Adaptive EDCF: Enhanced Service Differentiation
for IEEE 802.11 Wireless Ad-Hoc Networks
Lamia Romdhani, Qiang Ni, and Thierry Turletti
INRIA Sophia Antipolis, 2004 Route des Lucioles, BP-93, 06902 Sophia Antipolis, France
Email: {lromdhan, qni, turletti}@sophia.inria.fr
Abstract— This paper describes an adaptive service differentia-
tion scheme for QoS enhancement in IEEE 802.11 wireless ad-hoc
networks. Our approach, called Adaptive Enhanced Distributed
Coordination Function (AEDCF), is derived from the new EDCF
introduced in the upcoming IEEE 802.11e standard. Our scheme
aims to share the transmission channel efficiently. Relative
priorities are provisioned by adjusting the size of the Contention
Window (CW) of each traffic class taking into account both
applications requirements and network conditions. We evaluate
through simulations the performance of AEDCF and compare
it with the EDCF scheme proposed in the 802.11e. Results
show that AEDCF outperforms the basic EDCF, especially at
high traffic load conditions. Indeed, our scheme increases the
medium utilization ratio and reduces for more than 50% the
collision rate. While achieving delay differentiation, the overall
goodput obtained is up to 25% higher than EDCF. Moreover,
the complexity of AEDCF remains similar to the EDCF scheme,
enabling the design of cheap implementations.
I. INTRODUCTION
IEEE 802.11 wireless LAN specification defines two dif-
ferent ways to configure a wireless network: ad-hoc and
infrastructure mode. In infrastructure mode an Access Point
(AP) is needed to connect wireless stations to a distribution
system, whereas in ad-hoc mode all wireless stations are
distributed without access coordinator. In this paper, we focus
on ad-hoc networks since distributed random access control
are often preferred to centrally coordinated access control
[10], [11]. Distributed Coordination Function (DCF) is the
basic medium access mechanism of 802.11 for both ad-
hoc and infrastructure modes. It uses CSMA/CA (Carrier
Sense Multiple Access with Collision Avoidance) protocol.
In this mode, if the medium is found idle for longer than
a DIFS (Distributed InterFrame Space) then the station can
transmit a packet. Otherwise, a backoff process is started.
More specifically, the station computes a random value called
backoff time, in the range of 0 and CW (Contention Window)
size. The backoff timer is periodically decremented by one for
every time slot the medium remains idle after the channel has
been detected idle for a period greater than DIFS. As soon as
the backoff timer expires, the station can access the medium.
If no acknowledgment is received, the station assumes that
collision has occured, and schedules a retransmission by re-
entering the backoff process.
Quality of Service (QoS) support is critical to multimedia
applications. Time-bounded services such as audio and video
conference typically require some specified bandwidth, delay
and jitter guarantee, but can tolerate some losses. However, in
DCF all the stations in a Basic Service Set or all the flows
from the same station compete the resources and channel with
the same priority. There is not any differentiation mechanism
to guarantee packet delay and jitter to stations or flows sup-
porting time-bounded multimedia services. The performance
evaluation results in [3], [4] show that DCF suffers from
significant throughput degradation and high delay at high load
conditions, which are caused by the increasing time used for
channel access negotiation. Many medium access schemes
have been proposed for IEEE 802.11 WLAN to provide some
QoS enhancements for real-time traffics. Previous research
works mainly focus on the station-based DCF enhancement
scheme [1], [5], [9], [11]. Other recent works focus on queue-
based enhancement schemes [2], [7], [15] since they perform
more efficiently. More related works are detailed in [8].
In parallel the IEEE working group is currently working on
the support of QoS in a new standard, called IEEE 802.11e
[15]. A new access method called Hybrid Coordination Func-
tion (HCF) is introduced, which combines functions from the
DCF and Point Coordination Function (PCF) mechanisms.
Enhanced DCF (EDCF) is a contention-based HCF channel
access specified in IEEE 802.11e [7], [15]. The goal of this
scheme is to enhance the DCF access mechanism of IEEE
802.11 and to provide a distributed access approach that can
support service differentiation. The proposed scheme provides
capability for up to eight types of traffic classes. It assigns a
short CW to classes that should have higher priority in order
to ensure that in most cases, high-priority classes will be able
to transmit before the low-priority ones. Indeed, the CW
min
parameter can be set differently for different priority classes,
yielding higher priority classes with smaller CW
min
.For
further differentiation, in 802.11e different IFS (Inter Frame
Space) can be used according to traffic classes. Instead of
DIFS, an Arbitration IFS (AIFS) is used. The AIFS for a given
class should be a DIFS plus some (possibly zero) time slots.
Classes with the smallest AIFS will have the highest priority
as it is shown in Figure 1. Each Traffic Category (TC) within
the station behaves like a virtual station: it contends for access
to the medium and independently starts its backoff time after
sensing if the medium is idle for at least AIFS.
Per priority differentiation used by EDCF ensures better ser-
vices to high priority class while offering a minimum service
for low priority traffic. Although this mechanism improves
the quality of service of real-time traffic, the performance
obtained are not optimal since EDCF parameters cannot be
0-7803-7700-1/03/$17.00 (C) 2003 IEEE 1373
Authorized licensed use limited to: Brunel University. Downloaded on August 27, 2009 at 10:25 from IEEE Xplore. Restrictions apply.
Fig. 1. Some IFS relationships
adapted to the network conditions. In fact, since each TC is
implemented as a virtual station, the collision rate increases
very fast when the contentions to access the shared medium are
very high, which significantly affects the goodput, the latency
and thus, decreases the performance of delay-bounded traffic
[6]. This motivates us to propose a scheme that adapts the
CW parameter according to the network conditions. Our new
scheme called AEDCF aims to provide real-time support in
802.11 ad-hoc networks.
The remainder of this paper is organized as follows. In Sec-
tion II, we describe AEDCF in detail. Simulation methodology
and performance evaluation of our proposal are detailed in
Section III. Section IV concludes the paper by summarizing
results and outlining future works.
II. T
HE ADAPTIVE EDCF (AEDCF) SCHEME
Figure 2 compares the 802.11e architecture that supports
queue-based differentiation with the original one queue based
DCF access mechanism. To improve the performance under
different load rates and to increase the service differentiation
in EDCF-based networks, we propose a new scheme called
Adaptive EDCF (AEDCF). This scheme extends the basic
EDCF by making it more adaptive taking into account network
conditions.
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(AIFS)
(CW)
(PF)
backoff
(DIFS)
(15)
(2)
802.11e:
up to 8 independent backoff instances
one priority
legacy:
transmission
attempt
transmission
attempt
scheduler (resolves virtual collisions by granting TXOP to highest priority)
newold
P1 P2 P3 P4 P5 P6P0 P7
Fig. 2. Queue-based EDCF vs. basic DCF
We assume that n stations are sending packets through the
wireless media. The flows sent by each station may belong to
different classes of service with various priority levels. In each
station and for each class i, the scheme maintains: the current
contention window value (CW [i]), the minimum contention
window value (CW
min
[i]), and the maximum contention
window value (CW
max
[i]) . Note that i varies from 0 (the
highest priority class) to 7 (the lowest priority class).
A. Scheme Description
In order to efficiently support time-bounded multimedia
applications, we use a dynamic procedure to change the
contention window value after each successful transmission
or collision. We believe that this adaptation will increase the
total goodput of the traffic which becomes limited when using
the basic EDCF, mainly for high traffic load.
In the basic EDCF scheme for ad-hoc networks [15], the
CW
min
[i] and CW
max
[i] values are statically set for each
priority level. After each successful transmission, the CW [i]
values are reset to CW
min
[i]. We propose to reset the CW [i]
values more slowly to adaptive values (different to CW
min
[i])
taking into account their current sizes and the average collision
rate while maintaining the priority-based discrimination. In
other words, we ensure that at each instant, the highest priority
class has the lowest contention window value so that it has the
highest priority to access the media. The adaptive slow CW
decrease is a tradeoff between wasting some backoff time and
risking a collision followed by the whole packet transmission.
After each collision, the source has to wait for a timeout
to realize that the packet has collided, and then doubles its
CW to reduce the number of collisions [15]. We propose to
change the mechanism and differentiate between classes using
different factors to increase their CWs.
In the next sub-sections, we explain in detail how the
contention window of each priority level is set after each
successful transmission and after each collision.
1) Setting CW After Each Successful Transmission: After
each successful transmission, the basic EDCF mechanism
simply sets the contention window of the corresponding class
to its minimum contention window regardless the network
conditions. Motivated by the fact that when a collision occurs,
a new one is likely to occur in the near future, we propose to
update the contention window slowly (not reset to CW
min
)
after successful transmission to avoid bursty collisions. The
simplest scheme we can use to update the CW of each class
i is to reduce it by a static factor such as 0.5 ∗ CW
old
.
In the remainder of this paper, we denote this approach the
Slow Decrease (SD) scheme. However, a static factor cannot
be optimal in all network conditions. In our scheme, we
propose that every class updates its CW in an adaptive way
taking into account the estimated collision rate f
j
curr
in each
station. Indeed, the collision rate can give an indication about
contentions in a distributed network. The value of f
j
curr
is
calculated using the number of collisions and the total number
of packets sent during a constant period (i.e. a fixed number
of slot times) as follows:
f
j
curr
=
E(collisions
j
[p])
E(data
−
sent
j
[p])
, (1)
where E(collisions
j
[p]) is the number of collisions of station
p which occurred at step j, and E(data
−
sent
j
[p]) is the total
number of packets that have been sent in the same period j
1374
Authorized licensed use limited to: Brunel University. Downloaded on August 27, 2009 at 10:25 from IEEE Xplore. Restrictions apply.
by flows belonging to the station p. Note that the above ratio
f
j
curr
is always in the range of [0, 1].
To minimize the bias against transient collisions, we use
an estimator of Exponentially Weighted Moving Average
(EWMA) to smoothen the estimated values. Let f
j
avg
be
the average collision rate at step j (for each update period)
computed according to the following iterative relationship:
f
j
avg
=(1− α) ∗ f
j
curr
+ α ∗ f
j−1
avg
(2)
where j refers to the j
th
update period and f
j
curr
stands for
the instantaneous collision rate, α is the weight (also called
the smoothing factor) and effectively determines the memory
size used in the averaging process.
The average collision rate is computed dynamically in each
period T
update
expressed in time-slots. This period should not
be too long in order to get good estimation and should not be
too short in order to limit the complexity.
To ensure that the priority relationship between different
classes is still fulfilled when a class updates its CW , each class
should use different factor according to its priority level (we
denote this factor by Multiplicator Factor or MF). Keeping in
mind that the factor used to reset the CW should not exceed
the previous CW, we limit the maximum value of MF to
0.8. We have fixed this limit according to an extensive set of
simulations done with several scenarios. In AEDCF, the MF
of class i is defined as follows:
MF[i]=min((1 + (i ∗ 2)) ∗ f
j
avg
, 0.8). (3)
This formula allows the highest priority class to reset the
CW parameter with the smallest MF value (i.e., priority level
0, see P0 in Figure 2). After each successful transmission of
packet of class i, CW [i] is then updated as follows:
CW
new
[i]=max(CW
min
[i] ,CW
old
[i] ∗ MF[i]) (4)
The equation above guarantees that CW [i] is always greater
than or equal to CW
min
[i] and that the priority access to the
wireless medium is always maintained.
2) Setting CW After Each Collision : In the current version
of EDCF[15], after each unsuccessful transmission of packet
of class i,theCW of this class is doubled, while remaining
less than the maximum contention window CW
max
[i]:
CW
new
[i]=min(CW
max
[i] , 2 ∗ CW
old
[i]). (5)
In AEDCF, after each unsuccessful transmission of packet
of class i, the new CW of this class is increased with a
Persistence Factor PF[i], which ensures that high priority
traffic has a smaller value of PF[i] than low priority traffic:
CW
new
[i]=min(CW
max
[i] ,CW
old
[i] ∗ PF[i]). (6)
In fact, this PF parameter has been proposed in a previous
version of the draft, but it has been removed from draft [15].
In this paper, we introduce PF in our AEDCF scheme because
by this way we can reduce the probability of a new collision
and consequently decrease delay.
B. Evaluation of Complexity
Our mechanism is easy to implement, and needs very few
resources: it requires four registers to buffer the parameters
defined above: f
j−1
avg
, T
update
, MF[i] and α. To reset the
CW [i] values, f
j
avg
and MF[i] parameters defined in Equa-
tions 2 and 3, are updated only at the beginning of each new
update period T
update
. The calculation of MF[i] requires one
addition, two multiplications and one comparison for each
active class. Then, two multiplications and two additions are
required to compute f
j
avg
and one more division to obtain
f
j
curr
(which is defined in Equation 1 for all the active TCs).
One comparison and one multiplication are required to
compute MF[i] and to decide which value will be used to reset
the CW [i] (see Equation 4). Finally, during the update period,
we need two counters to increment collisions and data sent,
one comparison and one multiplication that are introduced in
Equation 6 to calculate the CW
new
[i] and to decide which
value will be used to reset the CW .
III. S
IMULATION METHODOLOGY AND RESULTS
We have implemented AEDCF in the ns-2 simulator, our ns
source codes are available in [12]. We report in this section
part of simulations we have done with different network
topologies and source characteristics. An analysis of perfor-
mance is presented in detail. In order to show advantages of
using an adaptive factor (MF[i]) to decrease the CW after
successful transmission, we also present the results of the static
Slow Decrease scheme.
A. Impact of T
update
and α parameters
As mentioned in Section II-A, our scheme uses an update
period (defined in number of time slots) after which it should
update the estimated collision rate (f
j
avg
). We have done
several set of simulations to observe the effect of the update
period on the delay and on the goodput performances [12].
Results obtained show that the update period has a slight
impact on the obtained total goodput. However, we have noted
that the delay significantly increases when T
update
is greater
than 6000 time-slots. In the following simulations we have
chosen T
update
equal to 5000 time-slots which provides a good
tradeoff between goodput and latency.
0
0
1
1
0
0
1
1
00
00
11
11
0
0
0
1
1
1
0
0
0
1
1
1
0
0
0
1
1
1
0
0
0
1
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
0
1
1
1
0
0
1
1
0
0
1
1
0
0
1
1
01
0
0
1
1
Node 2
Audio
Audio
Video
Video
Audio
Audio
Video
Audio
Video
Node 1
Video
Node n
Node i
Node i+1
BT
BT
BT
BT
BT
Fig. 3. Simulation Topology
Let us now analyze the impact of the smoothing factor
α on the performance of AEDCF. Note that this factor is
1375
Authorized licensed use limited to: Brunel University. Downloaded on August 27, 2009 at 10:25 from IEEE Xplore. Restrictions apply.
used to estimate the average collision rate defined in Equation
2. For this purpose, we use the topology shown in Figure
3, which consists of n stations indexed from 1 to n. Each
station generates the same traffic of three data streams, labeled
with high, medium and low, according to their priorities.
Station n sends packets to station number 1. Station i sends
to station i +1 three flows belonging to the three classes of
service: Audio (high priority), Video (medium priority), and
Background Traffic (denoted by BT for low priority). We use
CBR sources to simulate BT, video, and audio traffics.
In the following simulations, we assume that each wireless
station operates at IEEE 802.11a PHY mode-6 [14], see
network parameters shown in Table I.
TAB LE I
IEEE 802.11
A PHY/MAC PARAMETERS USED IN SIMULATION
SIFS 16µs
DIFS 34µs
ACK size 14 bytes
Data rate 36 Mbits/s
Slot time 9 µs
CCA Time 3µs
MAC Header 28 bytes
Modulation 16-QAM
Preamble Length 20µs
RxTxTurnaround Time 1µs
PLCP header Length 4µs
Table II shows the network parameters selected for the three
TCs.
TAB LE II
MAC
PARAMETERS FOR THE THREE TCS.
Parameters High Medium Low
CW
min
5 15 31
CW
max
200 500 1023
AIFS(µs) 34 43 52
PF 2 4 5
Packet Size(bytes) 160 1280 200
Packet Interval(ms) 20 10 12.5
Sending rate(Kbit/s) 64 1024 128
3
3.5
4
4.5
5
5.5
6
0 0.2 0.4 0.6 0.8 1
Latency(in ms)
Smoothing factor value
Effect of the smoothing factor
Fig. 4. Effect of the smoothing factor on mean delay
Figure 4 and Figure 5 show the average delay and goodput
as a function of the smoothing factor α, respectively. All the
results are averaged over 20 simulations. Twenty-five stations
are used to evaluate the effect of smoothing factor, which
1600
1650
1700
1750
1800
0
0.2 0.4 0.6 0.8 1
Total goodput (KBps)
Smoothing factor value
Effect of the smoothing factor
Fig. 5. Effect of the smoothing factor on goodput
corresponds to a load rate of 84.5%. We can note that choosing
a value of α in the range of [0.75, 0.9] achieves a good
tradeoff between goodput and mean delay. So in the following
simulation scenarios we set α to 0.8.
B. Effect of traffic load
To evaluate the performance of AEDCF, we investigate
in this section the effect of the traffic load and compare it
with the basic EDCF and SD schemes. Our simulations use
different types of traffics to evaluate service differentiation.
Three queues are used in each station. The highest priority
queue in each station generates packets with packet size
equal to 160 bytes and inter-packet interval of 20 ms, which
corresponds to 64 Kbit/s PCM audio flow. The medium traffic
queue generates packets of size equal to 1280 bytes each 10 ms
which corresponds to an overall sending rate of 1024 Kbit/s.
The low priority queue in each station generates packets with
sending rate equal to 128 Kbit/s, using a 200 bytes packet
size. To increase the load of the system, we gradually increase
the number of stations. All the stations are located within an
Independent Basic Service Set such that every station is able
to detect a transmission from any other station, and stations
are not moving in the simulation. The topology is shown in
Figure 3. We start simulations with two wireless stations,
then we increase the load rate by increasing the number of
stations by one every eight seconds. Figures 6 - 9 show the
average obtained over 5 simulations. We increase the number
of stations from 2 to 44 which correspond to load rates from
6.7% to 149%.
To evaluate the performance of the different schemes, the
following metrics are used:
• Gain of goodput: This metric stands for the gain (in
%) on the average goodput of new schemes (SD or
AEDCF), compared with the basic EDCF. It is calculated
as follows:
Gain
−
of
−
goodput =
AG
new
−AG
EDCF
AG
EDCF
∗ 100%
• Mean delay: It is the average delay of all the flows
that have the same priority in the different stations. This
metric is used to evaluate how well the schemes can
accommodate real-time flows. However, real-time flows
also require low average delay and bounded delay jitter.
• Latency distribution: Latency distribution allows to
trace the percentage of packets that have latency less than
the maximum delay required by the applications.
1376
Authorized licensed use limited to: Brunel University. Downloaded on August 27, 2009 at 10:25 from IEEE Xplore. Restrictions apply.
• Medium utilization (M
u
): Due to the scarcity of wireless
bandwidth, we also study the medium utilization of the
different schemes by computing the percentage of time
used for transmission of data frames:
M
u
=
TotalTxTime−CollisionT ime−IdleT ime
TotalTxTime
∗ 100%
• Collision rate: Collisions in wireless LAN cause addi-
tional delays by increasing the delay that stations should
wait, before initiating a new transmission attempt. This
rate is calculated as the average number of collisions that
occur per second.
Figure 6 shows the mean delay of the audio flow corresponding
to the high priority class. The AEDCF scheme is able to keep
the delay low even when the traffic load is very high, i.e., with
a large number of stations. We can see that the mean delay
of audio for AEDCF is 51% smaller than that for the basic
EDCF when the load rate is up to 88% (26 stations). Moreover,
the mean delay of audio of AEDCF is still 38% smaller than
that of the basic EDCF when the load rate reaches 149% (44
stations). Indeed, when the number of stations is more than
13, the delay obtained by the basic EDCF increases faster
than AEDCF and SD scheme, while AEDCF always keeps a
lower mean access delay less than 10 ms. We can also note
that AEDCF has a mean delay 30% less than the static slow
decrease scheme when the load rate reaches 149%.
In Figure 7, we plot the gain on goodput as a function of the
traffic load of AEDCF and SD. We observe that the goodput
gain of AEDCF increases when the traffic load increases. It
reaches about 28% when the load rate is about 119% (i.e.
for 35 stations). Moreover, the goodput of AEDCF is 10%
higher than the SD scheme when the load rate is 149%. Indeed,
AEDCF is much more efficient during high load rates.
0
2
4
6
8
10
12
14
16
18
5 10 15 20 25 30 35 40
Audio mean delay(in ms)
Number of stations
EDCF
SD
AEDCF
Fig. 6. Mean Delay
0
5
10
15
20
25
30
5 10 15 20 25 30 35 40
Gain on goodput (%)
Number of stations
SD
AEDCF
Fig. 7. Gain on goodput
10
20
30
40
50
60
70
5 10 15 20 25 30 35 40
Medium utilization(%)
Number of stations
EDCF
SD
AEDCF
Fig. 8. Medium utilization
0
200
400
600
800
1000
1200
1400
5 10 15 20 25 30 35 40
Collision number per second
Number of stations
EDCF
SD
AEDCF
Fig. 9. Collision rate
Figure 8 shows the medium utilization as a function of the
traffic load. For the three schemes the medium utilization de-
creases when the traffic load increases. However, our scheme
achieves better medium utilization than the basic scheme
whatever the number of stations.
The obtained collision rate is shown in Figure 9. The
collision rates achieved by both schemes are similar when
the traffic load is low, i.e. the number of stations is less
than 8. However, when the traffic load increases, AEDCF is
able to maintain a lower collision rate than EDCF and SD
schemes. We can explain this behavior by the fact that we
use an adaptive technique to change the contention windows
according to the collision rate. The reduction of collision rate
of AEDCF leads to significant goodput improvement and delay
decrease. Moreover, AEDCF achieves an efficient tradeoff
between collision decrease and idle time increase.
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0 20 40 60 80 100
Latency(in sec)
simulation time(in sec)
EDCF
Fig. 10. The audio delay for EDCF
We have done a different simulation to study performance
on delay and jitter. This new experiment has the same topology
than Figure 3, but the number of stations is increased from 2
(4 sec) to 25 (100 sec) and the simulation stops at 115 sec.
The delay variations of both EDCF and AEDCF schemes are
1377
Authorized licensed use limited to: Brunel University. Downloaded on August 27, 2009 at 10:25 from IEEE Xplore. Restrictions apply.
