Design and Analysis of an Adaptive Backoff Algorithm

for IEEE 802.11 DCF Mechanism

Mouhamad Ibrahim and Sara Alouf

INRIA – B.P. 93 – 06902 Sophia Antipolis – France

{mibrahim, salouf}@sophia.inria.fr

Abstract. This paper presents an adaptive backoff algorithm for the contention-
based Distributed Coordination Function (DCF) of the IEEE 802.11 standard.
Relying on on-line measurements of the number of sources, the algorithm, called
Adaptive BEB, judiciously sets the size of the minimal contention window to
adapt to the congestion level in the shared medium. The paper also provides an
extension to Adaptive BEB for enhancing its performance over noisy channels.
In this extension, a simple EWMA filter is used to derive a Packet Error Rate es-
timator. The performance evaluation of our proposal is addressed via simulations.

Keywords: DCF mechanism, optimal adaptation, performance evaluation, noise.

1 Introduction

Since its initial appearance in 1997, the IEEE 802.11 standard have engendered several
research activities due to wireless local area networks (WLANs) specific features. One
of the well-known drawbacks of IEEE 802.11 is the low performance of its MAC proto-
col in terms of throughput in congested networks. Several papers have pointed out this
problem and proposed solutions to counter it. However, and as will be shown later, the
performance of some of these degrades substantially under various scenarios.
In this paper, we aim at designing a new adaptive and robust backoff algorithm that
enhances the performance of the Distributed Coordination Function (DCF) of the IEEE
802.11 standard under a wide variety of conditions. This new algorithm is meant to be
simple and as close as possible to the standard. Our objectives are: maximum system
performance in terms of maximum system throughput and minimum packet delay, and
robustness to channel errors. Considering saturation conditions, i.e. active nodes have
always packets to transmit, we derive a simple expression relating the minimum con-
tention window size of the DCF backoff to the optimal transmission probability. The
latter probability has been computed in [1], and an alternative formulation can be de-
rived from [2, 3]. Both formulations require an estimate of the number of contending
stations. In this paper, we propose a novel method for estimating the number of sources,
which relies on counting signs of life coming from other stations to estimate their num-
ber. Despite many studies concerned with the optimization of the DCF in congested
networks, the presence of errors on the channel is often disregarded. Most proposals
rely on missed acknowledgments (ACK) from the destination to trigger the control of
the contention window. However, missed ACKs are due either to congestion or channel

F. Boavida et al. (Eds.): NETWORKING 2006, LNCS 3976, pp. 184–196, 2006.
c IFIP International Federation for Information Processing 2006
 Design and Analysis of an Adaptive Backoff Algorithm 185

errors. In this paper, we introduce a mechanism based on an Exponentially Weighted

Moving Average (EWMA) estimator of the Packet Error Rate (PER) seen on the chan-
nel to adjust the contention window, reducing thus the overhead introduced by the noise
while still avoiding collisions.
The rest of the paper is as follows: Sect. 2 briefly reviews the DCF and some related
work. The analytical background needed to design our algorithm is presented in Sect. 3.
Section 4 is devoted to the algorithm itself, motivating every step of it and describing its
design and implementation. A simple estimator of the number of contending nodes is
presented in Sect. 5 and an enhancement of the backoff algorithm in noisy environments
is presented in Sect. 6. Section 7 presents the simulation results and Sect. 8 studies the
fairness of our algorithm. Finally, Sect. 9 summarizes our work.

2 Preliminaries on the DCF Mechanism and Related Work

The Distributed Coordination Function (DCF) is the primary access protocol in 802.11
for sharing the wireless medium between active stations. In DCF, the stations listen to
the channel before transmitting: upon channel idleness for a duration greater than the
Distributed InterFrame Space (DIFS) period, the station transmits directly; otherwise it
waits for channel idleness. When the channel becomes idle again for a DIFS, the sta-
tion enters a deferring phase by selecting a random number of time slots in the range
[0, CW − 1] that is used to initialize a backoff counter [4]. Here CW refers to Con-
tention Window and it is an integer, set at the first packet transmission attempt to the
minimum contention window CW0 , and doubled as long as the transmission fails until
reaching a maximum size CWmax = 2m CW0 , where m denotes the maximum backoff
stage. The process of doubling the size of CW is called binary exponential backoff.
The backoff counter is a timer decreased as long as the channel is sensed idle. When
the backoff counter reaches 0, the station transmits directly, then waits for an ACK from
the destination station. If the source station has not received an ACK within a specified
ACK timeout, it will assume that the transmitted packet was lost due to a collision, and it
will start the backoff procedure again after doubling its contention window size. On the
other hand, if the packet is well received by the destination station, the latter will send
an ACK to the source station that upon its reception will reset its contention window
size to the minimal value CW0 and proceed to the next packet in the buffer. In addition
to this two-way handshaking technique, named basic access mechanism, DCF speci-
fies another optional four-way handshaking technique called Request To Send/Clear To
Send (RTS/CTS). The idea is to reserve the channel prior to a transmission by exchang-
ing the RTS and CTS control frames between the source and the destination before the
transmission of the packet.
The binary exponential backoff algorithm used in DCF has two major drawbacks.
First, the contention window is increased upon transmission failure regardless of the
cause of failure. Second, after a successful transmission of a packet, the contention
window is reset to CW0 , thus forgetting its knowledge of the current congestion level in
the network. These two points are at the basis of the inefficiency of the DCF mechanism.
As mentioned earlier, several research works have proposed modifications to IEEE
802.11 back-off algorithm to enhance its operation in congested networks. The earliest
186 M. Ibrahim and S. Alouf

proposal [5] is based on a unique contention window CW , whose size is updated after
each transmission attempt given an estimation of the number of active stations N . In
order to estimate N , the authors measure the number of busy slots observed during a
backoff decrease period. Another algorithm based on a unique contention window size
is given in [2]. The authors establish an analogy between the standard protocol and the
p-persistent IEEE 802.11 protocol, in which the backoff interval is sampled geometri-
cally with parameter p. By maximizing the analytical expression found for the through-
put of the p-persistent protocol, the authors derive the optimal p and subsequently the
optimal size of the backoff interval. An estimation of the number of actual active users
is required in this approach and the authors rely on measures of idle time to perform this
estimation. In [6], the authors propose an extension to the DCF mechanism in order to
achieve maximum throughput. Instead of directly transmitting when the backoff counter
reaches 0, the authors propose to differ the transmission with a certain probability that
depends, among other terms, on the slot utilization and the distribution of the packet
lengths (assumed to be geometric). For this algorithm to work, each station requires a
measure of the slot utilization and an inference of the packet length distribution. The
authors of [7] propose to multiplicatively decrease the contention window after a suc-
cessful transmission using a decrease factor δ in the range [0, 1] that is set constant for
all stations. The choice of δ greatly influences the performance of the algorithm. Maxi-
mum system throughput is achieved when δ = 0.9 at the cost of a high system response
time. A linear increase/linear decrease contention window algorithm is proposed in [8].
More precisely, upon a missed ACK, the current contention window is increased by a
constant value ω. Upon receiving an ACK, it is decreased by ω with probability 1 − δ,
and kept unchanged with probability δ. The parameters ω and δ are set heuristically to
constant values. To the best of our knowledge, only [9] proposes a mechanism to en-
hance the DCF mechanism for noisy environments. The proposed algorithm resets the
contention window when a failed transmission is assumed to be noise-corrupted. When
the RTS/CTS mechanism is used, it considers as noise-corruption the absence of ACK
and as collision the absence of CTS. When the basic access mechanism is used, the
assumed noise-corruption probability is adjusted linearly to approach the ideal trans-
mission probability, computed thanks to a count of the number of active stations. Last,
we would like to briefly mention that [10] proposes a Kalman filter to estimate the num-
ber of active users N and relies on channel sensing to measure the collision probability
(noisy channels are not considered in this study).

3 The Model
To design our algorithm, we adopt the model developed in [1]. In [1], it is assumed
that every source node has always packets to send; the channel conditions are assumed
to be ideal; the collision probability is constant among all sources and independent of
the past. Last, it is assumed that there is no limit on the number of retransmissions of
a lost packet. In [1], the author derives the following expression for the transmission
probability τ in terms of the protocol parameters
2(1 − 2p)
τ= (1)
(1 − 2p)(CW0 + 1) + p CW0 (1 − (2p)m )
 Design and Analysis of an Adaptive Backoff Algorithm 187

where m = log2 (CWmax /CW0 ) and p denotes the collision probability seen by a
transmitting source node. It is equal to 1 − (1 − τ )N −1 , where N denotes the number of
active stations. For a given N , τ and p can be computed using a fixed-point approach.
The author of [1] derives also an approximation of the optimal transmission probability,
where the optimality refers to maximizing the system throughput. Let Tc denote the
expected time taken by an unsuccessful transmission and σ denote the slot time length,
the approximate optimal transmission probability is then given by

1
τ∗ =  . (2)
N Tc /(2σ)

Insights on the Optimality of the Transmission Probability

The optimal transmission probability found in [1] will not only maximize the saturation
throughput but will minimize as well the system response time. In other words, the idle
time wasted over the wireless channel will be minimal. In this section, we will provide
some insights on the optimality of the transmission probability.
The time over the wireless channel can be partitioned into successive virtual trans-
mission times. A virtual transmission time initiates just after a successful transmission
over the wireless channel, and ends at the end of the next successful transmission, as
illustrated in Fig. 1. The wasted time in a virtual transmission time, denoted as wasteτ ,
is due to collisions and idle slot times and can be written

wasteτ = E [Nc Tc + (Nc + 1)Idleness] = E[Nc ]Tc + (E[Nc ] + 1)E[Idleness] (3)

where E[Nc ] and E[Idleness] respectively represent the average number of collisions
and the average length of an idle period in a virtual transmission time (Nc and Idleness
are statistically independent). In (3) it is assumed that all packets have the same size,
hence Tc is a constant. The first and second terms of (3) respectively account for the
total time wasted in collisions and the total idle time in a virtual transmission time.
Observe that the virtual transmission time is simply wasteτ +Ts (Ts being the expected
time taken by a successful transmission), so the normalized system throughput can be
expressed as the ratio E[payload]/(wasteτ + Ts ). To derive expressions for E[Nc ] and
E[Idleness] we look at the distribution, in a virtual transmission time, of the number of
collisions, Nc , and the idle period length. Let PI = (1 − τ )N , Ps = N τ (1 − τ )N −1
and Pc = 1 − PI − Ps be the probabilities of having an idle slot time, a successful
transmission and a collision, respectively. Let σ be the slot time duration. We have
 j
P [Nc = j] = Pc
Pc +Ps
Ps
Pc +Ps , P [Idleness = jσ] = PIj (1 − PI )

SIFS DIFS Tc DIFS Ts DIFS

Idle Idle Successful Idle
ACK slots Collision slots transmission ACK slots
j−1 σ virtual transmission time j SIFS j+1

Fig. 1. An illustration of the virtual transmission time

188 M. Ibrahim and S. Alouf

600 (a) Expected collisions, idle and wasted times (expressed in μs)

Idle, N = 50
500

Idle, N = 20
Idle, N = 10
400

Idle, N = 5
Collisions, N = 50
Collisions, N = 20
300

Collisions, N = 10
Collisions, N = 5
200

Waste, N = 50
Waste, N = 20
100

Waste, N = 10
Waste, N = 5
0

0 0.005 0.01 0.015 0.02 0.025 0.03

Transmission probability τ

Fig. 2. Expected wasted time, expected time spent in collisions and expected idle time (all are in
μs) in a virtual transmission time for different values of N (Tc = 4335μs and σ = 20μs).

for j ∈ IN, yielding E[Nc ] = Pc /Ps and E[Idleness] = σPI /(1 − PI ). After some
calculus, it comes that the total time wasted in collisions and the total idle time are
 
)N )N
E[Nc ]Tc = Tc N1−(1−τ
τ (1−τ ) N −1 − 1 , (E[Nc ] + 1)E[Idleness] = N τσ(1−τ
(1−τ )N −1 .

It can easily be proved that E[Nc ]Tc is a monotone increasing function of τ , whereas
(E[Nc ]+ 1)E[Idleness] is a monotone decreasing function of τ . In Fig. 2, both terms are
plotted against τ for different values of the number of contending stations N . Their sum,
wasteτ , is also plotted and its minimal value for each N has been marked. The minimal
values of wasteτ naturally correspond to the optimal transmission probability for each
N . As seen in Fig. 2, the minimal values of wasteτ correspond to the intersections
of both collisions and idle times. In other words, the optimality is achieved when the
wasted time is equally shared between collisions and idleness. This optimal operating
point has been identified in [11, p. 243] for the CSMA slotted Aloha protocol and in [3]
for the p-persistent IEEE 802.11 protocol.
Observe that having the system response time minimized at the optimal transmission
probability does not guarantee that the packet delay is minimized as well. This mini-
mization also relies on the protocol fairness and its ability to equally share the medium
among contending sources.

4 Adaptive Binary Exponential Backoff (BEB)

According to equation (2), optimal system performance can be achieved in different

network topologies by fine-tuning the transmission probability of the stations to the
optimal transmission probability τ ∗ for each network topology. For a given network
topology, transmitting at the optimal transmission probability can be achieved by well
adjusting the size of the minimum starting contention window CWmin . Note that a one-
to-one relation between these two parameters can be obtained by inverting equation (1),
and CWmin will then be expressed in terms of τ ∗ as follows:
 Design and Analysis of an Adaptive Backoff Algorithm 189

Algorithm 1. Adjustment algorithm

Require: An estimation N of the number of active nodes

Ensure: Sub-optimal values of CWmin and m
1: Set τ = τ ∗ = √ 1
N Tc /2σ

2: Set p = 1 − (1 − τ ∗ )N−1
∗ )(1−2p)
3: Compute cw = τ ∗(2−τ(1−p−p(2p)m )
4: Select j such that 2j CW0 is the closest to cw, j ∈ [0, m]
5: Set CWmin = 2j CW0
6: Set m = log2 CW max
CWmin
{CWmax is fixed}

(2 − τ ∗ )(1 − 2p)
CWmin = . (4)
τ ∗ (1
− p − p(2p)m )

Unfortunately, the parameter m will still depend on the previous value of CWmin as
follows m = log2 (CWmax /CWmin ). However, this practically will not impact the new
selected value of CWmin since the term p(2p)m can be neglected with respect to 1.
The operation of our algorithm can then be summarized as follows. After a failed
transmission, the behavior is the same as in the standard. However, after a successful
transmission, the initial window size of the binary exponential backoff mechanism is set
to an optimal value, hereafter denoted CWmin . The selection of this value is detailed in
Algo. 1. Given an estimation of N , the approximate optimal transmission probability
is computed using (2) (line 1), enabling the computation of the collision probability p
(line 2). The optimal minimal size of the contention window size is computed according
to (4) (line 3). Note that to be as close as possible to the standard, CWmin is imposed
to take only those values defined by the standard {2j CW0 , j = 0, . . . , m} (recall that
CW0 is the size of the minimal contention window in the standard) (lines 4–5). Last,
the value of m is updated (line 6). Observe that the values of CWmin and m will be
sub-optimal due to the fact that CWmin is made discrete.
The analytical throughput achieved by our adaptive algorithm, hereafter referred
to as “Adaptive BEB”, can be obtained by substituting the new values of CWmin
and m and the optimal value of p given in line 2 of Algo. 1 to compute the sub-
optimal value of τ , and therefore the throughput achieved by Adaptive BEB will be
T = E[payload]/(wasteτ + Ts ).

5 Estimation of the Number of Contending Stations

To adjust CWmin , we need an estimation of N , hereafter denoted N , that tracks its

time-evolution. However, since the value of CWmin is made discrete, a reactive and
adequately accurate estimator is sufficient. For instance, consider the case when N =
30. When applying Algo. 1, we would obtain the correct optimal value CWmin =
24 CW0 for N ∈ [20.45, 40.95].
As opposed to previously proposed methods (e.g. [2, 5, 10]) which estimate the cur-
rent number of sources by measuring the channel activity, we propose to estimate the
190 M. Ibrahim and S. Alouf

(a) N and its mean observations

before and after the correction (b) N and its estimation vs. time
50

08 100
Mean observation before correction Estimated N
Mean observation after corrected True N
40

True N
30

60
20

40
10

20
0

0
0 10 20 30 40 50 0 25 50 75 100 125
Number of active sources N Time (s)

Fig. 3. (a) Mean observations of N vs. N . (b) Estimation of N vs. time in sample simulation

number of active stations directly by counting as active each station from which the con-
cerned measuring station receives a sign of life, i.e. error-free data or RTS packets. The
measurement period will be the virtual transmission time of the station, i.e. the time be-
tween its consecutive successful transmissions. The idea can be summarized as follows:
since each active node is always filtering received packets whether they are destined to
it or not, it can thus keep trace of a large number of active nodes in a given time interval
by counting the number of distinct signs of life received in that time interval from these
nodes. Let N n denote the count of distinct signs of life at the nth measurement period
of a given station. For multiple reasons, these measurements cannot be used directly in
Algo. 1 as an estimation of N . First, the count of signs of life cannot account for stations
whose transmitted packets are corrupted, either by noise or collisions. Second, variable
and relatively small measurement periods result in counting only a variable portion of
all active stations. Therefore, based on the measurements {N n }n∈IN∗ , an unbiased es-
timator should be devised. Observe that measurements collected over relatively large
measurement times are more “accurate” than those collected over small measurement
times, and should be preferentially treated. By observing that the expected length of a
measurement period is reflected in cw, the value reached by the contention window at
the end of the measurement time, one can use this value toq−1
proportionally weight the cor-
responding measurement, so that the following filter i=0 cwn−i N n−i / q−1 cwn−i
i=0
can be used. Simulation analysis has shown that low values of the filter order q will suffice
for a convenient performance, so hereafter, we set q = 3 so as to heighten reactiveness.
However, and because of the fact that corrupted packets cannot contribute to the mea-
surements, the previous ratio, henceforth referred to as the observation, underestimates
N and needs thus to be corrected. To derive an appropriate correction on the observations,
we have performed several simulations with different values of N . Figure 3(a) depicts
the evolution of the observation against the actual number of sources. Obviously, a linear
correction is needed, resulting in the following estimator
q−1 q−1
 
Nn = a cwn−i N n−i / cwn−i + b (5)
i=0 i=0

with a = 1.35405, b = 1.75998 for q = 3. These latter values, which are independent
of the nodes distribution, need to be adjusted in the case of noise over the channel; this
 Design and Analysis of an Adaptive Backoff Algorithm 191

issue is left for future work. Figure 3(b) illustrates the estimation of N over a simula-
tion in which nodes arrivals are Poisson and activity time is exponentially distributed.
The estimator exhibits good reactiveness to changes in N at the cost of large fluctua-
tions, but these have only a limited impact on the performance of Algo. 1 because of the
discretization process.

6 Enhancement of the Backoff Algorithm in Noisy Environments

The assumption of ideal channel conditions is in general unrealistic due to the existence
of various “noise” factors (e.g. fading, shadowing, interference) that perturb the state
of the wireless medium. Whenever a frame is noise-corrupted, both the standard and
Adaptive BEB behave as if the loss is due to a collision, and the contention window
of the backoff algorithm will be accordingly increased. Clearly, there is a flaw in the
design of these algorithms due to the automatic invocation of the contention window
increase process in the absence of CTS/ACK. Both algorithms lack to identify the cause
of a missed CTS/ACK, and fail in adapting to error-prone environments. In the following,
we present an optional extension that can be applied to Algo. 1to enhance its performance
in noisy environments. In this extension, we propose a Packet Error Rate (PER) estimator
(Sect. 6.1), and a persistence mechanism that makes use of this estimator (Sect. 6.2). The
resulting algorithm will be denoted as “Adaptive BEB++ ”.

6.1 PER Estimation

To estimate the Packet Error Rate, we propose to infer it from the measured corruption
probability in a station virtual transmission time, and filter these inferred values through
a simple Exponentially Weighted Moving Average (EWMA) filter. This method works
well in both RTS/CTS and basic access modes. When a transmission occurs on the chan-
nel, all other stations within communication range receive the transmitted packet. Upon
receiving a packet, every station verifies first its Check Redundancy Code (CRC) so that
corrupted packets could be disregarded. Therefore, every station is capable of counting
how many received packets are corrupted out of all of them. The ratio between the two
counts, noted p̂cr , is nothing but a measure of the corruption probability, pcr , perceived
in a measurement time. The corruption probability can be written pcr = pc + pe (1 − pc )
where pe is the PER and pc is the collision probability seen on the channel by a mea-
suring station. We have pc = 1 − (1 − τ )N −1 − (N − 1)τ (1 − τ )N −2 . Using N
and τ ∗ we can infer pc ; the inferred value is denoted p̂c . Therefore, pe can be inferred
as follows p̂e = (p̂cr − p̂c )/(1 − p̂c ) and is used to feed the following EWMA filter:
p̄e,n = αp̄e,n−1 + (1 − α)p̂e,n , where p̄e,n and p̂e,n respectively denote the estimated
and inferred PER in the nth measurement time at a given station. As for α, we have per-
formed several simulations with N = 25 and an abruptly varying PER, as illustrated in
Fig. 4(d). We have computed both expectation and variance of the error pe − p̄e,n for
different values of α, and selected the value minimizing the mean error. This value is
α = 0.95 as can be seen from the table below. Observe in Fig. 4(d) how α = 0.95 yields
a highly reactive estimator.
192 M. Ibrahim and S. Alouf

α 0.9 0.95 0.99 0.995 0.999

E[pe − p̄e,n ] 0.0174836 0.0153285 0.0192210 0.0259682 0.0714458
Var[pe − p̄e,n ] 0.0494257 0.0498566 0.0479849 0.0453496 0.0282030

6.2 The Persistence Algorithm

Unlike the standard and the Adaptive BEB mechanisms that upon a missed CTS/ACK
blindly defer the transmission, we propose a persistence probability Pp such that the for-
mer behavior is undertaken only with probability 1 − Pp . Upon a missed CTS/ACK and
with probability Pp , the station retransmits the packet when the CTS/ACK timeout ex-
pires. In other words, Pp can be regarded as the probability of assuming that a failed
transmission is due to a noise-corruption, not a collision. Obviously, the choice of this
persistence probability is crucial for the algorithm to perform well. Pp should account for
p̄e , the backoff stage, and the count of persistent trials within the current backoff stage.
Clearly, it should increase with p̄e . Also, as a station increments its window size for a
given packet retransmission, the collision probability decreases. If, however, the retrans-
mission still fails, then it is more likely that it was due to a noise-corruption rather than
a collision, and therefore, Pp should increase with the actual backoff stage that is equal
to log2 (CW/CWmin ). Last, if, within the same stage, persistent retransmissions are re-
peatedly being unsuccessful, then it is more likely that the station is observing collisions
and not corruption by the noisy channel, and as such, the persistence probability should
decrease with trials, the count of persistent retransmissions within the same backoff
stage. In our implementation, we have used the following expression of the persistence
probability which exhibits all above-mentioned desired trends:
trials
1+log2 (CW/CWmin )
Pp = p̄e .

7 Simulation Results
In this section, we investigate the effectiveness of our proposal through simulations con-
ducted in ns-2 [12]. Simulated nodes are uniformly distributed in a 100m×100m square
and the power transmission is sufficiently high so that all the nodes are within commu-
nication range. All sources generate Constant Bit Rate (CBR) traffic. The protocol pa-
rameters are as follows: CW0 = 32, CWmax = 1024 and the slot size is σ = 20μs. The
expected collision duration Tc used in (2) corresponds to the maximum length of the data
packet delivered by the MAC layer to the PHY layer and depends on the data rate. The
proposed algorithms have been thoroughly tested considering both ad hoc and infras-
tructure modes of operation, fixed and abruptly changing number of sources as well as
Poisson source arrivals. Both error-free and error-prone channels have been considered.
Due to space limitations, we will discuss only 2 scenarios exhibiting the most relevant
properties of the proposed algorithms.

Scenario A: N nodes are uploading CBR traffic to an access point. Starting with 10
active nodes for a duration of 25s, two 20-node bursts join the network at instants 25s
 Design and Analysis of an Adaptive Backoff Algorithm 193

(a) Normalized system throughput vs. time (b) End-to-end packet delay (s) vs. time

7
0.8

Adaptive BEB

6
Standard
Additive
0.6

5
Multiplicative

4
0.4

3
Adaptive BEB
Standard

2
0.2

Additive
Multiplicative

1
0

0
0 25 50 75 100 125 0 25 50 75 100 125
Time (s) Time (s)
(c) Gain in normalized system throughput vs. PER (d) PER and EWMA estimators vs. time
1.4

N = 10 Adapt. BEB

0.6
N = 30 Adapt. BEB
1.3

N = 50 Adapt. BEB
N = 10 Adapt. BEB++

0.45
++
N = 30 Adapt. BEB++
1.2

N = 50 Adapt. BEB
0.3
1.1

Estimated PER, α = 0.9

Estimated PER, α = 0.95
0.15

Estimated PER, α = 0.99

1

True PER
0.9

0 0.1 0.2 0.3 0.4 0.5 0.6 0 25 50 75 100 125

Packet Error Rate Time (s)

Fig. 4. (a) Simulative system throughput and (b) end-to-end delay (in s) over time (scenario A).
(c) Throughput gain of Adaptive BEB and Adaptive BEB++ against the standard (scenario B).
(d) PER and EWMA estimators vs. time for several α values.

and 50s, and then leave the network consecutively at instants 75s and 100s. Each node
generates 1050B-packets every 5ms. Data rate at the physical layer is 2Mbps. There is
no error on the wireless channel and the basic access mechanism is used. We have in-
vestigated the performance of 3 algorithms beside the standard: the Adaptive BEB, the
additive increase/additive decrease algorithm [8] and the multiplicative slow decrease
algorithm [7] referred to “Additive” and “Multiplicative” respectively. Figures 4(a) and
4(b) respectively illustrate the system throughput and the packet delay (in seconds) over
time, when using the standard and the three adaptive algorithms. As can be seen, Adap-
tive BEB exhibits a steady performance over the simulation time whether in terms of
throughput or packet delay, and is the only one to rapidly recover after the abrupt in-
crease of N at times 25s and 50s. This can be explained by the fact that an optimal CWmin
is rapidly selected, consequently avoiding the high number of collisions that are expe-
rienced by the standard. The slow decrease approaches [7, 8] perform very well when
the number of sources is constant or changes smoothly. However, their performance de-
grades severely in the scenario at hand for 2 reasons. First, their control algorithms are
relatively slow compared to the important change in the network state. Second, in the in-
frastructure mode of operation, some congestion occurs at the access point buffer yielding
lost packets and subsequent missed ACKs. Buffer overflow is penalizing more the slow
decrease approaches than the Adaptive BEB as in these approaches the contention win-
dow after a successful transmission will be unnecessarily large. In contrast, the Adaptive
BEB controls the CWmin by using N , so that only retransmissions affect the system
performance.
194 M. Ibrahim and S. Alouf

Scenario B: There are 2N nodes in the network: N sources and N destinations op-
erating in ad hoc mode with a data rate at the physical layer set to 11Mbps. The basic
access mechanism is used. Each source generates 1000B-packets every 0.8ms. The error
on the channel is modeled as a Bernoulli loss process. Packets are noise-corrupted with a
probability PER set constant throughout each simulation runtime (100s). Two algorithms
are investigated: the Adaptive BEB and the Adaptive BEB++ . Figure 4(c) plots, for dif-
ferent values of N , the throughput gain (defined as the ratio between both throughputs)
achieved by each algorithm with respect to the standard versus the PER. Due to its per-
sistence mechanism, Adaptive BEB++ shows a high robustness to error on the channel,
and the gain obtained is always greater than 1 for different values of PER and N . For
instance there is as much as 19% more throughput with Adaptive BEB++ when N = 10
and PER = 0.6. The Adaptive BEB has previously shown better performance than the
standard. This is no longer true at high PER values. Actually, the Adaptive BEB is much
more penalized by high PERs since in this case the average backoff time will be much
higher than in the standard due to a larger starting contention window. There will be
simply too much idle times over the channel.

8 Fairness Analysis

Fairness describes the MAC protocol capability to distribute the available resources
equally among communicating terminals. There are a variety of fairness definitions in-
tended to support different QoS and service differentiation. In our study, fairness is simply
the equal partitioning of the bandwidth among all flows at hand. We have run simulations
when either one of the following algorithms is used: Adaptive BEB or the standard when
PER = 0, and Adaptive BEB++ or the standard when PER = 0.4. We considered 3
different values of N , namely, 10, 30 and 50. For each simulation, we have computed
 2   
I I 2
the index of fairness of Jain et al. [13] f (x1 , x2 , . . . , xI ) = i=1 xi / I i=1 xi
where xi denotes user i allocation, and I the number of users sharing the bandwidth. This
index returns the percentage of flows being treated fairly. To compute the index of fair-
ness, we use the Sliding Window Method (SWM) introduced in [14]. A small sliding

(a) error-free environments (PER = 0) (b) error-prone environments (PER = 0.4)

1

1
0.8

0.8
0.6

0.6

N = 10, Adapt. BEB N = 10, Adapt. BEB++

0.4

0.4

N = 10, Standard N = 10, Standard

N = 30, Adapt. BEB N = 30, Adapt. BEB++
N = 30, Standard N = 30, Standard
0.2

0.2

N = 50, Adapt. BEB N = 50, Adapt. BEB++

N = 50, Standard N = 50, Standard
0

10 100 1000 10000 10 100 1000 10000

Sliding window size Sliding window size

Fig. 5. Fairness index versus the sliding window size

 Design and Analysis of an Adaptive Backoff Algorithm 195

window size allows the study of the short-term fairness whereas large sliding window
sizes enable the long-term fairness one. The results are plotted in Fig. 5.
Observe that Adaptive BEB (respectively Adaptive BEB++ ) achieves better fairness,
i.e. higher index value, than the standard when PER = 0 (respectively when PER =
0.4), for any investigated value of N . The fact that our algorithm assigns to the contending
nodes optimal and relatively equal CWmin allows on one hand to minimize the number
of collisions, and consequently the risk of having largely different backoff intervals, and
on the other hand, to obtain equal opportunities of accessing the channel for the various
nodes.

9 Conclusion

In this paper, we have proposed and evaluated a simple, efficient and robust adaptive
mechanism that can be easily incorporated within the IEEE 802.11 DCF function in or-
der to optimize its performance. The proposed algorithm includes two mechanisms. The
first optimally adjusts the minimum contention window size to the current network con-
gestion level by using an on-line estimation of the number of active stations. The other
mechanism extends the first to enhance its performance in the presence of noisy channels.
Simulation results have shown that our proposed algorithm outperforms the standard in
terms of throughput, packet delay, fairness and robustness to noise, for different scenarios
and configurations. When compared to other proposals, our algorithm has shown better
performance under several configurations.

Acknowledgements. The second author wishes to thank I. Aad and R. van der Mei for
fruitful discussions at an early stage of this work.

References
1. Bianchi, G.: Performance analysis of the IEEE 802.11 distributed coordination function.
IEEE Journal on Selected Areas in Communications 18(3) (2000) 535–547
2. Cali, F., Conti, M., Gregori, E.: Dynamic tuning of the IEEE 802.11 protocol to achieve a
theoretical throughput limit. IEEE/ACM Trans. on Networking 8(6) (2000) 785–799
3. Cali, F., Conti, M., Gregori, E.: IEEE 802.11 protocol: design and performance evaluation of
an adaptive backoff mechanism. IEEE J. on Sel. Areas in Comm. 18(9) (2000) 1774–1786
4. IEEE Std. 802.11b, Part 11: Wireless LAN Medium Access Control (MAC) and Physical
Layer (PHY) Specification. (1999)
5. Bianchi, G., Fratta, L., Oliveri, M.: Performance evaluation and enhancement of the
CSMA/CA MAC protocol for 802.11 wireless LANs. In: Proc. of PIMRC ’96. (1996)
6. Bononi, L., Conti, M., Gregori, E.: Runtime optimization of IEEE 802.11 wireless LANs
performance. IEEE Transactions on Parallel and Distributed Systems 15(1) (2004) 66–80
7. Aad, I., Ni, Q., Barakat, C., Turletti, T.: Enhancing IEEE 802.11 MAC in congested environ-
ments. In: Proc. of IEEE ASWN ’04, Boston, Massachusetts. (2004)
8. Galtier, J.: Optimizing the IEEE 802.11b performance using slow congestion window de-
crease. In: Proc. of 16th ITC Specialist Seminar, Anvers, Belgique. (2004)
9. Nadeem, T., Agrawala, A.: IEEE 802.11 DCF enhancements for noisy environments. In:
Proc. of PIMRC ’04, Barcelona, Spain. (2004)
196 M. Ibrahim and S. Alouf

10. Bianchi, G., Tinnirello, I.: Kalman filter estimation of the number of competing terminals in
an IEEE 802.11 network. In: Proc. of IEEE INFOCOM ’03. (2003)
11. Bertsekas, D., Gallager, R.: Data Network. Prentice Hall (1992)
12. The network simulator, version 2.28 (2005) http://www.isi.edu/nsnam/ns/.
13. Jain, R., Hawe, W., Chiu, D.: Quantitative measure of fairness and discrimination for resource
allocation in shared computer systems. Technical Report DEC-TR-301, DEC (1984)
14. Koksal, C.E., Kassab, H., Balakrishnan, H.: An analysis of short-term fairness in wireless
media access protocols (extended abstract). Perf. Eval. Rev., 28(1) (2000) 118–119