period at each station in-
creases. Hence, throughput increases. • The ACKs are part of the data packets,
As far as the number of stations is con- and are not explicitly sent. No hand-
cerned, throughput is indirectly propor- shaking mechanism.
tional to it. That is, as the network While Bianchi’s paper makes use of two-
expands and grows, the throughput re- dimensional Markov chain models to ana-
duces. lyze the throughput, this paper uses a one-
• For the protocol making use of the dimensional Markov chain. The paper anal-
RTS/CTS mechanism: The paper yses the algorithm in the presence as well as
shows that throughput reduces with in- absence of a retry limit M, as opposed to pa-
crease in minimum contention window per [3] which ignores this limit. The limit
size. Furthermore, in this mechanism, dictates that a packet should be dropped if
the throughput remains largely unaf- it is involved in more than M collisions.
fected by the number of stations in the The findings from this paper are:
wireless network. • For EB without a maximum retry limit:
The paper finally settles upon the result It has been shown that throughput re-
that the handshaking mechanism performs mains largely independent of the num-
better than the basic method of acknowl- ber of nodes in the network. The
edgement in large networks. medium access delay, however, is di-
rectly proportional to the number of
Paper [9] paper analyses the performance of
nodes.
the Exponential Backoff (EB) algorithm, op-
posed to Binary Exponential Backoff (BEB) • For EB with a maximum retry limit M:
analysed in paper [3]. In addition to satura- As the number of stations in the net-
tion throughput, the paper also investigates work increases, the throughput begins
the medium access delay of this algorithm. to drop. The medium access delay in-
The delay is the time between its arrival at creases (not linearly) with increase in
the MAC layer to the time it is successfully number of nodes.

5
The algorithm setting a limit of M is usu- • The delay increases as the mini-
ally deployed in real-world implementations. mum contention window increases
The perk of this is that the other packets because retransmission will take
awaiting transmission will not be kept under longer now. This delay is more
indefinite hold. as compared to RTS/CTS. The de-
Paper [21] implements a two-dimensional lay increases as the minimum con-
Markov process to capture the Binary Ex- tention window increases. As the
ponential Backoff algorithm. In addition to channel bit rate increases, delay re-
deriving an expression for throughput, this duces. It also increases as the num-
paper accounts for the variation of access de- ber of stations increases, and this
lay with respect to the system variables. increase is more drastic as com-
pared to the below protocol.
The paper applies the Markov model to two
variations of the CSMA/CA protocol – one 2. For the protocol making use of the hand-
which uses an acknowledgement mechanism shake mechanism:
[ACK] for verifying successful transmissions,
and another which uses a handshaking mech- • The number of nodes in the net-
anism [RTS (Request to send)/CTS (Clear work does not have significant im-
to send)]. The performance parameters have pact on the throughput. The
been derived for both mechanisms. This pa- throughput begins to dip as the
per investigates the response of throughput minimum contention window size
and delay with respect to changes in the net- begins to increase. Throughput de-
work parameters such as number of nodes, creases as the channel bit rate in-
channel bit rates, and contention window creases.
size. The following are the outcomes of this
• The delay increases as the mini-
paper:
mum contention window increases.
As the channel bit rate increases,
1. For the protocol making use of the ACK
delay reduces. It further increases
mechanism:
as number of stations in the net-
• As the number of nodes in the net- work increases.
work increases, the throughput de-
creases.This is because, greater the We see that the results of this paper agree
number of nodes, greater is the with those obtained in paper [3] with regard
probability for collision. As the to throughput. Comparison for delay cannot
minimum contention window size be made as paper [3] does not delve into its
increases, so does the throughput. study.
Throughput decreases as the chan- The paper concludes its findings by stating
nel bit rate increases. that RTS/CTS has better throughput and

6
smaller delays as compared to the basic ac- transmission. When the network is under
cess mechanism when the network load is heavy load, this algorithm allows the net-
high. This agrees excellently with the out- work to decrease the number of collisions,
come of paper [3]. and thus improve performance. However,
when the network is under a light load, oc-
casional and even infrequent collisions may
3.2 Variations in backoff algo- cause the station to increase the window size
rithms exponentially, and require a larger period
of time for the window size to decrease
The standard defined by IEEE 802.11 for again. During this time the window size
media access control, is DCF, which uses is decreasing, the link may be idle. This
BEB algorithm to execute the necessary can reduce the performance of the network
backoff procedure. However, the binary under lightly loaded conditions.
backoff algorithm causes a station to reset
Paper [17] proposes an algorithm (EIED)
its contention window size after a successful
where the window size decreases expo-
transmission. This results in a largely varied
nentially as well following a successful
contention window size. It also results transmission. Compared to the linear
in a decrease in network performance, decreasing mechanism described by [1],
as stations would have to start from the
the EILD algorithm in [17] will be able to
minimum contention window size every time
quickly recover from infrequent collisions,
a successful transmission occurs. When
as the backoff window size decreases much
there a lot of contention in the network faster in EIED than in EILD. However, a
(many stations want to transmit over the problem of EIED algorithm is that it may
medium at the same time) this may result be unfair to stations in the network with
in more collisions, as each new packet at
small number of competing stations.
a station would wait a very small amount
of time before being sent, and would most
likely collide with another packet already
being transmitted. Therefore, in heavy
traffic conditions, the resetting used in the
3.3 Analysis in IEEE 802.15.4
BEB would not be beneficial to network The most important predecessor for this
performance. standard has been the IEEE 802.11. It must
In order to counter this, a new backoff be noted here that due to the difference
algorithm (EILD) is suggested in [1], where in the nature of situations where the stan-
the window size is increased exponentially dard is applied, the backoff algorithm has
(multiplied by a specified constant) follow- slight changed. For instance, the back off
ing a collision, but is decreased linearly algorithm is IEEE 802.15.4 decrements the
(subtracted by one) following a successful counter regardless of the state of the channel

7
while IEEE 802.11 does not perform that. which the use of superframes. However it
Another important aspect of this report is has not taken into consideration the inac-
the modelling of the backoff algorithm using tive time period of this superframe which is
Markov chains in IEEE 802.15.4. Here it is the major advantage and the reason to use
important to know that the parameters used them in IEEE 802.15.4 as they provide for
while designing the Markov model plays a lower power consumption by giving the op-
major role in differentiating them. For in- tion of switching off the power when not in
stance paper [7] and [13] have only one dif- use which is essential for all devices used in
ferentiating parameter i.e. paper [7] includes a personal network. We personally believe
the number of remaining Clear channel As- the highest improvement was done by paper
sessments (CCA) which is an important pa- [7] which takes into account the superframes
rameter which paper [13] does not take it in structure and the retransmission schemes in
account. And the change can be simultane- an unsaturated traffic condition. But this
ously seen in the performance metrics when paper does contain a few assumptions like
simulated. assuming no transmission errors etc., How-
Also as it has been mentioned before, the ever it must be mentioned here that , with
report focuses on all papers that suggest the understanding of the paper we achieved
, it can be stated that the assumption on pa-
a model for the backoff algorithm used in
per [7] do not affect the performance metrics
the specified IEEE standard. The authors
and much as the previous papers and tries to
have then followed their model by their sim-
give an accurate reading during it’s simula-
ulation while recording various performance
metrics such as performance or time delay. tions.
It also must be noted here that the models
are prescribed for slotted CSMA/CA. Papers
[12], [18], [14] and [10] have assumed satu-
rated traffic conditions. However the above
standard is prescribed for low power con-
sumption and low cost networks and hence 4 Analytical model
the traffic is less or in other words un-
saturated. Also these papers have consid-
As seen in the literature survey, the Markov
ered Contention Access Period to be infi-
models have been applied in the field of
nite which is an ideal more than practical
case. An improvement is seen in the pa- wireless connectivity. Furthermore, each
per [15] where they do consider the above field applies a variation of the model –
one-dimensional, two-dimensional, or multi-
mentioned period to be finite. Also this pa-
dimensional. In this section, we explain how
per does taken into account one of the most
Markov models have been used to model the
important aspects of PAN communication
backoff algorithms.

8
4.1 Bianchi’s model as given in • If the station transmits its frame without en-
tering the backoff procedure, it means that
paper [3]
the frame it has just sent out was success-
ful, and now the channel is free. Since p is
The aim of the paper is to apply a two- the probability that a collision occurred, 1- p
dimensional Markov model to the Binary is the probability of successful transmission.
Exponential Backoff algorithm, in order to Similarly 1- pb is the probability of an idle
determine the values of throughput and channel. Hence,
delay under saturation conditions.
Assume that the network has a finite
number of stations n. Since the operation
is carried out under saturation conditions, Markov model as the process {s(t), b(t)}.
each station, after the completion of a This model is discrete-time. {i, k} excel-
successful transmission, immediately has a lently captures the state of each node in
packet available for transmission. the network, where i indicates the chosen
Let b(t) denote the duration of the backoff random number and takes the values 0
window for a particular node at time t. through m, while k denotes the backoff
Thus, b(t) is stochastic, but it is not Marko- delay and takes the values 0 to Wi−1 in slot
vian because the future is not completely times.
independent of the past, that is, the number The following figure depicts the Markov
of collisions occurred previously influence model used in the paper. An unusual
the backoff window size. {−1, 0} state has been defined in the figure.
Say the maximum random number that A station invokes the backoff procedure if it
can be chosen is m. Then, there are m+1 has recently encountered a collision, or, if it
values to choose from. Let the random senses the channel as busy. There is a case
number chosen be i, and let s(t) be the when the backoff algorithm is not invoked.
stochastic process when the backoff stage is This is when the backoff counter equals zero,
i (i ∈ (0, m)). and the node detects an idle channel after
The value of the backoff counter is now waiting for a random period of time called
chosen in an unbiased manner from the the DIFS (distributed interframe space).
range (0,Wi−1 ), where Wi =2i Wmin . In this case, the packet is sent without the
node having to invoke the backoff procedure.
Let pb be the probability that the channel
Thus, state {1, 0} models the root state
is busy, and let p be the probability that
where no procedure has been summoned.
a collision occurs. It is imperative to note
that both of the above probabilities are
independent of the state s(t) of the station.
We can now express our two-dimensional

9
• The station transmits a new frame and re- After a successful transmission which oc-
sides in stage 0 of the backoff procedure if curs with probability 1-pc , a node enters
it observes a busy channel after a success- backoff state i=0 with contention window
ful transmission (= (1- p)pb ), or, if it detects size W0 . On the other hand, if a node is
that its transmitted frame has encountered in the backoff state i=k, and it experiences
a collision (= p). Hence, a failed transmission, then the node enters
backoff state i=k+1. The probability of this
happening is Pc .
Let Bk denote the k t h state that a node
Through a series of mathematical manip- resides in, in steady state. Then, Bk is a
ulation, the paper successfully derives a for- Markov chain with the transition probabili-
mula throughput of the system. ties pi ,j given as follows:
Explanation:
4.2 Model as used in paper [9] • A node can only enter the original
state 0 when there is a successful trans-
The aim of the paper is to apply a one- mission. The probability of successful
dimensional Markov model to the general transmission is 1- pc , and so the first
Exponential Backoff algorithm, in order to probability follows.
determine the values of throughput and de-
lay under saturation conditions. • The backoff stage for a node increases
The figure below depicts the one-dimensional by 1 (from i to i+1) only when it is in-
Markov model used to represent the backoff volved in a collision. The probability of
procedure. a collision is pc , and so the second prob-
ability follows.
A station can be in any one of the infinite
number of backoff states in steady state. In- • There are no other transitions possible,
finite because, in this case, we do not as- and so the probability is 0 for the re-
sume any limitation on the number of re- maining cases.
transmissions.
Let Pc denote the probability that the packet 4.3 Models of variations of
is involved in a collision. In backoff state i
backoff algorithm given in
(i=0, 1, 2. . . ) the size of the contention win-
dow for a node is: paper [20]
The paper presents a new Markov chain
Wi = ri W0 model to compute average delay in three dif-
ferent exponential backoff algorithms: the
where W0 is the size of the minimum con- BEB (Binary Exponential Backoff) algo-
tention window. rithm, EILD (Exponential Increase Linear

10
Figure 1: Two-dimensional Markov Model for the backoff algorithm

• The station chooses a backoff delay of next


• A stop is put to the counter when it is sensed stage i after an unsuccessful transmission at
that the channel is busy. present stage i-1.

• The station has reached the last stage of the


• The counter is reduced by one when an idle backoff procedure. It continues to stay there
channel is detected. after unsuccessful transmissions.

11
• If the current frame has been sent success-
fully (= 1-p) and it is sensed that the channel
is occupied (= pb ), then when a station tries
to transmit a new frame, it has to summon
a backoff delay of stage 0.

Figure 2: One-dimensional Markov chain


• The station enters into the original (1,0) model applied for the backoff algorithm
state if it sent out it’s packet successfully and
if it senses that the channel is free again

Decrease) and EIED (Exponential Increase


Exponential Decrease). Whereas in the BEB
algorithm, the size of the contention window are modelled using a new analytical Markov
in increased by a factor of two for each col- model that is one-dimensional. By using this
lision until it reaches a maximum size, the model, the average delay under overloaded
window size is reset following a successful conditions is calculated for each of the above
transmission. algorithms, and they are thus compared.
The rapid decrease in the backoff window In the Markov chain models used below, a
size may lead to a decrease in performance, common notation is used. Wi is the stage of
as it may increase the probability of collisions backoff, and acts as a counter that is decre-
in the period following a successful transmis- mented for every slot that the link is idle
sion. In order to counter this problem when and paused when the link is busy. When the
there is a large number of stations that want counter attains a value of zero, the station
to send, a slower reduction in the size of the can begin transmitting.
backoff window is suggested. This led to In paper [20], the probability p that a col-
the development of algorithms that increase lision will occur is taken to be a constant
the size of the backoff window after a col- value that is independent of the stage of
lision by doubling it, but decrease the size backoff (Wi ). By making this assumption,
of the window following a successful trans- the Markov chain models proposed by pre-
mission either exponentially (EIED) or lin- vious authors can be reduced from two vari-
early (EILD). EIED was originally proposed ables to one. This allows the use of a one-
in [17], and EILD was originally proposed in dimensional Markov model, where a state i
[1]. of the model corresponds to the backoff stage
In paper [20], the above mentioned algo- Wi . Also, the limit on the number of retries
rithms are explained and analysed. They is not considered. This means that the sys-

12
tem will stay in the final stage (Wm ) till the state.
packet has been transmitted without colli- The models do not take the finite retry limit
sion, where m is maximum backoff stage, or into account (the limit before the packet is
the maximum retry count. dropped). However, by modifying the model
The following figure 3 is the Markov chain such that once the system stays in the fi-
model for the EIED algorithm. nal state Wm once the retry limit has been
The W0 stage is the stage where the sta- reached, it can be ensured that the system
tion begins transmitting. The final stage is cannot retry sending the packet any longer,
the Wm . The transition probability is given thus modelling finite retry limits as well.
by p, (packet collision probability at the par-
ticular time of transmission). If the medium 4.4 Model as used in paper [7]
is sensed as idle (which has a p-1 probability)
the counter is decremented, and the system In paper [7], the authors try to establish
moves from state Wi to Wi − 1. an enhanced Markov chain model for slotted
CSMA/CA for WPAN’s which is mentioned
The main difference between this algorithm
in the IEEE standard explained above. They
and the BEB, is that in every state, once the
have then given a simulation results for the
packet has been successfully sent, the system
model prescribed by them in order to reveal
does not immediately move into the initial
the contrast between their model and mod-
state. In order to take this into considera-
els prescribed in previous papers. The an-
tion, the model is modified such that pack-
alytical Markov chain model for the above
ets can be sent not just from the initial state
specified mechanism considers a superframe
W0 , but also from any state in the system.
structures as well as retransmission schemes
This is done so that the history of conges-
which is extremely important. It also works
tion is not erased, and the station can still
on the assumption that the environment is
send packets. Note that the rate of increase
unsaturated which is more practical as it is
of the window size in this algorithm need not
applied to personal area networks which in
be the same as the rate of decrease.
general don’t contain many devices.They pa-
The following figure 4 is the model for the per starts off by briefly explaining to us the
EILD algorithm: IEEE 802.15.4 MAC protocol and then go on
This algorithm linearly decreases the back- to propose the analytical model of the slot-
off stage instead of exponentially decreasing ted CSMA/CA for the IEEE 902.15.4 MAC
it, as is done in EIED. The states are rep- protocol which is mainly implemented by the
resented the same way as they are in the Markov models. The major parameters of
EIED model, although the difference being the proposed Markov model are the stochas-
that this model has a larger number of states tic processes mentioned below :
owing to the linear decrease in state. In this
model also, packets can be sent from any • S(t) representing the backoff stage

13
Figure 3: EIED

Figure 4: EILD

• C(t) representing the count of backoffs 2. Assumption made by the authors is that
existence of errors by transmission or
• R(t) representing retransmission
channel sensing is null.
counter state
• W(t) representing the number of re- 3. The topology prescribed is star. Also
maining clear channel assessments transmitted frames can be lost only due
to collision as it is uplink data transmis-
They then go on to examine the perfor- sion.
mance through metrics like throughput ob-
tained while using the model explained and 4. In a situation that a device is transmit-
conclude it by verifying their analytical re- ting a frame, it is assumed that it does
sults through simulations. not take in new frames from the above
This paper considers a network consisting of layers.
a PAN coordinator. The number of sensor
nodes is fixed to be ‘n’. It is also taken 5. Data frames arrival is according to the
that the system is single hop system. The Poisson process. Also device is assumed
assumptions taken are: to be capable of storing only one frame.
1. Since the model is prescribed for per- 6. All nodes are homogeneous.
sonal uses, it is assumed that the de-
vices are within the range of successful For the analysis, a Markov model taking into
tranmission. consideration the key characteristics of the

14
IEEE 802.15.4 MAC protocols such as struc- compute. The paper then goes on to the
ture of the superframe, acknowledgment and derivations of the other four parameters us-
retransmission schemes is constructed. ing these known parameters that shall be ex-
Here the Markov chain states can be clas- plained later. This can then be extended to
sifies into three different varieties. The first every level in order to get the entire model.
variety is of the form {s(t), b(t), w(t), r(t)}
where s(t) establishes the number of back-
offs at time t. 5 Applications
Also b(t) represents the backoff counter’s Markov chains have been extensively used in
value at the time t, w(t) represents the num- the field of networks. It is predominantly
ber of clear channel assessments that are left applied in the World Wide Web, Queueing
to be done for the transmission at that point Network theory, and in the area of collision
and r(t) represents the value of retransmis- avoidance algorithms – details of which have
sion plane at time t. been already explored in the document.
In this state the parameters required by the In the World Wide Web, it has two ap-
Markov models are P1 and P11 which are the plications, namely, HTTP pre-fetching, and
probabilities of the channel being idle while the PageRank algorithm used by the Google
performing the first clear channel assessment search engine. In Queueing Networks,
and the second clear channel assessment re- Markov chains are used to determine perfor-
spectively. mance parameters of switching techniques,
The second type is to established the nonde- such as Packet switching and Circuit switch-
ferred and the postponed tranmissions in the ing.
s(t)th backoff stage on the r(t) transmission
plane at time t. Here the important param-
eters for the Markov model are Pcnd and Pcd
5.1 Queueing Networks
which denote the probabilities of the occur- In networking, delay in the network is caused
rence of a collision for the tranmissions men- due to a number of factors, and the aim is
tioned above. to minimize it. We should first understand
The final type is of the form {Ds(t), r(t)} what it is. Delay consists of the following
which represents the waiting state and helps components:
in computing the Pd which is the probabil-
ity that a postponed transmission can occur. 1. Transmission delay
An other important parameter is α is the
2. Propagation delay
probability of a situation where the device
is in transitioning states and a new frame 3. Queueing delay
is entering. As we can see from the brief
, two parameters i.e Pd and α are easy to 4. Processing delay

15
Figure 5: Markov model for the prescribed mechanism

Using a queuing model, we can accurately 5.1.1 Queueing theory


make quantitative predictions about the av-
erage packet delay in the network, as well as Networks of queues are systems with a num-
qualitatively understand the performance of ber of queues are inter-connected queues.
the network, without having to simulate all This means that entities can enter one queue,
the individual components of the network. receive a particular service, and then join
other queues, or leave the network. This can
be described using the example of customers
coming to a service center. Customers ar-
rive at the service center, wait in the queue

16
if necessary, receive service from the server,
and depart (if they have no other service to
receive). This presents a very basic queuing
model. Queueing systems are described by a
combination of letters (in the Kendall nota-
tion): Figure 6: Queue
A/B/C/D/E
• A = Describes the arrival process
• B = Service time distribution
• C = Number of servers
Figure 7: Markov Chain Model for the
• D = Maximum number of entities in the M/M/1 Queue
queue (capacity of the queue/system)
• E = Queue discipline (order in which In order to solve the continuous time mod-
entities are serviced in the queue) els, the balance equations associated with
each state must be solved.
Now, we consider how Markov Chains can
be applied to a queuing model in order to By doing so, we can determine various ex-
obtain more information about it. pected values for the network, including:
Consider a simple queuing system – M/M/1. 1. The probability of being in state k of
In this system, the arrival process, as well the Markov chain, ie, the probability of
as the service time is Markovian or mem- having k customers in the system:
oryless. This means that it is a Poisson
process with exponential distribution (effec-
tively random).
where ρ = the expected number of cus-
The M/M/1 queueing system consists of cus-
tomers being served.
tomers (following the example above) arriv-
ing randomly according to a Poisson process
2. Expected delay per customer:
with rate λ, and one server with a service
time per customer that is exponentially dis-
tributed with rate µ. We assume other values
(such as number of entities) to be infinite.
As the buffer may contain any number of
3. Expected waiting time in the queue
customers, the number of states is not fi-
nite. Also, the transitions occur in contin-
uous time.

17
4. Expected number of customers in the Consider the average packet size to be L
queue bits. The bandwidth of the link is µ bits
per second (bps). The average transit time
is equal to the packet size/bandwidth of the
link. Thus, average transit time is L/µ .
5. Expected throughput of the system Using the formulae described above we can
now determine various expected values in the
network.
1. Packet Switching
Application in network: The model de- In packet switching network, each ses-
scribed above can be directly applied to net- sion operates independently of the other
work processes. For example, packet switch- sessions, and there is no prior allocation
ing and circuit switching can both be mod- of link bandwidth. Therefore, the entire
elled as an M/M/1 queue system in order to bandwidth is shared by all the m ses-
calculate various expected network parame- sions.
ters. Expected delay per session:
Consider the arrival of packets in a network.

• The number of times a packet can ar-


Expected number of packets per session:
rive in a given time interval can be any
natural number, including 0, 1, 2 ...

• Arrival of one packet doesn’t affect the


arrival of another. Therefore, the arrival 2. Circuit Switching
of packets is an independent process. In circuit switching network, a link is
allocated to a session when it is set
If we assume that multiple packets don’t up, and is used by the session until it
arrive at the exact same time, and that the terminates. This means that each of
rate of arrival of packets in a network is a the m sessions are given only 1/m of
constant rate λ, then the necessary condi- the link bandwidth.
tions are satisfied. We can consider the ar-
rival of packets in a network to have Pois-
Expected delay per session:
son distribution, and we can model it using
a Markov Chain Model.
Consider the arrival rate of m sessions to be
λ/m. Packet sizes are also effectively ran- Expected number of packets (per ses-
dom, and have an exponential distribution. sion):

18
of which are impossible to go through. Thus,
a search engine will be of help only if it dis-
plays the results in order of relevance. Statis-
From the above, we can see that in cir- tics show that 91% of the users only browse
cuit switching, the expected number of pack- the links on the first page of the search re-
ets (N) is m times more than in packet sults. Around 5% take a look at Page 2, and
switching. The expected delay (T) is also the rest 4% go through the remaining pages.
m times more in circuit switching than in It is thus imperative for any search engine to
packet switching. order/rank their pages in a relevant manner.
This is a useful result that helps network
designers understand how the choice of a About PageRank algorithm designates
switching method affects network perfor- importance to each page by cumulating the
mance. number of links to it. It not only totals the
number of links, it also checks how credible
these links are. The credibility of a link is
5.2 Applications in the World
determined by the PageRank of the source
Wide Web page. PageRank makes this assumption
Markov chains have been used in the World because it is intuitive to see that the value
Wide Web in two major fields: link predic- of a page is high if a lot of other pages cite
tion, and search engine algorithms. it. Both number and quality matter. That
is, a page having, say, 20 links to it from
less credible sources is certain to rank lower
5.2.1 PageRank algorithm
as compared to a page that is cited by 5
History The PageRank algorithm was de- credible sources.
vised by Larry Page and Sergey Brin, To begin explaining the PageRank algo-
while they pursuing their Masters Degree rithm, we start with the random surfer
at Stanford University. They are the ones model. It states that any visitor is likely to
who founded Google, and consequently, the start their quest from any random page on
search engine implementes the PageRank al- the Web. Their journey on the Web how-
gorithm. To this date, this algorithm plays ever, continues mostly through hyperlinks
an important role in the Google search en- embedded in the pages. Selection of such
gine, setting it apart from the rest of its com- links is random, and thus is chosen with
petitors [namely Bing and Yahoo]. uniform probability.
Hence, if a page i has ni links embedded in
Importance With so many pages on the it, then the probability that one of them is
Internet, ordering of the search results is a clicked becomes 1/ni .
must. Search results yield millions of hits, all We define the quantity ‘PageRank’ for a

19
page i as the probability that the random Thus, the growth of x(k) can be expressed as
surfer will be at page i at any point of time.
Thus, PageRank is obtained by capturing
the random surfer model with a Markov x(k + 1) = Ax(k)
Chain. Iterating through the mathematical
expression several times leads to its station- The link matrix A = (aij )  Rnn is given
ary distribution, which is nothing but the by aij = 1/nj if (j, i)  E and 0 otherwise,
PageRank. where, nj is the number of outgoing links of
page j. Clearly, matrix A is the transition
To implement the PageRank algorithm, we
probability matrix of the random surfer. If
represent the World Wide Web in terms of a
we keep substituting x(k) with Ax(k-1), then
directed graph, G=(V,E), where V is the set
we finally get,
of vertices representing the webpages of the
Web, and E is the set of edges representing x(k+1) = Ak+1 x(0)
links between the pages. The node i is
connected to the node j by an edge, i.e., (i, As stated earlier, PageRank is the direct
j)  E, if page i has a link going to page j. result of the stationary distribution of this
Let xi (k) be the probability that the random Markov process. Let the PageRank vector
surfer is visiting page i at time k, and let be denoted by x*. Then,
x(k) be the vector containing all xi (k). That x* = limk→∞ x(k)
is, for every k, The value in this vector denotes the prob-
ability that the surfer would be at page i
at any random time. Thus, a page having
x(k) = [x1 (k), x2 (k), x3 (k). . . xi (k). . . xn (k)]
high PageRank shall have higher probability
and, ni=1 x(k) = 1
P
in this vector. It is the solution of the linear
(summation of probabilities is 1) equation:
(n is the overall number of webpages on the x* = Ax*
web) This is because of the definition of a station-
ary distribution: it does not change with the
The initial vector x(0) is a probability vector application of the transition matrix [that is,
in which xi (0) denotes the probability that x(k+1) = x(k)].
the random surfer starts his quest with page
In mathematics, this is equivalent to: 1x* =
i.
Ax* Where, 1 denotes the eigenvalue, and
After 1 step, the distribution becomes:
x* denotes the eigenvector.
x(1) = Ax(0) Definition: Strongly connected directed
graph: A directed graph is strongly con-
After 2 steps, the distribution becomes:
nected if there is a route between every pair
x(2) = Ax(1) of vertices.

20
Definition: A matrix is said to be irre- and qij = 1/N for every i,j.
ducible if and only if its associated directed β is known as the damping factor. In the
graph is strongly connected. An irreducible original PageRank algorithm,
matrix has a unique eigenvector for a given β = 0.15. That is, there is only a 15% chance
eigenvalue. that a webpage has been reached via a ran-
Considering the directed graph that repre- dom jump.
sents the World Wide Web, we see that it Since matrix M includes the term βQ, all
is not a strongly connected graph. This is the entries in M are positive. By Perron’s
because, the Web has billions of pages, with theorem, the eigenvector is unique for every
millions of them unrelated, and so it is im- eigenvalue in such matrices. We were only
possible to have a path between every pair of concerned by the uniqueness of the eigenvec-
nodes/pages. Thus, the link matrix A is not tor for eigenvalue 1, and now this holds for
reducible and thus the eigenvector x* is not the matrix M. Hence, we redefine the vector
unique. If x* is not unique, then each page x* by using M instead of A as follows:
can have multiple ranks depending on which x* = Mx*
eigenvector we use. This proves to be of no The matrix M is an N xN matrix [N = num-
help in giving accurate ranks to the pages.
ber of pages on the web], and hence its large
To resolve this issue, we need to modify the dimension makes x* difficult to compute.
random surfer model. This brings us to the There are several ways to compute x*, how-
Teleportation Model. The idea of the tele- ever, those methods are beyond the scope of
portation model is that the random surfer, at this report.
any time, may become disinterested and stop
hopping along hyperlinks. In such a case,
the surfer skips to another page not directly 5.2.2 Link Prediction
connected to the one currently visiting. This
Before getting into the concept of link predic-
page can in fact be completely unrelated in
the domains and/or the contents. If there tion, the concept of adaptive website needs
are N pages in the Web, then the probabil- to be reported. Adaptive websites are web-
ity to reach any page via this unpredictable sites that change themselves according to
the user patterns. One of the major actual
jump is 1/N.
use of adaptive websites are in e-commerce
To capture the two ways to reach a page – websites where other products are suggested
one via hyperlink (matrix A), while the other based on the present selected product. There
as a random jump (matrix Q), we define a are many ways to achieve this in a website.
new matrix M as follows: Another use of the link prediction is enhanc-
M = (1-β)A + βQ ing navigation. One of the most used effi-
cient methods used for navigating is search
where Q is an N xN matrix, with Q = (qij ) engines as of now. However these search en-

21
gines don’t help in giving us a set of related In this formulation, a Markov state can cor-
pages once a page is selected. There are cer- respond to an URL, HTTP request or any
tainly some alternate options available like action such as sending email. The matrix A
agent assisted navigation or tour generation. is being defined here as the following :
Agent assisted navigation is where the sys-
tem suggests what the user can do while tour A(s,s’) = C(s,s’)/Σs”C(s,s”)
generation is where the user is given a tour λ(s) = C(s) / Σs”C(s”)
of related pages. However the key here is
personalising the efforts done by the above where C(s,s’) is the count of the number of
techniques and this is what is achieved by times s’ follows s in the data. An element in
link prediction. Link prediction is basically the matrix A, say A(s,s’) can be interpreted
trying to pre-fetch the next links that are as the probability of transitioning from state
probable when the user is in the current link. s to s’ in one step. Similarly an element
of A*A will denote the probability of transi-
tioning from one state to another in two steps
Ways to achieve link prediction In this and so on. Given the link history of the user,
topic we will focus on how to achieve link we can represent each link as a vector with
prediction especially using Markov chains as the probability at that state for that time
that is our main topic. As we know every which is denoted by i(t-k), i(t-k+1 ). . . i(t-
markov model is characterised by transition 1). The Markov chain models estimation of
matrix and initial probability distribution of the probability of being in a state at a time
states. And this information can be used to t is shown as:
predict the next state in the link given a par- ŝ(t) = î(t-1)A
ticular state and the probabilities of the dif- This equation is then used to develop equa-
ferent transitions. And this is the major aim tions for transitions which are achieved
of link prediction. A discrete Markov chain through more than one steps. However that
model can be defined by the tuple (S,A, λ) shall be discussed in the final report.
where S corresponds to the state space, A is
the matrix representing the transition prob- Applications In this section we shall see
abilities from one state to another for all pos- four different application of link prediction
sible combinations and λ is the initial prob- using Markov chains in brief.
ability distribution of every state present in
S. As we have seen above the fundamental Web Server HTTP Request Predic-
property of the Markov model is the depen- tion The main aim for this application is to
dency on the previous state. So therefore if enhance the server performance.In the past
the vector s(t) denotes the probability vector the analysis of HTTP requests have always
for all the states at time t, then: included request patterns or statistical anal-
ŝ(t) = ŝ=(t-1)A ysis of file sizes. However here link modelling

22
can be applied and made user specific. Here the one above on path traversal patterns in
every client request can be processed by the order to list out a series of hubs that would
prediction system to list the probabilities of work for that particular user.
the next requests based on the history of re-
quests.
6 Conclusion
Adaptive Web navigation In this We hope to have established a fair account of
particular application, link prediction can be our understanding of the various aspects and
modelled for each user. Here once a user applications of Markov model. We aimed at
clicks on the link, the link predictor works giving a brief of how Markov models is used
on the server side and suggests a list of links to analyse performance aspects and also its
that the user can go next possibly and these practical application is areas such as adap-
links are ordered according to the probabil- tive websites and queueing networks. We
ity. This can be used to build a Navigation also want to conclude by stating that this
system that leads the user to different pages draft does not include the major derivations
based on the statistical data of the history. which we hope to give a clear view in the
final document.
Tour Generation For this particular
application, the Tour Generation Using
Markov Models algorithm has been imple- References
mented. Here the Markov model is used to
[1] V Bharghavan, A Demers, and S
predict a list of states that can be visited
Shenker. “MACAW: a media access
next. Every state should be marked visited
protocol for wireless LAN’s”. In: ACM
or unvisited in order to make sure there is no
SIGCOMM Computer Communication
cyclic actions. Also in case of same probabil-
Review 24.4 (1994), pp. 212–225.
ity, a method to choose the next best state
needs to be formulated. Here the first link [2] Giuseppe Bianchi. “IEEE 802.11-
with the longest matching prefix URL is cho- saturation throughput analysis”. In:
sen. And this process is implemented to get IEEE communications letters 2.12
the sequence of states ie URLs. (1998).
[3] Giuseppe Bianchi. “Performance anal-
Personalised Hub Hub refers to any ysis of the IEEE 802.11 distributed co-
link that can often be taken as good start- ordination function”. In: IEEE Jour-
ing states to find information. However these nal on selected areas in communica-
hubs need to be personalised in order to be of tions 18.3 (2000).
any use. Hence Markov models can be used
to generate an algorithm that is similar to

23
[4] G. Bolch et al. Queueing Networks and [11] David Malone, Ken Duffy, and Doug
Markov Chains: Modeling and Perfor- Leith. “Modeling the 802.11 dis-
mance Evaluation with Computer Sci- tributed coordination function in non-
ence Applications. John Wiley Sons, saturated heterogeneous conditions”.
2006. isbn: 9780471193661. In: IEEE/ACM Transactions on net-
[5] Cali et al. “IEEE 802.11 protocol: de- working 15.1 (2007), pp. 159–172.
sign and performance evaluation of [12] TR Park et al. “Throughput and en-
an adaptive backoff mechanism”. In: ergy consumption analysis of IEEE
IEEE journal on selected areas in com- 802.15. 4 slotted CSMA/CA”. In: Elec-
munications (2000), pp. 1774–1786. tronics Letters 41.18 (2005), pp. 1017–
[6] Nghia T. Dao and Robert A. Malaney. 1019.
“A new Markov model for non- [13] Park and Pangun. “A generalized
saturated 802.11 networks”. In: IEEE Markov chain model for effective anal-
Consumer Communications and Net- ysis of slotted IEEE 802.15. 4”. In:
working Conference (2008). 2009 IEEE 6th International Confer-
[7] Jung and Chang Yong. “Enhanced ence on Mobile Adhoc and Sensor Sys-
Markov chain model and throughput tems (2009).
analysis of the slotted CSMA/CA for [14] S. Pollin et al. “Performance analysis
IEEE 802.15. 4 under unsaturated traf- of slotted carrier sense IEEE 802.15.4
fic conditions”. In: IEEE Transactions medium access layer”. In: Proc. IEEE
on Vehicular Technology 58.1 (2009), GLOBECOM (2006), pp. 1–6.
pp. 473–478. [15] Ramachandran et al. “Analysis of
[8] Ching Wai Ki, Huang Ximin, and the contention access period of IEEE
Ng Michaelk. “Markov Chains: Mod- 802.15. 4 MAC”. In: ACM Transac-
els, Algorithms and Applications”. In: tions on Sensor Networks (TOSN) 3.1
(2013). (2007).
[9] Byung-Jae Kwak, Nah-Oak Song, [16] Ramesh R Sarukkai. “Link predic-
and Leonard E. Miller. “Performance tion and path analysis using Markov
analysis of exponential backoff”. In: chains”. In: Computer Networks 33.1
IEEE/ACM transactions on network- (2000).
ing 13.2 (2005). [17] N. Song et al. “Enhancement of IEEE
[10] Tae-Jin Lee, Hae Rim Lee, and Min 802.11 distributed coordination func-
Young Chung. “MAC throughput limit tion with exponential increase expo-
analysis of slotted CSMA/CA in IEEE nential decrease backoff algorithm”.
802.15. 4 WPAN”. In: IEEE Commu- In: The 57th IEEE Semiannual Spring
nications Letters (2006), pp. 561–563.

24
Vehicular Technology Conference Vol.4
(Apr. 2003), pp. 2775–2778.
[18] Zhifeng Tao et al. “Performance anal-
ysis and a proposed improvement for
the IEEE 802.15. 4 contention access
period”. In: IEEE Wireless Commu-
nications and Networking Conference,
2006 4 (2006).
[19] K. S Trivedi. Probability Statistics
with Reliability, Queuing and Com-
puter Science Applications. John Wi-
ley Sons, 2008. isbn: 9780471460817.
[20] Vukovic, Ivan N., and Natt Smavatkul.
“Delay analysis of different backoff al-
gorithms in IEEE 802.11”. In: Vehic-
ular technology conference, VTC2004-
Fall 2004 IEEE 60th Vol.6 (2004).
[21] Ziouva, Eustathia, and Theodore
Antonakopoulos. “CSMA/CA perfor-
mance under high traffic conditions:
throughput and delay analysis”.
In: Computer communications 25.3
(2002).

25
Members Contribution

You might also like