Professional Documents
Culture Documents
Sciences
Department of Electrical Engineering
Farah Nadeem
Degree of MS Electrical Engineering
2014
Advisor
Dr. Arshad Hussain
Certificate of Approval
It is certified that the research work presented in this thesis, entitled IEEE 802.11 EDCA
Steady State Modelling as Non-Preemptive Priority Queues: Delay and Throughput
Analysis was conducted by Farah Nadeem under the supervision of Dr. Arshad Hussain.
No part of this thesis has been submitted anywhere else for any other degree.
This thesis is submitted to the Department of Electrical Engineering in partial fulfillment of
the requirements for the degree of Master of Science in Electrical Engineering
at the
National University of Computer and Emerging Sciences
Lahore, PAKISTAN
June, 2014
Candidate Name: Farah Nadeem
Signature: ______________________
Examination Committee:
a) Name: Dr. Arshad Hussain
Signature: ______________________
Professor
Dept. of Electrical Engineering
National University of Computer & Emerging Sciences
b) Name: Mr. Shafiq-ur-Rahman
Signature: ______________________
Associate Professor
Computer Science Department
National University of Computer & Emerging Sciences
c) Name: Dr. Asim Loan
Signature: ______________________
Professor
Dept. of Electrical Engineering
University of Engineering and Technology, Lahore
Dr. Arshad Hussain_____________________________________________________
Head of Department of Electrical Engineering, National University of Computer and Emerging
Sciences, Lahore Campus
Authors Declaration
I, Farah Nadeem, Roll No. 12L-5102, certify that the work is original and has not been
previously submitted for assessment in any other course or institution, except where specifically stated.
Farah Nadeem
Roll Number: 12L-5102
ii
Plagiarism Undertaking
I, Farah Nadeem, Roll No. 12L-5102, certify that the ideas, experimental work, results
and conclusions set out in this thesis are entirely my own effort under the guidance of Dr.
Arshad Hussain, except where otherwise indicated and acknowledged.
Farah Nadeem
Roll Number: 12L-5102
iii
Abstract
In a bid to deliver the stringent quality of service (QoS) guarantees to real time multimedia applications, the IEEE 802.11e enhanced distribution channel access (EDCA) has
been incorporated in the original IEEE 802.11 standard. Both simulations and analytical
modeling have been extensively used by the research community for studying the service
differentiation mechanism of the EDCA. Provision of requisite delay and throughput for
real time traffic entails differentiated parameters including contention window size, nonuniform back-off and channel access time. Given these features, most analytical models
do not accurately capture EDCA performance. In addition, more accurate models are
complex, often employing techniques including fixed point iteration, which needs to be
evaluated from scratch for each new scenario. This thesis builds an analytical model that
accurately captures the differentiation mechanism of IEEE 802.11 EDCA. Each access category has been modeled as a single dimension M/G/1 non-preemptive priority queue. The
basic purpose of this model is to provide service delay. Including the effect of contention
free bursting, a framework has been developed for the medium access (MAC) layer delay
and further, the aggregate queuing delay for real time traffic in saturation and non-ideal
channel conditions. The throughput has also been calculated. The framework does not
require iterative solution. Results show close agreement between the analytical model
and simulation. This model has the capability to provide a solid foundation for effective
admission control, since accurate delay prediction, in addition to throughput, is required
for effective resource utilization when additional streams are admitted.
iv
Acknowledgments
It has been a privilege to work under the guidance of Dr. Arshad Hussain (Professor &
Head of Department of Electrical Engineering, NUCES, Lahore). I would like to thank
him for his invaluable guidance, patience and motivation.
I would like to extend my gratitude to Professor Shafiq ur Rahman (Head of Department of Computer Sciences, NUCES Lahore) who gave his valuable time for evaluation of
my work. I would also like to thank Dr. Bushra Anjum (Department of Computer Science,
NUCES Lahore) and Dr. Asim Loan (Department of Electrical Engineering, University
of Engineering and Technology, Lahore) for their insightful feedback. In addition, I would
like to acknowledge the valuable review and suggestions made by Dr. Imran Cheema
(Department of Electrical Engineering, NUCES, Lahore).My special thanks to my friends
and family, in particular my son, Mustafa. Their constant encouragement and support
has made this milestone possible.
Contents
Certificate of Approval
Authors Declaration
ii
Plagiarism Undertaking
iii
Abstract
iv
Acknowledgments
Notation
1 Introduction
2.1
Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2
Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1
3.1.2
3.2
3.3
3.4
Prior Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4.2
G/G/1 Queues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
18
4.1
4.2
4.3
4.4
4.3.1
E[Vbo ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3.2
E[Vbusy ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
vii
CONTENTS
4.5
Steady State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.5.1
4.6
E[Vbusy ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
29
5.1
Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2
Computational Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3
5.4
5.3.1
5.3.2
Retransmission Probability . . . . . . . . . . . . . . . . . . . . . . . 32
5.3.3
Vacation Time V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.3.4
Delay Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.3.5
Throughput Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3.6
5.3.7
5.4.2
Retransmission Probability . . . . . . . . . . . . . . . . . . . . . . . 38
5.4.3
Vacation Time V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4.4
Delay Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4.5
Throughput Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.4.6
5.4.7
5.5
5.6
. . . . . . . . . . . . . . . . . . . . . 43
5.6.1
System Throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.6.2
45
6.1
. . . . . . . . . . . . . . . . . . . . . . . 48
6.2
55
60
List of Figures
3.1
4.1
5.1
First Moment of Busy Time for VO AC: a)10ms intervals, b)20s intervals . . . . . 31
5.2
Second Moment of Busy Time for VO AC: a)10ms intervals, b)20s intervals . . . . 31
5.3
5.4
5.5
5.6
5.7
Throughput Results VO AC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.8
5.9
. . . . . 32
. . . . 34
5.10 First Moment of Busy Time for VI AC: a)10ms intervals, b)20s intervals . . . . . 37
5.11 Second Moment of Busy Time for VI AC: a)10ms intervals, b)20s intervals . . . . 37
5.12 Retransmission Probability for VI AC: a)10ms intervals, b)20s intervals . . . . . . 38
5.13 First Moment of Vacation Time: a)10ms intervals, b)20s intervals . . . . . . . . . 39
5.14 Second Moment of Vacation Time: a)10ms intervals, b)20s intervals . . . . . . . . 39
5.15 Delay Results: Analytical Vs Simulation: a)10ms intervals, b)20s intervals . . . . 40
5.16 Throughput Results for VI AC . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.17 Effect of Retransmission Probability on System Throughput . . . . . . . . . . . . 41
5.18 Jitter Results: Simulation: a)10ms intervals, b)20s intervals . . . . . . . . . . . . 41
5.19 Comparison with Tadayons Model . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.20 Effect of Retransmission Probability on System Throughput VO AC . . . . . . . 43
5.21 Effect of Retransmission Probability on System Throughput VI AC . . . . . . . . 44
5.22 Effect of Retransmission Probability on Delay and Jitter VO AC . . . . . . . . . 44
5.23 Effect of Retransmission Probability on Delay and Jitter VI AC . . . . . . . . . . 44
6.1
6.2
6.3
2
E[Tbusy
] under heavy loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.4
viii
List of Tables
3.1
5.1
MAC Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2
EDCA Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3
Traffic Specification
6.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
ix
Notation
AC
Access Category
ACK
Acknowledgment
AIF SV I
AIF SV O
BE
Best Effort
BK
Background
CAC
CBR
CDF
CSM A/CA
CWmax
CWmin
DCF
DIF S
DCF IFS
EDCA
EIF S
Extended IFS
F IF O
IF S
Inter-Frame Space
Ni
nT XOP
M AC
pV I
pV O
PCF
P DF
PHY
Physical
QoS
Quality of Service
Ri
NOTATION
RTP/UDP/IP
RTS/CTS
RV
Random Variable
Tbusy
TM AC
Tno ack
Tslot
Slot Length
TCP/IP
T XOP
Transmission opportunity
Vacation Interval
VV I
VV O
Vbo
Vbusy
VBR
VI
Video
VO
Voice
VoIP
Voice over IP
Wi
WmV I
WmV O
Yi
Chapter 1
Introduction
It is evident that wireless communication technology has enjoyed unprecedented success
both in terms of deployment and development. One of the most popular examples is Wi-Fi,
which is pervasive across professional and personal usage alike. Originally, the evolution
of Wi-Fi was to support data traffic, while voice, and later video traffic, were handled
by dedicated cellular systems. However, the focus has been shifting towards providing
support for heterogeneous traffic, both by cellular and Wi-Fi networks. This presents
challenges when it comes to providing adequate quality of service to users.
The Wi-Fi standard is also known as the IEEE 802.11. Given the mass deployment and
popularity of the IEEE 802.11, extensive effort has gone into providing a solid framework
for quality of service (QoS) provision for real time traffic. A major contribution towards
this end is the IEEE 802.11e [1]. Like the legacy IEEE 802.11, the 802.11e provides both
a centralized medium access mechanism, and a decentralized mechanism, the enhanced
distributed channel access (EDCA). Service differentiation is introduced to provide higher
priority to voice and video traffic in a bid to ensure greater throughput and smaller delays,
so as to provide better performance.
While the 802.11e provides a basic framework for QoS, to ensure acceptable performance, particularly under heavy loads, effective admission control is required. For this
purpose, accurate delay prediction is required for voice and video streams. While extensive
work has been put into analytical modeling for delay prediction, less focus has been on
developing tractable models for real time applications. Thus there is a dearth of simpler
models for run time admission control.
Given the nature of the delays that occur due to contention for channel access and
actual transmission, it is possible to model this process as a single dimensional queue.
The time for contention can be viewed as a reservation (vacation) interval from the perspective of classical queuing theory, which precedes each packet transmission. While the
differentiated channel access parameters are designed so as to provide higher priority to
voice and video, the vacation interval of each access categorys (ACs) (higher or lower
priority) queue overlaps the transmissions of other ACs; thus the priority is catered for
3
CHAPTER 1. INTRODUCTION
Chapter 2
2.1
Problem Statement
There has been a great focus on deriving analytical models to accurately capture the
service differentiation mechanism of the IEEE 802.11 EDCA and translate it to accurate
results for delay and throughput. The best case scenario has always been at making a
descriptive model based solely on the number of active stations within an area. This has,
however, lead to very complex models that become computationally intractable. Even at
very high complexity, these models fail to capture all results of the EDCA. This thesis has
addressed this gap in analytical models by providing a queuing theory based framework
for analysis of service delay, over all waiting time, and throughput results for real time
voice and video access categories. The model presents the access categories as single
dimensional non-preemptive priority queues. The model caters to both the channel access
differentiation parameters (CW and AIFS) as well as TXOP. Theoretically, this model
presents a tractable and complete framework for analysis of performance metrics at the
MAC layer of the enhanced distributed channel access.
2.2
Contributions
Most analytical models for throughput and delay results are based on three or four dimensional Markov Chain models, followed by standard queuing models. The alternate
approach has been to develop very complex probability distributions for the service time
to facilitate calculation of waiting time and delays. In this context, the main contribution
of this work has been to reduce the framework for delay and throughput analysis for EDCA
to single dimension queues. Although this approach presents a reduction in complexity
as compared to most existing models, the entire set of features of the EDCA is catered
5
for. Thus this model is novel in terms of decreased complexity and greater accuracy.
A novel contribution of this work has been the application of the memoryless property
to arrive at elegant expressions for vacation time for EDCA. While various other works
have applied the description of a memoryless process, it has been angled to arrive at a
Markov chain model. In this context, a recent work presented by Bianchi [2] indicates that
this may not be the accurate case. An alternate approach is presented here that offers a
comprehensive analysis utilizing the memoryless nature of the transmission/retransmission
process.
Real time traffic requires guarantees in terms of throughput, delay and jitter. The
model developed in this thesis helps to highlight the effect of erroneous transmission on
these metrics. From the developed mathematical framework, it becomes apparent that
beyond an error probability of 0.25, the expected value of delay becomes infinite. From
here, it can be extended to the case of jitter: beyond a value of error probability of 0.0625,
the expected value of jitter becomes infinite. This presents a limit on the probability of
packet retransmission caused by both collisions and channel impairments. While this fact
seems to predict severely limited service in case of heavy load, simulations show that this
is not the case. Since the protocol employed is carrier sense multiple access with collision
avoidance (CSMA-CA), the process of back-off window initiation ensures that the error
probability settles down after a short period of high packet transmission errors. Thus the
protocol ensures that the retransmission probability remains within bound; however, the
increasing back-off window size means that the delay increases.
Chapter 3
3.1
Background
The details of the MAC layer for IEEE 802.11 are presented in this section. Primary focus
is on the binary exponential back-off process of the collision avoidance process, which
lends itself to analytical modeling. The inherent nature of this protocol dictates that the
service time is probabilistic. This feature has been exploited for the purpose of creating
frameworks for analysis. In addition, the importance of the MAC layer delay is also
given, which is the aggregate of the service time and the queuing delay under saturation
conditions. Service delay at the MAC layer is defined as the time that a packet arrives at
the head of queue till it is either transmitted with a corresponding ACK (acknowledgment)
received, or it is dropped due to expiration of retry limit.
3.1.1
Being designed for the wireless medium, legacy DCF builds upon carrier sense multiple
access/ collision avoidance (CSMA/CA). Although a centralized medium access mechanism, point coordination function (PCF), is also specified by the original standard, the
decentralized mechanism has gained wider applicability. This is due to simplicity, ease
of implementation and support for flexible network configurations. Collision avoidance is
achieved through the back-off algorithm and the inter-frame space (IFS) inset [1].
The back-off algorithm is implemented as follows: when a packet arrives at the head
of queue, if the medium is sensed idle for a DCF IFS (DIFS), it is transmitted. If the
7
medium is sensed busy, the back-off process is initiated, whereby the station randomly
picks a back-off counter value from the interval [0,CWmin]. The counter is decremented
every time the medium is sensed idle for a DIFS, frozen otherwise. When the counter
attains a value of zero and the channel is idle, the packet is transmitted. The correct
reception of the packet is confirmed via an ACK from the receiving station. In case more
than one station transmits simultaneously, there is a collision on the medium. The collision
is sensed by all the listening stations by the erroneous check-sum. When the transmitting
station senses the collision, it doubles the contention window size and repeats the back-off
process. For each subsequent collision, the contention window is doubled till it reaches the
maximum contention window size. The packet is dropped if the transmission retry limit
is reached and the packet suffers a collision. To ensure prompt sending of ACK when a
packet is received correctly, the receiving station waits a short IFS (SIFS) before transmitting. This precludes channel contention. In addition to the ACK, a four way hand shaking
procedure Request to send/Clear to send (RTS/CTS) is also offered by the mechanism.
This is particularly useful for larger packet sizes, but induces greater overheads for smaller
packet sizes.
Here it is pertinent to mention that packet loss occurs from both collision and channel
impairments. From the perspective of a station, both are treated the same, and are not
differentiable. Thus, for the purpose of this analysis, the term error probability refers to
both channel outage and collision. The expected retransmission probability incorporates
both the probabilities of collision and channel imperfections.
p = 1 (1 pch )(1 pcoll )
(3.1)
Where p is the retransmission probability, pch is the channel outage probability and pcoll
is the expected collision probability.
The entire traffic originating from a station forms a single first out first in (FIFO)
queue. Hence all traffic streams have equal channel access priority. To provide real time
traffic flows with higher priority, EDCA introduces service differentiation based on the
nature of traffic; the arriving traffic is categorized into four access categories (ACs), VO
(voice), VI (video), BE (best effort) and BK (background), in descending order of priority.
Each AC then contends independently for the medium. Channel access parameters are assigned different values for each AC; thus each AC has a different channel access priority [1].
The parameters employed to this end are the arbitrary IFS (AIFS) instead of the
uniform DIFS, different minimum and maximum contention windows, and an additionally
defined parameter transmission opportunity (TXOP). TXOP, when set to a non-zero value,
allows the channel to transmit multiple packets in succession once it has gained medium
access. Channel access differentiation ensures delay requirements are met, while TXOP
additionally guarantees greater throughput. To increase the probability of greater channel
access to colliding stations immediately following a collision, an extended IFS (EIFS) is
defined. The concept of virtual collision caters to the scenario when two ACs within a
station simultaneously decrement their back-off counters to zero; the higher priority AC
is given channel access, while the lower priority AC undergoes the post collision back-off
process. The different EDCA intervals are illustrated in fig 3.1.
To achieve the necessary service guarantees, values of channel access parameters are
assigned so that higher priority ACs have a higher channel access probability than lower
priority ACs. These parameters are the arbitrary IFS (AIFS) instead of the uniform DIFS,
and differentiated CWmin and CWmax . In addition, TXOP is defined for each category.
This enables contention free bursting for an AC once it has obtained channel access, for
a duration that does not exceed the TXOP limit for that AC. The AC that has stringent
delay requirements, VO, is assigned smaller AIFS, CWmin and CWmax than other ACs.
Since the VI AC has the greatest throughput requirement, it is provided longer medium
access through larger TXOP limit. In addition, an extended IFS (EIFS) is defined by the
standard. The purpose is to provide the stations that were involved in a collision greater
likelihood to access the channel as compared to other stations in the post collision period.
When two ACs within the same station simultaneously decrement their back-off counters
to zero, a virtual collision occurs. Channel access is granted to the higher priority stream
and the lower priority stream undergoes the post-collision back-off process, with doubled
CW (up to CWmax ).
3.1.2
The MAC layer delay is the aggregate of service time and queuing delay. It is an important
performance metric, particularly from the perspective of real time multimedia applications.
Effective algorithms for admission control require that the delay for such applications is
kept within bound when new sessions are admitted within the existing resources. The
importance of the queuing delay is manifest. MAC frames transmit packets from upper
layers including Transmission Control Protocol / Internet Protocol (TCP/IP) or RealTime Transport Protocol/ User Datagram Protocol/Internet Protocol (RTP/UDP/IP).
Higher layer protocols are indifferent to the working of the MAC layer, and will observe
the total delay the packet is subjected to, but not the different causes associated with
it [3, Pp. 623]. Consequently, both the queuing delay and the medium access delay are
necessary for meaningful analysis of the delay of 802.11e EDCA.
10
In unsaturated cases, with less generated traffic, the average length of queue is usually
less than a packet, and the medium access delay is dominant. This case represents a
scenario where there are channel resources freely available to cater for the traffic, hence
ensuring QoS is not an issue. An addition to that, the medium access delay then is usually
in the order of few milliseconds, and is not significant in the context of overall end-to-end
delay experienced for Internet communication. When the channel becomes saturated,
features of 802.11e EDCA for service differentiation become significant. In cases when
saturation condition is reached, and the inter arrival time becomes comparable to the
mean service delay, the queuing delays begins to grow dramatically, while the service
times attains a somewhat constant value (dependent on the number of active stations) [4].
In this scenario, both the considerations of finite buffer size, and expiration of a packets
useful life due to extensive delay, come into consideration.
3.2
A two dimensional Markov chain model for binary exponential back-off in DCF was originally proposed by Bianchi [5], and forms the basis of subsequent models proposed for both
DCF and EDCA. Bianchi models the exponential back-off for the DCF using a discrete
time two dimensional Markov chain. One dimension caters to the back-off counter value,
which is non-Markovian since it is dependent on the number of retries suffered. Thus the
second dimension describes the back-off stage. A significant contribution of this model
is a constant collision probability, p which stems from the decoupling assumption. A
point of interest is the time scale adopted; the interval between consecutive time slots may
include the event of back-off counters freezing, this indicates that the interval is likely to
span an entire packet transmission. Thus the time between two consecutive time slots does
not relate directly to the system time. The model yields an elegant expression for transmission probability, since the constant collision probability allows for Markovian analysis.
In case the retry limit is not considered, as in this model, the resulting expression is closed
form. In actual DCF and EDCA, the probability of decrementing of back-off counter
and transmission is conditional on the state of the channel. This has not been explicitly
catered for in the model.
In [2], Bianchi presented a thorough and accurate framework for analysis of EDCA
throughput. Instead of a constant collision probability assumption, the modeling relies on
the evolution of the back-off counter with time. This ultimately leads to a bi-dimensional
stochastic process which models the back-off counter and the retransmission stage. This
is, however, not a Markov chain. Each AC [k = 0, 1, 2, 3] is modeled independently so as
to incorporate the service differentiation mechanism. The system is defined as having c
ACs, i.e. k = 1, ..., c. There are nk stations of each AC. The modeling for each AC starts
with the evolution of the back-off counter value, which is denoted by stochastic process
bk (.) Defining k as the difference between the kth AIFS (AIFS of AC k) and the minimum
11
(3.2)
Over time, the back-off counter value attains a steady state distribution, which is represented as:
Bk (j) = lim P bk (t) = j
t
(3.3)
Where the cumulative distribution function (CDF) of the back-off counter value is given
by:
(
k (i) =
0
Pi1
i0
s=0 Bk (s)
otherwise
(3.4)
The model hinges around the steady state distribution of the back-off counter value. From
the perspective of one particular station h, the probability (Qh (i)) that no other station
transmits before the ith time slot is given by:
Qc
Qh (i) =
k (i k )]nk
1 h (i h )
k=1 [1
(3.5)
Probability Th (i) that h sees the first transmission at time slot i is given by Th (i) =
Qh (i) Qh (i + 1).
From there on, the model moves to the bi-dimensional stochastic process representing
the retransmission stage sk (t), sk 0, 1, ?Rk (where Rk is the retransmission limit for AC
k) and the back-off counter value bk (t), bk 0, 1, ...CWSk (where CWSk is the contention
window at the Sk retransmission stage). Steady state transition probabilities expressed
in terms of Th (i)&Qh (i), as opposed to p in case of DCF. Defining the steady state
distribution of the bi-dimensional process sk (t), bk (t):
Y
k
(3.6)
This in turn is used to find the steady state distribution for the back-off counter:
Rk Y
X
Bk (j) =
(s, j)
(3.7)
s=0 k
The set of above equations is solved recursively using fixed point iteration until convergence is achieved. The model is then used to demonstrate that variation in the collision
probability. The model has been used for throughput performance. This work presents a
more accurate model for the EDCA with a computational complexity comparable to that
of multi-dimensional models. The assumptions made in this work include ideal channel
conditions, saturation condition, no internal collisions and TXOP disabled. The significant
contribution of this work has been to illustrate that the collision probability no longer remains constant, but changes significantly with each passing time slot after a channel event.
The results presented in the paper depict close agreement between the analytical model
and the simulated results for throughput performance.
3.3
12
Since the decentralized mechanism has enjoyed greater success, EDCA has been the focus
of several analytical models, which have been introduced to study its effectiveness in providing QoS support, particularly in terms of throughput and delay.
Extensive research has been conducted on the QoS support provided by the EDCA.
Since EDCA aims at providing support for real time traffic, the primary focus has been on
studying throughput and delay. A survey of models existing in literature provides insight
into the research directions and the limitations of existing works. This section explores
the prominent contributions made to date in the area of delay and throughput analysis
for the purpose of efficient admission control. This includes analytical modeling of EDCA
for throughput; modeling of service delay for DCF and EDCA and extensions to queuing
delay for overall MAC layer delay.
Various analytical models have been put forward to predict throughput in different
network scenarios; often these frameworks use bi-dimensional Markovian chain to model
EDCA. For the analysis of delay, which is particularly important for voice and video traffic, service time distribution for DCF and EDCA has also been modeled in literature. In
saturation conditions, however, the service delay solely does not depict the service provided, since queuing delay becomes a dominating factor as traffic loads increase. Hence
queuing analysis has been conducted based on service delay distribution, since it has direct
bearing on the QoS of delay sensitive applications. Primarily, the area of queuing delay
has been explored for DCF.
The main purpose behind these models is to provide a solid basis for analysis of the
EDCAs service differentiation mechanism under steady state conditions. Thus they are
theoretically rigorous and achieve a high level of accuracy. To accurately capture the
nature of the binary exponential back-off and the AIFS differentiation mechanism, existing models are usually very complex. The basis is generally multi-dimensional Markov
chain models [68] and bi-dimensional stochastic process [2]. Even with this basis, several
assumptions have to be made to make the analysis possible, the foremost being an assumption of a steady network with fixed number of active stations. Generally, in the case
of wifi networks, this is not the case, since the number of active stations changes frequently.
A comprehensive expression for the pdf of service delay has been derived in [9] for
DCF. The expression relies on a steady state analysis based on a bi-dimensional Markov
chain model, which makes it difficult to extend for EDCA. Also, under heavy traffic loads,
the delays are under-predicted; in admission control, this would yield unacceptable delays.
The basic analytical model has also been extended for overall delay using queuing
analysis. This work, while extensively done for DCF, has been less explored for EDCA.
A main contribution in this regard is the work presented in [10]. Comprehensive analysis
13
has been conducted, however, as most other analytical models, the underlying assumption
is that of steady state network. In terms of results, the model severely over-predicts the
delays.
The main thrust of analytical modeling of the performance of 802.11 EDCA is on
estimation of throughput, frame dropping rate and the average delay of MAC layer. In
EDCA, AIFS differentiation leads to collision probabilities that differ during different time
slots following any channel event (transmission, collision etc.). In order to capture this
effect, the constant collision probability can no longer be applied. However, this then takes
away the desirable tractability of the Markovian analysis. In order to retain this analysis,
EDCA has been modeled along the same lines as the bi-dimensional Markov model, incorporating more states to model the differentiation parameters [7, 8, 11]. These models are
increasingly complex, and build upon an assumption that is violated in actual EDCA. A
significantly different approach has been adopted in Bianchis model presented in section
3.2.
The service time (delay) represents a probabilistic distribution. Attempts have been
made to model this distribution; building upon analytical frameworks, service delay has
been studied in several works [9, 10, 12]. In [12], the delay distribution for DCF has been
derived based on the Markov chain model. A subsequent conclusion drawn in this work
has been to approximate the service time as a Poisson distribution.This work has been
extended in [10] for EDCA. In a recent work by Tadayon et al. on DCF [9] the probability generating function (PGF) of service time distribution has been derived, which in
turn yields the first and higher moments of service delay, incorporating imperfect channel
conditions. However, both the first and second moments of service delay, as predicted by
the analytical model, fall behind the simulation results under high traffic conditions.
While Markov chain models are abundant in literature, there has been less focus on
predicting the queuing delay. Given four different service time distributions and four different inter arrival processes, the EDCA represents a multi-class system. Existing queuing
models in literature treat each AC as an independent single server queue. While extensive work has not been done for EDCA, queuing delay of DCF has been derived using
different queuing disciplines. These include, in order of increasing complexity: M/G/1
[1214], M/GI/1/K [15], G/M/1 [10] and G/G/1 [8, 10, 16]. The disciplines M/G/1 and
G/G/1 interpret the 802.11 MAC as a single server which follows a general distribution
(in terms of service time). For meaningful queuing analysis, either the service distribution
itself is needed, or estimates of the first two moments of the distribution are needed. An
alternative approach is to approximate the service time with a distribution that is easier
to analyze, particularly Poisson [17]. An alternate approach for finding the service delay
entails finding the average of the back-off counter [9], and modeling the behavior of the
back-off counter as it evolves over time. The average value of the back-off counter represents, in the steady state, the number of non-transmitting slots that a station undergoes
14
The most comprehensive queuing analysis for EDCA has been in conducted by Chen
et al. [10]. Being one of the few results presented for EDCA queuing analysis, this model
sets a benchmark. The probability distribution for the service delay has been derived
from the Markov chain model. The model relies on constant collision probability, and
approximates the service distribution by treating the model as a signal flow graph. Based
on an earlier work on DCF [12], the service time distribution has also been modeled as
a Poisson distribution, in addition to a general distribution. Since the service time has
either been taken as general or approximated using Poisson, this has led to either the
G/G/1 or G/M/1 discipline. However, in both cases, the results diverge significantly from
the simulation results, particularly in the case of higher traffic loads. This presents a contradiction in terms of the approximation used; the approximation should become exact as
traffic load increases, instead of diverging. The key contributions and assumptions of the
above models are summarized in table 3.1.
Problem Addressed
Assumptions
Methodology
Results
Saturation
conditions;
Imperfect
channel;
Constant
collision probability
Both
saturation
and non-saturation
conditions; Perfect
channel;
Constant
collision probability;
Infinite buffer size;
TXOP disabled; No
virtual collisions
First and second delay of service delay
from bi-dimensional
Markov
chain;
G/G/1 and G/M/1
queuing disciplines
Analytical
model
significantly underpredicts first and
second moment of
service delay under
heavy traffic
Both
analytical
models significantly
over predict delay, particularly at
higher traffic loads
15
The focus on steady state results has left a gap in terms of analytically tractable
models for the purpose of real time admission control. This has provided motivation
for developing a simpler model for delay analysis, which is based on a single dimensional
queue for each access category (AC). The idea is that a simpler model, that does not make
limiting assumptions about the network, can provide a better framework for estimating
performance metrics such as delay and throughput in real deployment scenarios.
3.3.1
Prior Research
In the context of analytical modeling of EDCA and call admission control for the same,
two works have been previously conducted [17, 18] (Appendix A and B). The analytical
model presented in [18] is based on a multidimensional Markov chain model, and aims at
capturing AIFS differentiation at less complexity. The results are comparable with models
of greater intricacy; this can be explained based on the fact that even four dimensional
Markov chain models can only go so far in capturing all the features of the EDCA. Thus
added states only cause a very small increase in analytical accuracy. In this context, the
existing work aims to simplify the problem to a more tractable single dimensional queuing
model.
The measurement based call admission control algorithm presented in [17] relies on
the instantaneous error probability to make decisions regarding adding new streams to
the system. While this simple algorithm provides favorable results in terms of balance
between network utilization and QoS, the system still admits greater flows than optimal.
In this regard, there is a need to introduce better criteria for admission control, particularly
in terms of the bound on probability of error/retransmission for real time access categories
and data traffic. The existing work builds on these deficiencies, incorporating other factors
that effect system performance in addition to error probability. What is of particular
interest is the bound on error probability that guarantee acceptable quality of service for
real time traffic such as delay, jitter and throughput. While in [17], the upper limit on
error probability have been picked as a qualitative measure, an effort has been made in
this work to quantify these limits based on analytical results.
3.4
The purpose of queuing analysis is to model the delays and lengths of queues. In the context of data networks, this provides a powerful tool for predicting delays. For the queuing
model to be accurate, traffic models are required which effectively describe the statistical
nature of the actual network traffic. Thus the overall results of the queuing analysis are
dependent on the selection of appropriate traffic models. Selection of traffic models, in
turn, is dependent on the nature of the traffic and the type of network under consideration
[19].
16
The two main parameters describing network traffic are: packet length and packet
inter-arrival distributions, the latter of which is of prime importance.Commonly used
traffic models for real time and best effort traffic include [20]:
Video traffic: Variable Bit Rate (VBR) video traffic is bursty, and shows strong
correlation. Applied models include On/Off traffic model, Markov models including
Markov modulated Poisson process.
Web Traffic: On/Off sources, Poisson, fluid models.
Voice traffic: Constant Bit Rate (CBR)
While voice traffic presents the simplest case, being constant bit rate, video and data
traffic pose a greater challenge. However, as indicated in [19, 20], Poisson arrival lends
well to analysis in queuing systems, and for most cases, gives a fairly accurate model for
most types of internet traffic.
Kendalls notation [21, Pp. 17] is used to describe and classify the queue. The standard
description using three factors is A/S/c, where A denotes the inter-arrival time, S denotes
the service time distribution and c denotes the number of servers. Additional factors may
be used to denote the buffer size, size of population to be served. When just three factors
are specified, then it is assumed that the buffer size and population being served are being
served are infinite and the queuing discipline is FIFO.
For the purpose of our analysis, the most important arrival and service distributions
are Poisson (or Markovian), denoted by M and general, denoted by G, and the system is a
single server system, i.e. c=1. The key parameters of interest in queuing analysis are the
mean waiting time in queue (W), total waiting time, i.e. queuing delay and service time
(T) and number of waiting packets in queue (N) [22, Pp. 153].
These parameters are defined in terms of the mean traffic arrival time E[t] = 1 , service
time with mean E[X] =
1
,
utilization factor = . In addition, two factors that are also employed are the coefficient
of variance of the service time, Cb2 =
b2
E[X 2 ]
a2
E[t2 ]
A fundamental relation is established between the waiting time and queue length by the
Littles Theorem [22, Pp. 152] N = T .
3.4.1
The classic queuing theory utilizes the M/G/1 and the G/M/1 queues abundantly. The
assumption is that either the arrival process or the service time is Markovian; the other
distribution is general. These queues are duals of each other, thus the same analysis with
transformation is applicable to both queues. The Pollaczek-Khinchin (P-K) Formula is the
best descriptor for these queues. Coming to the M/G/1 queue, where the arrival process is
17
Markovian, and the service distribution is general (with a finite first and second moment),
the mean waiting time is given as:
E[T ] =
E[X 2 ]
2(1 )
(3.8)
From Littles theorem and 3.8, the mean number of waiting packets in the system is
given by:
E[N ] = +
2 E[X 2 ]
2(1 )
(3.9)
From the above equation, we see that there is a linear increase in N with variance of
service time. The average time spent in system by a data packet is given as:
E[W ] = E[X] +
E[X](1 + Cb2 )
2(1 )
(3.10)
3.4.2
G/G/1 Queues
When both the arrival times and service times are general independent and identically
distributed, the relevant queuing discipline is G/G/1. Let I denote the random variable
(RV) representing the idle time (I =
(1)
,
a2 + b2 + E[t2 ](1 2 ) I 2
2E[t](1 )
2I
(3.11)
An approximation to the above result is given by the Kingmans formula for the G/G/1
Queue [24], where the waiting time in the queue (Wq ) is given by:
E[Wq ] E[X]
Ca2 + Cb2
(1 )
2
(3.12)
Another commonly used approximation under heavy traffic is given in [22, Pp. 206],
i.e. as 1, approximation to the waiting time is:
E[W ] =
a2 + b2
2(1 )
(3.13)
This represents the upper bound on the waiting time in queue for heavy traffic loads.
Chapter 4
4.1
The service time for all access categories is difficult to characterize, since it is a function
of the number of active stations of each AC, the individual data rates, the channel condition etc. The best fit for this situation is a general distribution. Coming to the arrival
process, voice traffic is CBR, which is deterministic. Video traffic is variable in nature,
and is dependent on the codec used. While it is not exactly a Poisson process, surveys
[20, 25] show that this is a reasonable approximation; in most cases the variability within
the traffic arrival will conform to that of a Poisson process. In addition, the Markovian
approximation offers high computational tractability, thus it has remained a well used
analysis tool. Since the system is time shared, the server is inherently 1. Thus the most
appropriate queuing discipline is the M/G/1 for the EDCA.
The basis for the analytical framework is established by viewing each access category
(VO, VI, BE and BK) as an independent single dimensional queue. Each queue is a lim18
19
ited service, partially gated system with vacation/reservation. This implies that after each
packet arrives at the head of queue, the server immediately goes on vacation. This vacation accounts for the randomly selected back-off counter value, other possible collisions
and transmissions within this interval and the collisions that the station it self suffers and
the back-off counter picked from the subsequently doubled contention window.
For any arriving packet, the waiting time in queue is the sum of the residual time
Ri (the time remaining for the currently ongoing transmission/vacation), the service time
of all the preceding packets
Ni
,
well as its own vacation time Yi . In case of T XOP being enabled, the vacation interval
does not precede each packet transmission, but several consecutive packet transmissions,
as dictated by the T XOPlimit .
From [22, Eq 3.63, Pp 198], the expected waiting time for the ith arriving packet Wi
is given by:
E[Wi ] = E[Ri ] +
E[Ni ]
+ E[Yi ]
(4.1)
As i the system attains steady state. From Littles theorem, N = W and using
=
(4.2)
A major focus of the delay analysis is on developing an expression for the vacation time
that precedes each transmission. The problem of finding the queuing delay necessitates
using the first two moments of the vacation time. However, the PDF of the vacation time
is not required. This indicates that complex models, as presented in [9], although necessary for accurate analysis of the protocol, are not required for delay analysis in particular.
The main advantage of this approach is that the model remains analytically tractable,
and robust enough to be applied in a dynamic environment, where the retransmission
probability and the channel busyness is constantly changing. This is not possible for most
other analytical models, which start with an assumption of steady state, implying that a
new station may not appear on the station, or an existing station may not stop transmitting. Furthermore, in current models, channel conditions, if taken into account [9], are
held to be constant, which is also inaccurate for the wireless environment.
4.2
The parameters such as arrival distribution, service time, vacation etc., will differ for voice
and video ACs. To consider the relation between the two ACs, it is established, based
on contention window and AIFS differentiation, that voice will almost always have higher
20
priority. However, when any packet is being transmitted from say voice AC, the other
queues, whether voice or video, will be undergoing vacation. Similarly, when a video
packet is being transmitted, all other queues, both voice and video, will be undergoing vacation. Thus the transmission interval of one queue will overlap with the vacation interval
of all other queues. Thus the priority is being catered for within the vacationing.
For the mean residual time, consider figure 4.1.
We will derive the general expression for queuing delay, which can be used for voice
and video by using the relevant parameters plugged into equations (4.1-4.2).
R=
E[X 2 ] (1 )E[V 2 ]
+
2
2E[V ]
(4.3)
The expected time due to vacation considering a partially gated system, is given by:
Y =
(1 + )E[V ] (1 )
(E[V ])2
2
2E[V ]
(4.4)
Since the system under consideration is a limited service system, the waiting time in
eq (4.2) is extended by a factor E[V ]W since each packet causes a vacation interval.
21
(4.5)
E[R] + E[Y ]
1 E[V ]
Plugging the values of E[V ] and E[R] from eq 4.3 and 4.4 into 4.5, we get the result
for the waiting time:
E[W ] =
1
2(1 E[V ])
E[X ] +
+ (1 + )E[V ]
(4.6)
For each AC (VO, VI, BE, BK) the parameters , E[V ], E[V
2]
and
E[X 2 ]
will be differ-
1
nT XOP
E[V ]E[W ]
(4.7)
1
2(1
1
nT XOP
E[V ])
E[X ] +
+ (1 + )E[V ]
(4.8)
The main result of this analysis is given by 4.8. This caters for the transmission opportunity (T XOP ) explicitly, and AIFS and CW differentiation within the vacation interval.
The main results for overall delay each packet suffer is summarized by 4.6 and 4.8. This
model presents a significant difference from most existing models, particularly Markov
chain based frameworks, in that iterative solution is not required.
In the above derived expression for waiting time, the service time and arrival rates are
known; the only unknowns remaining are the first and second moment of vacation time.
The following section deals with these two expressions for dynamic network conditions;
the subsequent section details the work for steady state conditions.
22
4.3
After each packet arrives at the head of queue, contention for channel access starts, which
continues until the station has gained medium access, and successfully transmits the
packet. In this model, this time constitutes the reservation or vacation interval that precedes each transmission. This includes, in addition to the time spent during the back-off
process, the time spent due to collision and subsequent unsuccessful transmissions. Thus
the vacation time is dependent on the number of transmitting stations, and the probability of re-transmission, which is the combined effect of both channel outage and collision.
The total vacation/reservation interval (V ) preceding transmission of each data frame
can be attributed to two reasons, increasing back-off window following each failed transmission attempt (Vbo ) and other transmissions/collisions occurring within the back-off
period (Vbusy ).
E[V ] = E[Vbusy ] + [Vbo ]
(4.9)
(4.10)
4.3.1
E[Vbo ]
The IEEE 802.11 MAC layer protocol dictates that the back-off window doubles after each
unsuccessful transmission attempt, starting at the minimum contention window size (Wm ,
which will be the integer value of the contention window times Tslot ), until the maximum
contention window size is reached, or the retry limit (R) is exceeded. The vacation due
to back-off (Vbo ) is the number of idle slots that the station traverses before successful
transmission. Since the number of slots traversed increases in case of failed transmission
attempt, Vbo is a function of the retransmission (collision) probability (p).Since the collision probability p, and the minimum contention window size Wm , are different for VO
and VI ACs, both will have dissimilar vacation due to back-off.
In case the retry limit is reached, and the packet suffers another collision, the packet
is dropped. However, in usual traffic load, this event occurs very rarely, as is validated
by the analytical model presented in [2]. Thus practically, this case can be ignored; this
leads to a simpler analysis. The vacation time due to back-off Vbo is a memoryless process;
the probability of success (or failure) remains the same after each retransmission. In
case where the retry limit is not considered, the first moment (or expectation) of Vbo
can be derived in terms of (p) and the minimum contention window size Wm . In case of
a successful transmission, the expected waiting time is the same as that of the average
23
contention window size. In case of a collision/failed attempt, the expected vacation time
doubles from that of the previous attempt. This can be derived as follows for the generic
case, which can be tailored for both VO and VI (and also for BE and BK):
E[Wm ]
1 2p
(4.11)
2
E[Vbo2 ] = (1 p)E[Wm
] + pE[(Wm + 2Vbo )2 ]
1
2
E[Wm
] + 4pE[Wm ]E[Vbo ]
=
1 4p
E[Vbo2 ] =
2]
E[Wm
(E[Wm ])2 4p
+
,
1 4p
(1 2p)(1 4p)
4p < 1
(4.12)
The case of retry limit can similarly be considered; the above geometric distribution is
then replaced by a truncated geometric distribution. However, in actual network scenarios,
the probability of the retry limit being reached is negligibly small. Hence analysis without
inclusion of the retry limit is close to the actual results.
4.3.2
E[Vbusy ]
Besides the exponentially increasing back-off window, the factors that contribute to the
vacation interval are the other channel events (N ) that occur during the back-off counter
decrement, i.e. other successful transmissions and collisions, where the stations back-off
counter remains frozen. The total time that the channel is busy due to these events (Tbusy )
is a random variable, dependent on the number of other transmitting flows.
If the packet does not suffer a re-transmission, the vacation Vbusy will consist of the time
that the channel is busy due to transmission from other stations, plus the necessary AIFS
(TM AC ). In case the packet undergoes a retransmission, the vacation will consist of the
AIFS, the time to transmit the packet and establish that a collision has occurred (Tno ack ),
as well as the entire duration (Vbusy ) that will occur due to a retransmission attempt. Just
as in the case of the vacation due to back-off, the process Vbusy is memoryless and lends
well to a closed form result. Following the same principle as adopted for vacationing due
to back-off, E[Vbusy ] depends on the number of transmission attempts, increasing in case
of failed transmission. Defining the probability of these events in terms of p, we have the
first moment of Vbusy given by (4.13):
24
1
p
(TM AC ) +
Tno ack
1p
1p
(4.13)
Equation (4.13) is intuitively descriptive of the vacation due to channel events and retransmissions. The wait includes the time that the channel is busy due to other stations
transmissions (Tbusy ), the AIFS initiated for each transmission attempt, which follows a
geometric distribution (events till the successful transmission); and the time to establish
a failed transmission up until the last successful attempt, which also follows a geometric
distribution (events before the successful transmission).
Here the main parameter to be determined is E[Tbusy ]. Under stable conditions, i.e.
when the delays are bounded, both the first and second moment exist for Tbusy . Just
like the retransmission probability (p), these two parameters from the perspective of any
station, can be maintained and updated at runtime.
Assuming that the first and second moment of Tbusy exist, and are know, using (4.9)
and the subsequent results derived in (4.13), the first moment of the vacation time (when
retry limit is not considered) is given as:
E[V ] =
E[Wm ]
1
p
+ E[Tbusy ] +
TM AC +
Tno ack
1 2p
1p
1p
(4.14)
For the second moment, the same result can be derived by putting in the requisite
values in (4.10).
E[V 2 ] =
1
2
2
(1 + p)(TM
AC + pTno ack ) + 2pTM AC Tno ack
2
(1 p)
2]
E[Wm
2E[Wm ]
1
(E[Wm ])2 4p
+
(TM AC + pTno ack ) +
+
1 2p
1p
1 4p
(1 2p)(1 4p)
1
E[Wm ]
2
+ E[Tbusy ] + 2E[Tbusy ]
(TM AC + pTno ack ) +
1p
1 2p
(4.15)
Equations (4.14-4.15) completely describe the vacation interval for the purpose of queuing analysis.
4.4
We can use (4.14 4.10) to study several special cases, i.e. the expected vacation time
when the retransmission probabilities are small. This occurs when there are few transmit-
25
When there are very few contending stations, the retransmissions are caused by channel
conditions, rather than collisions. In this case it can be assumed that 1 p 1, and the
above equations simplify as:
E[V ] = E[Tbusy ] + TM AC + pTno ack + E[Wm ]
(4.16)
2
2
E[V 2 ] = TM
AC + pTno ack + 2pTM AC Tno ack + 2E[Wm ](TM AC + pTno ack )
2
2
+ E[Wm
] + (E[Wm ])2 4p + 2E[Tbusy ](TM AC + pTno ack + E[Wm ]) + E[Tbusy
]
(4.17)
When there is a moderate number of contending flows, the retransmissions occur both
due to channel outage and collision. However, if p remains small, such that p2 0, then
the expressions for first moment remains the same as (4.14), however the second moment
simplifies as:
E[V 2 ] =
2E[Tbusy ]E[Wm ]
1
2
(p(TM AC + Tno ack )2 + TM
AC ) +
1 2p
1 2p
2
2E[Wm ]
E[Wm ] (E[Wm ])2 4p
+
(TM AC + pTno ack ) +
+
1 3p
1 4p
1 6p
2E[Tbusy ]
2
+
]
(TM AC + pTno ack ) + E[Tbusy
1p
(4.18)
In the case when the retransmission probability is very small, and it can be assumed
to be zero, we arrive at very simplified expression for both the first and second moment:
E[V ] = E[Tbusy ] + TM AC + E[Wm ]
(4.19)
2
2
E[V 2 ] = TM
AC + 2E[Wm ]TM AC + E[Wm ]
(4.20)
2
+ 2E[Tbusy ](TM AC + E[Wm ]) + E[Tbusy
]
In case where the number of active stations is very small, E[Tbusy ] approaches zero.
In this case, when the packet arrives at the head of queue, and the medium is sensed free
for an AIFS, the packet will be transmitted without back-off. Thus E[V ] = TM AC ; this
case will occur with very small probability, and is catered for within this model, since
E[Wm ] caters for the case where the randomly picked back-off counter is zero. The above
expressions are generic, and will differ for VO and VI, and can be found by plugging in
the requisite values of pV O and pV I .
Here the vacation time has been defined as the total time that a packet waits at the
26
head of queue until initiation of successful transmission. The conventional approach has
been to estimate the collision probability (or its average) based on complex analytical
models. Here it has been proposed that the retransmission probability can be estimated
and updated at runtime, which can then be used to periodically update the delay results
for decisions to allow or reject new flows. Another factor that effects the transmission
time is number of other stations transmitting during the back-off interval. Again, from
the perspective of multi-dimensional models, this represents a complex expression based
on the per station transmission probabilities and the number of contending higher and
lower priority flows. However, just as in the case of retransmission probability, this time
too, can be estimated at runtime.
4.5
Steady State
The above model has been derived for actual, dynamic conditions where the IEEE 802.11
is employed. For the sake of completeness, the analysis is extended to steady state and
saturation conditions. Here steady state refers to a stationary process; saturation implies
that each queue will always have a packet to transmit.
In case the assumption of steady state is made, i.e. the number of active stations
remains constant, we can derive a purely analytical expression in terms of expected transmission and the retransmission probability p. These parameters can be calculated from
the rigorous analytical model for EDCA developed by Bianchi [2].
4.5.1
E[Vbusy ]
In this case, the expression for Vbusy will be developed in terms of a random number
of random events. The factors that contribute to the vacation interval are the other
channel events (N ) that occur during the back-off counter decrement, i.e. other successful
transmissions and collisions. Again, defining the probability of these events in terms of
the retransmission probability, we have the first moment given by (4.21):
1
p
(TM AC ) +
Tno ack
1p
1p
(4.21)
Here the main parameter to be determined is N . One possible way of solving this
problem is to determine E[Nb ], the number of expected events during the minimum backoff window. Using Nb and borrowing from eq 4.11-4.12 above, E[N ] and E[N 2 ] can then
be derived as:
E[N ] =
E[Nb ]
1 2p
(4.22)
27
E[Nb2 ]
(E[Nb ])2 4p
+
,
1 4p (1 2p)(1 4p)
4p < 1
(4.23)
This analysis is valid for steady state, since the number of events tend to increase as
the number of retransmission attempts increase, as all contending stations remain similarly active at all times.
Thus the number of other channel events is a random sum of random variables. From
[26, Eq 7.48, Pp 255], the second moment of this sum is given as:
E[(N Tbusy )2 ] = (E[Tbusy ])2 E[N 2 ] + T2busy E[N ]
(4.24)
Using 4.9 and the subsequent results derived, the first moment of the vacation time
(when retry limit is not considered) is given as:
E[V ] =
E[Wm ]
E[Nb ]
p
1
+
E[Tbusy ] +
TM AC +
Tno ack
1 2p
1 2p
1p
1p
(4.25)
(4.26)
2]
E[Wm
(E[Wm ])2 4p
+
)
1 4p
(1 2p)(1 4p)
E[N ]E[Tbusy ]
E[Wm ] 1
p
+2
(
TM AC +
+
Tno ack )
1 2p 1 p
(1 2p)
1p
2
p + p2 2
(E[Nb ])2 4p
1+p 2
2 E[Nb ]
+
T
+
(E[T
])
(
+
)
+ 2 TM
busy
AC
p
(1 p)2 no ack
1 4p (1 2p)(1 4p)
E[Nb ]E[Tbusy ]TM AC
E[Nb ]
p
+ T2busy
+2
+2
TM AC Tno ack
1 2p
(1 p)(1 2p)
(1 p)
p
E[Nb ]
+2
E[Tbusy ]Tno ack
2
(1 p) (1 2p)
E[V 2 ] = (
E[V 2 ] =
(E[Tbusy ])2
1
(E[Wm ])2 4p
(E[Nb ])2 4p
2
(E[Wm
]+
)+
(E[Nb2 ] +
)
(4.27)
1 4p
1 2p
1 4p
1 2p
E[Nb ]
1
2
2
+
((1 + p)(TM
+ T2busy
AC + pTno ack ) + 2pTM AC Tno ack )
1 2p (1 p)2
2E[Wm ] E[Nb ]
1
+
(
E[Tbusy ] +
(TM AC + pTno ack ))
1 2p 1 2p
1p
2E[Nb ]E[Tbusy ]
+
(TM AC + pTno ack )
(1 2p)(1 p)
In case p is very small, such that p2 0, the above equation (4.15) simplifies as:
E[V 2 ] =
28
(E[Tbusy ])2
(E[Wm ])2 4p
(E[Nb ])2 4p
)+
(E[Nb2 ] +
)
1 4p
1 2p
1 4p
1 2p
E[Nb ]
1
2
+ T2busy
+
(p(TM AC + Tno ack )2 + TM
AC )
1 2p (1 2p)
2E[Wm ] E[Nb ]
1
+
(
E[Tbusy ] +
(TM AC + pTno ack ))
1 2p 1 2p
1p
2E[Nb ]E[Tbusy ]
+
(TM AC + pTno ack )
(1 2p)(1 p)
2
2
(E[Wm
]+
First and second moments for voice and video ACs will differ, and are defined from
the above equations (4.14-4.15) by using NbV O , pV O , WmV O , NbV O and NbV I , pV I , WmV I ,
NbV I .
4.6
Throughput
Since the model has been designed to compute the expected delay, it is redundant to go
through the same methodology to find the system throughput. It is an easier approach to
find the expected value of throughput from Littles Theorem. If we denote the per station
throughput by T h, per station delay by D, packet size by Psize , the average throughput
as a function of delay is given by:
Th =
1
Psize
D
(4.28)
This equation can be used to find the throughput once the delay has been calculated.
Thus the complete performance characterization can be obtained from the queuing framework.
Chapter 5
5.1
Simulation Setup
The IEEE 802.11g standard has been chosen for the simulation, since it is the currently
employed PHY standard. The parameters for MAC specifications [1, table 17-15] are listed
in table 5.1. The corresponding EDCA parameters are given in table 5.2 [1, table 7-37].
Table 5.1: MAC Parameters
Slot Time
SIFS
DIFS
Data Rate
Tno ack
Max MPDU
CWmin
CWmax
9 s
16 s
34 s
54Mbps
46s
4095 bytes
15
1023
One access point was simulated, serving as a sink for the traffic generated by stations.
Each station transmits all three traffic streams. The number of transmitting stations increase at regular intervals, going from one up to 40 stations (3-120 flows). Each increment
follows after a fixed interval, giving enough time to study the network.
Each station computes the time the channel remains busy due to transmissions from
other stations between two successive successful transmissions of its own. The collision/retransmission probability for each AC is also maintained and updated every 10ms.
29
30
AC (Priority)
AIFSN
CW min
CW max
TM AC
M SDU life
Voice
3
2
3
7
34s
300ms
Video
2
2
7
15
34s
500ms
Voice
104
160
CBR
Video
384
220 (average)
VBR
Background
200
128
CBR
The nature of the voice, video and background traffic has been chosen to reflect actual
real time data over the Internet. Voice and background data have been simulated as CBR,
while video data has been simulated as VBR. The details are shown in table 5.3 [25].
5.2
Computational Efficiency
The results obtained from the simulation have been tabulated in the following sections.
Within the simulation, the results for the error probability and busy time have been
continuously updated. The analytical results have also been applied for every changing
network condition. A point of interest is the efficiency of the analytical system. Since the
calculations do not involve fixed point iterations, the system is far more tractable than
multi-dimensional Markov chain models. As an example, using an Intel core i7rprocessor,
if the data for a simulation with a fixed setup is analyzed for delay and throughput, the
total computational time for one set of results is 0.023 seconds.
The only hurdle that might appear is the constant collection of statistics at the terminals. However, this is not an issue since the data collection is limited to each node itself;
there needs to be no centralized data and delay analysis. For facilitation of admission
control algorithms, the delay results per AC can be communicated to the AP only when
the system parameters, e.g. retransmission probability, change beyond a certain threshold.
5.3
First the voice access category is considered. The results, for the simulation and the
analytical model, have been presented in the form of graphs.
31
5.3.1
For a changing network scenario, the busy time has been measured between each successive
packet transmission. From this data, the first and second moment of the busy time have
been tabulated.
E[Tbusy ]
The following plot (figure 5.1) shows the values for expected busy time, E[Tbusy ] plotted
against increasing number of stations.
x 10
x 10
3.5
5
3
4
2.5
2
ms
ms
1.5
2
1
1
0.5
0
0
1
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.1: First Moment of Busy Time for VO AC: a)10ms intervals, b)20s intervals
The above figures indicate that the busy time increases with the number of transmitting
stations; beyond a certain point the busy time undergoes significant increment. Also, from
figure 5.1 it is evident that the busy time shows considerable variation when not averaged.
This can be attributed to the dynamic nature of the network, as well as the fact on of
the traffic sources in the network, i.e. video, is VBR. Thus the number of active stations
within any time span vary.
2
E[Tbusy
]
Figure 5.2 shows the second moment of busy time, E[Tbusy ] plotted against increasing
number of stations, as obtained from the contending VO ACs.
4
4
x 10
x 10
2.5
3.5
3
2
Seconds2
Seconds2
2.5
1.5
1.5
1
0.5
0.5
0
0
1
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.2: Second Moment of Busy Time for VO AC: a)10ms intervals, b)20s intervals
32
The second moment of busy time shows variation similar to that of the first moment.
It is notable that the values of the second moment of busy time is comparatively small in
the context of QoS provisioning. Another fact that becomes apparent is that the second
moment shows a decrease in the heavy load scenario. This can be attributed to the
fact that as the traffic increases beyond a certain point, the busy time starts remaining
uniformly large, thus the variation decrease, and correspondingly the second moment.
5.3.2
Retransmission Probability
0.12
0.012
0.1
0.01
0.08
0.008
Probability
Probability
Figure 5.3 shows the retransmission/collision probability against the number of stations.
0.06
0.006
0.04
0.004
0.02
0.002
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.3: Retransmission Probability for VO AC: a)10ms intervals, b)20s intervals
The figures indicate that the retransmission probability shows significant variation,
particularly when new stations start transmitting. It can be observed that in the instantaneous case, the error probability goes up to 0.5. However, from figure 5.3, it can be seen
that the average retransmission probability never grows out of bound, i.e. the average does
not cross 0.05. There is no distinct trend that increasing transmitting flows necessarily
means increasing retransmission probability. This can be explained due to initiation of
back-off process:
When new flows are suddenly added to the system, collisions momentarily shoot up;
Following this, post collision back-off counters are doubled, enabling the system to
settle down into steady state whereby the collision probability goes down.
An important conclusion that can be drawn is that the collision/retransmission probability does not solely indicate the delays that will occur in transmission.
33
5.3.3
Vacation Time V
Based on the collected data for busy time and the error probability, the first and second
moment of vacation time has been calculated.
E[V ]
Figure 5.4 and ?? shows the first moment of vacation time, as computed by equation
2 ] from the above data set.
(4.14), using the values of p, E[Tbusy ] and E[Tbusy
3.5
5
3
4
2.5
2
ms
ms
1.5
2
1
1
0.5
0
0
1
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.4: First Moment of Vacation Time: a)10ms intervals, b)20s intervals
E[V 2 ]
The second moment (4.15) is shown in figure 5.5 (instantaneous) and ?? (averaged).
4
4
x 10
x 10
6
2
Seconds2
Seconds2
0
1
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.5: Second Moment of Vacation Time: a)10ms intervals, b)20s intervals
Comparing the two sets of figures (5.1-5.2) and (5.4-5.5), we see that the first and
second moments of vacation time closely follow the first and second moment of busy time.
This can be explained by the fact that , as long as p remains within bound, the major
contributors in 4.14 and 4.15 are the first and second moments of Tbusy respectively.
5.3.4
Delay Results
Using the results presented above in equation (4.6), the analytical results for delay have
been calculated and compared to the simulation results, shown in figure 5.6 (instantaneous)
34
0.045
E[W] Simulation
0.04
E[W] Analytical
0.025
0.035
0.02
seconds
Seconds
0.03
0.025
0.02
0.015
0.01
0.015
0.01
0.005
0.005
0
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.6: Delay Results: Analytical Vs Simulation: a)10ms intervals, b)20s intervals
It can be seen that at high traffic loads, the results show close agreement with the
simulation results. The apparent discrepancy at lower loads can be explained by the fact
that the analytical model assumes that the contention window keeps on doubling every
failed transmission attempt, In actual fact, the window for VO AC doubles just once. At
larger delays, this issue no longer remains as significant in the presence of delays caused
by busy time. In the context of actual delay requirements for voice, the over prediction is
not significant, since it runs in the order of few milliseconds.
5.3.5
Throughput Results
Using Littles Theorem, the delay results have been used to tabulate throughput results.
The comparison is shown in fig 5.7.
4
11
x 10
Simulation
Analytical
10
bits/sec
10
15
20
25
Number of Stations
30
35
40
It can be seen that the results under predict the per station throughput at lower loads,
since the delay is slightly over predicted. At higher loads, the results show close agreement
with the simulation.
Presented results indicate that the delay results depend largely on the vacation times,
which in turn are function of the busy time and the error probability. However, it can be
35
seen that even under heavy traffic loads, the error probability remains within bounds.
An interesting feature that comes to light is that since the vacation time follows the
busy time (both first and second moment), and the probability of error always remains
within a small bound, the approximate delay results (equation 4.26, ??) are very close
to the actual analytical results. This offers an even more tractable solution to the delay
model.
36
5.3.6
It is of particular interest to note how the system throughput for one particular access
category suffers as the average transmission probability goes up. This data has been taken
from the simulation alone. Figure (5.8) depicts the effect of the retransmission probability
on the decrease in system throughput that is primarily brought on by addition in the
expected values of the busy time. It becomes evident that as the probability goes beyond
0.2, the system throughput falls bellow the required average. Within the actual simulation,
these instances occur very rarely.
7
System Throughput
Acceptable Throughput
0.05
0.1
0.15
0.2
0.25
0.3
Retransmission Probability
0.35
0.4
0.45
0.5
5.3.7
The following figure (5.9) shows the simulation jitter results for voice access category.
4
4
x 10
1.2
x 10
2
Jitter (Second2)
Jitter (Second2)
0.8
0.6
0.4
0.2
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Just as in the case of delay, we see a steady increase in jitter with increasing the number
of stations. However, even under heavy loads, the jitter remains within acceptable limits.
37
5.4
The results for the video AC show a close similarity to those of voice AC. These results
have been tabulated and presented bellow.
5.4.1
E[Tbusy ]
The following plot (figure 5.10) shows the values for expected busy time, E[Tbusy ] plotted
against increasing number of stations.
3
x 10
0.014
0.012
2.5
0.01
Seconds
Seconds
2
0.008
0.006
1.5
1
0.004
0.5
0.002
0
1
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.10: First Moment of Busy Time for VI AC: a)10ms intervals, b)20s intervals
The above figures indicate that the busy time increases with the number of transmitting
stations. Again, as in the case of voice, it is evident from figure 5.10 that the busy time
shows considerable variation.
2
E[Tbusy
]
Figure 5.11 shows the second moment of busy time, E[Tbusy ] plotted against increasing
number of stations, as obtained from the contending VI ACs.
4
3
x 10
x 10
1.8
1.6
1.4
2
Seconds2
Seconds2
1.2
1
0.8
0.6
0.4
0.2
0
1
10
15
20
25
30
35
40
10
15
20
25
Number of Station
30
35
40
Figure 5.11: Second Moment of Busy Time for VI AC: a)10ms intervals, b)20s intervals
The second moment of busy time shows variation similar to that of the first moment.
The values of second moment of busy time are small in the context of acceptable delays.
38
5.4.2
Retransmission Probability
Figure 5.12 shows the retransmission/collision probability against the number of stations.
0.04
0.016
0.035
0.014
0.03
0.012
0.01
Probability
Probability
0.025
0.02
0.008
0.015
0.006
0.01
0.004
0.005
0.002
0
1
10
15
20
25
Number of Station
30
35
40
10
15
20
25
Number of Station
30
35
40
Figure 5.12: Retransmission Probability for VI AC: a)10ms intervals, b)20s intervals
As in the case of voice, the figures indicate that the retransmission probability shows
significant variation. From figure 5.12, it can be seen that the average retransmission
probability never grows out of bound, i.e. the average does not cross 0.1. This bound that
the error probability remains within can be attributed to the initiation of the back-off
process. The retransmission probability is not the sole metric of interest in predicting
delays and throughput.
39
5.4.3
Vacation Time V
E[V ]
Figure 5.13 shows the first moment of vacation time, as computed by equation (4.14),
2 ]. The first figure shows instantaneous values,
using the values of p, E[Tbusy ] and E[Tbusy
3
x 10
3.5
2.2
x 10
1.8
1.6
Seconds
Seconds
2.5
1.5
1.4
1.2
1
0.8
0.6
0.5
0.4
0
10
15
20
25
Number of Station
30
35
0.2
40
10
15
20
25
N
) umber of Station
30
35
40
Figure 5.13: First Moment of Vacation Time: a)10ms intervals, b)20s intervals
E[V 2 ]
The second moment (4.15) is shown in figure 5.14.
4
4
x 10
x 10
6
1.8
5
1.6
1.4
Seconds2
Seconds2
1.2
1
0.8
0.6
0.4
1
0.2
0
0
1
10
15
20
25
Number of Station
30
35
40
10
15
20
25
Number of Station
30
35
40
Figure 5.14: Second Moment of Vacation Time: a)10ms intervals, b)20s intervals
Comparing the two sets of figures (5.10-5.11) and (5.13-5.14), the first and second
moments of vacation time closely follow the first and second moment of busy time. As
long as p remains within bound, the major contributors in 4.14 and 4.15 are the first and
second moments of Tbusy respectively.
5.4.4
Delay Results
Using the results presented above in equation (4.6), the analytical results for delay have
been calculated and compared to the simulation results, shown in figure 5.15.
40
0.03
E[W] Simulation
E[W] Analytical
0.05
E[W] Simulation
E[W] Analytical
0.025
0.02
seconds
Seconds
0.04
0.03
0.015
0.02
0.01
0.01
0.005
10
15
20
25
Number of Station
30
35
40
10
15
20
25
Number of Stations
30
35
40
Figure 5.15: Delay Results: Analytical Vs Simulation: a)10ms intervals, b)20s intervals
AS was the case in voice, the results show closer agreement at higher loads then at
less traffic. In the case of video, however, the divergence is less. This further supports the
theory regarding the discrepancy. In case of video, the contention window size doubles
only once, however, the largest contention window size for video is double that of voice,
lending for a better approximation.
5.4.5
Throughput Results
Per station throughput has been calculated from the delay using 4.28. The results are
shown in fig 5.16.
5
x 10
Simulation
Analytical
3.5
bits/sec
2.5
1.5
0.5
1
10
15
20
25
Number of Stations
30
35
40
The results are indicative of the same trends as that for voice AC; the busy time
steadily increases at higher traffic loads, as does the average error probability. The later
remains within bound of 0.02. The delay results closely follow the simulation at higher
traffic loads.
41
5.4.6
The following figure (5.17) depicts the effect of increasing retransmission probability on
system throughput for video access category. In this case, we see that the system throughput falls bellow the acceptable minimum at lower error probability than the video AC.
This can be attributed to the greater throughput requirements for video.
11
Minimum Acceptable Throughput
System Throughput
10
9
8
7
6
5
4
3
2
0.1
0.2
0.3
0.4
Retransmission Probability
0.5
0.6
5.4.7
The following figure (5.18) shows the simulation jitter results for voice access category.
4
4
x 10
x 10
1.6
1.4
Jitter (Second2)
Jitter (Second2)
1.2
1
0.8
0.6
0.4
0.2
10
15
20
25
Number of Stations
30
35
40
10
15
20
25
Number of Stations
30
35
40
Just as in the case of delay, we see a steady increase in jitter with increasing the number
of stations. However, even under heavy loads, the jitter remains within acceptable limits.
42
5.5
For the sake of a complete picture in terms of delay prediction, it is meaningful to compare
the presented model with existing analytical models for delay prediction. For this purpose,
the work presented by Tadayon et. al [9], hereon referred to as Tadayons Model. This work
specifically deals with delay analysis. Since it is for DCF, a new simulation was conducted
in which only one type of traffic was generated; The parameters were set to that of best
effort traffic. A relatively light load was generated, with a total of 20 transmitting stations.
The delay results obtained from the simulation as well as both models were plotted against
probability of retransmission. The results are shown in figure 6.3.
0.01
Tadayons Model
Presented Model
Simulation Results
0.009
0.008
0.007
seconds
0.006
0.005
0.004
0.003
0.002
0.001
0
0.05
0.1
0.15
0.2
0.25
0.3
Retransmission Probability
0.35
0.4
0.45
0.5
AN inspection shows the results presented both by [9] and the existing work substantiated. The current work offers greater accuracy in terms of delay prediction for higher
loads, or when the retransmission probability is higher. Tadayons work shows high accuracy in cases of lower load, or when the collision probability is lower. The practice of
categorizing the overall system load in terms of retransmission probability has been adhered to in this analysis for the sake of consistency with the work compared with. From
the perspective of delay analysis in its entirety, both analytical frameworks offer distinct
advantages. However, in the case of heavy loads, the presented work has a clear edge.
This scenario is of the most significance for the purpose of admission control and network
planning.
43
5.6
Given the analytical modeling, the impact of error or retransmission probability poses limitations on system performance for delay, throughput and jitter if it grows beyond certain
bounds.
To analyze this phenomena, a simulation was conducted with very high traffic load (60
Mbps over a 54Mbps channel) with high path loss. The resulting system performance for
voice and video access category is summarized in the following graphs.
5.6.1
System Throughput
Figures 5.20 5.21 illustrate the impact of increasing error probability on overall system
through put for voice and video ACs respectively. The results show that there is a sharp
drop in throughput once the error probability crosses 0.25. This corroborates the results
derived for the analytical framework.
6
5.5
x 10
4.5
4
3.5
3
2.5
2
1.5
1
0
0.1
0.2
0.3
0.4
Retransmission Probability
0.5
0.6
0.7
5.6.2
Per station delay as a function of increasing error probability is analyzed for voice and video
ACs respectively. Here the impact of error probability increasing beyond 0.25 becomes
more pronounced. A comparison of both per station delay and jitter is shown in 5.22 and
5.23. In the case of voice AC, the jitter starts increasing rapidly once the retransmission
probability increases beyond 0.05. This effect is also present in VI AC, although it is
less marked. The reason for this may be attributed to the larger packet size for video at
the same data rate as voice; thus the number of transmissions tend to be more evenly
distributed.
44
x 10
5.5
5
4.5
4
3.5
3
2.5
2
1.5
1
0.05
0.1
0.15
0.2
0.25
0.3
Retransmission Probability
0.35
0.4
0.45
0.5
0.12
Per Station Delay
Jitter
0.1
Seconds
0.08
0.06
0.04
0.02
0.1
0.2
0.3
0.4
Retransmission Probability
0.5
0.6
0.05
0.045
0.04
Second
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
0.05
0.1
0.15
0.2
0.25
0.3
Retransmission Probability
0.35
0.4
0.45
0.5
Chapter 6
46
vacation time (figure (6.1)) this becomes apparent. This is particularly applicable for the
current PHY standards of the IEEE 802.11, e.g. g and n. Thus effectively we can model
the vacation time with the busy time.
E[V ] E[Tbusy ]
(6.1)
2
]
E[V 2 ] E[Tbusy
(6.2)
This makes a considerable impact on the calculation efficiency, further bringing down
the computation expense by a sizable margin.
3
Delay Results
x 10
E[V]
E[V2]
E[Tbusy]
E[Tbusy2]
8
7
6
5
4
3
2
1
0
10
15
20
25
Number of Stations
30
35
40
An interesting effect observed was the behavior of busy time under heavy traffic loads.
Two points came to light, the first being that the increase in busy time slows down with
increasing contending stations; this can be explained by the doubling in contention window
47
sizes with every collisions. Most stations are under going back-off instead of transmitting.
The second is that the variability decreases; delays due to other transmissions become
uniformly large.
x 10
E[Tbusy]
4.5
4
Seconds
3.5
3
2.5
2
1.5
1
0.5
0
10
15
20
25
30
35
Number of Stations
40
45
50
55
60
55
60
2.5
x 10
E[Tbusy]
Seconds
1.5
0.5
10
15
20
25
30
35
Number of Stations
40
45
50
2
Figure 6.3: E[Tbusy
] under heavy loads
48
stantaneous events that this bound is exceeded, the simulations show that the throughput
drops drastically to an unacceptable limit. In case this bound is exceeded, real time communication will no longer be possible due to unacceptable delay and jitter performance,
as illustrated in figure 6.4.
0.12
Per Station Delay
Jitter
0.1
Seconds
0.08
0.06
0.04
0.02
0.1
0.2
0.3
0.4
Retransmission Probability
0.5
0.6
6.1
Referring back to table 3.1, the developed model can be compared to the existing works.
In comparison with [2], the proposed work offers a far more tractable solution, with an
extension to throughput results. There is a many fold enhancement in computational
efficiency; the model in [2] requires fixed point iteration for convergence to a solution for
a fixed network scenario, if the scenario changes by even a single transmitting station, the
entire set of calculations have to be repeated again. The presented model adapts to the
changing environment far more quickly since there are no iterative calculations involved.
The model presented by Bianchi et. al, although accurate as compared to other Markov
chain models, does not cater for transmission opportunity parameter; this issue has also
been dealt with in the current work. In addition, one of the most limiting conditions,
that of stationary process, is no longer required. This allows for a model that adapts to
changing conditions and traffic loads without having to reformulate the problem.
Coming over to the delay results presented in [9], we see that, again, the presented
model is far more analytically tractable. Although important from a purely theoretical
perspective, the probability density function of the service time is not of extreme importance when analyzing from the perspective of system performance. Just the first two
moments of service time are required, or in this case referred to as the vacation time.
However, a factor more significant than the complexity is the accuracy. The framework
presented by Tadayon et. al under predicts the delay under heavy traffic loads. This can
cause unacceptable delays when applied for admission control. In contrast, the presented
49
work offers far greater match with simulation results at higher traffic loads. In addition,
since the model was developed for DCF, it does not cater for transmission opportunity,
which is factored in within this work. A point to note is the range of error probabilities over which the analytical model in [9] seems to predict acceptable performance; this
seems to be an overly optimistic result, since the analytical model developed in this work
indicates that the delay/throughput performance becomes unacceptable when the error
probability goes beyond 0.25, while jitter has a far smaller tolerance, i.e. 0.0625.
Coming over to the queuing model presented in [10], there are two areas that make for
meaningful comparison. The first is the development of the service time distribution; the
basis is a Markov chain, which assumes a stationary process. This again necessitates a set
of fixed point iteration based calculation for every new network scenario. This severely
impairs computational efficiency when it comes to predicting performance metrics in real
time systems, as opposed to the proposed model. The second aspect is that of accuracy;
the results presented in [10] substantially over predict the delay at higher traffic loads. The
presented model offers a far greater agreement between simulation and analytical results
at higher traffic loads. A comparison with existing work (table 3.1) is provided in table 6.1.
Bianchis
Model
for
EDCA Throughput [2]
Presented Work
Problem Addressed
Assumptions
Results
50
Models
6.2
51
In this work, delay and throughput analysis of the IEEE 802.11 EDCA has been extensively studied. Existing models aim to comprehensively capture the features of the
enhanced distributed channel access to arrive at accurate analytical frameworks for performance metrics, primarily delay and throughput. The nature of the back-off process
coupled with the service differentiation parameters make this a complex problem. Based
on this premise, the model developed in this thesis aims to address both the issue of complexity and accuracy. A novel approach of measuring actual network conditions including
error probability and the time the channel is busy, resolves both the issues at hand. The
effect of all active stations within the network are captured by these parameters; the ensuing analysis is then a simple application of non-preemptive priority queuing. The effect
of TXOP, AIFS and CW are all catered for, which is a great step up from most other
models. The results show a high level of accuracy, comparable with the most complex
existing models.
The most significant aspect of this work in comparison with other existing analytical results is that of high applicability to dynamic network conditions. The environment
changes due to the channel conditions, or the number of active stations increasing or decreasing; the model provides updated results without lengthy recursive calculations. This
gives an edge over models that are based on the assumption of stationarity, employ iterative solutions, or both. Borne out by the comparison presented above in this section, this
feature offers an edge both in terms of accurate analysis and run time application.
An inherent limitation of the analytical model presented in this work stems from that
the fact that it based on averaging. This accounts for the decreased accuracy under light
traffic loads. The system performance is not adequately captured in this scenario. This
limitation can be addressed by incorporating the retry limit and the maximum contention
window size into the analytical work. However, this leads to cumbersome analysis, and
from the perspective of admission control, this is not necessary; at lower loads, the system
throughput is less in comparison with the system capacity and stringent admission control
is not required. The analytical tractability of this work comes from run time estimation
of both the error/retransmission probability and the time that the channel stays busy between successive successful transmissions of any station. In this context, the model alone
is not enough to predict the delay based solely on the number of active stations. This is
an inherent limitation of this work, and form a good area for future exploration.
This work can be extended to the analytical formulation of jitter, which is a very
important, and largely overlooked, service metric for real time traffic. Another important
direction for future consideration is the extension of this work to the centralized channel
access mechanism, the hybrid coordination channel access (HCCA). This would ensure
that the entire system is analytically modeled, thus providing a complete picture of the
achievable service by the system under changing network conditions.
Bibliography
[1] IEEE standard for information technology-part 11: Wireless lan medium access control (MAC) and physical layer (phy) specifications, IEEE Std 802.11-2012 (Revision
of IEEE Std 802.11-2007), pp. 12793, 2012.
[2] I. Tinnirello and G. Bianchi, Rethinking the ieee 802.11 e edca performance modeling
methodology, Networking, IEEE/ACM Transactions on, vol. 18, no. 2, pp. 540553,
2010.
[3] A. Tanenbaum and D. Wetherall, Computer Networks. Pearson Prentice Hall, 2010.
[Online]. Available: http://books.google.com.pk/books?id=I764bwAACAAJ
[4] P. E. Engelstad and O. N. sterb, Queueing delay analysis of ieee 802.11 e edca,
IFIP WONS, 2006.
[5] G. Bianchi, Performance analysis of the IEEE 802.11 distributed coordination function, IEEE Journal on Selected Areas in Communications, vol. 18, no. 3, pp. 535547,
2000.
[6] N. C. Taher, Y. G. Doudane, B. El Hassan, and N. Agoulmine, First step towards
an efficient admission control: A complete analytical model for 802.11 e EDCA for
throughput and delay prediction, in IM09. IFIP/IEEE International Symposium
onIntegrated Network Management-Workshops, 2009. IEEE, 2009, pp. 217224.
[7] M. Gas, K. Kosek-Szott, M. Natkaniec, and A. Pach, 3D Markov chain-based saturation throughput model of IEEE 802.11 EDCA, Electronics letters, vol. 47, no. 14,
pp. 826827, 2011.
[8] J. Y. Lee and H. S. Lee, A performance analysis model for IEEE 802.11 e EDCA
under saturation condition, IEEE Transactions on Communications, vol. 57, no. 1,
pp. 5663, 2009.
[9] N. Tadayon, E. Askari, S. Assa, and M. Khabazian, A novel analytical model for
service delay in ieee 802.11 networks, 2012.
[10] X. Chen, H. Zhai, X. Tian, and Y. Fang, Supporting qos in ieee 802.11 e wireless
lans, Wireless Communications, IEEE Transactions on, vol. 5, no. 8, pp. 22172227,
2006.
52
53
BIBLIOGRAPHY
[14] P. E. Engelstad and O. Osterbo, Analysis of the total delay of ieee 802.11 e edca
and 802.11 dcf, in Communications, 2006. ICC06. IEEE International Conference
on, vol. 2.
[15] M. Ozdemir
and A. B. McDonald, An m/mmgi/1/k queuing model for ieee 802.11 ad
hoc networks, in Proceedings of the 1st ACM international workshop on Performance
evaluation of wireless ad hoc, sensor, and ubiquitous networks. ACM, 2004, pp. 107
111.
[16] O. Tickoo and B. Sikdar, Queueing analysis and delay mitigation in ieee 802.11 random access mac based wireless networks, in INFOCOM 2004. Twenty-third annualjoint conference of the IEEE computer and communications societies, vol. 2. IEEE,
2004, pp. 14041413.
[17] A. Hussain, F. Nadeem et al., Measurement based call admission control (cac) to
improve qos for ieee 802.11 e wlan, in New Technologies, Mobility and Security
(NTMS), 2014 6th International Conference on.
Wiley
Chichester, 1998.
[22] D. P. Bertsekas, R. G. Gallager, and P. Humblet, Data networks.
Prentice-Hall
54
BIBLIOGRAPHY
[24] J. F. C. Kingman, On queues in heavy traffic, 1962.
[25] C. Lewis and S. Pickavance, Selecting Mpls Vpn Services.
[26] A. Papoulis and S. U. Pillai, Probability, random variables, and stochastic processes.
Tata McGraw-Hill Education, 2002.
[27] NS-3. Available online at http://www.nsnam.org.
Appendix A
55
56
Najeeb ul Hassan
Farah Nadeem
Member, IEEE
Department of Electrical Engineering
National University of Computer
& Emerging Sciences
Lahore, Pakistan
Email: arshad.hussain@nu.edu.pk
I. I NTRODUCTION
With the advent of wireless technology, portable communication devices have become increasingly pervasive, as have
real time multimedia services. Supporting these applications
pose formidable challenge because of their quality of service
(QoS) requirements in terms of throughput and delay in an
inherently dynamic environment.
The IEEE 802.11e [1] provides a basic framework for
QoS provisioning. This set of amendments to the medium
access control (MAC) layer specifies a distributed access
approach, called enhanced distributed channel access (EDCA).
The purpose is to support priority based service differentiation
in the MAC layer. However, this framework can successfully
provide QoS only if accompanied by effective admission
control; admission of new flows has to be restricted to prevent
overloading in the QoS basic service set (QBSS). Otherwise,
the system undergoes performance degradation, particularly
in terms of delay and effective throughput. A call admission
control (CAC) algorithm ensures that a new traffic stream will
not degrade the QoS requirements of already existing streams,
while connections that can be handled within existing resources are accommodated; thus achieving a balance between
optimal resource utilization and service provisioning.
57
delay bound limit associated with this category. For the higher
priority ACs, the standard mandates the implementation of
CAC.
The IEEE 802.11e specifies a sample CAC algorithm [1].
When a mobile node in a QBSS initializes real-time multimedia traffic, it seeks admission from the QAP by sending
it an add traffic stream (ADDTS) request frame. The request
frame contains the traffic specification information (TSPEC),
including average data rate (), nominal MSDU size (L),
physical rate (R) and surplus bandwidth allowance (BW s ).
The node calculates the BW s (which is the time required
for retries) based on the retransmission probability, pe , and
probability of any given frame being dropped (Pdrop ). The
QAP also computes the medium time, tm , which is the time
required by the requesting AC.
The QAP maintains variable network utilization characteristic for EDCA (N U CEDCA ). This is defined as the time utilized for EDCA per unit time, which, along with the medium
time, determines whether a new stream can be admitted. A
TXOP for the new connection in beacon time (tbnc ) is allowed
as long as the following relation holds:
tbnc > N U CEDCA + tm .
(1)
Me [i]
M [i] + Me [i]
(2)
58
TABLE I
EDCA PARAMETERS FOR S IMULATION
AC (Priority)
AIFSN
CW min
CW max
TXOP Limit (ms)
Admission Required
MSDU Life Time (ms)
Voice
3
2
7
15
3
Yes
50
Video
2
2
15
31
5
Yes
100
TABLE II
S UMMARY P ERFORMANCE C OMPARISON OF THE CAC A LGORITHMS
Background
0
7
31
1023
0
No
No limit
Admitted flows
Throughput (kb/s)
Delay (ms)
Percentage Drop
(3)
Type
Voice
Video
Voice
Video
Voice
Video
Voice
Video
IEEE
23
21
1
3
20
42
17
27
Proposed
18
17
1
3
17
39
7
12
59
Fig. 2. Mean delay for video flows when different CAC algorithms are applied
Fig. 3. Percentage drop for video flows when different CAC algorithms are
applied
VI. C ONCLUSION
In this paper, we have analyzed the importance of the
admission control mechanism in WLAN for providing QoS
to real-time flows. We have extensively analyzed the performance of the IEEE reference 802.11e CAC algorithm via
simulation and observed that under heavy loads, it degrades
the performance of the system by introducing high delay
and large percentage drop for frames. To remedy this at
lower computational cost than existing CAC algorithms, we
Appendix B
60
61
Abstract The IEEE 802.11e EDCA is designed to provide quality of support for real time applications with
stringent latency and throughput requirements. Theoretical
frameworks for analysis of throughput performance of
wireless LAN employing exponential back-off exist
extensively. Several models rely on simplification
assumptions that preclude their direct applicability to the
enhanced distributed coordination access (EDCA) which
uses heterogeneous protocol parameters, while other
models are exceedingly complex to analyze. In this paper, a
tractable analytical model is proposed for saturation
throughput of the IEEE 802.11e EDCA. The prioritization
through channel access parameters including the AIFS and
contention window is catered for within a three dimensional Markov chain. The integration of back-off counter
freezing and retry limit enhance the models precision. Its
validation is done by simulation on NS-2. Practical applicability of the model is established based on accuracy and
computational efficiency. The model is utilized for
throughput analysis of the EDCA under saturated traffic
loads.
Keywords IEEE 802.11e EDCA Wireless networks
Performance analysis Saturation throughput
M. F. Usman
Mentor Graphics, Lahore, Pakistan
e-mail: fahad_usman@mentor.com
Notations
Steady state probability of the state representing
aki;j
back-off stage i, back-off counter j, and previously
sensed medium state k
CWmax Maximum contention window size
CWmin Minimum contention window size
m
Maximum back-off stage after which the value of
contention window is not increased
nh
Number of contending higher priority flows
nl
Number of contending lower priority flows
p0
Probability that the medium is busy after after an
idle slot
p1
Probability that the medium is busy after after a
busy slot
Pc
Probability that a given time slot contains a
collision
Pi
Probability that a given time slot is idle
Ps
Probability of a successful transmission in a time slot
PL
Length of payload in bits
r
Retry limit
S
Saturation throughput
TXOP Transmission opportunity
Tc
Time spent due to a collision
Ts
Time spent due to a successful transmission
Wi
Back-off window at the ith back-off stage
r
Time slot duration
h
Probability that a higher priority flow accesses
s0
the medium after a busy slot
Probability of medium access by a lower priority flow
sh1
1 Introduction
A. Hussain
e-mail: arshad.hussain@nu.edu.pk
123
62
Int J Wireless Inf Networks
communication networks. As they have become ubiquitous, so have real time multimedia applications over
wireless medium. This poses substantial challenges, since
real time applications have inherently greater sensitivity to
delay and throughput than best effort services. The original
IEEE 802.11 standard, designed for best effort services, is
fundamentally lacking in mechanisms to support quality of
support (QoS); thus providing priority based service differentiation is an essential requirement while dealing with
real time applications.
Although the IEEE 802.11 defines several PHY layer
specifications, e.g. the 802.11a, 802.11b and 802.11g, the
medium access control (MAC) layer is same for all. To
introduce a framework that facilitates service guarantees,
enhancements have been suggested at the MAC layer. The
focus is on providing decentralized or distributed prioritization mechanisms as a first step towards furnishing real
time applications including voice over IP and live high
definition TV (HDTV) streaming with requisite latency and
throughput. To this end, the IEEE 802.11e [1] standard has
been proposed which has now been incorporated in the
IEEE 802.11. For the IEEE 802.11e, the distributed channel access mechanism is the enhanced distributed coordination access (EDCA), falling under the ambit of
prioritization via channel access parameters. The extent to
which EDCA succeeds in providing QoS support remains
an ongoing research area. In this respect, theoretical
framework for performance analysis is crucial. Generally,
essential levels of service are provided via admission
control algorithms, which determine whether new stations
can be handled while maintaining the required access
delay, jitter and throughput for existing real time flows of
the stations. These algorithms require performance prediction in terms of metrics to operate, which are provided
by analysis models. Since the throughput, defining successful data transmission per unit time, is an important
metric for ensuring whether an application will receive
services consistent with the given quality criterion, the
central goal of the analysis is to predict throughput under
both saturated and unsaturated traffic loads. In this context,
several models have been put forth in literature; however,
these models often rely on simplification assumptions that
fail to take into account salient features of the standard. In
the cases where attempts have been made to factor in all
details of the standard, the resulting models are exceedingly complex, which precludes extensions and run-time
analysis.
To address these issues, this paper proposes an analytical model to accurately capture service differentiation
mechanism of the IEEE 802.11e EDCA at less computational cost, making it applicable to call admission control
and runtime optimization. The fundamental framework is a
three dimensional (3-D) discrete time Markov model, an
123
63
123
64
123
i 2 f0; r g; j 2 f0; Wi 3g
Pfi; j; 1ji; j; 0g p0
1 p0
Pf0; j; 1ji; 0; 0g
W0
p0
Pfi; j; 1ji 1; 0; 0g
Wi
i 2 f0; r g; j 2 f0; Wi 2g
i 2 f0; r 1g; j 2 f0; W0 1g
i 2 f1; r g; j 2 f0; Wi 1g
i 2 f0; r g; j 2 f0; Wi 2g
Pfi; j; 1ji; j; 1g p1
1 p1
Pf0; j; 1ji; 0; 1g
W0
p1
Pfi; j; 1ji 1; 0; 1g
Wi
i 2 f0; r g; j 2 f0; Wi 1g
i 2 f0; r 1g; j 2 f0; W0 1g
i 2 f1; r g; j 2 f0; Wi 1g
1
W0
65
a1i;j
Wi 1 jp0 1 1
ai;0 i 2 f0; rg j 2 f1; Wi 1g
1 p1
1
a0i;j Wi 1 ja1i;0
i 2 f0; rg
j 2 f1; Wi 2g
a1i;0
i
Y
p1
x1
p0
Wx1 1 a10;0
Wx Wx
i 2 f1; rg
j0
j0
123
66
Int J Wireless Inf Networks
123
67
Pfi; j; 0ji; j 1; 0g 1 p0
i 2 f0; r g; j 2 f0; Wi 2g
Pfi; j; 1ji; j; 0g p0
1 p0
Pf0; j; 1ji; 0; 0g
W0
p0
Pfi; j; 1ji 1; 0; 0g
Wi
i 2 f0; r g; j 2 f0; Wi 1g
i 2 f0; r 1g; j 2 f0; W0 1g
i 2 f1; r g; j 2 f0; Wi 1g
r
P
i 2 f0; r g;
i 2 f0; r g;
j 2 f0; Wi 1g
j 2 f0; Wi 1g
j 2 f0; W0 1g
1 p1 p0 Wi 1 j p1 Wi 2 j 1
ai;0
1 p1 1 p0
i 2 f0; rg
j 2 f1; Wi 2g
5
Similarly,
Wi j Wi 1 jp1 1
a0i;j
ai;0
1 p0
j 2 f0; Wi 2g
sh1
a1i;0
10
1 Pi
i0
a0i;0
11
Pi
Where a0i;0 are the states defined for the lower priority
flows. From a particular stations perspective, a time slot
is busy if at least one of the other stations transmits in
that time slot. The probabilities of a busy time slot, after
an idle or busy time slot, p0 and p1 , respectively, are
given by:
p0 1 1 sh0 nh 1 sl nl
p1 1 1
Pi
i 2 f0; rg
i0
12
sh1 nh
13
1 p1
1 p0 p1
14
6
Table 1 System parameters
Parameter
name
Value
Parameter name
Value
Packet payload
8184 bits
Slot time r
50 ls
MAC header
224 bits
ACK timeout
300
ls
PHY header
192 bits
CWmin (higher
priority)
ACK
112 bits
CWmax (higher
priority)
16
RTS packet
CWmin (lower
priority)
16
CTS packet
CWmax (lower
priority)
32
Channel bit
rate
6 Mbit/s
AIFSN (higher
priority)
Propagation
delay
1 ls
AIFSN (lower
priority)
DIFS
34 ls
SIFS
16 ls
i1
i 2 f1; rg
j0
123
68
Int J Wireless Inf Networks
123
69
123
70
Int J Wireless Inf Networks
15
nl sl 1 sl nl 1 1 sh0 nh Pi
The probability that a time slot contains a collision is given
by:
Pc 1 Pi Ps
16
Ps EfPLg
Pi r Pc Tc Ps Ts
17
123
5 Model Validation
The system of non-linear equations formed by (9)(17)
have been numerically solved using Fixed Point Iteration
[9] in order to obtain the value of saturation throughput in
different network scenarios. These results are henceforth
referred to as Fahads model. The numerical results
obtained from our mathematical model have been validated
by comparing them with the results of computer simulation
using the EDCA implementation in [2] on the Network
Simulator 2 (NS-2).
The simulation topology in NS-2 consists of one access
point and a varying number of transmitting stations. Since
virtual collisions are not catered for, each station generates
just one type of traffic, either high or low priority. Saturation state has been obtained using constant bit-rate flows.
Throughput has been calculated by observing the combined
good-put of all the flows once saturation state is obtained.
Both the access mechanisms, basic and RTS/CTS, have
been simulated, with system parameters listed in Table 1.
71
123
72
Int J Wireless Inf Networks
6 Conclusion
Throughput and delay performance are important performance parameters for wireless networks. In this paper, a
3-D Markov chain based analytical model has been proposed for evaluating saturation throughput of the IEEE
802.11e EDCA. The model has been designed to cater for
123
References
1. (2005) IEEE Std. 802.11e-2005, Part 11: Wireless LAN Medium
Access Control (MAC) and Physical Layer (PHY) specifications,
Amendment 8: Medium Access Control (MAC) Quality of Service Enhancements. Tech. rep.
2. Ansel P, Ni Q, Turletti T (2004) An efficient scheduling scheme
for IEEE 802.11 e. In: Proc. Modeling and Optimization in
Mobile, Ad Hoc and Wireless Networks, pp 2426.
3. Bianchi G (2000), Performance analysis of the IEEE 802.11
distributed coordination function. IEEE Journal on Selected
Areas in Communications 18(3):535547.
4. Foh CH, Tantra JW (2005), Comments on IEEE 802.11 saturation throughput analysis with freezing of backoff counters. IEEE
Communications Letters 9(2):130132.
5. Gas M, Kosek-Szott K, Natkaniec M, Pach A (2011), 3D Markov
chain-based saturation throughput model of IEEE 802.11 EDCA.
Electronics letters 47(14):826827.
6. Huang CL, Liao W (2007) Throughput and delay performance of
ieee 802.11 e enhanced distributed channel access (edca) under
saturation condition. Wireless Communications, IEEE Transactions on 6(1):136145.
7. Inan I, Keceli F, Ayanoglu E (2007a) Modeling the 802.11 e
enhanced distributed channel access function. In: Global Telecommunications Conference, 2007. GLOBECOM07. IEEE,
IEEE, pp 25462551.
8. Inan I, Keceli F, Ayanoglu E (2007b) Performance analysis of the
ieee 802.11 e enhanced distributed coordination function using
cycle time approach. In: Global Telecommunications Conference,
2007. GLOBECOM07. IEEE, IEEE, pp 25522557.
9. Jacques I, Judd C (1987) Numerical Analysis. Chapman and Hall:
London.
10. Lee JY, Lee HS (2009), A performance analysis model for IEEE
802.11 e EDCA under saturation condition. IEEE Transactions on
Communications 57(1):5663.
11. Lin Y, Wong VW (2006) Saturation throughput of ieee 802.11 e
edca based on mean value analysis. In: Wireless Communications
and Networking Conference, 2006. WCNC 2006. IEEE, IEEE,
vol 1, pp 475480.
12. Taher NC, Doudane YG, El Hassan B (2009a) A complete and
accurate analytical model for 802.11 e edca under saturation
conditions. In: Computer Systems and Applications, 2009.
73
13.
14.
15.
16.
123