ieee transactions on mobile computing, vol. 7,...
TRANSCRIPT
Airtime Fairness for IEEE 802.11Multirate Networks
Tarun Joshi, Student Member, IEEE, Anindo Mukherjee,
Younghwan Yoo, Member, IEEE, and Dharma P. Agrawal, Fellow, IEEE
Abstract—Under a multirate network scenario, the IEEE 802.11 DCF MAC fails to provide airtime fairness for all competing stations
since the protocol is designed for ensuring max-min throughput fairness. As such, the maximum achievable throughput by any station
gets bounded by the slowest transmitting peer. In this paper, we present an analytical model to study the delay and throughput
characteristics of such networks so that the rate anomaly problem of IEEE DCF multirate networks could be mitigated. We call our
proposal Time Fair CSMA (TFCSMA) which utilizes an interesting baseline property for estimating a target throughput for each
competing station so that its minimum contention window could be adjusted in a distributed manner. As opposed to the previous work
in this area, TFCSMA is ideally suited for practical scenarios where stations frequently adapt their data rates to changing channel
conditions. In addition, TFCSMA also accounts for packet errors due to the time varying properties of the wireless channel. We
thoroughly compare the performance of our proposed protocol with IEEE 802.11 and other existing protocols under different network
scenarios and traffic conditions. Our comprehensive simulations validate the efficacy of our method toward providing high throughput
and time fair channel allocation.
Index Terms—Analytical model, contention window, fairness, max-min, multirate, packet errors, simulation, throughput.
Ç
1 INTRODUCTION
OVER the last six years, the IEEE 802.11 [1] WorkingGroup has come up with physical-layer enhancements
to support data rates of the order of 54 megabits per second(Mbps) and is currently exploring transmission rates of over120 Mbps. Further, the standard also supports Dynamic RateAdaptation (DRA), which allows stations to dynamicallyadapt the transmission data rate to their channel conditions.Propriety algorithms are frequently used by vendors toadaptively adjust the data rate with an intention tomaximize throughput performance.
While, intuitively, there appears to be a direct relationshipbetween the data rate and the throughput, the recent seminalwork by Heusse et al. [2] suggests otherwise. Surprisingly,under multirate networks,1 the long term throughput ofeach station becomes largely independent of its own datarate; rather, it gets bounded by the lowest data rate peer. Anobvious implication of this phenomenon restricts theachievable benefits of any rate adaptation mechanism. In
particular, the presence of a low data rate link down-equalizes the aggregate throughput, thereby restricting anybenefits of higher bit rates used by the peers. Thisphenomenon is better explained by the example topologyin Fig. 1. Station 1 has a two-hop route to station 3, withstation 2 being the intermediate forwarder. Both the linksare assumed to be perfect in the sense that they can transmitat the highest data rate of 11 Mbps. Station 2 shares thechannel with another link R (station 4 to station 5). Fig. 2depicts the variation of the end-to-end throughput and thechannel occupancy ratio (COR) for station 2 with thetransmission rate of link R. The throughput decreases by63 percent as the rate for link R varies from 11 to 1 Mbps.We notice that both the links have approximately the samethroughput, even though station 2 always transmits at11 Mbps. This can be explained by the ratios of COR for thetwo competing stations, which is roughly proportional totheir respective data rates. This clearly suggests that slowerlinks occupy more channel airtime to ultimately transmitthe same number of packets as faster links.
This phenomenon has been quoted as the “rate anomaly”problem of IEEE Distributed Coordination Function (DCF)networks. It is caused by the fairness philosophy of 802.11,which ensures long-term equal-channel access probability. Thisimplies that, if similar-sized packets are used under similarchannel conditions, then each station achieves roughly thesame throughput, irrespective of its own transmission rate.Now, if all the competing stations employ similar data rates,802.11 automatically guarantees equal time shares as well.However multirate networks penalize faster stations, sincethey invariably need to wait for their slower peers tocomplete their transmissions or retransmissions.
Although several researchers have suggested solutionsto this problem [2], [4], [5], all solutions rely upon a centralbase station or an access point, which will perform thecoordination. Decentralized solutions to the problem have
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008 513
. T. Joshi is with Microsoft, 15325 Redmond Way #E124, Redmond, WA98052. E-mail: [email protected].
. A. Mukherjee is with Google Inc., 3475 Granada Ave #334, Santa Clara,CA 95051. E-mail: [email protected].
. Y. Yoo is with the School of Computer Science and Engineering, PusanNational University, 30 Jangjeon-dong, Geumjeong-gu, Busan, 609-735,Republic of Korea. E-mail: [email protected].
. D.P. Agrawal is with the Department of Computer Science, University ofCincinnati, 833 Rhodes Hall, PO Box 210030, Cincinnati, OH 45221-0030. E-mail: [email protected].
Manuscript received 4 Dec. 2006; revised 21 May 2007; accepted 26 June2007; published online 27 July 2007.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-0329-1206.Digital Object Identifier no. 10.1109/TMC.2007.70740.
1. We define multirate networks as networks where different stationsmay employ different transmission data rates.
1536-1233/08/$25.00 � 2008 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
not been adequately analyzed in the literature. Suchscenarios often arise in ad hoc networks, which lack acentral controller.
This paper addresses the behavior of stations undermultirate scenarios in a distributed manner. The novelcontributions of this work are listed as follows:
1. Analytical formulation for characterizing the link delayand throughput of stations in a multirate environment.The model is particularly useful in providing a mathe-matical formulation of the rate anomaly problem.
2. Time-fair carrier sense multiple access (TFCSMA), whichis an online algorithm for ensuring long-term airtimefairness, thereby mitigating the performance anomalyproblem of the IEEE multirate networks. The algorithm isentirely distributed in nature and is completely compa-tible with the IEEE 802.11 DCF.
The underlying idea behind TFCSMA is to allocate equaltime shares to all competing stations in a distributedmanner such that stations employing higher data ratesreceive greater transmission opportunities, thereby increas-ing the aggregate throughput. We propose a simple yetefficient mechanism by which a station can maintain arunning estimate of its packet error rate (PER). Each stationuses the estimated PER and the employed transmission rateto dynamically determine a target throughput. Thiscalculation is based on the baseline property [4], whichguarantees airtime fairness in the network if all competingstations meet their target throughput. Every stationattempts to meet its respective target throughput bydynamically adjusting its minimum contention windowCWmin and, hence, controlling its transmission opportu-nities. Previous works [6], [7] assume equal packet lengthsand scale the contention window in proportion to the datarate (relative to the absolute maximum possible data rate).In contrast, TFCSMA directly alters the transmissionprobability based on the channel occupancy, therebyproviding greater transmission opportunities for links,which promise greater throughput. We thoroughly evaluatethe performance of TFCSMA by means of simulations andcomparisons to existing approaches, showing better per-formance in terms of fairness and throughput.
The rest of this paper is organized as follows: Section 2discusses the IEEE 802.11 DCF. An analytical model for thelink delay of multirate networks is presented in Section 3.The fairness objectives for multirate networks are discussedin Section 4. The inefficiencies in the existing solutions forthe rate anomaly problem are highlighted in Section 5.
Section 6 describes the functioning of TFCSMA. Weevaluate the performance and provide comparisons inSection 7. This paper concludes in Section 9 with someinteresting directions for future work.
2 IEEE 802.11 DISTRIBUTED COORDINATION
FUNCTION
The IEEE 802.11b standard [1] provides for both adistributed access mechanism and a contention-free arbitra-tion access method for media access control (MAC).Although the DCF uses CSMA with collision avoidance(CSMA/CA) for channel access, the Point CoordinationFunction (PCF) provides for contention-free access via anarbitrator that resides in access points. The standard allowsboth the methods to coexist: a contention period followed bya contention-free period. The DCF access mechanism followsthe CSMA/CA principle: A station wishing to transmitsenses the channel prior to transmitting. If the channel ismonitored idle for a period of Distributed Interframe Space(DIFS), the station transmits. Otherwise, it continues tomonitor the channel until it is sensed idle for a period ofDIFS. It then enters the collision-avoidance phase, where itgenerates a random backoff interval (exponential backoffscheme) to minimize the probability of collision with othercontending stations. For every packet transmission, thebackoff period is uniformly chosen in the range ð0; w� 1Þ.The value of w is called the contention window and itsvalue is contingent on the number of failed transmissionattempts for the packet. For the first packet transmission,w isset equal to a value ofCWmin called the minimum contentionwindow. With every unsuccessful transmission, m isdoubled till a maximum value ofCWmax ¼ 2mCWmin. Duringthe collision-avoidance phase, the backoff counter isdecremented as long as the channel is sensed idle. If atransmission is detected on the channel, the counter is“frozen” and reactivated when the channel is sensed idleagain. Once the counter reaches 0, the station transmits.
A successful packet reception is marked by the transmis-sion of a positive acknowledgment (ACK) by the destina-tion station after another fixed period of time (ShortInterframe Space (SIFS)). The failure of reception of anACK frame at the transmitter is assumed as a collision at thereceiver. Fig. 3 illustrates the contention process for threestations attempting to transmit using the DCF procedure.Station A selects a backoff window of nine slots. However,it has to freeze at slots 7 and 2, since at these instants,
514 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
Fig. 1. Example topology. Fig. 2. Throughput and COR versus data rate.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
station B and station C complete their respective backoffsand, hence, transmit. Station A transmits the packet whenthe backoff counter reaches 0.
The above-described two-way handshaking mechanismis popularly known as the Basic Access technique. DCFdefines an additional four-way handshaking procedureknown as the request-to-send/clear-to-send (RTS/CTS)mechanism. This access mechanism strives to reserve thechannel for the entire packet transmission duration and,hence, minimizes the impact of collisions. The additionalcost is a slightly increased transmission overhead (that is,RTS/CTS frame exchange). Fig. 4 depicts the RTS/CTSmechanism. Notice that the channel in the transmitter’svicinity is reserved via an RTS and CTS packet exchange.
3 ANALYTICAL MODELING OF MULTIRATE
NETWORKS
Performance evaluation of IEEE 802.11 DCF has beenwidely studied in [8], [9], [10], [11], and [12]. However, allthese works assume a single transmission rate by allcompeting stations. In practice, however, stations typicallyuse different transmission rates. This leads to the perfor-mance anomaly problem, which was discovered experi-mentally and studied with a simple analysis in [2].However, neither a mathematical formulation of theproblem nor mechanisms to remedy it have been consid-ered. This section provides a simple yet accurate analyticalframework to study the performance of multirate networks.A shorter version of this section is published in [20].
We begin by deriving the link-delay characteristics formultirate networks following the IEEE 802.11 DCF as a MACprotocol. We consider an IEEE 802.11b wireless network,which supports four transmission rates: 1, 2, 5.5, and11 Mbps. Each station is assumed to be in a saturatedcondition.2 First, we classify the stations according to their
transmission rates and assign a class identifier i ¼ 1, 2, 3,
and 4 to each class. In general, the total number of classes
can be N (N ¼ 4 for 802.11b, N ¼ 7 for 802.11a, and N ¼ 9 for
802.11g). Let nðiÞ be the number of stations in class i and Lavbe the average packet length transmitted by all stations. The
minimum contention window, transmission, and collision
probabilities for each station in class i is denoted by CWðiÞmin,
� ðiÞ, and pðiÞ, respectively. Last, to simplify our analysis, we
enumerate all the classes in the order of increasing data rates
(Fig. 5). Classes i and ðiþ 1Þ are comprised of stations
operating at a transmission rate of RdðiÞ and Rdðiþ 1Þ,respectively, such that RdðiÞ < Rdðiþ 1Þ.
As discussed in Section 2, if the channel is sensed idle by
a station for the duration of DIFS, it transmits the packet
after completing a random backoff. If a collision is detected,
this backoff time is doubled. In general, the station waits for
rj backoff slots after the jth collision, where rj is uniformly
chosen as follows [14]:
rj ¼ U 0; 2jCWðiÞmin � 1
� �j � m; ð1Þ
rj ¼ U 0; 2mCWðiÞmin � 1
� �j > m: ð2Þ
Here, U is the uniform distribution function.
Let k be the number of collisions before a station in class i
successfully transmits a packet. Then, the delay dðiÞ
experienced by the packet can be represented as
dðiÞ ¼ kT ðiÞc þ rTðiÞb þ T ðiÞs ; ð3Þ
where the value of the total number of backoff slots r isobtained as
r ¼ r1 þ r2 þ . . . . . . rk: ð4Þ
In addition, T ðiÞc , TðiÞb , and T ðiÞs are the average times for
collision, backoff, and transmission, respectively. We note
that TðiÞb is the average backoff duration and is different from
the backoff slot time � since the backoff timer freezes
whenever any transmission or collision is detected on the
network. It has been assumed in this derivation that the
delays encountered during each collision are independent
of each other. A similar assumption has been made in [8],
and past observations have shown this to be fairly accurate.
To obtain TðiÞb , we observe that the backoff slot time
comprises three components. First is the mandatory time
that is equal to �. In addition to this, the backoff slot time
may be incremented by T ðiÞc or T ðiÞs , depending on whether
a collision or a successful transmission had occurred
during that slot time, respectively. Since the respective
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 515
Fig. 3. IEEE 802.11 DCF basic access method.
Fig. 4. IEEE 802.11 DCF RTS/CTS method.
Fig. 5. Classification of stations on the basis of data rates.
2. A station that always has another frame waiting for transmission oncea frame has been successfully transmitted is said to be in a saturatedcondition.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
probabilities for these events are pðiÞc and pðiÞs for class i, we
obtain TðiÞb as
tðiÞb ¼ �þ
�pcT
ðiÞc þ
�psT
ðiÞs : ð5Þ
Here, the products pc �T ðiÞc and ps �T ðiÞs may be viewed as theexpected intervals for a collision or successful transmissionin the network when a station from class i is in the backoffstage. The successful packet transmission time T ðiÞs for atransmitting station in class i, including all the overheadand ACK packet time, is given by
T ðiÞx ¼ HPHY þLav þHMAC
RdðiÞ ; ð6Þ
T ðiÞs ¼ DIFS þ T ðiÞx þ SIFS þHPHY þ112
RcðiÞ : ð7Þ
Here, RdðiÞ and RcðiÞ are the data and control rates,respectively, for class i. At this point, we observe that theduration of a collision depends on the data rate of the sloweststation among all the colliding stations. Therefore, thecollision interval will equal T ðiÞx if no station in class jðj < iÞtransmits, and at least two stations in classkðk >¼ iÞ transmit.Further, we define � ðiÞ as the transmission probability ina given slot, and pðiÞ as the conditional collision probabilityof a station (belonging to class i). From [8], we have
� ðiÞ ¼ 2
1þ CW ðiÞmin þ pðiÞCW
ðiÞmin
Pmj¼0
2pðiÞð Þj; ð8Þ
pðiÞ ¼ 1�1� � ðiÞ� �nðiÞ
1� � ðiÞYNk¼1k 6¼i
1� � ðkÞ� �nðkÞ
: ð9Þ
Now, ps �T ðiÞs is the expected time that a station in class iwould wait if a valid transmission was to occur during oneof its backoff slots. Also, with the assumption that allpackets are of the same length, this duration will be equal toclass k’s packet transmission time if no other stations aretransmitting.
To derive pc �T ðiÞc , we observe that the length of this
duration will be equal to class k’s packet transmission time if
no other stations belonging to class jðj < kÞ are transmitting.
On the other hand, if stations belonging to class jðj > kÞtransmit, the collision time will still be equal to pc �T ðiÞc .
Thus, mathematically,
pcTðiÞc ¼
XNk¼1
T ðkÞx
nðkÞ
1
!� ðkÞ
Yk�1
j¼1
1� � ðjÞ� �nðkÞ !
1�
QNj¼k
1� � ðjÞ� �nðjÞ1� � ðkÞð Þ
0BBB@
1CCCA;
ð10Þ
psTðiÞs ¼
XNk¼1
T ðkÞsnðkÞ
1
� �� ðkÞ 1� � ðkÞ� �nðkÞ�1YN
j¼1j 6¼k
1� � ðkÞ� �nðjÞ
:
ð11Þ
We now estimate the value of the average collision time.The average collision duration for a station of class i can
either be T ðiÞs or T ðjÞx ðj < iÞ. A packet from class i may collidewith another packet of 1) its same class, 2) class jðj < iÞ, or3) class kðk > iÞ. In all the three cases, the collision durationfor a packet of class i can never be less than thetransmission time for that class. For this reason, we accountfor the ACK time when the packet collides with a packet ofthe same or a class with stations at higher transmission rates(since ACK time-out will occur) and neglect the time whenthe collision is with a packet of a slower class (assumingthat the ACK time-out will occur within the collisioninterval). Hence, T ðiÞc can now be expressed as
T ðiÞc ¼ T ðiÞs 1�
QNj¼i
1� � ðjÞ� �nðjÞ1� � ðiÞð Þ
0BBB@
1CCCA
Yi�1
j¼1
1� � ðjÞ� �nðjÞ !
þXi�1
j¼1
T ðjÞx 1� 1� � ðjÞ� �nðjÞ� �Yj�1
k¼1
1� � ðkÞ� �nðkÞ
:
ð12Þ
The IEEE 802.11 Standard specifies a long retry limitfor control packets, for example, RTS and a short retrylimit (for DATA packets) of seven and four, respectively,before a packet (control or data) is discarded. However,in our analysis, we consider unlimited retrials. Althoughthis assumption is an approximation, it does not skew theresults significantly since the probability of more thanm retrials is small, even for saturated traffic conditions.We also note that the majority of the research on the IEEE802.11 MAC analysis considers infinite retrials [8], [9],[10], [11], [12], [13].
The CDF for dðiÞ [13] can now be obtained as
FdðiÞ ðxÞ ¼P dðiÞ � x� �
¼X1j¼0
P kðiÞ ¼ j� �
P dðiÞ � xjkðiÞ ¼ j� �
¼X1j¼0
P kðiÞ ¼ j� �
P r� x� kTðiÞc � T ðiÞsTðiÞb
jkðiÞ ¼ j !
;
ð13Þ
where kðiÞ is the number of collisions in class i.We know that the probability mass function (pmf) of the
sum of two uniform random distributions is obtained by theconvolution of their distributions. Therefore, the pmf for r,given k collisions, is obtained as
DkðrÞ ¼UCWmin
� U2CWmin� . . .� U2kCWmin
ðrÞ k � mUCWmin
� U2CWmin� . . .� U2mCWmin
� Uk�m2mCWmin
ðrÞ k > m;
(
ð14Þ
where � is the convolution operator and UYX represents
Y convolutions of a uniform distribution between 0 toX � 1. A similar formulation has been derived by Tickooand Sikdar [15]. We can now easily derive the distributionfor the number of backoff slots r:
P ðr � Rjk ¼ jÞ ¼FrðRÞ
¼XRr¼0
DjðrÞ:ð15Þ
516 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
Now, substituting (15) in (13) and observing that PðkðiÞ ¼jÞ ¼ ð1� pðiÞÞ:ðpðiÞÞj [8], we can obtain the distribution for
dðiÞ as
FdðiÞ ¼X1j¼0
1� pðiÞ� �
pðiÞ� �ðjÞ
Frx� jT ðiÞc � T ðiÞs
TðiÞb
!: ð16Þ
The corresponding pmf PdðiÞ ðxÞ is obtained as
PdðiÞ ðxÞ ¼ Limh�>0
FdðiÞ ðxþ hÞ � FdðiÞ ðxÞh
: ð17Þ
The average throughput can now be calculated by
SðiÞav ¼LavPdmax
x¼0
xPdðiÞ ðxÞ: ð18Þ
4 FAIRNESS OBJECTIVES
Since the 802.11 MAC implicitly provides equal transmis-sion opportunities, slower stations tend to occupy moreairtime as compared to faster stations, provided that equal-sized packets are used by all stations. Clearly, a policyguaranteeing equal throughput is not appropriate for suchnetworks. Such policies fail to guarantee airtime fairnesswhen multiple stations with different data rates competewith each other. This neutralizes any advantage of employ-ing higher rates and thereby adversely lowers the aggregatethroughput. We note that maximizing the throughput isalso not a viable option since that will allow only the highrate stations to transmit, thereby leading to gross unfairnessand possible starvation of the slower stations.
Reference [3] recommends proportional fairness as theobjective of resource allocation to strike a balance betweenfairness and throughput for multirate networks. Whenmultiple stations contend for a single resource, proportionalfairness tries to allocate the resource in proportion to astation’s individual capacity. If channel bandwidth is takenas the resource in question, then proportional fairness shallstrive to ensure that each station achieves an individualthroughput proportional to its capacity.
This idea has been further explored by Tan and Guttag[4] in the context of time-based fairness. They suggest abaseline property, the consequence of which is time fairness.The property states that under equal time shares for allstations, the long-term throughput that a station wouldachieve while competing against n stations, each operatingat different data rates, is identical to the throughput thatthe station would achieve if it were competing againstn stations, all operating at their own data rates. Althoughthe property was validated through simulations, no formalmathematical analysis or proof was provided for the same.
Jiang and Liew [16] demonstrate the equivalence of
proportional and time-based fairness for wireless local area
networks (WLANs). They have shown that airtime fairness
is a natural consequence of the more fundamental property
of proportional fairness. Based on their analysis, they show
that, under airtime fairness, the ratio of throughputs of two
stations i and j is proportional to their respective data rates
Ri and Rj; that is, SiSj �Ri
Rj. However, their analysis neglected
protocol overheads and channel errors and, hence, the
result stands as a weak approximation.
If the effective good channel time is denoted by T and
assuming that this time is equally shared by all n stations,
then Si / TnTsi
, where Si is the throughput and Tsi is the
same as in (7). Therefore, we have SiSj� Tsj
Tsi. If the baseline
property has to be true, then Sei Tsi ¼ SejTsj , where Sei is the
throughput of station i if all other stations were operating at
Ri. Using [8], we can compute this ratio (say, rij) as follows:
rij ¼Sei TsiSejTsj
¼1�Ptrð Þ�Tsj
þ PtrPs þPtr 1�Psð ÞTcj
Tsj
1�Ptrð Þ�Tsi
þ PtrPs þPtr 1�Psð ÞTci
Tsi
: ð19Þ
Here, Ptr is the probability of at least one transmission in
a slot, and Ps is the probability of a successful transmission
in a slot. Clearly, rij must approach 1 for any i and j if the
baseline property has to hold. Fig. 7 plots rij for i ¼ 1 Mbps
and j ¼ 11 Mbps, respectively, against the total number of
stations for different payload sizes. The choice of 11 and
1 Mbps is obvious since these rates reflect the maximumpossible disparity. Several interesting observations manifest
from this plot. First, we observe that rij increases with
increasing payloads. This implies that the accuracy of the
baseline property increases with data payload size. Second,
we observe that, for lesser number of stations (20-40), rij is
close to 0.98 for almost all data payload sizes. Therefore, the
baseline property is sufficiently correct for a contention
region of around 40 stations. This is an interesting
observation since a neighborhood cardinality of 40 is a
reasonable assumption for a wireless network (both
WLANs and Ad Hoc). Last, we note that rij decreases with
increasing stations in the network. For a packet size of
1,500 bytes and 200 stations, rij is observed to be around
0.93. Fig. 8 plots rij for i ¼ 1 Mbps and j ¼ 2 Mbps,
respectively. For a neighborhood size of 200 stations and a
packet size of 1,500 bytes, rij is observed to be around 0.99.The baseline property is an interesting result for
developing an airtime-fair extension of 802.11. It allowsderiving an expected performance benchmark for a station,provided that n is known or can be estimated. A simplecontrol algorithm can then be developed to dynamicallyadjust the CWmin (and, hence, the transmission probability)to achieve this target in a distributed manner.
5 THE CONTENTION WINDOW SCALING PROTOCOL
(CWSP)
We now discuss the CWSP proposed in [6] to rectify the rateanomaly problem. The authors claim that the phenomenacan be cleanly resolved by setting the initial contentionwindow size to be inversely proportional to the bit rates. Asimilar idea is also proposed in [7]. Both of these methodspropose that the ratio of the contention windows be equal tothe ratio of their respective rates. Thus,
CW
CWc¼ Rmax
R: ð20Þ
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 517
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
In this case, if a station transmits at the highest availablerate Rmax, it uses a predefined minimum contentionwindow CWc. If a station transmits at a lower rate R, itmodifies the minimum contention window by scaling itwith respect to the maximum data rate. In [7], Heusse et al.propose that CWc be computed by using an algorithm thatmeasures the number of idle slots on the channel.
Although CWSP is a very simple mechanism formitigating the rate anomaly problem, it suffers from severaldrawbacks. First, it can lead to a considerable under-utilization of the channel. Consider a scenario where thereare only stations A and B operating at 1 and 2 Mbps,respectively. Such a scenario may exist when the channelconditions in a neighborhood are poor due to multipath orpropagation characteristics. If CWc is taken to be 31 for11 Mbps (the maximum data rate), then stations A and Bshall have a CWmin of 341 and 171, respectively, resulting insubstantial backoff. Notice that the anomaly could havebeen resolved with window sizes of 62 (station A) and 31(station B). In other words, the contention window needs tobe scaled with respect to the fastest transmitting peer andnot the absolute maximum data rate. This assumes greaterimportance, especially when the ratio between the slowestand the fastest possible rates is sufficiently large.
Table 1 compares the aggregate throughput for a two-station network topology (described above) for differentmaximum allowed data rates. The results are obtained viaOPNET simulations. We notice that an Optimal ScalingProtocol ((OSP), which is a theoretical protocol that adjusts thecontention window with respect to the fastest peer) alwaysoutperforms CWSP. The aggregate throughput for CWSPdrops by 28 percent (compared to IEEE 802.11 DCF) whenthe maximum allowable transmission rate is 54 Mbps. Thesuboptimal values of CWmin (1,674 and 837 for stations 1and 2, respectively) are primarily responsible for a largenumber of idle slots on the network bringing down theaggregate throughput. Further, we observe that, for all thescenarios, OSP increases the throughput by 12 percent withrespect to the IEEE 802.11.
Second, we argue that the ratio in (20) should indeedbe equal to Tsx
Tsmax(ratio of the effective good channel time)
instead of Rmax
R . The justification for this follows from thefact that the ratio of the contention windows should takeinto account any protocol overheads (Physical layer,ACKs, and so forth) and payload sizes. This is betterunderstood in Fig. 9, which plots the throughput of eachindividual flow for the topology in Fig. 6. Airtime fairnessis achieved at a ratio of around 7 instead of 11. Noticethat the throughput at this point is almost equal to thebaseline throughputs, validating the exact accuracy of theproperty for smaller number of stations.
Third and most importantly, CWSP ignores the time-varying properties of a wireless channel. The quality of thereceived signal at a receiver may vary considerably over ashort period of time due to noise, interference, multipath, orDoppler effects (due to mobility). This can lead to bit errorsand packet corruption at the receiver. The PER increaseswith higher transmission rates since higher rates require ahigher signal-to-noise ratio (SNR) to maintain the sameerror rates. Consider a scenario where there are twotransmitting stations, A and B, operating at 1 and 11 Mbps,respectively. Let PERa and PERb be the error rates at A andB’s receivers, respectively. With all other factors remainingthe same, we can assume that PERa � PERb. Now, CWSPattempts to achieve proportional throughput fairness bymodifying the transmission probability of a station inproportion to its data rate, with the implicit assumptionthat the error probability is equal for all rates. This may not
518 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
TABLE 1Aggregate Throughput (Kbps)
Fig. 6. Three-station example topology.
Fig. 7. Baseline property ði ¼ 1 Mbps; j ¼ 11 MbpsÞ.
Fig. 8. Baseline property ði ¼ 1 Mbps; j ¼ 2 MbpsÞ.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
be true for a practical scenario. Clearly, under suchcircumstances, scaling the contention window of station A11 times with respect to station B is incorrect. Fig. 10 depictsthe dependence of time fairness on different error rates. Asthe PER for the 11-Mbps (station B) link increases, the ratioof the minimum contention windows for stations A and Bshould increase in order to ensure a time-fair allocation ofthe channel. This is intuitive since the IEEE 802.11 DCFmandates an exponential backoff, even when a packet lossis due to link bit errors. As a result, erroneous links demandgreater transmission opportunities, and hence, the ratioincreases. Therefore, the idea of using a fixed ratio (as inCWSP) will clearly not meet the objective of airtime fairnesswhen different links have different error rates.
6 THE TIME FAIR CARRIER SENSE MULTIPLE
ACCESS PROTOCOL (TFCSMA)
6.1 Protocol Outline
This section describes our proposed protocol: Time FairCarrier Sense Multiple Access (TFCSMA). TFCSMA is anonline totally distributed extension of IEEE 802.11 DCFfor ensuring airtime fairness under multirate scenarios.TFCSMA utilizes the baseline property to dynamicallyderive a performance benchmark (target throughput) for astation based on its error and transmission data rates. Itthen uses a simple control algorithm to dynamically adjustthe CWmin (and, consequently, the transmission probability)to achieve this target in a distributed way. The protocol iscompletely compatible with IEEE 802.11 and requires nochanges to the basic access mechanism. It requires the use ofRTS-CTS since that facilitates the calculation of PERs. Inorder to account for scenarios where RTS-CTS is not used(for example, if the packet sizes are small), an alternatemethod will be required to estimate the PER. Such methodsmay involve temporal averaging of packet PER values andso forth. We intend to explore such methods in futureenhancements to our protocol.
TFCSMA operates at each individual station and hasthree basic modules: 1) PER Estimator (PERE), 2) TargetEstimator, and 3) Minimum Contention Window Controller(MCWC). The following sections describe their roles ingreater detail.
6.2 Packet Error Rate Estimator (PERE)
The Packet Error Rate Estimator (PERE) maintains a moving
average of the PER. TFCSMA mandates the use of an RTS-
CTS exchange before a data packet is transmitted. This is
necessary for decoupling collisions from packet errors due
to bit errors in the channel. RTS-CTS control packets reserve
the channel around the transmitter and the receiver,
respectively, and thereby minimize the probability of
collision of DATA packets. Therefore, a failure of an ACK
reception after a data packet transmission following an
RTS-CTS exchange can be assumed to be a result of link
error. Of course, we note that this mechanism is not
completely accurate. Collisions can still occur after an RTS-
CTS exchange [17]; however, the probability of its occur-
rence is minimal. Therefore, for the purpose of this work,
we neglect any packet losses due to collisions after an RTS-
CTS exchange.
PERE constantly updates a moving average of the PER
for different transmission rates and packet sizes. It main-
tains a 2D table3 T , where T ½i�½j� indicates the PER at rate
index i and packet size index4 j. PERE continuously counts
the number of ACKs � received within a window of
transmitted packets � (using the statistics report5) and
maintains T by the following:
perði; jÞ ¼ � � �
�
� �þ ð1� �Þ � perprevði; jÞ: ð21Þ
Here, � denotes the forgetting factor for the moving
average estimate. An example typical PER table for an IEEE
802.11b station using four different packet sizes is depicted
in Table 2.
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 519
Fig. 9. Throughput versus minimum CW. Fig. 10. Dependence on error rates.
3. For simplicity, we assume a single receiver PER transmitting station(as in WLANs) and, hence, propose a 2D table. In an ad hoc scenario, astation may have multiple receivers. In that case, we would require a3D table so that the PER can be maintained for every receiver/link.
4. While maintaining tables, we denote different transmission rates andpacket sizes by integers. For example, rates 1, 2, 5.5, and 11 Mbps aredenoted by rate indices 0, 1, 2, and 3, respectively.
5. After every successful packet transmission (ACK reception) or after apacket has been dropped (its retransmission numbers exceed the retrylimit), a statistic report is generated by the MAC. This provides completeinformation about the number of retrials, the end-to-end delay, backoff slotduration, and so forth for the current packet transmission.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
6.3 Target Throughput Estimator (TTE)
The TTE is a module that selects the target throughput St
for a station based on its current transmission rate, PER, andpacket size. It maintains a three-tuple table Tbaseline, whereTbaseline½i�½j�½k� denotes the saturation throughput if therewere k competing stations all operating at data rate index iand packet size index j. This table is computed offline byusing the saturated throughput expression under error-prone channel conditions, as derived in [18, Section 4.2,Equation (27)]. Since this expression was calculated forBasic IEEE 802.11, we have slightly modified the expressionby accounting RTS-CTS exchange time in collision time andfor successful and unsuccessful packet transmission. Table 3depicts a snapshot of a truncated offline table. All values arecomputed by using the expression described above.
TTE exploits the baseline property for achieving airtimefairness. Each station assumes that the transmission rate,packet sizes, and packet errors encountered by every otherstation in the network are the same as its own. Based onthese three parameters, the TTE selects St from thetable Tbaseline. Under similar data rates and packet sizes,802.11 guarantees both time and throughput fairness.Hence, if every station is able to achieve its target, thenairtime fairness is automatically guaranteed by virtue of thebaseline property.
It is assumed that every station is aware of the number ofcompeting stations in the network. The authors agree thatthis estimation is not a trivial task, particularly in thecontext of multirate networks. Most of the existing works inthis area assume a single-rate network with no channelerrors. The most notable among these are [21] and [22]. Theformer approach shows that the number of contendingstations can be expressed as a function of the collisionprobability and proposes a Kalman Filter algorithm forestimation based on runtime measurements. This mechan-ism fails for practical networks since it attributes packetlosses to only collisions. The latter mechanism advocatesbroadcasting backlogged packet status so that the activityratio is maintained at each neighboring station. This mayfail for multirate networks, where wireless links can beasymmetric, since a station operating at a higher rate has amuch smaller transmission range as opposed to a station ata lower rate. It is one of our ongoing works to dynamicallycompute the number of active stations in a typical multiratescenario. We also assume that each station operates under asaturated traffic condition and has a packet to send at anygiven instant.
TTE works in conjunction with PERE for dynamicallydetermining St. Just before a DATA packet transmission,TTE verifies with the PERE the current PER for thetransmission rate and packet size in question. It then
indexes into Tbaseline to compute the target throughput.Referring to Table 3, we note that Tbaseline uses a givenrange of the PER (for example, 0 to 10 percent) forcomputing the offline targets. In other words, the offlinethroughput for a data rate of R Mbps and P packet sizeremains the same for a PER of 0 to 10 percent (computed byusing the mean value, that is, 5 percent in this example).Therefore, St does not fluctuate unless the error ratechanges substantially or the station decides to change itstransmission rate or packet payload. For the presentimplementation of TFCSMA, a variation of more than 0.1in the PER forces a new target throughput.
6.4 Minimum Contention Window Controller(MCWC)
A control algorithm is required to track the measuredthroughput and dynamically adjust the minimum conten-tion window (and, consequently, the transmission prob-ability) to make it converge to the target throughput in adistributed manner. The MCWC is responsible for guidinga station to its respective target throughput. The transmis-sion probability is inversely proportional to the minimumcontention window. Also, the measured throughput isproportional to the transmission probability. By combiningthese two ideas, we arrive at our control algorithm. TheMCWC studies the statistic reports for a window ofaSamplePackets6 successful packet transmissions andadjusts the CWmin for the next window of transmissionsby the following control equation:
CWnewmin ¼ CW
prevmin �
Sm
St; ð22Þ
where Sm is the measured throughput computed from thelink delay obtained from the statistic reports. If themeasured throughput is observed to be greater in compar-ison to the target throughput, a station increases itsminimum contention window, thereby reducing its trans-mission probability and consequently pushing down thethroughput. Fig. 11 shows a block diagram of an IEEE802.11 MAC, which uses TFCSMA for airtime fairness.
520 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
TABLE 2PER Table
TABLE 3Snapshot of Offline Throughput Table
6. Measuring throughput on a PER-packet basis may result in erroneousfluctuations. Hence, the average throughput over a sample number ofpackets is used as the test criterion for the minimum contention window inuse. For the present implementation of TFCSMA, aSamplePackets was takento be 10. The test criterion is abandoned (and a new test started) if the targetthroughput changes within this test window.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
7 PERFORMANCE EVALUATION
The performance evaluation section is divided into two
subsections: 1) Analytical Model and 2) Simulation Results.
The former verifies the mathematical formulation of the rate
anomaly problem, whereas the latter compares the perfor-
mance of TFCSMA with CWSP and IEEE 802.11.
7.1 Analytical Model
The first part of our performance evaluation presents
results from our analytical model for multirate networks.
We analytically demonstrate the rate anomaly problem for
such network scenarios and the dependence of throughput
on the minimum contention window of the contending
stations.We consider a network topology composed of two
stations in class 4 (11 Mbps) and varying number of
stations in class 1 (1 Mbps). We note that our analysis is for
the basic IEEE DCF and, hence, the throughputs obtained
are higher than that for an RTS-CTS mechanism. The
number of stations in class 1 range from 2 to 15. The
payload size is fixed at 1,500 bytes.
Table 4 summarizes the parameters of IEEE 802.11b usedin the performance evaluation. We also assume that thenetwork is overloaded; that is, each station is in a saturatedcondition.
7.1.1 Throughput Dependence on CWmin
Figs. 12 and 13 illustrate the improvement in the aggregatethroughput through the differentiated minimum contentionwindow. Class 4 stations use the default initial CWmin ¼ 32,whereas class 1 (1-Mbps stations) uses larger values ofCWmin ranging from 32 to 352 while maintaining themaximum backoff window as 1,024. The upper value of 352represents a ratio of 11 between the CWmin of class 1 andclass 4 stations. When the CWmins are the same for bothclasses ð¼ 32Þ and there are two stations each in both, thethroughputs of classes 1 and 4 are both 1.6 Mbps. Therefore,the saturation throughput of each station in class 4 becomes0.8 Mbps (equal to the saturation throughput of each stationin class 1), even though stations in class 4 transmit data11 times faster. This exhibits the behavior of a performanceanomaly in multirate networks. Moreover, we observe thatthe throughput of stations in class 4 decreases as thenumber of stations in class 1 increases.
The performance anomaly phenomenon starts diminish-ing as we increase the CWmin of the stations in class 1 from32 to 352. Therefore, with CWmin being 32 and 352,
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 521
Fig. 11. Block diagram.
TABLE 4IEEE 802.11b Parameters
Fig. 12. Throughput dependence on CWmin.
Fig. 13. Aggregate throughput.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
respectively, and the total number of stations being four
(two in each class), we observe that the throughput for
stations in class 4 increases dramatically from 0.8 to 4 Mbps.
The throughput for class 1 stations expectedly drops down
to 0.35 Mbps. With differentiated minimum contention
window sizes for different classes, the total throughput
increases almost three times from 1.5 to 4.4 Mbps.The reason for this result is obvious. A higher contention
window leads to fewer transmission opportunities and
thereby increases the performance of stations that contend
with smaller window size. It is expected that the perfor-
mance of class 4 stations will keep on increasing (and,
consequently, class 1 stations’ throughputs keep decreas-
ing) with an increasing CWmin of class 1 stations. Therefore,
it is important to determine the optimal ratio of the
minimum contention window so that the throughput for
both flows is proportionally fair. Hence, with CWmin of 256
(which is close to a ratio of 8), we observe that stations in
classes 1 and 4 achieve an individual throughput of 0.4 and
3.5 Mbps, respectively, which is similar to the baseline
performance of both of these flows (that is, the throughput
that each station will achieve if all other stations are
operating at the same rate). This ratio also corresponds to an
airtime-fair channel allocation. Ideally, we would like our
proposed protocol to operate around this point.
7.1.2 Throughput Dependence on Packet Size
This section discusses the impact of packet payload size. Wefix the payload size for class 4 (11 Mbps) stations to1,500 bytes and vary the payload size of class 1 (1 Mbps)stations from 256 to 1,500 bytes. In addition, both classesuse the same value for the minimum contention window of32. The throughput of high-rate stations slightly increases asthe size of packet payload decreases for the low-ratestations. This is depicted in Figs. 14 and 15. When thereare four stations in the network, the throughput of eachstation in class 4 increases from 0.8 to 2.5 Mbps as thepayload size for class 1 decreases from 1,500 to 256 bytes.Smaller payload sizes correspond to lesser airtime. There-fore, although stations from both classes have equaltransmission opportunities (since CWmin is the same),class 1 stations occupy lesser airtime, which benefitsstations of class 4.
7.1.3 Link-Delay Characteristics
Figs. 16 and 17 plot the CDF of the link-delay characteristicsfor class 1 and class 4 stations, respectively. The number ofstations was fixed to five stations in each class. We vary theCWmin of class 1 stations from 32 to 352 while maintainingthe CWmin of class 4 stations at 32. We observe similartrends to those in Section 7.1.1. As the CWmin of class 1
522 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
Fig. 14. Throughput dependence on packet size.
Fig. 15. Aggregate throughput.
Fig. 16. Link-delay characteristics (1 Mbps).
Fig. 17. Link-delay characteristics (11 Mbps).
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
increases, the link delay for class 4 stations decreases. Thedelay characteristics are similar when the minimumcontention windows are the same, illustrating the rateanomaly problem.
7.2 Simulation Results
To evaluate the performance of our proposed protocol, wehave extended the IEEE 802.11 implementation in OPNET[19] that implements TFCSMA and CWSP. We presentresults for an IEEE 802.11b MAC which uses Direct-Sequence Spread Spectrum (DSSS) as the physical-layertechnology. We note that our results are representative ofany IEEE 802.11x suite of MAC protocols. The parametersfor TFCSMA are � ¼ 5 and aSamplePackets ¼ 10. Asimulation runs for a duration of 300 seconds and eachvalue is averaged over five different runs. Each stationoperates under saturated traffic load conditions.
7.2.1 Throughput
The goal of TFCSMA is to maximize throughput whileproviding airtime fairness. In this part, we evaluate thethroughput of our proposed protocol, but we do notexpect an extraordinary increase. The network topologycomprises two stations in class 4 and a varying number ofstations in class 1 (from 2 to 14). In addition, the PER forstations in class 4 is varied from 0 to 0.75 in steps of 0.25.The theoretical range of PER can be from 0 to 1, though inmost practical situations, it is typically not as high as 0.5
or 0.75. Such high values of PER are used for proof of
concept for our scheme. We compare the throughput
achieved by 802.11b, CWSP, and TFCSMA (Figs. 18, 19, 20,
and 21). Unless stated otherwise, the throughput of a class
refers to the individual throughput of a station in that
particular class.The throughput of class 4 stations decreases with an
increase in the number of stations in class 1 for all the three
protocols. Consequently, a reverse trend is observed for
class 4 throughputs. This is obvious, since the throughput
of class 1 stations is bound to increase with an increase in
their total number.Next, we observe for all PER ¼ 0:5 and 0.75 that CWSP
provides greater class 4 throughput and lower class 1throughput in comparison to TFCSMA. On the other hand,for PER ¼ 0:5 and 0.75, TFCSMA is better. This is becauseCWSP uses a fixed contention window ratio of 11 betweenclasses 1 and 4. Hence, CWSP ends up favoring class 4flows, thereby increasing their throughput and decreasingthe throughput of class 1 stations. However, this differencediminishes as the PER varies from 0 to 0.5. In fact, at a PERof 0.5, we note that the class 4 throughput of TFCSMAsurpasses that of CWSP. This can be attributed to thesuboptimal fixed CWmin ratio, which fails to providesufficient transmission opportunities to class 4 stations(under extremely poor link conditions), thereby restrictingtheir performance.
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 523
Fig. 18. Throughput comparison—PER ¼ 0 percent.
Fig. 19. Throughput comparison—PER ¼ 25 percent.
Fig. 20. Throughput comparison—PER ¼ 50 percent.
Fig. 21. Throughput comparison—PER ¼ 75 percent.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
7.2.2 Fairness
As stated above, the primary objective of TFCSMA is toprovide airtime fairness under multirate scenarios. Tomeasure airtime fairness, we define a fairness index (FI) as
FI ¼Smavg � St��� ���
St; ð23Þ
where Smavg is the average throughput of a flow over thesimulation duration. St is the target or the baselinethroughput for a flow. FI provides an extent of deviationfrom the target throughput and thereby gives a normalizedmeasure of airtime fairness in the network. Therefore, anFI of close to 0 represents airtime-fair channel allocation,whereas a value of nearly 1 (or greater than 1) indicatesgross unfairness in the network.
To evaluate and compare the fairness of TFCSMA, weuse a similar topology described in Section 7.2.2. Figs. 22,23, 24, and 25 plot the FI for varying PERs. Table 5 denotesthe target (or baseline) throughput for individual flows.Almost under all the scenarios, TFCSMA outperforms both802.11 and CWSP. It consistently provides an FI close to 0for all the flows in the network. As expected, 802.11 is theleast airtime-fair protocol among all the three protocols.
7.2.3 Channel Utilization
We discussed in Section 5 that CWSP can lead to aconsiderable channel underutilization since it scales the
CWmin with respect to the absolute maximum data rateinstead of the fastest transmitting peer. To evaluate theperformance of TFCSMA under such scenarios, we considera network topology composed of an equal number ofstations from both class 1 (1 Mbps) and class 2 (2 Mbps). Wevary the total stations from 4 to 10. Table 6 presents thecomparison of the aggregate throughput between 802.11,CWSP-2, CWSP-11, CWSP-54,7 and TFCSMA. TFCSMAoutperforms all other protocols and provides the maximumaggregate throughput under all scenarios. Tables 7 and 8compare the airtime fairness for both flows. We can see thatTFCSMA provides much better airtime fairness in compar-ison to all other protocols. 802.11 and CWSP-54 are the leastfair, exhibiting gross unfairness.
7.2.4 Convergence
The convergence speed of the TFCSMA protocol is the nextaspect to be evaluated. It is of paramount importance thatTFCSMA quickly restores airtime fairness whenever net-work conditions change. We consider the followingscenario. In the beginning, there are two stations: STA-1and STA-2. STA-1 uses a data rate of 1 Mbps and has a PERof 0. We conduct two different experiments. In the firstexperiment, the PER for STA-2 is varied while maintaininga transmission rate of 11 Mbps. The second test focuses on
524 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
Fig. 22. Fairness comparison—PER ¼ 0 percent.
Fig. 23. Fairness comparison—PER ¼ 25 percent.
7. CWSP-k refers to a network where the maximum permissible data rateis K Mbps. This implies that class 1 stations use a CWmin ¼ 32�K, whereasclass 2 stations use a CWmin ¼ 16�K.
Fig. 24. Fairness comparison—PER ¼ 50 percent.
Fig. 25. Fairness comparison—PER ¼ 75 percent.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
varying the transmission rate of STA-2 while maintaining a
PER of 0. It can be observed in Fig. 26 that TFCSMA adapts
rapidly to changing network conditions. An average
convergence period of less than 1.5 seconds was observed
for both the experiments. Fig. 27 shows the variation of the
CWmin with time for the first experiment. We can observe
that STA-1 steadily changes its window with varying error
rates for STA-2. Whenever the error rate increases for STA-2
(Table 9), it experiences more backoffs. This causes STA-1 to
start achieving a greater throughput than its required target
(because of the relatively more transmission opportunities).
In response, the MCWC of STA-1 increases the CWmin,
thereby indirectly helping STA-2 to achieve its targetthroughput.
7.2.5 Performance under Unsaturated Loads
Finally, we evaluate TFCSMA under unsaturated trafficload conditions. We again consider two flows in thenetwork: one at 1 Mbps (STA-1) and the other at 11 Mbps(STA-2). Fig. 28 plots the variation of the throughput of eachflow with the offered load.
Several interesting observations are in order based onthe plot. For an offered load of up to 600 Kbps, both CWSPand TFCSMA provide an equivalent throughput. When the
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 525
TABLE 5Target Throughput Table (Flows versus Number of Stations)
All values are in Mbps.
TABLE 6Aggregate Throughput Results
TABLE 7FI (Class 1 Flows)
TABLE 8FI (Class 2 Flows)
Fig. 26. Convergence of TFCSMA.
Fig. 27. CWmin variation for TFCSMA.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
offered load varies from 600 to 800 Kbps, TFCSMA exhibitsa dip in the performance of the 1-Mbps flow. Since theSTA-1 flow is also unsaturated, we would expect a greaterthroughput than 425 Kbps (the throughput underTFCSMA). However, TFCSMA computes its target withthe assumption that all its contending stations also operateunder saturation. Therefore, the MCWC of STA-1 pushesdown any incremental increase that would have beenotherwise possible.
The benefits of TFCSMA surface again when theoffered load for the STA-2 is increased beyond 2 Mbps.Again, the fixed window ratio in the case of CWSP beginsfavoring the 11-Mbps flow, thereby decreasing the 1-Mbpsflow below its baseline value (425 Kbps). As usual, IEEE802.11 continues to provide throughput fairness for bothflows, thereby severely impacting the performance of the11-Mbps flow.
8 COMPARISON WITH OPPORTUNISTIC AUTORATE
(OAR)
One of the earliest works on temporal fairness in IEEE802.11 networks was the OAR protocol [23] proposed bySadhegi et al. OAR is based on the assumption that thechannel coherence time typically exceeds multiple packettransmission times. Therefore, a station opportunisticallytransmits multiple back-to-back data packets via the IEEEFragmentation procedure whenever it experiences goodchannel conditions. The number of back-to-back packetstransmitted, Nback, is simply taken as the ratio of thetransmission rate to the base minimum rate. We observetwo major flaws in this mechanism:
. The protocol makes a strong assumption about thechannel coherence time. Although this may be truefor IEEE 802.11b, where the maximum value of Nback
can be 11, we highly doubt that such a mechanismwill work for 802.11g or networks where themaximum ratio can be as high as 54. Such anapproach is clearly not scalable with the maximumallowed data rate getting higher and higher.
. OAR, like CWSP, implicitly assumes that the PER is0 for a transmission rate in use. As discussed inSection 5, such an assumption may not be suitablefor ensuring time fairness.
In many ways, OAR shares the same philosophy asCWSP, except that OAR advocates back-to-back packettransmission, whereas CWSP provides Nback times moretransmission opportunities by varying the contentionwindow. Therefore, we believe that the flaws pointed outfor CWSP also apply to OAR. Although we do not comparethe performance of TFCSMA with OAR, we believe thatTFCSMA is a more robust mechanism for achieving airtimefairness based on the above arguments. A potential futurework in this area can be a comparison of TFCSMA withboth OAR and CWSP on a wireless testbed.
9 CONCLUSION AND FUTURE DIRECTIONS
We have presented TFCSMA, an online extension of theIEEE 802.11 DCF that dynamically adapts the minimumcontention window of contending stations to provideairtime fairness. This also mitigates the rate anomalyphenomenon observed in multirate networks. In addition,we have also presented an analytical model for evaluatingthe link-delay characteristics and throughput of stationsoperating in a multirate scenario.
Our method relies on the baseline property to dynami-cally derive a target throughput for a station based on itserror and transmission data rates. It uses a simple controlalgorithm to adjust the minimum contention window ofstations to achieve this target. It is fully distributed anddoes not mandate the need of a centralized point ofcoordination. Moreover, the protocol is completely compa-tible with the IEEE 802.11 and does not require any changesin the basic access procedure. Our simulation results onOPNET show very encouraging results in terms ofthroughput and fairness.
Future investigations will focus on the behavior ofTFCSMA in multihop scenarios. An undesirable conse-quence of airtime fairness under multihop networks can bean increased congestion at layer 2. This may happen when alow-rate link follows a high-rate hop in a route. New routemetrics, which pick routes with similar data rate links, maythen be required for satisfactory performance.
Another research direction can be the study of theeffect of hidden terminals in multirate scenarios. Trans-mission range decreases with an increase in data rates.This can lead to several asymmetric links in a multirate
526 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 7, NO. 4, APRIL 2008
TABLE 9Simulation Setup
Fig. 28. Results for unsaturated traffic conditions.
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.
network. Therefore, it is our intuition that rate disparitycan lead to an increase in the number of hidden terminalsin a network. We plan to address these issues in ourfuture work.
REFERENCES
[1] IEEE 802.11b/d3.0 Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specification, IEEE, Aug. 1999.
[2] M. Heusse, F. Rousseau, G. Berger-Sabbatel, and A. Duda,“Performance Anomaly of 802.11b,” Proc. IEEE INFOCOM, Apr.2003.
[3] B. Radunovic and J.-Y. Le Boudec, “Rate Performance Objectivesof Multi-Hop Wireless Networks,” Proc. IEEE INFOCOM, 2003.
[4] G. Tan and J. Guttag, “Time-Based Fairness Improves Perfor-mance in Multi-Rate WLANs,” Proc. Usenix Ann. Technical Conf.,June 2004.
[5] G. Tan and J. Guttag, “The 802.11 MAC Protocol Lead toInefficient Equilibria,” Proc. IEEE INFOCOM, 2005.
[6] H. Kim, S. Yun, I. Kang, and S. Bahk, “Resolving 802.11Performance Anomalies through QoS Differentiation,” IEEEComm. Letters, July 2005.
[7] M. Heusse, F. Rousseau, R. Guillier, and A. Duda, “Idle Sense: AnOptimal Access Method for High Throughput and Fairness inRate Diverse Wireless LANs,” Proc. ACM Ann. Conf. Applications,Technologies, Architectures, and Protocols for Computer Comm.(SIGCOMM ’05), 2005.
[8] G. Bianchi, “Performance Analysis of the IEEE 802.11 DistributedCoordination Function,” IEEE J. Selected Areas in Telecomm.,wireless series, vol. 18, no. 3, Mar. 2000.
[9] B.L. and R. Battiti, “Analysis of the IEEE 802.11 DCF with ServiceDifferentiation Support in Non-Saturation Conditions,” Proc. FifthInt’l Workshop Quality of Internet Services (QofIS ’04), 2004.
[10] M. Carvalho and J.J. Garcia-Luna-Aceves, “A Scalable Model forChannel Access Protocols in Multihop Ad Hoc Networks,” Proc.ACM MobiCom, 2004.
[11] M. Sasaki, K. Sakakibara, and J. Yamakita, “Performance Analysisof Back Off Algorithms under Unsaturated Conditions,” Electro-nics and Comm. in Japan, Dec. 2004.
[12] M. Carvalho and J.J. Garcia-Luna-Aceves, “Delay Analysis of IEEE802.11 in Single-Hop Networks,” Proc. 11th IEEE Int’l Conf.Network Protocols (ICNP ’03), 2003.
[13] Y. Wang and J. Garcia-Luna-Aceves, “Performance of CollisionAvoidance Protocols in Single-Channel Ad Hoc Networks,” Proc.10th IEEE Int’l Conf. Network Protocols (ICNP ’02), 2002.
[14] A. Mukherjee, W. Li, and D.P. Agrawal, “Performance Analysis ofIEEE 802.11 for Multi-Hop Infrastructure Networks,” Proc. IEEEGlobal Telecomm. Conf. (GLOBECOM ’05), 2005.
[15] O. Tickoo and B. Sikdar, “Queueing Analysis and DelayMitigation in IEEE 802.11 Random Access MAC-Based WirelessNetworks,” Proc. IEEE INFOCOM, 2004.
[16] L. Jiang and S.C. Liew, “Proportional Fairness in WLANs andAd Hoc Networks,” Proc. IEEE Wireless Comm. and NetworkingConf. (WCNC ’05), Mar. 2005.
[17] K. Xu, M. Gerla, and S. Bae, “How Effective Is the IEEE 802.11RTS/CTS Handshake in Ad Hoc Network?” Proc. IEEE GlobalTelecomm. Conf. (GLOBECOM ’02), Nov. 2002.
[18] Q. Li, T. Li, T. Turletti, and Y. Xiao, “Saturation ThroughputAnalysis of Error-Prone 802.11 Wireless Networks,” Wiley WirelessComm. and Mobile Computing, vol. 5, no. 8, pp. 945-956, Nov. 2005.
[19] OPNET Technologies, http://www.opnet.com, 2008[20] T. Joshi, A. Mukherjee, and D.P. Agrawal, “Analytical Modeling
of the Link Delay Characteristics for IEEE Multi-Rate WLANs,”Proc. IEEE Canadian Conf. Electrical and Computer Eng. (CCECE ’06),2006.
[21] G. Bianchi and I. Tinnirello, “Kalman Filter Estimation of theNumber of Competing Terminals in an IEEE 802.11 Network,”Proc. IEEE INFOCOM, vol. 2, pp. 844-852, 2003.
[22] J.B. Kim, J.H. Kim, Y.A. Song, and J.G. Kim, “A Self-TuningContention Resolution Scheme for IEEE 802.11 Wireless LAN,”Proc. Eighth IEEE Int’l Conf. Advanced Comm. Technology (ICACT’06), 2006.
[23] B. Sagdehi, V. Kanodia, A. Sabharwal, and E. Knightly, “OAR: AnOpportunistic Autorate Media Access Protocol for Ad HocNetworks,” Proc. ACM MobiCom, 2002.
Tarun Joshi received the BS degree in technol-ogy in computer science from the Indian Instituteof Technology, Roorkee, in 2002 and the PhDdegree in computer science and engineeringfrom the University of Cincinnati, Ohio. From2004 to 2005, he was a systems engineer internwith Motorola Inc. He is currently a softwaredesign engineer with the Entertainment andDevices Division, Microsoft Corp. His researchinterests include media access control (MAC)
protocols over directional antennas, broadcasting and routing issues withdirectional antennas, link adaptation for 802.11 devices, mobilitymanagement, and security in infrastructure wireless networks. Hereceived the honorable Outstanding Doctoral Dissertation Award fromthe University of Cincinnati in 2006-2007. He is a student member of theIEEE and the IEEE Computer Society.
Anindo Mukherjee received the BS degree intechnology from the Indian Institute of Technol-ogy, Roorkee, in 2000 and the PhD degree incomputer engineering from the University ofCincinnati in 2006 under the guidance of Dr.Dharma P. Agrawal. From 2000 to 2002, he wasa software engineer with Cisco Systems, India,and Tejas Networks, India, where he designedand developed embedded software. He is cur-rently with Google as a software engineer.
Younghwan Yoo received the BS and MSdegrees in computer engineering in 1996 and1998, respectively, and the PhD degree inelectrical engineering and computer science in2004 from Seoul National University. From 2004to 2006, he was a postdoctoral researcher withthe OBR Center for Distributed and MobileComputing, University of Cincinnati. He iscurrently an assistant professor with the Schoolof Computer Science and Engineering, Pusan
National University, Korea. He has served as a reviewer for a variety ofjournals and conferences. His research interests include routing andsecurity issues in ad hoc and sensor networks, wireless mesh networks,wavelength-division multiplexing (WDM), asynchronous transfer mode(ATM), and the Internet. He is a member of the IEEE.
Dharma P. Agrawal is the Ohio Board ofRegents Distinguished Professor and the found-ing director of the Center for Distributed andMobile Computing, Department of ComputerScience, University of Cincinnati, Ohio. He hasbeen a faculty member with Carnegie MellonUniversity (on sabbatical, Fall 2006 and Winter2007), North Carolina State University, Raleigh(1982-1998), and Wayne State University, De-troit (1977-1982). He has been a Fulbright senior
specialist for five years. He has served as an editor for many reputablejournals and has been the program chair or general chair of manyinternational conferences and meetings. He has also given severalkeynote talks and plenary speeches at major international conferences.His current research interests include resource allocation and security inmesh networks, efficient query processing and security in sensornetworks, and heterogeneous wireless networks. He has published morethan 500 papers in international journals, magazines, and conferenceproceedings. He is a coauthor of a widely accepted textbook on wirelessand mobile computing that has been reprinted in China and India and hasbeen translated to Korean and Chinese and a coauthor of a best-sellingbook on ad hoc and sensor networks published in 2006. He holds fivepatents in wireless networking. He is a fellow of the IEEE, the ACM, theAAAS, and the World Innovation Foundation and is a member of the IEEEComputer Society. He has received numerous certificates and awardsfrom the IEEE Computer Society, received the Third Millennium Medalfrom the IEEE for his outstanding contributions, and was named as anInstitute for Scientific Information (ISI) Highly Cited Researcher incomputer science.
JOSHI ET AL.: AIRTIME FAIRNESS FOR IEEE 802.11 MULTIRATE NETWORKS 527
Authorized licensed use limited to: UNIVERSITY OF COLORADO. Downloaded on March 10, 2009 at 18:14 from IEEE Xplore. Restrictions apply.