802.

11e EDCA Protocol Parameterization: A Modeling and Optimization Study

loannis Koukoutsidis* Vasilios A. Siris

National & Kapodistrian University of Athens FORTH-ICS/University of Crete
Dept. Informatics and Telecommunications P.O. Box 1385, 711 10 Heraklion, Greece
Ilissia, 157 84 Athens, Greece [email protected]
I.Koukoutsidis @ di.uoa.gr

Abstract WLAN, using a CSMA/CA medium sharing mechanism

The IEEE 802.]]le MAC protocol specifies an enhanced [4]. In a heterogeneous environment with multiple traffic
distributed channel access (EDCA) mechanism with ad- classes having different service requirements, it is appro-
justable protocol parameters, providing differentiated ac- priate to provide service differentiation through the channel
cess to wireless stations. The EDCA parameters are: CWmin eDC mechanism.
access mechanism [5]With this intent, the pEte
a hnumberof intol IEEE 802.11le
EDCA mechanism [5] a number of protocol parameters can
and CWax (mnimum nd maimum
Wm,,(miimumand ontenton widow),
axium cntenionwindw), be set differently for each class of wireless stations.
AIFS (arbitration interframe space), and TXOP (transmit Fo a given iclass u eDCA paatersa
F
opportunity). In this paper we study thejoint setting ofthese
parameters to maximize capacity and optimize performance * CWmitn,i, CWmax,i : The minimum and maximum
under QoS constraints. Wefirst revisit the analytical model- contention windows. Before each transmission attempt
ing in saturation and non-saturation conditions andprovide and in order to reduce the probability of a collision,
more detailed approximations of the mean channel access a station enters a "backoff stage", which consists of
delay and throughput. Considering two classes of wireless waiting a random number of time slots, wherein the
stations with higher and lower QoS demands, and optimiz- medium is sensed idle. This number is initially se-
ing with respect to average measures and constraints, we lected uniformly in a minimum "contention window"
provide Pareto optimal pairs for the number of supported interval [0, CWmin,i -1]. In the event of a colli-
stations from the two classes for different parameter set con- sion, a new backoff stage is entered and a larger con-
figurations and load conditions. Further; we examine opti- tention window is selected. The protocol suggests
mal parameter selection for a class of elastic traffic, in the doubling the window until a maximum stage mi, for
presence of a delay-sensitive class whose parameters are which 2mi CWmin,i = CWmax,i. The number
fixed. Our findings show that drastic service differentia- of backoff slots is in general chosen in the interval
tion can undermine capacity of a WLAN. They also demon- [0, 2(kAmi) CWmin,i - 1], where k = O...rni is the
strate the optimality ofjointly setting high values for TXOP current backoff stage.
and AIFS in order to maximize the throughput of the elastic
traffic class while guaranteeing delay of the delay-sensitive * AIFSi : The arbitration inter-frame space. It is the
class. We also reveal different optimal settings for CWmin amount of time a station must sense the channel idle
for different load conditions. We summarize ourfindings as before decrementing its backoff counter, or attempting
guidelines for the setting of 802.1]e parameters in a sce- a transmission. Assigning different values of AIFS to
nario with data and real-time service classes. different classes can distribute contention for the chan-
nel and prioritize access. For two stations of classes k
1. Introduction and j, with class k having a lower priority, the value
(AIFSk - AIFSj), yielding the AIFS separation in
The IEEE 802.11 series gives protocol specifications for slots, is of practical significance.
the interconnection of telecommunications equipment in a
___________________ ~~~~*TXOPi : The transmit opportunity. It is used to
*This work was done while I. Koukoutsidis was a research associate permit consecutive frame transmissions by a station.
with FORTH-ICS. The work was supported by the General Secretariat for It is deie as th aiu nevloiedrn
Research and Technology of Greece, through the project 05-AKMON-80
within Action 4.2 of the Operational Programme Competitiveness - 3rd which a station has the right to initiate a sequence
Community Support Programme. of frame transmissions uninterrupted by others, after

1-4244-0992-6/07/$25.OO © 2007 IEEE

 it has gained access to the channel through the con- anteeing delay of the delay-sensitive class. Further, we re-
tention mechanism. It is a basic building block of the veal that CWmin should be set differently for different load
enhanced protocol, and every contention win is consid- conditions, and primarily adopt small values not only when
ered to grant a TXOP. The duration in which the TXOP stations are unsaturated, but also when both classes are sat-
holder maintains uninterrupted use of the medium in- urated. It is also shown that having a drastic service dif-
cludes the last frame's ACK reception time, therefore ferentiation in saturation conditions may seriously under-
it refers to an integral number of frames. mine capacity of a WLAN when both classes have QoS
constraints. In contrast, since constraints are easier satisfied
Tchclase {iscWmled W
each class i is called the EDCA ar,aIrs,Tn
EDCAbparame
XOP
tr
tese and istas
in non-saturation,
deterioration it istoeasier
in order to create a tolerable
improve service
capacity. To our knowl-
The quainathe
mTte veaond
frameuantitavesspoiffernti(aP).
quantitative differentiation acedge, these are the first general conclusions regarding ca-
qualitative and accom- paiynderom ceptizininhersneofbh
plished through EDCA parameters is fairly known from ex- day-seny sipetv and eaticztior allE paramete
e

isting modeling and simulation works [8-10]. The ultimate and different loads.
goal, however, is to find sets of parameters that satisfy given This paper is structured as follows. In Section 2 we de-
QoS constraints for each class and optimize its performance scribe basic elements of the mathematical modeling of an
- or equivalently its capacity, defined as the maximum num- 802.11 e WLAN. Performance measures of interest, namely
ber of supported stations - under these constraints, the average channel access delay and throughput, are calcu-
Previous research in this direction tries to accomplish lated in Section 3 for saturation conditions. The approach
this by evaluating performance or capacity when modify- taken for modeling non-saturation conditions is presented in
ing a subset of the parameters in a given scenario. We dis- Section 4. In both Sections 3 and 4 we introduce several re-
cuss some representative works. In [3] the voice capacity finements compared to previous works. The considered op-
of a WLAN is examined, with and without the integration 5.
of data traffic, for different settings of the CW , CWax,Secon
Finally, guidelines on the setting of 802.1 le parameters for
voiceaation atosatestting imprvma
and AIFS separation; the authors attest the improvement in
performance of voice stations by setting small CWmin and
performance and an elastic and real-time class are given in the concluding
Section 6.
CWmax values, as expected for unsaturated sources. In [10]
the authors argue that, in view of an unsaturated real-time
class and a saturated data traffic class, the AIFS separation 2. Basic mathematical model
should be tuned to satisfy delay guarantees of the real-time
class, and the CWmin of the data class should be adjusted Valuable references for the mathematical modeling of
to maximize its throughput. Finally, in [2] the authors ex- the basic 802.11 protocol are the paper of Bianchi [1] as
amine a traffic scenario with TCP data stations, prioritizing well as the work of Kumar et al. [7]. Appropriate modifica-
the bottlenecked AP by setting for it a small CWmin and in- tions for 802.1 le were given in [8,9] for saturation condi-
creasing the AIFS of data stations. tions, and presented in the next subsections. Non-saturation
As a general remark, results so far have been limited to conditions will be treated separately in Section 4.
a subset of parameters in specific test cases, and no con- We consider more generally a number of Ni contending
sistent optimization study has been done. In this paper we stations belonging to service class i (i = 1, . . ., K). The
attempt such a study, covering all EDCA parameters and in stations will be assumed to send traffic to an AP (and not to
representative ranges of loads from low traffic to saturation each other), whose sole purpose is to send back ACKs. Fur-
conditions. We look at a scenario with two service classes, thermore, we do not consider different traffic classes within
one of which has tighter QoS constraints than the other. the same station. This elementary multiservice model is
Loosely, these refer to delay-sensitive (e.g., voice, stream- appropriate for our purposes. Nevertheless, extensions to
ing video) and elastic (data) traffic. Constraints are in terms cover the inclusion of an access point and multiple service
of the mean channel access delay and throughput, which are classes within one station are straightforward.
also the measures to be optimized. We provide Pareto opti-
mal pairs for the number of supported stations from the two 2.1. Solving for attempt and collision prob-
classes when modifying all parameters. Further, we exam- abilities
ine optimal parameter selection for the class of elastic traffic
when parameters of the delay-sensitive class are fixed. The key element of the mathematical model is the so-
From an applications' viewpoint, our main contributions called "decoupling approximation" [7]. It purports that the
are the following. We demonstrate the optimality of jointly probability of an attempt by a station at each idle-sensed
setting high values for TXOP and AIFS in order to maxi- slot, or equivalently the probability of a collision at an at-
mize the throughput of the elastic traffic class while guar- tempt is the same throughout the time evolution of the sys-
tem. As a consequence, the (re)-transmission processes of i.e. the sum of collision probabilities for class i in zones
the different stations are mutually independent (decoupled). j > i, weighted by the probability of being in each zone
This approximation becomes more accurate as the num- when there is an attempt. Along with (1), we have again a
ber of stations increases. Practically, it works well even for system of 2K nonlinear equations.
a number of stations as low as 2 (see [1]).
Since the collision probability is dependent on the at- 3. Performance measures
tempt rate and vice-versa, appropriate expressions can con- The major performance measures we are preoccupied
struct a system of equations to solve for these values. with are the channel access delay and throughput, for each
Denote the collision and attempt probabilities of a station class station. The analysis follows previous works, mainly
in class i (i = 1, . . ., K) by ci, pi respectively. From a [3]. It also includes the use of TXOP, which has not
stochastic analysis (either a Markovian analysis as in [1], or
been covered in many previous works on 802.1 le (includ-
a renewal theory analysis as in [7]), the attempt rate can be
expressed~~~~~~~.
as a fucto oftecliinpobblt n h ng [2,3,8,10]). We only consider basic channel access (i.e.,
without the use of RTS/CTS [4]), without loss of generality
backoff parameters CWmin,i, mi. Assuming for simplicity for our results.
that there does not exist a limit on the number of retries to The calculation of performance measures is confined to
send a packet (see also Remark 2. 1), we have
the case of 2 service classes A and B, of which class A will
2(1 - 2ci) be favored by service differentiation. The following refine-
(CWmin,i- 1)(1 - 2ci) + CWmin,iCi(l - (2Ci)mi) ments in the analytical model are introduced. We provide
(1) a correction in the calculation of the mean channel access
Then, in the frame of the decoupling approximation, delay (and subsequently, throughput) in the case of AIFS
considering a population Ni for each class i, we write differentiation, to include the whole delay a disadvantaged
class-B station faces until it is allowed to attempt or perform
Ci = 1- (1-pi)Ni- 1 17(1-pj)Nj (2) backoff. Additionally, we calculate the average throughput
ji#i seen by a station rather than the one seen by the system,
For K classes, we end up with a system of 2K nonlinear quantities which as we show may differ substantially, espe-
equations, through which pi, ci can be derived numerically. cially in non-saturation conditions.
Remark 2.1 In the single-class case, the parameter config- 3.1. Channel access delay
uration which minimizes the average time between success-
ful transmissions also yields the highest expected number The channel access delay is defined as the delay a frame
of successes in any bounded interval (0, t], and hence is op- experiences from the time it arrives at the head of the trans-
timal in any limited or unlimited-retry case. For multiple mission queue until it is transmitted successfully, and its
classes a minor influence can be expected, becoming negli- transmission is acknowledged. Thus it is (deliberately) set
gible as the retry-limit increases. to be a little more than just the "access" time.
We consider a station of class i (i = A, B) has a MAC
2.2. Contention zones packet size oui and transmits at rate Ri. Let also d be the
propagation delay, TRXTX the time for the transceiver to turn
Enhancing the idle sensing time of some stations by con- around, TPLCP the time to transmit the PLCP preamble and
trolling the AIFS parameter is a basic tool for setting prior- header (adjoined by the physical layer, see [4]) and ack the
ities. Consider classes indexed according to "service privi- size of a MAC level acknowledgement. The TPLCP is fixed
lege" order, i.e., AIFS1 < AIFS2 < < AIFSK. This for each physical layer configuration. Further, ACK packets
creates similarly indexed "zones" of channel activity, where are always transmitted at a (lower) basic service rate Rb. In
in zone i only classes j < i are allowed to contend. addition, a SIFS (short inter-frame space) interval is used
Denote by wi the stationary probability that the system between the transmission of a frame and the sending of an
is in zone i. This can be easily derived by Markov chain acknowledgement, to allow the MAC layer to receive the
analysis (see e.g. [8] for the case of 2 classes). Then the packet and subsequently the transceiver to turn around.
collision probability of a station of class i is According to the protocol, the duration the medium is
busy because of a successful transmission of class i -
Ci K
,K- K
(1 (1 P1) ... (1 _p)Ni l) + including the reception of the acknowledgement - is
i
_

± ~-K (l (l pl)1..*(l pi)Ni 1.*(l pK)NK), TSUCC =AIFS -TRxT x PLCP Ri 4

Ej=i 7rj ~~(3) +d+TPLCP +SIS+ Rb
 The time the medium is busy during a collision of class-i if it were in zone A or zone B, respectively. Also, the ex-
stations is pected duration of a collision is calculated by considering
T,coll = A-IFS,
TCOII -= AIFSi
TRTX+
TRxTx + d ++ TPLCP +
TPLCP R+
the max(TA°t, TBll) if it is among a class-A and class-B
station, or TW°11 or TW11, weighted by their respective prob-
abilities.
We first derive the mean channel access delay for a sta- Starting from (6), we can approximately calculate the
tion of the priority class A. Following the analysis in [3], variance of the channel access delay. We finally get for the
for a station of class i = A the channel access delay can be second moment that
expressed as:

A, Uk + - + (6)
E[(D acc)2] _E[S?] Qi + Qi(Qi - 1)E2[Si]
+ ci(1 + cj) E[(7coll)2] (Tucc)2
+

ZZS>j= 1T(1c)2(10)
+2E[Tic°t"]E[Si]6i + 2E[Si]QiTrUCC
where Ai is the number of channel attempts for the given 2c-
frame, and Uk is a random variable uniformly distributed + E[T9oll]Tsucc
in [0, 2 (k)AmiCWmin,i - 1]. Under the decoupling ap-
proximation, Ski are i.i.d random variables representing where Qi, 0i are defined in (12), (13), respectively. Rela-
the duration of each backoff decrement cycle as seen by the tion (10) is an approximation because we have considered in
station, where in such a cycle a successful transmission or the calculations that E[Q2] = E2 [Qi]. Finally, the variance
collision may follow the idle period. This is referred to as is computed as Var[Dacc] = E[(Dacc)2] - E2[Dacc].
the duration of a generic slot. We now extend the analysis to derive the mean channel
The number of transmission attempts follows a geomet- access delay of class-B stations, when there is AIFS differ-
ric distribution with Pr{Ai = k} = (1 - ci)ch1. Packet entiation. In this case, after each busy medium end a station
sizes and transmission rates are deterministic, however the of class B will go through a number of slots where it is not
duration of a collision depends on the type of stations impli- entitled to transmit.
cated. Considering the mean number of backoffs and colli- To include this period, we consider the AIFS separation
sions, we have: f = AIFSB - AIFSA and the discrete-time Markov chain
shown in Fig. 1, with probability qA = (1 - PA)NA and an
ED aCcc =ESE E 1Uk>A U1 absorbing state T. The state of the Markov chain represents
[ i ][ i [L,.~ l {Uk .i}I the number of idle slots that have elapsed since the last busy
k=1 j=1 (7) medium end, as indicated by the carrier sense mechanism of
+ 1-c E[T i] + Tis a station.

qA qA qA qA qA
Using that ' ett,.- V° ( 0 ;***; 1
2k-1CW"',"',_j ~~0 2 -1
1~~~rfUk>1
uk
Uk 1-q

E[ l{U1k>ji}]
j=1 j=1
Pr{U: j}
q

2k lCWmin,i - 1
=2 ' Figure 1. Discrete-time transitions in zone A.
after some calculations we obtain:
We are interested in the mean number of visits to states
E[Da00] -E[S-1 [CWmin,i 1 - (2ci)i + (2c )rmi 0 ,1 ... ,£ - 1 prior to absorption in state e. Denote by
-
i L[ 2 1- 2ci 1-cJ] [pa)] (0 K i,j .< ) the nth power of the transi-
1 c tion matrix and its elements. Then, starting from state 0 the
2(1 - ci)] + 1 r, E[Tc°ll] + Tr8ucc (8) mean number of visits to O is1
00
The expected duration of a backoff decrement cycle is mO = (n)
E[Si] = frA * E[S? ] +wB.E[Si ]: (9) POO
1For computations, m0 and ml .e- 1 are found merely by invert-
where E[S'J, E[SfB] are the expected generic slot durations ing (I -Q), where Q is the restriction of P to the set of transient states
for the class i =A station, when it is performing backoff, {0, 1,..., - 1} and I is an identity matrix with the same dimensions.
 A AiUk (Wini1 (2ci)m' (2ci)m' 1
E[E E 1{Uk>j}] = + (1)
- (12)

AE[5(A,
A i
1{U>j u] (CWmin,i 1)Ci CWmin) [Ci -[Ci + (Tni -21) ( -Ci)] (2Ci) +
+2(1 ci) 1(m -1)(1
-

0i 1 2(1 - ci) - [c

2ci(1 - ci)(1 - (2ci)mi- 1) 1 2milCW (C7 i - i± )(mi - mici + Ci) + ci+1(1 ci) ci(1 + ci)
(1 - 2Cj)2 CJ ± m, (1 -C)3 2(1-C)2
(13)

while the mean number of visits to states 1,... . -1 is randomly chosen frame of class i is found with probabil-
ity i/rhj to be at the head of the transmission queue prior
rn _j =
,1 +.0.(n)+
p (n) + 1-<, PolPo(n)
. ,- P2t- 1 (14) to a TXOPframes
following grant, that
and will - 1)/r1i
(rfi be to be one
transmitted of the rj It- is1
successively.
clear that the frame at the head of the queue will undergo
We know that a visit to one of states 1,... . - 1 lasts the standard transmit procedure, and thus suffer a mean
one idle slot time, while that to state 0 lasts either the time access delay obtainable from the previous analysis, which
of a successful transmission or of a collision. Therefore we call E[D"'C71j = 1]; on the other hand, the remaining
the mean channel access delay for a station of class B will frames benefit by having a much smaller delay, equal to
become dl = SIFS + 6 + TPLCP + o-i/Ri. The expected delay of
a class-i frame is then
E[D [E[SB] + T510t * -i + TbJsy ino ] rac 1 1
AB Uk acc E[Dacc 1] + (17)
A
E[J, X, 1tUk >j}] + E[TWl ] + TB Crl
E[D -EBE[[Dl1Ticc~7 /
d (17
k=1 j=1 1 For the calculation of generic slot times, note that the
(15) duration of a successful transmission which includes the
ACK reception is now increased to AIFS - TRXTX +
In the above expression, Tslot is an idle slot time, while +C rcp+ iP +ow P SiFesed + )A ±(Si-1TR,TFs
Usy!o is the mean duration of a busy slot in zone A. The
z
R1 Rb
latter writes as On the other hand, a collision is supposed to occur on the
first transmitted frame, and therefore the expression in (5)
TzOne slot
A NAPA(1 - PA) A TSUCC is unchanged.
IUsY 1 (1 - PA)NA A
-
Finally, we can also employ (10) to calculate the delay
1- NAPA(1 - PA)NA ( T)
(11-PA) variance in the case where TXOP is used.
+
+
NAl -~~PA)
1- (1-PA)NA
1-P)A

TA
c

3.2. Throughput
In the calculation of E[SB] above one must set AIFSB
AIFSA, since the AIFS separation is now accounted for The evaluated throughput is the rate of successfully
when calculating the mean time to absorption. transmitted MAC-level information per unit of time. De-
Finally, it is noted that in this case it is extremely difficult note it by 'Yi for a station of class i. The usual way to derive
to compute the delay variance, since second moments of the this (e.g., in [1, 3, 9]) is to consider each end of a generic
number of visits to a state prior to absorption are unknown. slot as a renewal epoch and calculate the mean amount of
successfully transmitted information over the mean generic
3.1.1. Use of TXOP slot duration,
As mentioned before, wireless stations can carry out mul- Yi =(WAUTAPA(1 - PA)NA + 7FBUBPB(l - PB)B-
tiple frame transmissions, taking advantage of advertised
TXOPs. Let a station of class i transmit ryi frames succes-
(1 - PA)'A)j(7AE[SA] + 7FBE[SB])
(18)
sively (as derived from the TXOPi limit), where Thi C N.
Transmissions are separated by a SIFS period, to allow However, this can only be characterized as the individ-
transceivers to turn around from ACK receptions [5]. A ual throughput seen by the system. To calculate the actual
throughput of a station we have to take the total channel station after a generic slot in the OFF period. We consider
access time as the renewal period, hence these constant and approximate them for Poisson arrivals as
-i (19) rN =eAiE[DVj (21a)
E[Dac] i (19 OFF
ri -AiE[Si] ~~~~~~~~~~~~~~~~~~~
This further allows to include the corrected calculation of Eq. (21a) represents a low load approximation where con-
class-B access delay in case of AIFS differentiation. secutive frame transmissions are less frequent; this is be-
cause we confine the probability of no arrival in [0, E[D acc])
4. Non-Saturation conditions (for consecutive transmissions it should be greater).
It is clear that this model describes a regenerative pro-
cess, as the end of an OFF period is a regeneration epoch.
In constructing a model for non-saturation conditions, The throughput for class i is calculated as the mean MAC-
we would like to include the traffic arrival characteristics in level information transmitted in a regeneration cycle, over
its parameters, and simply extend the set of nonlinear equa- the duration of this cycle:
tions. The key modeling assumption here is to consider a ON
constant busy station probability at each idle-sensed slot,
equal to the load of the station envisaged as a single server
_ _ _ __N_i
E[Dacc,ilj$N + E[Si]j(l- rFF)
(22)
queue, as shown in [3].
Consider an arrival rate Ai for each station in class i, and 400
the queue load pi. The collision probability now is ON/OFF model
350 -'Seen by system'-------
Simulation
~ ~pi)N~i-
1=- (1 -
Ci ~ ji7ilk-(20)
N_ 1 -jpj) N0
i 300
250
A.

where the probability of an attempt by class i is conditioned Q 200

on having a packet to transmit and hence given by the same 2 150
function of ci as in the saturation case (1). Along with pi F 100
Ai E[D acc], for K service classes we now have a system of 50
3K nonlinear equations to solve for pi, ci, Qi.
The mean channel access delay is then calculated follow- 10 100 1000
ing the same approach, which yields a sufficiently good ap- A (packets/sec)
proximation [6]. To calculate the average throughput expe-
rienced by a station, we would have to consider a sequence Figure 3. Comparison of different throughput
of alternating ON/OFF periods, where the OFF period is ge- approximations in a 1-class case.
ometrically distributed with parameter pi (and may take the
value 0). Treating this as a regenerative process, we would
have 7Yi = oui/(E[D acc] + (1 - pi)jpi E[Si]). This turns Fig. 3 shows an example of this throughput approxima-
out to be extremely inaccurate, yielding throughput results tion in a 1-class case with CWmin 32, m 5 and
ofabout1obe oxtrdereo mnagnitudegreater.
yieldingthroughputN = 10 stations, in 802.1 la at 6 Mbps, for different Poisson = =

th, w onsitdegreatore involvedON/OF

Toaboutacklrder arrival rates until saturation. Results are compared against a
discrete-event simulator written in C++. It can also be seen
model where a station may send 1 or more frames succes-
that the corresponding individual throughput "seen by the
sively in the ON period (Fig. 2). system" is largely inaccurate for non-saturation conditions
1st collision 2nd collision success 1st collision success (even if the 'kbps' throughput scale is a coarse scale when
OFF |= ON =| OFF protocol timings are in ,us). Also shown on the graph is
that the approximation by A uo is extremely accurate for a
1 st fram 2nd frame very large lange of loads, which reflects the simple fact that
regeneration cycle in non-saturation conditions where collisions are fewer, the
throughput is almost equal to the information arrival rate.
Figure 2. Sample path evolution of the sys-
tem with ON and OFF periods. 5. Optimization

Define TYN: the probability of an empty station after a We examine two optimization problems in 802.11le. We
frame transmission and T9FF: the probability of an empty first assess the capacity of the network for different param-
eter settings, and secondly we attempt to jointly optimize CWmin,B = 16, mA = mB = 5, rlA = rlB = 1, and also
parameters for the two classes A and B, transmitting delay- AIFSB - AIFSA = 0. Subsequently we modify each pa-
sensitive and elastic traffic, respectively. In the first prob- rameter separately - favoring class A - and derive Pareto
lem, we consider delay and throughput constraints for both pairs. We say a capacity improvement exists in the changed
classes; in the second, a delay bound for class A, while for parameter configuration if the Pareto pairs lie "above" those
class B a better-than-best-effort behavior, in terms of the of the balanced case (i.e., in a vector inequality sense). Each
best achievable throughput. modified parameter is shown in the title of each subgraph,
The results are based on the analytic evaluation of mean and Pareto optimal pairs are depicted by 'o' in the unbal-
performance measures in the previous sections. Results in anced cases, and by '+' in the balanced ones.
all cases are for 802.1 la MAC and PHY layer character- The general goal is to increase capacity of the system
istics. Values of related parameters used in our analyti- by service differentiation in favor of class A, since class-B
cal formulas are summarized in Table 1. The transceiver stations can tolerate lower quality. An overall capacity im-
turnaround time and propagation delay are negligible and provement can be seen in saturation conditions (Fig. 4(a))
are omitted. in cases where QoS deterioration for class B is more toler-
able (cases where mB = 10 and AIFSB - AIFSA=2).
Table 1. 802.11 a MAC and PHY parameters. However, a noteworthy observation is that a more drastic
service differentiation can have an adverse effect: when a
Parameter Value number of inferior-class stations transmits in the system and
Tsiot 9 ,us requires a certain - even inferior - QoS, capacity of the fa-
SIFS 16 ,us vored class should largely decrease to accommodate these
mini AIFSi 34 ,us stations in the network. For instance in the case where
TPLCP 20 /S CWmin,B = 32, when no class-B stations exist in the
ack 14 bytes system, the maximum number of allowed class-A stations
is 10. When at least 1 class-B station should be able to
We focus on EDCA parameter differentiation and do not transmit, the capacity of class A drops to 6. These phe-
introduce other biases. In this sense we set physical trans- nomena are more likely to occur in a saturated network:
mission rates equal, RA = RB = 6 Mbps. The basic ser- in non-saturation conditions constraints are easier satisfied
vice rate is also chosen as Rb = 6 Mbps. Moreover, the size and service deterioration usually effects an overall increase
of MAC packets is set equal to 160 bytes for all stations.2 in capacity, even with the same constraints for both classes
(Fig. 4(b)).
5.1. Pareto optimal pairs It is worth stressing that the influence of backoff pro-
tocol parameters on performance reflects also their influ-
We seek the maximum number of stations from each ence on capacity. Further results attest that in the protocol,
class that can be admitted in the system subject to CWmin, AIFS and TXOP are more influential parameters
QoS constraints. Since we have contending stations than CWmax, a parameter which may have no effect at all
with contradicting performance objectives, we shall de- in non-saturation conditions (see (Fig. 4(b))). The smaller
rive Pareto optimal pairs3. Different parameter settings influence of CWmax was also witnessed in [10].
{CWmin,A, CWmin,B, mA, mB, X, rlA, rlB } are examined.
Optimal pairs are shown in Fig. 4 for some cases of satura- 5.2. Optimal parameter selection
tion and non-saturation conditions. In the saturation case,
constraints are set as follows: for delay, E[DAcc] < 5 ms, Parameter design involves searching for performance-
E[Da~cc]
EB. 2 < 0k10 ins, for throughput,
s. Inandtheon urati cas7YA .w 300consider
kbps, optimizing parameters in a space of allowed values. Here
TYB > 200 kbps. In the non-saturation case we consider we confine ourselves to the finding of optimal parameters
more tight constraints, set to be the same for both classes for the class of elastic traffic, when the delay-sensitive class
i = A, B: E[Di0] < 1 is, and -Yj. 100 kbps. has its parameters fixed. This simplifies the automated
The arrangement for this set of results is as follows. search and allows for clearer conclusions. Also, in a prac-
We take a "balanced" configuration where EDCA param- tical WLAN scenario stations transmitting delay-sensitive
eters are the same for both classes. These are CWmin,A = traffic are usually fewer in number and unsaturated, and
2Packet sizes are set to be the same for both classes for simplicity, as hence little is to be gained further by optimizing their pa-
this does not change the essence of the results. The data size can be seen rameters.
as an average, bearing in mind that the majority of packets sent over theAnuprdlybndisefocas-stin.Frth
Internet (ranging from 44 to 1500 bytes) are also of small size.
3Avector (Ni, NB) is said to be Pareto optimal iff any other vector data transmitting stations no constraints are imposed, yet
(NA, NB) in the feasible set has NA < NA or NB . NB. we aim at maximizing their throughput. This formulation
 CW,in,B=32 mB=10 AIFSB AIFSA=2 A=3
10b 10 10 - -e- 12
8 + 8 +X 8 + 10
6
6 o
+
()1 '.
6- - G
D
6-
l1 \I6
N
NA e NA 0 NA NAe
4 + e 4- 4 A e CD
0+ 2 0 0+ 1
2 +C)~~~ 2- + 2-+C 0 __ __
0 0 0 11~~~~- 1
00

0 2 4 6 8 10 12 0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 0 2 4 6 8 10 12
NB NB NB NB
(a) Saturation conditions

CW,in,B=32 mB=10 AIFSB AIFSA=2 A=3

18 18 18 ) 25
16 f 16 e 16 e
14
C
: 14 - l 14 I 20
12 12 12 . + C

015 0; ;;+-a 9
NA10 A1 N15
o L; ; o 0
A8
0 NA 10 A08- 0( 0
+ D8 - + NA
6 ~~~~ED ~~~~6 6 +- 10

0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
NB NB NB NB
= 102 , AB =2
(b) Non-saturation conditions, 2A 102 packets/sec

Figure 4. Pareto optimal pairs for various parameter sets.

is consistent with intrinsic QoS demands of delay-sensitive Hence the exploitation of transmit opportunities is cru-
and elastic traffic. Depending on the number of class-A and cial for maximizing the throughput of data traffic.
class-B stations, we shall find the optimal protocol parame- .
Performance degradation of delay-sensitive traffic is
ters to set for class B.
at the same time counteracted by increasing the AIFS
We must consider a limited space in our search. In the separation. We see that this is adjusted to higher values
example we solve, this is the Cartesian product of the value for stricter delay constraints (Table 2(a)) or when class-
sets that follow, for each parameter:
B stations increase in number, in order to prevent the
CWmin,B: {4, 8, 16, 32, 64, 128, 256} access delay of class A from degenerating inappropri-
mB : 2, 3, 4, 5, 6, 7, 8, 9, 10 ately. Hence in an optimal configuration, depending on
{2, 3 5, 6, 7, 8, 9, 1O}
4, , load conditions and the tightness of constraints of the
{OU, 1, 2, J, w,9 ~delay sensitive class, setting high TXOP value should
r/B: {1, 2, 3, 4, 5, 6, 7, 8, 9, 10} be supplemented by appropriately increasing AIFS.
Class-A parameters are fixed at values CWmin,A = 16, The setting of CWmax, on the other hand, shows no clear
mA = 5, and nqA = 1. Results are presented in Table 2, for trend, which can be expected since it has a smaller influ-
a case where both classes are saturated, as well as for a case ence on performance. Finally, a useful observation is that
where only class-B stations are saturated, a scenario likely the aggregate throughput of elastic traffic increases (with a
to be encountered in practice. The behavior, with regard to decreasing rate) as the number of stations increases.
the most influential parameters, is as follows. Additional results when also class-A parameters are opti-
mally selected in the same range showed the same trends for
*Forthecituasionresemblestheoe
where only
class-Be
sa tede class-B parameters, and only small improvements in class-
thsexisituati
class relsemble theu utilizatin
exists (class-Aque
one wherezationas abotabos 30%
onl
B throughput. Finally, in cases where both classes are un-
saturated,
it is intuitive that the of TXOP is setting higher
inse,case,
there is an itaerm2(b)).
theresults of
in
intermediate value of CWmin,B which
is optimal. However, when both contending classes
1-cl
Henc, as
C
the
less important (since a station has fewer packets to send)
and contention windows can be reduced, since there is less
congestion in the network.
are saturated, CWmin,B should be adjusted to much
smaller values, in order to avoid successive channel
occupations by the contending class.
* The number of successive transmissions riB assumes Overall the results in this paper have shown that while in
the maximum value in all cases shown in the table. the single-class case, given a certain TXOP, an optimal se-
 Table 2. Optimal parameter sets
(a) Saturation conditions
Numbers of constraint x Optimal parameters for class B E[DAC'] max mB max NB7B
stations (E[D§C0]
A x is) CWmin,B mB £ riB (Ms) (kbps) (Mbps)
NA 5,NB 1 5 8 3 2 10 4.760 2774.62 2.775
NA 5, NB 10 5 4 10 3 10 4.933 333.17 3.332
NA 5, NB 20 5 16 5 3 10 4.964 173.43 3.469
NA 5,NB 30 5 8 10 3 10 4.981 117.38 3.521
NA 5,NB 1 3 8 3 4 10 2.915 2074.26 2.074
NA 5,NB 10 3 8 8 5 10 2.867 303.11 3.031
NA 5,NB 20 3 8 8 5 10 2.946 163.25 3.265
NA 5,NB 30 3 8 9 5 10 2.958 112.06 3.362
(b) Class B: saturated, Class A: unsaturated, AA - 102 packets/sec
NA 5,NB 1 3 16 2 3 10 2.629 3451.29 3.451
NA 5, NB 5 3 32 10 5 10 2.493 724.57 3.623
NA 5,NB 10 3 32 10 5 10 2.794 371.61 3.716
NA 5, NB 20 3 128 2 4 10 2.871 188.05 3.777

lection of CWmin practically suffices to achieve best capac- References

ity and performance, in a multiple-class case with different
class objectives additional settings are in order. [1] G. Bianchi. Performance analysis of the IEEE 802.11 dis-
In the most important scenario with an elastic and real- tributed coordination function. IEEE J. Select. Areas Com-
time class, the following general guidelines can be de- mun., 18(3):535-547, Mar. 2000.
duced: Throughput maximization of the elastic class can [2] P. Clifford, K. Duffy, J. Foy, D. Leith, and D. Malone. Mod-
be achieved by exploiting TXOPs, while delay constraints eling 802.1 le for data traffic parameter design. In Proc.
class can be met by
..
'
b increasing
~~~~~~IEEE
WiOpt
ip 2006, otn MA,
06 Boston, A USA,S,ArApr. 2006.
06

real-timewclas clanbes het

Of the real-time
separationof
the
i s
elastiF
the
the AIFS
t[3]
separation of the two classes. The CWmin of the elastic traf-
i
N. Hegde, A. Proutiere, and J. Roberts. Evaluating the voice
capacity of 802.11 WLAN under distributed control. In
fic class can be set to its respective 1-class optimal value Proc. IEEE LANMAN 2005, Chania, Greece, Sept. 2005.
if it faces light real-time traffic, while for increased real- [4] IEEE Computer Society. ANSI/IEEE Std 802.11, 1999.
time traffic it should be adjusted to smaller values. Specific [5] IEEE Computer Society. IEEE Std 802.11e-2005, 2005.
values can be investigated for each problem configuration, [6] I. Koukoutsidis and V. Siris. Modeling approximations for
based on the analytic modeling presented. Finally, CWmax an IEEE 802.11 WLAN under Poisson MAC-level arrivals.
is the least influential parameter and may be fixed at a con- In Proc. IFIP Networking'07, Atlanta, GA, USA, May 2007.
stant value. [7] A. Kumar, E. Altman, D. Miorandi, and M. Goyal. New in-
First,
The following critical issues should be considered: First,
following critical issues should be considered:
The
sights from a fixed-point analysis of single cell IEEE 802.11
WLANs. In Proc. IEEE Infocom '05, pages 1550-156 1, Mi-
that increasing the number of frames sent successively via ami, FL, USA, Mar. 2005.
TXOP increases delay variance, since the initial access de- [8] V. Ramaiyan, A. Kumar, and E. Altman. Fixed point analy-
lay of a frame is comparatively very large. A large trans- sis of single cell IEEE 802.1 le WLANs: Uniqueness, multi-
mission variance can result in losses due to retransmission stability, and throughput differentiation. In Proc. ACM Sig-
timeouts, even in the presence of a higher-layer protocol metrics '05, pages 109-120, Banff, Canada, June 2005.
with dynamic flow control (e.g., the self-clocking mecha- [9] J. Robinson and T. Randhawa. Saturation throughput analy-
sis of IEEE 802.1
IEEE J. le enhanced distributed coordination func-
nism of TCP). The use of TXOPsls more appropriate for tion. Select. Areas Commun., 22(5):917-928, June
the transmission of larger upper-layer segments, that would 2004.
anyway be fragmented by the 802.11 MAC and can now [10] I. Tinnirello, G. Bianchi, and L. Scalia. Performance eval-
be transmitted successively. Secondly, as it was shown in uation of differentiated access mechanisms effectiveness in
Section 5.1, setting drastic service differentiations should 802.11 networks. In Proc. IEEE Globecom '04, volume 5,
be avoided, since it may seriously undermine the overall pages 3007-3011, Dallas, TX, USA, Nov. 2004.
capacity of the network.