538 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED …yzchen/papers/papers/newpapers/qos_1hop.pdf · Many issues are involved to provide quality of service (QoS) for ad hoc networks,

Differentiation, QoS Guarantee, andOptimization for Real-Time Trafficover One-Hop Ad Hoc NetworksYang Xiao, Senior Member, IEEE, and Yi Pan, Senior Member, IEEE

Abstract—Nodes having a self-centrically broadcasting nature of communication form a wireless ad hoc network. Many issues are

involved to provide quality of service (QoS) for ad hoc networks, including routing, medium access, resource reservation, mobility

management, etc. Previous work mostly focuses on QoS routing with an assumption that the Medium Access Control (MAC) layer can

support QoS very well. However, contention-based MAC protocols are adopted in most ad hoc networks since there is no centralized

control. QoS support in contention-based MAC layer is a very challenging issue. Carefully designed distributed medium access

techniques must be used as foundations for most ad hoc networks. In this paper, we study and enhance distributed medium access

techniques for real-time transmissions in the IEEE 802.11 single-hop ad hoc wireless networks. In the IEEE 802.11 MAC, error control

adopts positive acknowledgement and retransmission to improve transmission reliability in the wireless medium (WM). However, for

real-time multimedia traffic with sensitive delay requirements, retransmitted frames may be too late to be useful due to the fact that the

delay of competing the WM is unpredictable. In this paper, we address several MAC issues and QoS issues for delay-sensitive real-

time traffic. First, a priority scheme is proposed to differentiate the delay sensitive real-time traffic from the best-effort traffic. In the

proposed priority scheme, retransmission is not used for the real-time traffic, and a smaller backoff window size is adopted. Second,

we propose several schemes to guarantee QoS requirements. The first scheme is to guarantee frame-dropping probability for the real-

time traffic. The second scheme is to guarantee throughput and delay. The last scheme is to guarantee throughput, delay, and frame-

dropping probability simultaneously. Finally, we propose adaptive window backoff schemes to optimize throughput with and without

QoS constraints.

Index Terms— Distributed medium access control, IEEE 802.11, quality of service, real-time transmission, ad hoc networks.

�

1 INTRODUCTION

AD hoc wireless networks consist of a collection ofmobile stations without a fixed infrastructure. In ad

hoc wireless networks, peer-to-peer nodes conduct initi-alization, organization, and administration of networks.Many challenges must be overcome to obtain the practicalbenefits of ad hoc networks, including routing, mediumaccess control (MAC), mobility management, power man-agement, security, and quality of service (QoS) issues [1].The nodes of an ad hoc network communicate directly withone another in a peer-to-peer fashion, and each node mustfunction as a router. Power capacity and transmission rangeare further limited by the mobility of these nodes. Due tothe mobility, the network topology is dynamically changed.Furthermore, the limited bandwidth of wireless channelsand hostile transmission characteristics impose additionalconstraints. For ad hoc networks with a contention-basedMAC layer, the nature of contentions further imposeschallenges for QoS support.

Many researches [1], [2], [3], [4], [5] on ad hoc networkshave been reported, and real-time transmissions forEthernet and cellular networks have been well studied [6],

[7], [8], [9], [10]. However, most researches [1], [2], [3], [4],[5] focus on ad hoc routing protocols under the assumptionthat some underline MAC protocols can provide goodservices to higher layers. In this paper, we focus ondesigning good MAC mechanisms for real-time transmis-sions in ad hoc networks. Without the MAC layer’s support,QoS support solely in higher layers is not possible.Carefully designed distributed medium access techniquesmust be used for channel resources so that mechanisms areneeded to recover efficiently from the inevitable framecollisions [1].

We are particularly interested in ad hoc networks withthe underneath IEEE 802.11 distributed MAC since it isavailable. The IEEE 802.11 MAC employs a mandatorycontention-based channel access function called DistributedCoordination Function (DCF), and an optional centrallycontrolled channel access function called Point Coordina-tion Function (PCF) [11]. The popularity of the IEEE 802.11market is largely due to the DCF, whereas the PCF is barelyimplemented in current products due to its complexity andinefficiency for normal data transmissions. The DCF adoptsa carrier sense multiple access with collision avoidance(CSMA/CA) with binary exponential backoff. Functions ofthe DCF and the PCF determine when a station/node,operating within a Basic Service Set (BSS) or IndependentBSS (IBSS), is permitted to transmit. There are two types of802.11 networks: Infrastructure Network (BSS) in which anaccess point (AP) is present and ad hoc Network (IBSS) inwhich an AP is not present. In this paper, we areparticularly interested in ad hoc networks formed bymultiple IBSSs in which no AP is present. There have been

538 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 6, JUNE 2005

. Y. Xiao is with the Computer Science Division, The University ofMemphis, 373 Dunn Hall, Memphis, TN 38152.E-mail: [email protected].

. Y. Pan is with the Department of Computer Science, Georgia StateUniversity, University Plaza, Atlanta, GA 30303.E-mail: [email protected].

Manuscript received 29 Dec. 2003; revised 16 Sept. 2004; accepted 19 Sept.2004; published online 21 Apr. 2005.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TPDS-0243-1203.

1045-9219/05/$20.00 � 2005 IEEE Published by the IEEE Computer Society

many performance studies for the DCF. Bianchi [13]proposed a simple and accurate analytical model tocompute saturation throughput. Bianchi and Xiao enhancedBianchi’s original model in [18]. Calı̀ et al. [20] studied anoptimization method for a p-persistent WLAN MAC.

Several priority studies have been reported in theliterature for the DCF. Deng and Chang [12] proposed apriority scheme by differentiating the backoff window: thehigher priority class uses the window ½0; 2jþ1 � 1� and thelower priority class uses the window ½2jþ1; 2jþ2 � 1�, where jis the backoff stage. Aad and Castelluccia [21] proposed apriority scheme achieved by differentiating interframespaces (IFS). Ahn et al. [19] proposed priority schemes bydifferentiating the initial backoff window size and themaximum window size. Pallot and Miller [22] proposed aprioritized backoff time distribution mechanism, in whichthe backoff time is chosen in the current window range withdifferent distributions for different priorities. All thepriority schemes [12], [19], [21], [22], [23] were based onsimulations. Related recent work also includes Sheu et al.[25] about priorities for Ad hoc networks, and Ge and Hou[26] about priority analysis for p-persistent WLANs.

In the DCF, error control adopts positive acknowl-edgment and retransmission to improve transmissionreliability in the wireless medium (WM). In other words,every transmitted frame needs a positive acknowledgment.If an acknowledgement for a transmitted frame has notbeen received for a timeout period, the transmitted frame isassumed corrupted, and the frame will be retransmitted formany times until a positive acknowledgement is received orthe number of retransmissions reaches a limit. In the latercase, the frame is dropped. Therefore, the DCF is a veryrobust protocol for the best-effort service in the WM.However, the current DCF is unsuitable for real-timeapplications with QoS requirements. In the DCF a stationmight have to wait an arbitrarily long time to send a frameso that real-time applications such as voice and video maysuffer [12]. Furthermore, for real-time multimedia trafficwith sensitive delay requirements, retransmitted framesmay be too late to be useful due to the unexpected delay.

In this paper, we consider two classes of traffic: delaysensitive real-time (RT) traffic and best-effort (BE) traffic.The RT traffic can be voice or video. The BE traffic is normaldata transmission. A priority scheme is proposed todifferentiate RT and BE classes. For the RT class, retrans-mission is not used and a smaller backoff window size isadopted. The BE class still follows the original DCF. Such ascheme is similar to video/multimedia over UDP. Therationale is the same as the rationale of UDP. An analyticalmodel for the proposed priority scheme (differentiating theRT and the BE priority classes) is proposed to evaluatesystem performance, and validated via simulations. Theproposed real-time differential mechanisms are a littlesimilar to recently IEEE 802.11e draft [24]. However, theproposed mechanisms are much simpler than IEEE 802.11eas well as other priority schemes in the literature, and morelikely to be used in real products in which the IEEE 802.11ehas not been implemented. In other words, they can beimplemented in the original IEEE 802.11a/.11b/.11g withvery little effort.

We further consider several QoS and optimization issues.We propose several schemes to guarantee QoS requirements:the first scheme is to guarantee frame-dropping probability

for the real-time traffic; the second scheme is to guaranteethroughput and delay; the last scheme is to guaranteethroughput, delay and frame-dropping probability simulta-neously. Finally, we propose adaptive window backoffschemes to optimize throughput with and without QoSconstraints, based on the fact that there exists an optimalinitial backoff window size for a fixed traffic load.

The rest of the paper is organized as follows: The IEEE802.11 CSMA/CA is introduced in Section 2. Servicedifferentiation, its analytical model, and results are studiedin Section 3. QoS guarantee mechanisms are presented inSection 4. Optimization schemes with adaptation and QoSguarantee are proposed in Section 5. We conclude thispaper in Section 6.

2 CSMA/CA IN THE DCF

The IEEE 802.11 MAC employs a mandatory DCF and anoptional PCF. In a long run, time is divided into repetitionintervals called superframes. Each superframe starts with abeacon frame, and the remaining time is further dividedinto a contention-free period (CFP) and a contention period(CP). The DCF works during the CP and the PCF worksduring the CFP. If the PCF is not active, a superframe willnot include the CFP.

In the DCF, a station with a frame to transmit monitorsthe channel activities until an idle period equal to adistributed interframe space (DIFS) is detected. Aftersensing an idle DIFS, the station waits for a random backoffinterval before transmitting. The backoff time counter isdecremented in terms of slot time as long as the channel issensed idle. The counter is stopped when a transmission isdetected on the channel, and reactivated when the channelis sensed idle again. The station transmits its frame whenthe backoff time reaches zero. At each transmission, thebackoff time is uniformly chosen in the range ½0; CW � 1�,where CW is the current backoff window size. At the veryfirst transmission attempt, CW equals the initial backoffwindow size CWmin. After each unsuccessful transmission,CW is doubled until a maximum backoff window sizevalue CWmax is reached. Once it reaches CWmax, CW shallremain at the value CWmax until it is reset. CW shall be resetto CWmin after every successful attempt to transmit, or theretransmission counter reaches the retry limit Lretry. In thelater case, the frame will be dropped. After the destinationstation successfully receives the frame, it transmits anacknowledgment frame (ACK) following a short interframespace (SIFS) time. If the transmitting station does notreceive the ACK within a specified ACK Timeout, or itdetects the transmission of a different frame on the channel,it reschedules the frame transmission according to theprevious backoff rules.

The above mechanism is called the basic access mechan-ism. To reduce the hidden station problem, an optionalfour-way data transmission mechanism called Request-To-Send (RTS)/Clear-To-Send (CTS) is also defined in DCF. Inthe RTS/CTS mechanism, before transmitting a data frame,a short RTS frame is transmitted. The RTS frame alsofollows the backoff rules introduced above. If the RTS framesucceeds, the receiver station responds with a short CTSframe. Then, a data frame and an ACK frame will follow.All four frames (RTS, CTS, data, and ACK) are separated byan SIFS time. In other words, the short RTS and CTS frames

XIAO AND PAN: DIFFERENTIATION, QOS GUARANTEE, AND OPTIMIZATION FOR REAL-TIME TRAFFIC OVER ONE-HOP AD HOC... 539

reserve the channel for that data frame transmission whichfollows.

3 SERVICE DIFFERENTIATION

In this section, we study service differentiation between thedelay sensitive real-time (RT) traffic and the best-effort (BE)traffic. We first introduce the priority scheme; then wepropose an analytical model to study performance ofservice differentiation; and, finally, we present simulationand numerical results.

3.1 A Priority Scheme to Differentiate RT and BE

The original IEEE 802.11 DCF does not support any priorityscheme. In this paper, we consider two classes of traffic: thedelay sensitive RT traffic and the BE traffic. The RT trafficcan be voice or video, and the BE traffic is normal datatransmission. The BE class still follows the original DCF.For the RT-traffic, retransmission is not used no matterwhether the previous transmission successful or not.Furthermore, The RT class has a smaller backoff windowsize than the BE class: CWmin;RT < CWmin;BE , whereCWmin;RT and CWmin;BE are the initial backoff windowsizes for the RT class and the BE class, respectively. LetCWmax;RT and CWmax;BE be the maximum window sizes forthe RT class and the BE class, respectively. We haveCWmax;RT ¼ CWmax;BE .

To implement the mechanism, we can simply let theretransmission limit for the RT traffic become zero. There-fore, retransmission mechanism will be automaticallydisabled. Such a scheme is similar to video/multimediaover UDP. The rationale is the same as the rationale of UDP.

3.2 Analytical Model

In this section, based on Bianchi’s model [13], an analyticmodel for the proposed priority scheme under high trafficcondition is proposed. The advantage of our model is toprovide priorities of the RT class and the BE class, whereasBianchi’s model [13] is for the original DCF. Furthermore,compared to Bianchi’s model, many aspects are improvedsuch as underline assumptions and a very accurate delaymodel. We assume that each station belongs to one and onlyone priority class and always has frames ready to send.

3.2.1 An Analytical Model

For convenience, let class 0 denote the RT class and class 1

denote the BE class. Let W0;0 and W1;0 denote CWmin;RT and

CWmin;BE , respectively. For the RT class, the backoff stage j

can only be zero, and the retry-limit Lretry is zero also. For

the BE class, let jðj ¼ 0; 1; . . . ; LretryÞ denote the jth backoff

stage and let W1;j denote CW in the jth retry/retransmis-

sion (or the jth backoff stage). The relationships among

W1;j, CWmin;BE , CWmax;BE , and Lretry are given as follows:

W1;j ¼2jW1;0 for j¼0; 1; . . . ;m�1; if Lretry>m

2mW1;0 ¼ CWmax;BE for j¼m; . . .Lretry; if Lretry>m

2jW1;0 for j¼0; 1; . . . ; Lretry; if Lretry�m;

8><>:

ð1Þ

where m ¼ log2ðCWmax;BE=W1;0Þ and W1;0 ¼ CWmin;BE .To understand the relationship of Lretry and m, we give

the following example: In the IEEE 802.11 MAC, the defaultvalue of Lretry for short packets is 7, however, the m value,the maximum backoff stage, may be larger than Lretry orsmaller than Lretry depending on the values of CWmin;BE

and CWmax;BE . If CWmax;BE ¼ 1; 024 and CWmin;BE ¼ 32, wehave m ¼ 5 < 7 ¼ Lretry; on the other hand, if CWmax;BE ¼2; 048 and CWmin;BE ¼ 8, we have m ¼ 8 > 7 ¼ Lretry.

For a given station in the priority i class ði ¼ 0; 1Þ, bði; tÞis defined as a random process representing the value of thebackoff counter at time t, and sði; tÞ is defined as therandom process representing the backoff stage j, where j ¼0 for class 0 and 0 � j � Lretry for the BE class. The values ofthe backoff counter bð0; tÞ and bð1; tÞ are uniformly chosenin the ranges ½0; 1; . . .W0;0 � 1� and ½0; 1; . . .W1;j � 1�, respec-tively. Let pi denote the probability that a transmitted framecollides, and pi also equals the probability that a station inthe backoff stage senses the channel busy. The bidimen-sional random process fsði; tÞ; bði; tÞg is discrete-timeMarkov chain under the assumptions that pi is independentto the backoff procedure. Therefore, the state of each stationin the priority i class is described by fi; j; kg, where j standsfor the backoff stage, and k stands for the backoff delay intimeslots.

The state transition diagram for the priority 0 class isshown in Fig. 1. As illustrated in Fig. 1, the retry limit iszero. The state transition diagram for the priority 1 class isshown in Fig. 2. As illustrated in the Fig. 2, in the Lretrythbackoff stage, the frame is dropped if a collision occurs. In


Fig. 1. State transmission diagram for the RT class (Class 0).

Fig. 2. State transmission diagram for the BE class (Class 1).

Fig. 1, the non-null transition probabilities for class 0 arelisted as follows: We have the following transitionprobabilities between states:

Prfð0; 0; kÞjð0; 0; 0Þg ¼ 1=W0;0; for 0 � k � W0;0 � 1

Prfð0; 0; kÞjð0; 0; kÞg ¼ p0; for 1 � k � W0;0 � 1

Prfð0; 0; kÞjð0; 0; kþ 1Þg ¼ 1� p0; for 0 � k � W0;0 � 2:

In Fig. 2, the non-null transition probabilities for class 1are listed as follows:

Prfð1; 0; kÞjð1; j; 0Þg ¼ 1� p1W1;0

for 0 � k � W1;0 � 1

and 0 � j � Lretry

Prfð1; 0; kÞjð1; Lretry; 0Þg ¼ 1=W1;0; for 0 � k � W1;0 � 1

Prfð1; j; kÞjð1; j; kÞg ¼ p1 for 1 � k � W1;j � 1


Prfð1; j; kÞjð1; j; kþ 1Þg ¼ 1� p1 for 0 � k � W1;j � 2


Prfð1; j; kÞjð1; j� 1; 0Þg ¼ p1W1;j

for 0 � k � W1;j � 1

and 1 � j � Lretry:

To understand how to obtain the above equations, we givethe following example: the output of the state ð1; Lretry; 0Þwillgo into the state ð1; 0; kÞ randomly in Fig. 2 and, therefore, wehave Prfð1; 0; kÞjð1; Lretry; 0Þg¼1=W1;0. Let bi;j;k¼ limðt�>1Þ Prfsði; tÞ ¼ j; bði; tÞ ¼ kg be the stationary distribution ofthe Markov chain, where i ¼ 0; 1. In steady state, we canderive following relations through chain regularities. Forclass 0, we have

b0;0;k ¼W0;0 � k

W0;0

1

1� p0b0;0;0 for 1 � k � W0;0 � 1 ð2Þ

XW0;0�1

k¼0

b0;0;k ¼ 1: ð3Þ

Plugging (2) into (3), we have

b0;0;0 ¼1

1þ 11�p0

PW0;0�1

k¼1W0;0�kW0;0

¼ 2ð1� p0Þ2ð1� p0Þ þ ðW0;0 � 1Þ :

ð4Þ

For class 1, we have

b1;j;0 ¼ pj1b1;0;00 � j � Lretry ð5Þ

b1;j;k ¼W1;j � k

W1;j

1

1� p1b1;j;0; 0 � j � Lretry; 1�k� W1;j�1 ð6Þ

XLretry

j¼0

XW1;j�1

k¼0

b1;j;k ¼ 1: ð7Þ

From (5), (6), and (7), we have

b1;0;01PLretry

j¼0 1þ 11�p1

PW1;j�1k¼1

W1;j�kW1;j

h ipj1

: ð8Þ

Let �iði ¼ 0; 1Þ be the probability that a station in thepriority i class transmits during a generic slot time. A

station transmits when its backoff counter reaches zero, i.e.,

the station is at any of states fi; j; 0g.

�0 ¼ b0;0;0 ð9Þ

�1 ¼XLretry

j¼0

b1;j;0 ¼ b1;0;01� p

Lretryþ11

1� p1: ð10Þ

Let niði ¼ 0; 1Þ denote the number of stations in the

priority i class. The probability pi that a station in the

backoff stage for the priority i class senses the channel busy

is given

p0 ¼ 1� ð1� �oÞn0�1ð1� �1Þn1 ð11Þp1 ¼ 1� ð1� �0Þn0ð1� �1Þn1�1: ð12Þ

Note that pi is also the probability that a transmitted

frame collides when one more station also transmits during

a slot time. Substituting (1) and (11) and (12) to (9) and (10),

we can solve unknown parameters numerically. Then, we

can calculate pi from (11) and (12). Let pb denote the

probability that the channel is busy. It happens when at

least one station transmits during a slot time. Please note

that pb is different from pi. Therefore, we have

Pb ¼ 1� ð1� �0Þn0ð1� �1Þn1 : ð13Þ

3.2.2 Saturation Throughput

Let ps;iði ¼ 0; 1Þ denote the probability that a successful

transmission occurs in a slot time for the priority i class, and

let ps denote the probability that a successful transmission

occurs in a slot time. We have

ps;0 ¼ n0�0ð1� �0Þn0�1ð1� �1Þn1 ð14Þps;1 ¼ n1�1ð1� �1Þn1�1ð1� �0Þn0 ð15Þps ¼ ps;0 þ ps;1: ð16Þ

Let Siði ¼ 0; 1Þ denote the normalized throughput for the

priority i class. Let �, TEðLÞ, Ts, and Tc denote the duration of

an empty slot time, the time to transmit the average

payload, the average time that the channel is sensed busy

because of a successful transmission, and the average time

that the channel has a collision, respectively. The prob-

ability that the channel is idle for a slot time is ð1� pbÞ, and

the probability that the channel is neither idle nor successful

for a slot time is ½1� ð1� pbÞ � ps� ¼ ðpb � psÞ. We have,

Si ¼Eðpayload transmission time in a slot time for the i classÞ

Eðlength of a slot timeÞ

¼ps;iTEðLÞ

ð1� pbÞ� þ psTs þ ½pb � ps�Tc:

ð17Þ

Let TH , TACK , SIFS, L�, and TEðL�Þ denote the time to

transmit the header (including MAC header, PHY header,

and/or tail), the time to transmit an ACK, SIFS time, the

length of the longest frame in a collision, and the time to

transmit a payload with length EðL�Þ, respectively. For the

basics access method, we have


Tbasics ¼ TH þ TEðLÞ þ SIFS þ TACK þDIFS ð18Þ

Tbasicc ¼ TH þ TEðL�Þ þDIFS: ð19Þ

Please refer to [14] for calculating TEðL�Þ. Let TRTS andTCTS denote the time to transmit an RTS frame and the timeto transmit a CTS frame, respectively. For the RTS/CTSaccess model, we have

Trts=ctss ¼ TRTS þ SIFS þ TCTS þ SIFS þ TH þ TeðLÞ

þ SIFS þ TACK þDIFS ð20ÞTrts=ctsc ¼ TRTS þDIFS: ð21Þ

3.2.3 Frame Dropping Probability

Let Pi;dropði ¼ 0; 1Þ denote the frame-dropping probabilityfor the priority i class. From Fig. 1, we observe that a framecan be dropped only in state f0; 0; 0g if a collision occurs.From Fig. 2, we observe that a frame can be dropped only instate f1; Lretry; 0g if a collision occurs. In other words, aframe can be dropped when the retransmission counterreaches the retry limit Lretry. Therefore, we have

P0;drop ¼ p0 ð22ÞP1;drop ¼ p

Lretryþ11 : ð23Þ

3.2.4 Saturation Delay

Saturation delay is the average delay under the saturationcondition, and it includes the medium access delay (due tobackoff, collisions, etc.), transmission delay, and interframespaces (such as SIFS). The average backoff delay dependson the value of a station’s backoff counter and the durationwhen the counter freezes due to others’ transmissions. LetXiði ¼ 0; 1Þ denote the random variable representing thetotal number of backoff slots, which a frame encounterswithout considering the case when the counter freezes, forthe priority i class. For class 1, the probability that the frameis successfully transmitted after the jth retry (which is theðjþ 1Þth transmission) is pj1ð1� p1Þ. The average number ofbackoff slots after the jth retry is

Xj

h¼0

ðW1;h � 1Þ=2:

Note that only successful transmissions are considered.

EðX0Þ ¼ð1� p0Þ1� p0

W0;0 � 1

2¼ W0;0 � 1

2ð24Þ

EðX1Þ ¼XLretry

j¼0

pj1ð1� p1Þ1� p

Lretryþ11

Xj

h¼0

W1;h � 1

2: ð25Þ

Let Biði ¼ 0; 1Þ denote the random variable representingthe total number of slots when the counter freezes, which aframe encounters, for the priority i class. The portion of idleslots is ð1� piÞ, which is used to decrease EðXiÞ. We have

EðBiÞ ¼EðXiÞð1� piÞ

pi: ð26Þ

We can treat EðXiÞ and EðBiÞ as the total number of idleslots and the total number of busy slots, which the frame

encounters during backoff stages, respectively. Let EðNretryÞdenote the average number of retries for the priority 1 class.The number of retries for the priority 0 class is zero. Wehave

EðNretryÞ ¼XLretry

j¼0

jpj1ð1� p1Þ1� p

Lretryþ11

: ð27Þ

Let Diði ¼ 0; 1Þ denote the random variable representingthe frame delay for the priority i class. Let T0 denote thetime that a station has to wait when its frame transmissioncollides before sensing the channel again. Let TACK timeout

and TCTS timeout denote the duration of the ACK timeoutand the duration of the CTS timeout, respectively. Note thatEðNretryÞ is one less than the number of transmissions. The

average slot lengths are �, ½Tsps=pb þ Tcðpb � psÞ=pb�, Tc þ To,and Ts for a idle slot at states fi; j; kg ðk > 0Þ, a busy slot atstates fi; j; kg ðk > 0Þ, a failed transmission slot for thisstation at states fi; j0g, and a successful transmission atstates fi; j; 0g, respectively. We have

EðD0Þ ¼ EðX0Þ� þ EðB0Þpspb

Ts þpb � ps

pbTc

� �þ Ts ð28Þ

EðD1Þ ¼ EðX1Þ� þ EðB1Þpspb

Ts þpb � ps

pbTc

� �

þEðNreryÞðTc þ ToÞ þ Ts ð29ÞTbasico ¼ SIFS þ TACK timeout ð30Þ

Trts=ctso ¼ SIFS þ TCTS timeout: ð31Þ

3.3 Results

In this section, we first validate the analytical model withsimulation results. Then, we study the proposed priorityscheme for delay sensitive real-time transmissions and best-effort transmissions.

We use IEEE 802.11a as an example. IEEE 802.11aparameters can be found in [14], [16], as well as how tocalculate TH þ TEðLÞ accurately [14]. The data rate is 6Mbpsand the control rate is 6Mbps. The frame size is fixed as1,024 bytes. As assumed in the analytical model, stationsalways have frames ready to send.

3.3.1 Simulation Validations

In this section, we conduct simulations to validate theproposed analytic model. The IEEE 802.11a simulationmodels had been developed based on IEEE 802.11astandard [16] and OPNET Wireless LAN simulation model8.0A (for IEEE 802.11b DCF). Furthermore, we adapt oursimulation model with similar assumptions as those in the

analytical model. The data rate is 6Mbps and the controlrate is 6Mbps. The frame size is fixed as 1,024 bytes. All thesimulation results have over 95 percent confidential inter-vals. The simulation executing time is long: about 2 hoursper run in order to get long term results, and multiple runsper case to calculate the confidential intervals.

Table 1 shows the simulation results vs. numerical results

when adopting the following parameters: Lretry ¼ 7, W0;0 ¼16, W1;0 ¼ 64, n0 ¼ n1, and CWmax;BE ¼ 1; 024. As illustratedin the table, analytical results and simulation results match


pretty well for both saturation throughput and saturation

delay.

3.3.2 Service Differentiation

Fig. 3 has the following parameters: Lretry ¼ 7, W0;0 ¼ 16,

W1;0 ¼ 64, and CWmax;BE ¼ 1; 024. Fig. 3a shows saturation

throughputs for the real-time class and the best-effort class

over the number of stations for both classes ðn0 ¼ n1Þ. The

real-time traffic has a much better saturation throughput

than the best-effort traffic. When the number of stations for

both classes ðn0 ¼ n1Þ changes from 5 to 45 (the total

number changes from 10 to 90), the saturation throughput

of the real-time traffic changes from 0.655333 to 0.394999,

and the saturation throughput of the best-effort traffic

changes from 0.071036 to 0.013877.Fig. 3b shows saturation delays for the real-time class and

the best-effort class over the number of stations for both

classes ðn0 ¼ n1Þ. The real-time traffic has a much less

saturation delay than the best-effort traffic. As the number

of active stations increases, the real-time traffic has a relatively

constant delay, whereas the best-effort traffic has a very high

delay. When the number of stations for both classes ðn0 ¼ n1Þchanges from 5 to 45 (the total number changes from 10 to 90),

the saturation delay of the real-time traffic changes from

3672.392221 � seconds to 28830.233173 � seconds, and the

saturation delay of the best-effort traffic changes from39284.808428 � seconds to 2061107.845170 � seconds.

Fig. 3c shows frame-dropping probabilities for the real-time class and the best-effort class over the number of

stations for both classes ðn0 ¼ n1Þ. The real-time traffic has amuch worse frame-dropping probability than the best-efforttraffic. When the number of stations for both classes ðn0 ¼n1Þ changes from 5 to 45 (the total number changes from 10to 90), the frame-dropping probability of the real-timetraffic changes from 0.325103 to 0.760000, and the frame-dropping probability of the best-effort traffic changes from0.000389 to 0.123574. We observe that the frame droppingprobability is high for the real-time traffic. We will showhow to control the frame dropping probability in a later

section.Fig. 3 also indicates that the proposed priority is very

effective in terms of service differentiation, and the RT classhas a better saturation throughput and delay, however theRT class scarifies a very high frame-dropping probability.

Fig. 4 has following parameters: n0 ¼ n1 ¼ 5, Lretry ¼ 7,W1;0 ¼ 200, and CWmax;BE ¼ 1; 024. Fig. 4a shows saturationthroughputs for the real-time class and the best-effort classover the initial window size of the RT class, W0;0, which

changes from 12 to 164. The real-time traffic has a muchbetter saturation throughput than the best-effort traffic. Aninteresting observation is that as the initial window size ofthe RT class increases, the throughput of the RT classincreases first and then decreases. The reason is that whenthe initial window size is small, the collision ratio is high sothat the throughput is low; as the initial window sizeincreases, the collision ratio goes down so that thethroughput increases; however, as the initial windowfurther increases, the throughput decreases since it takes

longer time to transmit. Another interesting observation isthat the total throughput increases a little! The reason maybe that the overall system performance can sometimes beimproved by using a contention protocol that assignsdifferent access probabilities to different stations [17].


TABLE 1Simulation Results versus Analytical Resutls Legend:ST (Saturation Throughput); SD (Saturation Delay)

(� Seconds); S (Simulation), A (Analytical)

Fig. 3. Throughput, delay (� seconds), and frame dropping probably

(FDP) versus number of active stations of both classes.


(FDP) versus the initial window size of RT.

Fig. 4b shows saturation delays for the real-time classand the best-effort class over the initial window size of theRT class, W0;0, which changes from 12 to 164. The real-timetraffic has a much less saturation delay than the best-efforttraffic. As the initial window size of the RT class increases,the delay of the BE class decreases a lot, whereas thesaturation delay for the real-time class remains almost thesame. In other words, the best-effort traffic is more sensitiveto W0;0.

Fig. 4c shows frame-dropping probability for the real-time class and the best-effort class over the initial windowsize of the RT class, W0;0, which changes from 12 to 164. Thereal-time traffic has a much higher frame-dropping prob-ability than the best-effort traffic. We observe that the framedropping probability for the best-effort class is very smallsince its initial window size is large, i.e., 200.

Fig. 5 has following parameters: W0;0 ¼ 16, CWmax;BE

¼ 1; 024,W1;0 ¼ 64,n0 ¼ n1 ¼ 10, andLretry ¼ 2; 3; 4; 5; 6; 7; 8;9; 10; 11; 12. Fig. 5a shows saturation throughputs for the real-time class and the best-effort class over the retry limit of thebest-effort class, Lretry. The real-time traffic has a much bettersaturation throughput than the best-effort traffic. As illu-strated in the figure, as the retry limit of the BE class increases,the saturation throughput of the BE class decreases since alarger retry limit indicates a larger backoff window size andlonger delay to access the channel. On the other hand, thesaturation throughput of the RT increases since the prob-ability of collisions decreases as the retry limit of the BE classincreases. We also observe that the throughput of the RT (BE)class increases (decreases) more when the retry limit is small.The reason is that when the retry limit is large enough, thenumber of frames that needs a larger number of retries issmall. Therefore, although the retry limit becomes larger, theeffect of frames needing a larger number of retries becomesless important.

Fig. 5b shows saturation delays for the real-time classand the best-effort class over the retry limit of the best-effortclass, Lretry. The real-time traffic has a much less saturationdelay than the best-effort traffic. As illustrated in the figure,

the saturation delay of the best-effort class increases as Lretry

increases.Fig. 5c shows frame-dropping probabilities for the real-

time class and the best-effort class over the retry limit of thebest-effort class, Lretry. The real-time traffic has a largerframe-dropping probability than the best-effort traffic. Asillustrated in the figure, both probabilities of the BE classand the RT class decrease as the BE class increases Lretry.The reason that the frame-dropping probability of the RTclass decreases little since the collision probability decreasesas the retry limit of the BE class increases. On the otherhand, the frame-dropping probability of the BE classdecrease much more and goes near zero when the retrylimit is near 12.

4 QOS GUARANTEE

In this section, we study some QoS guarantee mechanismsfor the real-time traffic. In Section 4.1, we propose a simplescheme to provide an upper bound on frame droppingprobability; in Section 4.2, we propose an admission controlscheme to provide guaranteed throughput and/or delay. InSection 4.3, we propose an algorithm to guarantee through-put, delay, and/or frame dropping probability. All thealgorithms in this section do not consider the throughputoptimization problem, which will be discussed in the nextsection. The methods proposed in this section can beapplied to both the real-time traffic and the best-efforttraffic, and we adopt the real-time traffic as an example.

4.1 An Upper Bound on the Frame DroppingProbability

One concern for the RT traffic may be that the framedropping probability may be very large since frames aretransmitted only one time, and dropped if failed. In thissection, we solve the issue, and we introduce a QoSparameter, i.e., a predefined upper bound for the framedropping probability, PQoS . Our goal is to use PQoS boundsthe frame dropping probability. We have

P0;drop ¼ 1� ð1� �0Þn0�1 � pQoS: ð32Þ

Plugging (22) and (11) into (32), we have

W0;0 �ðn0�1Þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1� pQoS

p2ð1� p0Þ

1�ðn0�1Þffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1� pQoS

p þ 1: ð33Þ

The above equation indicates that in order to guarantee(32), the initial window size should be chosen in such a waythat (33) holds. In other words, the initial window sizeshould be large enough, based on (33). Intuitively (alsoconfirmed in results), the frame dropping probabilitydecreases as the initial window size increases. Therefore,we can always choose the initial window size large enoughso that (32) holds. However, a too large value of the initialwindow size will degrade throughput and delay. We willdiscuss these issues in the next subsection. In this section,we only consider how to put a bound on the framedropping probability. To be realistic, (33) does not reallygive us an solution since (33) depends on both n0 and p0,although it provides us some intuitions. The authors in [15]address very well on how to obtain the value of n0.



(FDP) versus Lretry.

Therefore, we will adopt the same approach in [15] toobtain the value of n0. However, in (33), p0 also depends onthe initial window size. Therefore, we propose the follow-ing algorithm for choosing the initial window size:

1. Estimate the value of n0.2. Calculate the frame dropping probability with this

model; if the frame dropping probability satisfies(32), go to step 4; otherwise, go to step 3.

3. Increase the initial window size by 1, go to step 2.4. For all the stations, the initial window sizes are set to

the obtained value via following ways. If an AP ispresent, all the calculations from 1 to 3 are done bythe AP, and the AP distributes the value of the newinitial window size in a beacon frame to otherstations. If there is no AP, the station that isresponsible of sending beacon frames performscalculations from 1 to 3, and distributes the valueof the new initial window size in a beacon frame toother stations.

Next, we perform some experiments. Let PQoS be the pre-defined upper bound for the frame dropping probability.

We adopt PQoS as 0.5, 0.3, 0.1, and 0.05 for differentexperiments. Fig. 6a shows the obtained (required) initialwindow size versus number of stations under different QoSrequirements, i.e., upper bounds on the frame dropping

probability with PQoS . We observe that as the number ofactive stations increases, the required initial window sizeincreases to guarantee the PQoS bound. As the PQoS

decreases, the required initial window size increases.Fig. 6b shows that frame dropping probability is bounded

by PQoS . Fig. 6 indicates that the QoS requirements areguaranteed, and a stringent QoS requirement needs a largerinitial window size. Therefore, frame dropping probabilityfor the real-time traffic can be controlled with a reasonablevalue.

4.2 Admission Control to Guarantee Throughputand Delay

In the previous section, we propose a method to guaranteethe frame-dropping probability by increasing the initial

window size. However, increasing the initial window sizewill degrade throughput and delay. In this section, wepropose an admission control mechanism to guaranteethroughput and delay based on the fact that throughputand delay all degrade if the number of active stationsincreases.

We refer the AP if it is present or the station in charge ofsending beacon frames if no AP is present as to theadmission control coordinator (ACC). Any station mustobtain permission from the ACC before transmitting real-time traffic flow. The procedure works as follows:

1. A station sends a request for real-time transmissionto the ACC with a required throughput and/or arequired delay (the station knows the ACC addressvia beacon frames).

2. The ACC predicts the performance if the station isallowed to transmit the real-time traffic flow:

. The ACC estimates the value of n0.

. The ACC calculates throughput and delay basedon the model in the previous section with n0 þ 1.

. If the throughput (divided by n0 þ 1) and/ordelay satisfy the requirements, the ACC decidesto accept the request; otherwise the ACCdecides to reject the request.

3. The ACC notifies the station the decision.4. The station begins transmissions if the decision is

acceptance.

In the above algorithm, the value of n0, in fact, is not thenumber of real-time flows, but the equivalent number ofactive stations in term of saturation status. For example, thenumber of real-time flows is 10, but n0 may equal to 4. Inother words, the number of real-time flows is not smallerthan n0. In this sense, the above algorithm is veryconservative by using n0 þ 1 since the traffic of a saturationstation is normally larger than a real-time flow, however,based on our numerical results and simulation results, theabove algorithm works very well. We believe the followingphilosophy, “a good approach is always an approach of notpushing too hard.” Note that in this admission controlscheme, when a real-time flow finishes, admission controldoes not perform additional actions.

We perform experiments for the above algorithm. Sincewe are interested in the real-time traffic, we try to guaranteedelay requirement in these experiments. Let the requireddelay change from 40ms to 400ms for different experiments.Fig. 7 shows the number of accepted stations versus therequired delay under different initial window sizes. Weobserve that when the required delay is loosen (the requireddelay increases), more stations are accepted. For a fixedrequired delay, the scheme with a larger initial window sizecan accommodate more stations.

4.3 Admission Control for GuaranteedFrame-Dropping Probability, Throughput,and Delay Requirements

The approaches proposed in Section 4.1 and Section 4.2seem to be conflicting. However, based on the fact thatframe-dropping probability increases if the number ofactive stations increases, we can further enhance theapproach in Section 4.2 as follows:


Fig. 6. Obtained initial window size and frame dropping probability (FDP)

under different QoS requirements.

When the station sends the request to the ACC, a requiredframe-dropping probability (upper bound) is also included.In Step 2, the ACC also check the frame-dropping probabilitywhen making a decision of acceptance/rejection.

Fig. 8a shows the obtained initial window size range [lowbound, upper bound] versus number of active stations whenthe required delay is 4ms and PQoS ¼ 0:1. Within the range,QoS requirements are guaranteed. The range becomessmaller as the number of active stations increases. The rangeis also called admission control region. Fig. 8b shows theobtained initial window size range [low bound, upperbound] versus number of active stations when the requireddelay is 8ms and PQoS ¼ 0:2. Compared with Fig. 8a, Fig. 8bhas loose QoS requirements, and therefore the range is muchlarger.

5 OPTIMIZATION WITH ADAPTATION AND

QOS GUARANTEE

In the previous section, we discuss how to guarantee someQoS parameters, i.e., required throughput, required delay,and/or an upper bound of frame dropping probability. Inthis section, we will study how to optimize the throughputand to guarantee the above QoS parameters via adaptationof initial window size. In Section 5.1, we study how tooptimize the throughput without QoS considerations. InSection 5.2, we study how to optimize the throughput withQoS considerations. The methods proposed in this sectioncan be applied to both the real-time traffic and the best-effort traffic, and we adopt the real-time traffic as anexample.

5.1 An Adaptive Backoff Window Scheme

According to our numerical results and simulation results,we observe that for fixed number active stations or fixedamount of traffic load, when the initial window size of allstations is small, the throughput is small since there are a lotof collisions; as the initial window size increases, thethroughput increases until reaching a maximum value; andthen as the initial window size further increases, thethroughput decreases. In other words, there is a maximumthroughput for a fixed traffic load. In this section, our goalis to optimize the throughout via an adaptive backoffwindow scheme.

By observations from the previous section, the optimiza-tion problem, i.e., to optimize the throughput with the

initial window size, is to solve a group of nonlinear

equations. It is difficult to obtain the optimal value. Let

½CWleft; CWright� be the range of the initial window size,

where CWleft ¼ 2 and CWright is chosen relatively large so

that the optimal value falls in ½CWleft; CWright�. One

approach is to find the optimal value via enumerating

finite values as follows: Calculate and compare all the

throughputs for all the integer initial window sizes in

½CWleft; CWright� to obtain the optimal initial window size.

Another approach is to use the binary search approach as

follows:

1. The ACC estimates the value of n.2. Let CWleft ¼ 2 and CWright be a very large number.

Denote the throughput (T ) as a function initialwindow size: T ¼ T ðCWminÞ calculated using themethod in the previous section.

3. Letmleft¼dCWleftþðCWright�CWleftÞ=4c andmright¼dCWright � ðCWright � CWleftÞ=4c. Calculate and com-pare T ðCWleftÞ, T ðmleftÞ, T ðmrightÞ, and T ðCWrightÞusing the method in the previous section.

4. If the largest value is T ðmleftÞ, let CWright ¼ mright.5. Else if the largest value is T ðmrightÞ, let CWright ¼

mright.6. Else if the largest value is T ðCWrightÞ, let CWleft ¼

mright.7. Else CWright ¼ mleft.8. If jCWleft� CWrightj � 1, go to 9; otherwise go to 3.9. If T ðCWrightÞ < T ðCWleftÞ, the optimal value is

CWleft; otherwise, the optimal value is CWright.10. The ACC distributes the optimal value to all the

stations via a beacon frame.

Fig. 9 shows optimal initial window sizes vs. number of

stations. The optimal initial window size increases as the

number of active stations increases. The optimal initial

window size is 496 when the number of active stations is 30.Fig. 10 compares throughputs of the adaptive scheme

and the fixed scheme. In the fixed scheme, the initialwindow sizes are 76, 160, 244, 328, 412, and 496. As


Fig. 7. Number of accepted stations versus required delay.

Fig. 8. Accepted initial window size interval versus number of stations.

(a) Required delay = 4ms, and PQoS ¼ 0:1. (b) Required delay = 8ms,

and PQoS ¼ 0:2.

illustrated in the figure, the adaptive scheme achieves theoptimal throughput that bounds the fixed scheme.

5.2 An Adaptive Backoff Window Scheme withQoS Constraints

In the previous section, an adaptive scheme is proposedwithout QoS considerations. In this section, we also need toconsider QoS guarantee. Let ½CWleft; CWright� be the rangeof the initial window size, where CWleft ¼ 2 and CWright ischosen relatively large. Basically, it needs two steps to findthe optimal initial window size with QoS guarantee, i.e.,throughput, delay, and/or frame-dropping probability. Thefirst step is to find a set (denoted as A) of initial windowsizes in which required throughput, required delay, and/orrequired frame-dropping probability can be guaranteed,where the set A is a subset of ½CWleft; CWright�. The secondstep is to find an optimal initial window size with themaximum throughput in the set A. If the set A is the emptyset, it means that the optimal window size does not exist,and the solution is either to loosen QoS requirements or tostop some stations. The latter way is not practical. If theapproach is used for admission control, the request will berejected. The approach is listed as follows. Step 1: 1) TheACC estimates the value of n. 2) Try every integer pointwithin ½CWleft; CWright� as the initial window size to seewhether QoS constraints are satisfied. Collect all points

satisfying QoS requirements to form the set A. Step 2: Findthe optimal initial window size in the set A with themaximum throughput.

If the above approach is used to be admission control, theACC will notify the requested station. One advantage of theproposed admission control is adaptation: the initialwindow size is adaptively changed in order to accommo-date more real-time traffic flows.

Fig. 11 shows the results of optimization via adaptationwith QoS constraints, i.e., the required delay is 4ms andPQoS ¼ 0:1. As illustrated in the figure, the optimal initialwindow size increases as the number of active stationsincreases. The throughput is pretty high, about 81.4 percent to81.9 percent. The delay is bounded by the QOS requirement,i.e., 4ms. The frame-dropping probability is bounded by 0.1.Therefore, QoS requirements are guaranteed.

6 CONCLUSION

In this paper, we have studied service differentiation, QoSguarantee, and optimization issues for the real-time traffic.

In service differentiation, we have proposed a simplepriority scheme for real-time traffic with a smaller initialwindow size and no retransmission. An analytical model isproposed to evaluate system performance. Simulations areconducted to validate the analytical results. Our studiesshow:

. The proposed scheme can provide a good servicedifferentiation.

. Smaller window sizes might cause lower throughoutand higher delay.

. The frame dropping probability for the best-effortclass is very small, and the frame dropping prob-ability for the real-time class is relatively large, butwith some proper mechanisms, it can be controlledin a reasonable range.

. Both the initial window size and the retry limit aregood parameters to differentiate classes. If one of theparameters in one class increases, the collisionprobability will decrease, and therefore, the perfor-mance will increase for another class. The totalthroughput may increase a little.


Fig. 9. Optimal initial window size versus number of stations.

Fig. 10. Comparison between the adaptive scheme and static scheme.

Fig. 11. Optimal throughput with QoS constraints (required delay = 4ms,

and PQoS ¼ 0:1).

. Both frame-dropping probabilities of the BE class andthe RT class decrease as the BE class increases the retrylimit. The frame-dropping probability of the RT classdecreases little whereas the frame-dropping prob-ability of the BE class decreases much more and goesnear zero when the retry limit is near 12.

. Simulation results match pretty well with theanalytical results.

We have proposed and studied some QoS guarantee

mechanisms for the real-time traffic:

. We have proposed a simple scheme to provide anupper bound on frame dropping probability. Ourresults show that the QoS requirements (a bound onthe frame dropping probability) can be guaranteed,and a stringent QoS requirement needs a largerinitial window size. Therefore, frame droppingprobability for the real-time traffic can be controlledwith a reasonable value.

. We have proposed an admission control scheme toprovide guaranteed throughput and/or delay. Wehave observed that when the required delay isloosen (the required delay increases), more stationsare accepted. For a fixed required delay, the schemewith a larger initial window size can accommodatemore stations.

. We have proposed an algorithm to guaranteethroughput, delay, and/or frame dropping prob-ability. We have obtained the admission controlregion, in which QoS can be guaranteed.

We have further studied how to optimize the throughput

with and without guaranteeing QoS parameters via adapta-

tion of initial window size:

. We have studied how to optimize the throughputwithout QoS considerations. The optimal initialwindow size increases as the number of activestations increases. The optimal initial window size is496 when the number of active stations is 30. Theadaptive scheme achieves the optimal throughputthat bounds the fixed scheme.

. We have studied how to optimize the throughputwith QoS considerations. Results show that QoSrequirements are guaranteed and throughput isoptimized.

ACKNOWLEDGMENTS

The authors are grateful to the three referees for theircareful reading and suggestions which have greatlyimproved the readability of the paper, and to the associateeditor, Professor Wei Zhao, for timely and professionalhandling of our manuscript. Yi Pan’s research wassupported in part by the National Natural Science Founda-tion of China (NSFC) under Grant No. 60440420451 (“twobase” project).

REFERENCES

[1] S. Chakrabarti and A. Mishra, “QoS Issues in Ad Hoc WirelessNetworks,” IEEE Comm. Magazine, pp. 142-148, Feb. 2001.

[2] J. Wu, “Extended Dominating-Set-Based Routing in Ad HocWireless Networks with Unidirectional Links,” IEEE Trans.Parallel and Distributed Systems, vol. 13, no. 9, pp. 866-881, Sept.2002.

[3] Z. Cai, M. Lu, and X. Wang, “Distributed Initialization Algorithmsfor Single-Hop Ad Hoc Networks with Minislotted CarrierSensing,” IEEE Trans. Parallel and Distributed Systems, vol. 14,no. 5, pp. 512-528, May 2003.

[4] X. Li, G. Calinescu, P. Wan, and Y. Wang, “Localized DelaunayTriangulation with Application in Ad Hoc Wireless Networks,”IEEE Trans. Parallel and Distributed Systems, vol. 14, no. 10,pp. 1035-1047, Oct. 2000.

[5] K. Alzoubi, X. Li, Y. Wang, P. Wan, and O. Frieder, “GeometricSpanners for Wireless Ad Hoc Networks,” IEEE Trans. Parallel andDistributed Systems, vol. 14, no. 4, pp. 408-421, Apr. 2003.

[6] A. El-Kadi, S. Olariu, and H. Abdel-Wahab, “A Rate-BasedBorrowing Scheme for QoS Provisioning in Multimedia WirelessNetworks,” IEEE Trans. Parallel and Distributed Systems, pp. 156-166, Feb. 2002.

[7] A. Malla, M. El-Kadi, S. Olariu, and P. Todorova, “A Fair ResourceAllocation Protocol for Multimedia Wireless Networks,” IEEETrans. Parallel and Distributed Systems, vol. 14, no. 1, pp. 63-71, Jan.2003.

[8] J. Hou, J. Yang, and S. Papavassiliou, “Integration of Pricing withCall Admission Control to Meet QoS Requirements in CellularNetworks,” IEEE Trans. Parallel and Distributed Systems, pp. 898-910, Sept. 2002.

[9] Y. Kwok, V. Lau, and K.N. Lau, “A Novel Channel-AdaptiveUplink Access Control Protocol for Nomadic Computing,” IEEETrans. Parallel and Distributed Systems, pp. 1150-1165, Nov. 2002.

[10] M.A. Hiltunen, R.D. Schlichting, X. Han, M.M. Cardozo, and R.Das, “Real-Time Dependable Channels: Customizing QoS Attri-butes for Distributed Systems,” IEEE Trans. Parallel and DistributedSystems, pp. 600-612, June 1999.

[11] IEEE 802.11: Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specification, 1999.

[12] D.-J. Deng and R.-S. Chang, “A Priority Scheme for IEEE 802. 11DCF Access Method,” IEICE Trans. Comm., vol. E82-B, no. 1,pp. 96-102, Jan. 1999.

[13] G. Bianchi, “Performance Analysis of the IEEE 802. 11 DistributedCoordination Function,” IEEE J. Selected Areas in Comm., vol. 18,no. 3, pp. 535-547, Mar. 2000.

[14] Y. Xiao and J. Rosdahl, “Throughput and Delay Limits of IEEE802.11,” IEEE Comm. Letters, vol. 6, no. 8, pp. 355-357, Aug. 2002.

[15] G. Bianchi and I. Tinnirello, “Kalman Filter Estimation of theNumber of Competing Terminals in an IEEE 802. 11 Network,”Proc. IEEE INFOCOM ’03, 2003.

[16] IEEE 802.11a: Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specification: High-speed Physical Layer inthe 5GHz Band, Sept. 1999.

[17] A.S. Tanenbaum, Computer Networks, fourth ed. Prentice Hall,2002.

[18] G. Bianchi and Y. Xiao, “Modeling Saturation Performance of theIEEE 802. 11 MAC,” Wireless Comm. and Mobile Computing(WCMC) J., Sept. 2003.

[19] G.-S. Ahn, A.T. Campbell, A. Veres, and L.-H. Sun, “SupportingService Differentiation for Real-Time and Best Effort Traffic inStateless Wireless Ad Hoc Networks (SWAN),” IEEE Trans. MobileComputing, vol. 1, no. 3, pp. 192-207, July 2002.

[20] F. Calı̀, M. Conti, and E. Gregori, “Dynamic Tuning of the IEEE802. 11 Protocol to Achieve a Theoretical Throughput Limit,”IEEE/ACM Trans. Networking, vol. 8, no. 6, pp 785-799, Dec. 2000.

[21] I. Aad and C. Castelluccia, “Differentiation Mechanisms for IEEE802. 11,” Proc. IEEE INFOCOM ’01, 2001.

[22] X. Pallot and L.E. Miller, “Implementing Message Priority Policiesover an 802. 11 Based Mobile Ad Hoc Network,” Proc. IEEEMILCOM ’01, 2001.

[23] X. Yang and N. Vaidya, “Priority Scheduling in Wireless Ad HocNetworks,” Proc. MobiHoc ’02, 2002.

[24] IEEE 802.11e: Wireless Medium Access Control (MAC) andPhysical Layer (PHY) Specifications: Medium Access Control(MAC) Enhancements for Quality of Service (QoS), IEEE 802.11e/D2.0, Nov. 2001.

[25] J.-P. Sheu, C.-H. Liu, S.-L. Wu, and Y.-C. Tseng, “A Priority MACProtocol to Support Real-Time Multimedia Traffic in Ad HocNetworks,” Wireless Networks, pp. 61-69, Jan. 2004.

[26] Y. Ge and C.-J. Hou, “An Analytical Model for ServiceDifferentiation in IEEE 802. 11,” Proc. IEEE Int’l Conf. Comm. ’03,2003.


Yang Xiao received the PhD degree in compu-ter science and engineering from Wright StateUniversity, Dayton. He had been a softwareengineer, a senior software engineer, and atechnical lead working in the computer industryfor five years in the early 1990s. He worked atMicro Linear-Salt Lake City Design Center as anMAC architect involving the IEEE 802.11 (Wire-less LAN) standard enhancement work beforehe joined The University of Memphis as an

assistant professor of computer science. Dr. Xiao is a voting member ofthe IEEE 802.11 Working Group, a senior member of the IEEE and theIEEE Computer Society, and a member of the ACM. He is an associateeditor of EURASIP Journal on Wireless Communications and Network-ing, and he currently serves on the editorial boards of the (Wiley) Journalof Wireless Communications and Mobile Computing, the InternationalJournal of Wireless and Mobile Computing, and the International Journalof Signal Processing. He serves a lead guest editor for the (Wiley)Journal of Wireless Communications and Mobile Computing, specialissue on mobility, paging, and quality of service management for futurewireless networks in 2004, a lead guest editor for the InternationalJournal of Wireless and Mobile Computing, special issue on Mmdiumaccess control for WLANs, WPANs, ad hoc networks, and sensornetworks in 2004, and an associate guest editor for the InternationalJournal of High Performance Computing and Networking, special issueon parallel and distributed computing, applications and technologies in2003. He serves as a symposium cochair for the Symposium on DataBase Management in Wireless Network Environments in IEEEVTC’2003 Fall. He serves as a TPC member for many conferencessuch as ICC, GLOBECOM, ICDCS, WCNC, ICCCN, PIMRC, WMASH,etc. Dr. Xiao’s current research interests include wireless local areanetworks, wireless personal area networks, and mobile cellular net-works. He has published many papers in major journals and refereedconference proceedings related to these research areas, such as theIEEE Transactions on Mobile Computing, IEEE Transactions onWireless Communications, IEEE Transactions on Parallel and Distrib-uted Systems, IEEE Transactions on Vehicular Technology, IEEECommunications Letters, IEEE Communications Magazine, IEEEWireless Communications, ACM/Kluwer MONET, etc.

Yi Pan received the BEng and MEng degrees incomputer engineering from Tsinghua University,China, in 1982 and 1984, respectively, and thePhD degree in computer science from theUniversity of Pittsburgh in 1991. Currently, he isa chair and a full professor in the Department ofComputer Science at Georgia State University.Dr. Pan’s research interests include parallel anddistributed computing, optical networks, wirelessnetworks, and bioinformatics. Dr. Pan has

published more than 80 journal papers with 26 papers published invarious IEEE journals. In addition, he has published more than 90 papersin refereed conferences (including IPDPS, ICPP, ICDCS, INFOCOM, andGLOBECOM). He has also coedited 13 books (including proceedings)and contributed several book chapters. His pioneer work on computingusing reconfigurable optical buses has inspired extensive subsequentwork by many researchers, and his research results have been cited bymore than 100 researchers worldwide in books, theses, journal andconference papers. He is a coinventor of three US patents (pending) andfive provisional patents, and has received many awards from agenciessuch as the US National Science Foundation, AFOSR, JSPS, IISF, andMellon Foundation. His recent research has been supported by the USNational Science Foundation, NIH, AFOSR, AFRL, JSPS, IISF, and thestates of Georgia and Ohio. He has served as a reviewer/panelist formany research foundations/agencies such as the US National ScienceFoundation, the Natural Sciences and Engineering Research Council ofCanada, the Australian Research Council, and the Hong Kong ResearchGrants Council. Dr. Pan has served as an editor-in-chief or editorial boardmember for eight journals, including three IEEE Transactions and a guesteditor for seven special issues. He has organized several internationalconferences and workshops and has also served as a programcommittee member for several major international conferences such asINFOCOM, GLOBECOM, ICC, IPDPS, and ICPP. Dr. Pan has deliveredmore than 50 invited talks, including keynote speeches and colloquiumtalks, at conferences and universities worldwide. Dr. Pan is an IEEEDistinguished Speaker (2000-2002), a Yamacraw Distinguished Speaker(2002), a Shell Oil Colloquium Speaker (2002), and a senior member ofthe IEEE and the IEEE Computer Society. He is listed in Men ofAchievement, Who’s Who in Midwest, Who’s Who in America, Who’sWho in American Education, Who’s Who in Computational Science andEngineering, and Who’s Who of Asian American.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

538 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED …yzchen/papers/papers/newpapers/qos_1hop.pdf · Many issues are involved to provide quality of service (QoS) for ad hoc networks,