Computer Networks 56 (2012) 762–779

Contents lists available at SciVerse ScienceDirect

Computer Networks

journal homepage: www.elsevier .com/locate /comnet

Cross-layer opportunistic adaptation for voice over ad hoc networks

Suhaib A. Obeidat a, Abraham N. Aldaco b, Violet R. Syrotiuk b,⇑a Bennett College for Women, Greensboro, NC 27401, United Statesb Arizona State University, Tempe, AZ 85287-8809, United States

a r t i c l e i n f o a b s t r a c t

Article history:Received 8 June 2011Received in revised form 3 November 2011Accepted 6 November 2011Available online 13 November 2011

Keywords:Cross-layer designOpportunistic protocolAdaptationVoiceAd hoc networksPerformance

1389-1286/$ - see front matter � 2011 Elsevier B.Vdoi:10.1016/j.comnet.2011.11.002

⇑ Corresponding author.E-mail addresses: sobeidat@bennett.edu (S.A.

asu.edu (A.N. Aldaco), syrotiuk@asu.edu (V.R. Syroti

The support of voice communication is fundamental in the deployment of an ad hoc net-work for the battlefield or emergency response. We use the QoS requirements of voiceto identify factors influencing its communication, and validate their significance throughstatistical analysis. Based on the results, we propose an opportunistic protocol within across-layer framework that adapts these factors at different time scales. Hop-by-hop adap-tation exploits the PHY/MAC interaction to improve the use of the spectral resourcesthrough opportunistic rate-control and packet bursts, while end-to-end adaptationexploits the LLC/application interaction to control the demand per call through voice cod-ing and packet size selection. Our objective is to maximize the number of calls admittedwhile minimizing loss of quality. We evaluate the performance of the protocol in simula-tion with real audio traces using both quantitative and mean opinion score (MOS) audioquality metrics, comparing to several standard voice codecs. The results indicate that: (i)compression and packet-size selection play a critical role in supporting QoS over ad hocnetworks; (ii) header compression is needed to limit the overhead per packet especiallyover longer paths; (iii) good voice quality is achieved even in strenuous networkconditions.

� 2011 Elsevier B.V. All rights reserved.

1. Introduction therefore, our focus is on call admittance and survival with

Voice over IP (VoIP) is one of the fastest growing appli-cations in networking [1]. The rate at which wireless accesspoints are spreading only increases the importance of VoIPover wireless [2]. Supporting voice over ad hoc networks ispart of realizing an all-IP goal.

The wireless channel introduces many challenges forsupporting voice. These include the inherent broadcastnature of the channel, temporal response variability dueto fading and absorption, and sensitivity to noise and inter-ference. Ad hoc networks also suffer from a scarcity ofresources and a lack of centralized control. When com-bined, these challenges make supporting voice in thesenetworks a formidable task. Our interest is in supportingvoice in the battlefield, or in emergency situations;

. All rights reserved.

Obeidat), aaldacog@uk).

acceptable quality as opposed to providing the quality wehave come to expect in wireline telephony.

Experience in cellular networks has shown thatadaptive applications are resilient and robust [3–5]. Inaddition, cross-layer design, where performance gains areaccomplished through exploiting the dependence betweenprotocol layers, gives better performance compared to tra-ditional approaches [6]. However, increasing the numberof layers involved in a cross-layer design does not alwaystranslate into better performance. Kawadia and Kumarshow that, if not used carefully, unintended cross-layerinteractions may have undesirable consequences on over-all system performance [7].

Combining the merits of both adaptation and cross-layerdesign, while cognizant of the care required, we propose anopportunistic adaptive protocol within a cross-layer frame-work for supporting VoIP over ad hoc networks. Weincorporate three of the seven approaches to cross-layerdesign identified by Srivastava and Motani [8]: explicit

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 763

notification from one layer to another, directly setting aparameter of a different layer, and vertical calibrationacross different layers of the protocol stack.

We tackle the time-variant channel quality and capacityby introducing adaptive modulation to maximize channelutilization. We also minimize the amount of real-time traf-fic introduced in the network by using adaptive voice com-pression. A side effect of using adaptive compression is toalso vary the audio packet size used.

Adaptation of three factors, namely modulation, com-pression, and packet size, requires collaboration of threelayers of the protocol stack: the physical, link, and applica-tion layers. In terms of time scale, adaptation of modulationoccurs on a hop-by-hop basis as channel quality varies fromone hop to another and occurs at a fast pace. Adaptation ofcompression and packet size, on the other hand, occur on anend-to-end basis as this depends on the path quality andtherefore occurs on a longer time scale. Having the protocolwork at two different time scales combines the benefits ofhaving an accurate picture of both local and end-to-endconditions, and reduces protocol overhead.

This paper makes the following contributions:

� A cross-layer architecture for voice over ad hoc net-works is presented that combines the use of modulation,compression, and packet size spanning three layers ofthe protocol stack: physical, link, and application.� An adaptive protocol is proposed that operates at two

time scales, on a hop-by-hop basis and an end-to-endbasis, capturing local channel quality and end-to-endnetwork statistics, respectively.� A high fidelity simulation model is used that includes

the simulation of packetization delay and physical layerdetails, playout buffers, among others.� Both quantitative and mean opinion score (MOS) audio

quality metrics are evaluated using real audio traces,with comparisons to several standard voice codecs.

The rest of this paper is organized as follows. We iden-tify the factors whose adaptation is important in providingacceptable voice quality in Section 2. Using the selected fac-tors, we propose an opportunistic adaptive protocol in Sec-tion 3. In Section 4 we describe the simulation set-up, anddefine the quantitative degradation in voice quality (DVQ)and the qualitative subjective mean opinion score (MOS)performance metrics. Through simulation with real audiotraces the performance of our protocol is evaluated for bothstatic topologies and mobile scenarios in Section 5 compar-ing to non-adaptive protocols using standard voice codecs.An analysis bounding the maximum voice capacity for ourprotocol is presented in Section 6. In Section 7 we overviewrelated work and contrast our contributions. Finally, weconclude and propose future work in Section 8.

2. Factors influencing voice

The quality-of-service (QoS) requirements of voice are:

(1) A 0–150 ms end-to-end delay is acceptable for mostapplications [9].

(2) Voice can tolerate a packet loss on the order of 10�2–10�4 [10].

(3) Delay variations of less than 75 ms give good quality[11].

End-to-end delay is the time from when a frame isgenerated at the caller until it is played at the callee. Thereare five components to end-to-end delay: (1) Packetizationdelay is the delay at the caller to collect all bits that com-pose a packet. (2) Queuing delay is the time a packet spendswaiting to be forwarded. (3) Transmission delay is the timeit takes to first transmit a packet, while (4) propagationdelay is the time for it to propagate through a link. Finally,(5) play-out delay is the time a packet spends in the bufferof the callee for smooth play out. The delay budget refers tothe total end-to-end delay beyond which packets are con-sidered stale.

For one-way transmission time the ITU-T G.114 recom-mendation is that a 0–150 ms delay is acceptable for mostapplications but a delay above 400 ms is unacceptable [9].For highly interactive tasks, quality may suffer at a delay of100 ms.

Voice can tolerate a small amount of packet discard.Either the decoder uses sequence numbers to interpolatefor lost packets, or the encoder adds redundancy in thesent packets [12]. These techniques work well when thelosses are isolated. For compressed voice, packet loss con-cealment is used by most codecs and involves the calleeproducing a replacement for a lost packet. This is possiblebecause of the short-term self-similarity in audio data [13].If bursty losses take place then gaps occur and the qualityof voice suffers.

Delay variation (or jitter) is the difference between theminimum and the maximum delay that packets encoun-ter in a single session, and it results from variable queue-ing delays. It is important for voice traffic to be played atthe callee at a rate matching the rate generated at thecaller [14]. Buffering is used to overcome jitter. Oncethe callee starts receiving packets, it buffers them for atime equal to the delay variation, and then starts playingthem out. When packets arrive late some packets in thebuffer are consumed, while early arrival results in thebuffer growing.

From these QoS requirements, we see that delay is thekey quality impairment for voice. Fraleigh et al. haveshown that the availability of bandwidth can limit theimpact of delay [15]. This suggests that we should choosefactors that control the ratio of offered load to the availablebandwidth in our study. One way to increase the availablebandwidth is by introducing adaptive modulation wherethe spectral efficiency changes depending on the currentchannel conditions. Another is to control the real-time traf-fic within the network. Adaptive voice compression com-presses a real-time stream in light of the current channeland network conditions.

In VoIP over wireless, a packet has substantial overheadconsisting of headers from four protocols: the real-timeprotocol (RTP), the user datagram protocol (UDP), the inter-net protocol (IP), and the medium access control (MAC)protocol. While it is important to maximize the payloadper packet, a large payload results in high packetization

764 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

delay which may impact the perceived quality at thecallee. This suggests that for adaptive compression to bebeneficial, the level of compression has to be selectedjointly with packet size.

Together, these motivate our selection of three factorsfor our study: modulation, compression, and packet size.There are many trade-offs to consider in their adaptation.We have used statistically designed screening experimentsto validate that these factors and interactions among themare influential on delay. See [16] for a complete descriptionof the experiments and the associated results.

3. Adaptation architecture and protocol

Reinforced by the results of the statistical analysis, wedesign a cross-layer opportunistic protocol; Fig. 1 showsthe architecture of the adaptive protocol. The protocolcombines hop-by-hop and end-to-end adaptation eachworking at a different time scale. Cross communication be-tween the physical (PHY) and medium access control (MAC)layers takes place at every hop along the path from thecaller to the callee and enables adaptive modulation. Whilewe use the opportunistic auto rate (OAR) protocol over IEEE802.11b to make use of the multi-rate capability of the PHYlayer [17], the architecture we propose is generic and canwork with any multi-rate PHY/MAC. Cross communicationbetween the logical link control and application (LLC/APP)layers, on the other hand, takes place only at the callerand enables adaptive selection of compression rate andpacket size.

The dynamics of the cross-layer communicationbetween the PHY/MAC layers is as follows: At every hop,when a node receives a request-to-send (RTS) packet, itanalyzes the signal quality and extracts the signal-to-noiseratio (SNR) information to select the transmission rate. Thedecision involves determining the highest achievabletransmission rate from the current channel conditions;

Fig. 1. System architecture: hop-by-hop and end-to-end adaptation.

higher transmission rates require a stronger receivedsignal [17]. Once the receiver chooses the most suitablemodulation for the packet transmission, it piggybacks itsdecision in the clear-to-send (CTS) packet. Upon receivingthe CTS, this information is extracted and communicatedto the PHY layer.

Compression and packet size selection depend on theend-to-end feedback regarding the network conditionsexpressed in terms of the packet loss ratio and averagepacket delay. Fig. 2 shows the end-to-end protocol dynam-ics at a high level. An epoch-length is the duration of timethe callee waits before sending feedback to the caller.Whenever it receives a packet, the callee updates its statis-tics for packet loss and average packet delay for the currentepoch. Average delay is first calculated by subtracting thetime stamp of every arriving packet from its arrival time.The total delay of all packets arriving within an epoch isthen divided by their number. Packet loss is calculated bymonitoring the packet identifiers and logging the numbermissing.

At the end of every epoch the callee sends a 12 bytestatistics report, containing 6 byte fields of loss and delaystatistics, to the caller. On receipt of the statistics report,

Fig. 2. The end-to-end protocol dynamics.

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 765

the caller invokes the adaptive protocol to calculate boththe packet size and the compression level.

3.1. Packetization delay, packet size, and compression levelcalculations

The adaptive protocol selects the packet size to maxi-mize the payload per packet and limit the overhead perpacket, and minimize the contribution of packetizationdelay to the total end-to-end delay to improve the voicequality experienced.

When the network is lightly loaded and end-to-enddelay is low, most of the delay budget is directed to thepacketization delay component maximizing packet sizewithout compromising quality experienced by the user.When load conditions are high, the maximum packetiza-tion delay that can be allocated without contributing toend-to-end delay is equal to the time the packet has towait in the local LLC buffer before getting transmitted overthe channel. A pipelining opportunity is created wherepacket size is maximized without contributing to end-to-end delay.

The protocol starts by querying the LLC layer regardingthe average delay in the local buffer. Using both the localbuffer delay and the end-to-end delay and loss statistics,the protocol starts by calculating the packetization budget.This is the greater of the local delay, and the delay budgetminus the end-to-end delay. This way, the contribution ofpacketization delay to the accrued end-to-end delay isminimized.

For example, consider a network experiencing light loadconditions with a network delay of 70 ms. If the delay bud-get for our application is 150 ms then there is up to150 � 70 = 80 ms that can be used toward packetization.This way, with high likelihood, the packet reaches the cal-lee on time while the payload is maximized. However,since the network delay is an average value, a safety-margin is used. In our experiments, we assume a fixed va-lue of 20 ms for the safety-margin.

On the other hand, consider a network experiencingheavy load conditions with an average delay of 140 ms. Ifthe delay budget is 150 ms then the remainder of the delaybudget is too small to use for packetization. However, if theaverage delay of the local buffer is 30 ms then we can usethis value for packetization as producing a packet anyearlier than 30 ms does not reduce the end-to-end delay.This is because the packet must wait 30 ms in the local buf-fer. This way, the protocol does not add to the total delaywhile, at the same time, the packet size is maximized.

One more factor that contributes to the packetizationdelay, and hence the packet size to select, is the currentloss ratio. If the loss ratio crosses a maximum threshold,the protocol cuts the packetization budget by a predefinedpercentage. The reason is to avoid sending packets with alarge payload because losing large packets has a greatimpact on quality.

Following the approach of Chen et al. [18], in our exper-iments we assume that half of the losses are due to channelerrors, since there is currently no way to differentiate lossdue to congestion from one due to channel noise inwireless networks.

The protocol then calculates the compression rate touse. If the loss ratio is higher than a maximum threshold,the compression rate is cut to half of the current value. Ifthe current average delay crossed a maximum threshold,the protocol again cuts the compression rate by half. If nei-ther of these two conditions is true and both the loss ratioand average delay are less than some predefined minimumthresholds, the protocol increases the compression rate tothe next rate within the available set of compression rates.In this approach, the protocol reacts quickly to ‘‘bad news’’and conservatively to ‘‘good news.’’

Next, the protocol makes sure that the compression rateand the packetization delay calculated do not fall outsidethe allowed ranges. The protocol also ensures that packet-ization delay is within the limits of the minimum andmaximum thresholds to prevent sending very small or verylarge payloads. As a last step, the protocol calculates thepacket size based on the packetization budget and the cho-sen compression rate. It then ensures that the calculatedpacket size is an integer multiple of the frame size of thegiven compression rate.

In cases where the caller fails to receive a statistics re-port for a number of epochs equal to feedback-timer-length,the protocol reacts as follows. To start, the protocol cutsthe compression rate in half as a way of mitigating any net-work congestion that may be preventing the arrival offeedback from the callee. Next, the protocol queries theLLC layer for the local buffer delay and uses this value asthe packetization delay. As before, the protocol makes surethat the compression rate and the packetization delay cal-culated do not fall outside the allowed ranges, calculatesthe packet size based on the packetization budget andthe chosen compression rate, and makes sure the calcu-lated packet size is an integer multiple of the frame sizeof the given compression rate.

The thresholds that the protocol uses depend on theapplication. If the application requires stringent qualityrequirements, the thresholds may be adjusted to producehigh quality. Likewise, if the main goal is to communicateeven if quality is reduced, thresholds may be relaxed toproduce acceptable quality.

4. Simulation set-up

We use the ns-2 network simulator [19] release2.1b7a to evaluate the performance of our opportunisticadaptive protocol. We move from simple to more sophisti-cated static topologies in order to attribute cause to obser-vations, and then consider mobile scenarios.

4.1. Static topologies

We start with a line topology with i hops, 1 6 i 6 5,where node 1 is the caller and node i + 1 is the callee. Thistopology minimizes MAC-layer contention and physical-layer co-channel interference and thus gives an idea aboutthe upper-bound performance of our protocol. Thedistance between nodes is set to 150 m for two reasons.The first is to allow the different modulation schemes tobe used whenever channel conditions allow. The second

1

2

3 4 5 6

One-hop

7

Two-hopThree-hop

Four-hopFive-hop

150 m 150 m 150 m 150 m

200m

200

m

Fig. 3. Variant of the line topology.

Fig. 4. Grid topology.

766 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

is that when nodes are closer the interference effect on oneanother is higher.

To consider the impact of MAC layer contention, wenext use a variant of the line topology shown in Fig. 3.The total load generated is divided between callers 1 and2 and is communicated to the callee. In addition to theadded contention, node 3 is a bottleneck as both nodes 1

and 2 need to pass their traffic through 3 to the rest ofthe network; this is ensured by placing nodes 1 and 2 a dis-tance of 200 m away from node 3.

We then consider a 5 � 5 grid topology shown in Fig. 4.The distance between a node and each of its horizontal andvertical neighbours is 150 m. We consider two concurrentflows to introduce co-channel interference. We vary the

Table 1Characteristics of group applications and mobility model parameters.

Application and model Characteristics N �s ðm=sÞ Ds (m/s) �p ðsÞ Dp (s) r (m) Dr (m)

Event, nomadic Walking speedLong pauses

40 0.5 0.5 60 60 0 10

March, column Walking speedNo pauses

50 1.0 1.0 0 0 10 5

Pursuit, pursuit Vehicle high speedNo pauses

10 20.0 10.0 0 0 0 5

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 767

intensity of interference by varying the distance betweenthe two flows. We start with a low-interference trafficpattern with flow1 from caller node 3 to callee node 11and flow2 from caller node 15 to callee 23. For the high-interference traffic pattern, we move the caller of flow2 tonode 8 and its callee to node 16. The distance betweenthe two flows results in co-channel interference and maycause packets not to be routed on the direct 2-hop path(we observed many different 3-hop paths taken).

Following the approach of Singh et al. [20], we thenintroduce irregularity in the grid topology by uniformlyvarying the placement of each node within a square of side40 m centered at the grid point. This way, the networkremains connected while at the same time link qualitydepends on the distance between nodes. We vary theplacement of nodes from one simulation run to another.For the irregular-grid, the low-interference traffic patternconsists of two concurrent flows, while the high-interfer-ence traffic pattern selects four concurrent flows, withcaller–callee pairs selected at random. Similar to the grid,a route in the irregular-grid may use a variable numberof hops.

Even though the topologies described so far are static,we use the Ad hoc On-demand Distance Vector (AODV)routing protocol [13] to establish the caller–callee pathsbecause routes may vary over time due to interferenceand other physical layer effects.

4.2. Mobile scenarios

We also study the impact of mobility on the perfor-mance of our protocol. The scenarios where we envisionour protocol to be employed involve team work where agroup is coordinating its actions in the battlefield or anemergency situation. Therefore, we focus on three groupapplications: an event, a march, and a pursuit modelledby a nomadic, a column, and a pursuit mobility model,respectively. Table 1 summarizes these applications, theircharacteristics, and the parameters used to model them. �srefers to the average speed of a node, Ds is the range inwhich speed changes, �p refers to the average pause timeof a node, and Dp is the range in which pause timechanges.

A nomadic mobility model captures the collectivemovement of a group of nodes from one point to another.Nodes within a group follow a reference point aroundwhich they move freely. When the reference point moves,all nodes move to the new location where they move freelyagain. In a column mobility model nodes move around acertain line which is moving ahead. A pursuit mobility

model captures the movement of a group of nodes chasinga target.

To derive the movement pattern for each of thesemobility models, we use the implementation of the refer-ence point group mobility (RPGM) generic model [21]. Thethree mobility models can be derived from this model byvarying two parameters: r, the reference point separation,and Dr, the node separation from the reference point. Thereference point separation refers to the pace at which thegroup center moves while node separation from the refer-ence point defines the coupling of the group, i.e., how farnodes are from their reference point. For these parameters,we use the values summarized in Table 1 which are takenfrom [22] and are chosen because the movement tracesthey represent are appropriate for our applications. N isthe number of nodes in the group.

We consider two, four, and eight concurrent flows forevent and march applications, and up to three concurrentflows for the pursuit application.

4.3. Wireless channel model

We use a Ricean fading model of the wireless channel.The ns-2 wireless extensions of fading [23] are based ona simple and efficient approach first proposed by Punnooseet al. [24]. Even though the channel modelling extensionsaccurately simulate the wireless channel for each individ-ual flow, fading components of channels for different flowsare identical, which is unrealistic. A way to solve this prob-lem was suggested in [17]. We use the modified model inour simulations.

4.4. Simulation parameters

Table 2 summarizes the simulation parameters. Weconsider two delay budgets to account for a spectrum ofapplications. For applications that require a high level ofinteraction, we use a delay budget of 150 ms. For moreelastic applications, we use a delay budget of 300 ms.Any packet arriving at the callee past its delay budget isconsidered late and is counted as stale.

Each packet consists of headers, and a payload segmentconsisting of an integral number of audio frames. Theheaders total 56 bytes. Thus if the payload is 100 bytes,what is transmitted is a 156 byte packet. To make sure thatthere is a reasonable number of voice frames in a packet,we do not transmit a packet with less than 50 ms of voice.

As a way of mitigating the high overhead per packet, weuse the robust header compression (ROHC) protocol [26].Rein et al. [27] show that communicating GSM speech with

Table 2Simulation and adaptive protocol parameters.

Simulationparameter

Value

Simulator ns-2.1b7a

Simulation hardware Intel Core 2 Quad CPU Q9550 at 2.83 GHz,8 GB RAM

Simulation time 1000 sSimulation warm-up

time500 s

Audio stream audio book in mono, WAVE format8000 samples/s, quantized at 16 bits

Audio streamcompression

Speex [25]

Static topologies Line, line-variant, grid, and irregular-gridMobile scenarios See Table 1

Transmission Range 250 mChannel rates 2, 5.5, and 11 MbpsFading model Ricean with K = 10 dB with flow dependent

fading [17]

Protocol parameter Value

Routing protocol AODV [13]MAC protocol OAR over IEEE 802.11b [17]Compression levels 3, 5, 7, 8, 12, 16, 24, and 32 KbpsROHC enabled and disabledOverhead per packet 56 bytes (ROHC disabled), 32 bytes (ROHC

enabled)Buffer size 100 packets, drop-tail queueing policyDelay budget 150 ms and 300 ms

epoch-length 1 sfeedback-timer-

length3 s

min-loss-thresh 1%max-loss-thresh 10%perc-chnl-contrib 50%

min-pack-delay 50 msmax-pack-delay 100 msmin-delay-budget 50 msmax-delay-budget 130 msSafety margin 20 ms

Statistics report size 12 bytes

768 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

the optimistic variant of ROHC results in an average headersize of 6 bytes. If the UDP checksum is turned off, the aver-age header size is reduced further to 4 bytes. In a separatestudy, Seeling et al. show similar performance resultswhen communicating high quality video with optimisticROHC enabled [28]. We adopt these results compressingthe UDP/IP header from 28 to 4 bytes. Each experiment isrun with ROHC disabled and then enabled.

In all cases, we run at least 50 replicates of eachexperiment.

4.5. Quantitative degradation in voice quality (DVQ) metric

We gather both quantitative and qualitative metrics ofvoice quality. The degradation in voice quality (DVQ) is aquantitative metric [29] defined as:

DVQ ¼ plost þ plate

ptotal;

where plost is the number of packets lost, plate is the numberof packets arriving after their delay budget, and ptotal is the

total number of packets sent. As a result, 0 6 DVQ 6 1 andgives the percentage of lost and late packets.

Since adaptive compression and packet size selectionare used, measuring the amount of speech by countingthe number of packets is inaccurate because the amount ofspeech per packet depends on the compression level. This isbecause packets that are the same size may carry differentamounts of voice payload. Therefore, in the computation ofDVQ, rather than counting packets, we extract the amountof speech per packet.

4.6. Qualitative mean opinion score (MOS) metric

While the smaller the DVQ the better, how DVQ corre-lates to perceived voice quality is unclear. To this end weuse a subjective metric, the mean opinion score (MOS)[30]. MOS is expressed by the scale shown in Table 3 withrange from 1 (bad) to 5 (excellent), providing a numericalindication of the listening quality of the received audiostream.

All our simulations use real voice traces as input to thesimulation. Raw recorded speech, in the form of audiobooks stored in mono, WAVE-format, serves as input tothe simulation. The audio book consists of 8000 samples/swith each sample quantized at 16 bits. This stream is thenmodified according to the dynamics of the adaptiveprotocol.

The audio stream compression is achieved using theSpeex open source audio compression format [25]. Speexis part of the GNU project and is based on code excited lin-ear prediction (CELP). It has the capability to compressvoice at bit rates ranging from 2 to 44 Kbps. The coderhas many functionalities including voice activity detection,packet loss concealment, echo cancelation, and noisesuppression.

The received stream is compared with the originalaudio stream of the same duration (no larger than 2 min)using the methodology in [31]. The perceptual evaluationof speech quality (PESQ) [32] algorithm measures speechquality comparing an original speech reference with thecallee’s version, which has a known correlation to MOS.

4.7. Non-adaptive protocols used for comparison

We compare our adaptive protocol to non-adaptive ver-sions of the protocol in which the modulation and packetsize are fixed to standard settings of voice codecs, andthe MAC protocol is IEEE 802.11b DCF used at a fixed datarate of 2 Mbps. We also experimented with a data rate of11 Mbps but because all of the results show a similar trendto the results at 2 Mbps we do not present them here.Table 4 shows the codecs, and their ITU-T or ETSI standardsettings.

5. Simulation results

We first present simulation results for the static topol-ogies and then for the mobile scenarios. We plot the DVQand the MOS as a function of the number of calls per flow,however when we tabulate the number of calls supported

Table 3MOS listening-quality scale.

Quality of speech Score

Excellent 5Good 4Fair 3Poor 2Bad 1

Table 4Standard audio/voice codec attributes.

Codec Bit rate(Kbps)

Payload(bytes)

Framing interval(ms)

G.711 [33] 64 80 10160 20240 30

G.729 [34] 8 10 1020 2030 30

G.723 [35] 6.3 8 1016 2024 30

GSM-EFR 6.60[36]

12.4 31 20

GSM-FR 6.10[37]

13.2 33 20

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 769

per flow we only count calls in which the listening qualityis at least fair, i.e., the MOS P 3. If MOS < 3, we considerthe quality of the voice to be too poor for our applicationsof interest, i.e., voice communication in the battlefield orfor emergency response.

5.1. Results for line and line-variant topologies

Fig. 5 shows the DVQ and MOS for our adaptive protocolas a function of number of calls for line topologies with

0

1

2

3

4

5

10 20 30

MO

S

Numbe

0

0.2

0.4

0.6

0.8

1

DVQ

Fig. 5. DVQ and MOS as a function of number of calls per flow for line topo

1 6 i 6 5 hops, with a delay budget of 150 ms, and no head-er compression employed; all results are summarized inTable 5. The DVQ and MOS almost appear as mirror imagesof each other. Overall, longer line topologies support fewervoice calls with fair listening quality. This is expected aslonger paths result in longer delay due to more queueingat intermediate hops, resulting in more lost and late pack-ets. The delay also increases because a node cannot bothsend and receive at the same time with a half-duplextransceiver. For example, in a four-hop path, node 3 cannotreceive from node 2 and send to node 4 concurrently.

We repeat the experiment with a relaxed delay budgetof 300 ms and with header compression enabled. These re-sults are given in Fig. 6. Not surprisingly, more calls withfair quality can be supported with a less stringent delaybudget. Since this is true for all topologies we considered,henceforth we only present our results for the stricter de-lay budget of 150 ms.

Now, we repeat the experiments for the line topologiesusing the non-adaptive protocol with standard voicecodecs; all of these results are included in Table 5. Fig. 7shows the DVQ and MOS as a function of the number ofcalls per flow for the settings yielding the highest perfor-mance; this occurs when the framing interval is the longest.Interestingly, when the DVQ is zero the corresponding MOSfor each codec is different; this confirms prior observations[38]. The highest MOS of 4.19 is achieved by the G.711codec with a framing interval of 30 ms while the G.723 ob-tains the lowest MOS of 3.27 with the same framing inter-val. The highest MOS does not correspond to the highestvoice capacity of 16 calls; this is achieved by the G.723 witha 20 ms framing interval. In all cases, the adaptive protocoloutperforms the non-adaptive protocol, often supporting atleast five times the number of calls.

The line variant topologies introduce MAC layer conten-tion between the two callers. Fig. 8 shows the DVQ and

40 50 60 70r of Calls

Scenario 1

Scenario 2

1 Hop2 Hops3 Hops4 Hops5 Hops

logies using the adaptive protocol (150 ms delay budget, no ROHC).

Table 5Number of calls supported per flow with at least fair MOS (i.e., MOS P 3) by line and line-variant topologies for a 150 ms delay budget and no ROHC. Lineartopologies establish one flow, while line-variant topologies establish two flows. The calls are multiplexed over the flows.

Number of calls per flow

1-Hop 2-Hops 3-Hops 4-Hops 5-Hops

LineAdaptive protocol 64 27 14 11 10Non-adaptive G.711 10 ms 4 2 1 1 1

20 ms 8 4 2 2 230 ms 10 5 3 2 2

Non-adaptive G.729 10 ms 5 2 1 1 120 ms 10 5 3 3 230 ms 15 8 5 4 4

Non-adaptive G.723 10 ms 5 3 1 1 120 ms 10 5 3 3 230 ms 16 8 5 4 4

Non-adaptive GSM-EFR 6.60 20 ms 10 5 3 2 2Non-adaptive GSM-FR 6.10 20 ms 10 5 3 2 2

Line-variantAdaptive protocol 19 9 5 4 4Non-adaptive G.711 30 ms 5 2 1 1 1Non-adaptive G.729 30 ms 5 4 2 1 1Non-adaptive G.723 30 ms 7 4 2 2 2Non-adaptive GSM-FR 6.10 20 ms 5 2 1 1 1

0

1

2

3

4

5

10 20 30 40 50 60 70 80

MO

S

Number of Calls

0

0.2

0.4

0.6

0.8

1

DVQ

1 Hop2 Hops3 Hops4 Hops5 Hops

Fig. 6. DVQ and MOS as a function of number of calls per flow in line topologies using the adaptive protocol (300 ms delay budget, ROHC).

770 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

MOS as a function of the number of calls per flow achievedby the adaptive protocol in the line-variant topologiesusing a 150 ms delay budget and no ROHC. The resultsare tabulated in Table 5 on a per flow basis. Because eachcaller establishes a flow, the total number of calls is twicethat tabulated. Hence, between the channel contention andthe bottleneck node, the number of calls supported in theline-variant topologies ranges from about 59% to 80% ofthe corresponding line topologies. The non-adaptive proto-col, using the settings yielding the highest performance per

codec, supports approximately 20–50% of voice capacity ofthe adaptive protocol.

5.1.1. Changes in compression over call lifetimeIn order to better understand the behaviour of the adap-

tive protocol in terms of the speed of adaptation and thequality experienced over the lifetime of a call we showthe changes in compression rate of a call for two differentscenarios in Fig. 9. We select scenarios 1 and 2 of Fig. 5 tofocus on the details of a call’s behaviour. Scenario 2 is one

0

1

2

3

4

5

0 2 4 6 8 10 12 14 16 18

MO

S

Number of Calls

0

0.2

0.4

0.6

0.8

1

DVQ

G.723 6.3Kbps 30msG.729 8Kbps 30ms

G.711 64Kbps 30msGSM6.10 13.2Kbps 20ms

Fig. 7. DVQ and MOS as a function of number of calls per flow in line topologies for the non-adaptive protocol using standard voice codecs (150 ms delaybudget, no ROHC).

0

1

2

3

4

5

5 10 15 20 25 30 35 40

MO

S

Number of Calls

0

0.2

0.4

0.6

0.8

1

DVQ

1 Hop2 Hops3 Hops4 Hops5 Hops

Fig. 8. DVQ and MOS as a function of number of calls per flow in line-variant topologies using the adaptive protocol (150 ms delay budget, no ROHC).

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 771

call out of 69 multiplexed calls over a one-hop path andhas a MOS = 2.19. Scenario 1 is a better situation of one callout of 60 multiplexed calls; this call has a MOS = 3.78. AsFig. 9 shows, scenario 2 experiences more frequent fluctu-ations in compression as it keeps adjusting its rate inresponse to the changes in network load and channel

conditions. Scenario 1 only adjusts its rate a few times.When conditions are stable and fewer calls are multiplexedin a flow, callers experience good listening quality. Whentrying to support more calls and conditions fluctuate, theprotocol keeps looking for the current best achievablequality which may result in poor listening quality.

2

3

4

5

6

7

8

400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 2600

Cop

mre

ssio

n Le

vel

Packet Identifier

Fig. 9. Changes in compression over the call lifetime.

0

1

2

3

4

5

5 10 15 20 25

MO

S

Number of Calls

0

0.2

0.4

0.6

0.8

1

DVQ

Low-interference Traffic PatternHigh-interference Traffic Pattern

Fig. 10. DVQ and MOS as a function of number of calls per flow in grid topologies using the adaptive protocol (150 ms delay budget, no ROHC).

772 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

5.2. Results for grid and irregular-grid topologies

We next study the performance of our adaptive proto-col for the grid topologies. This topology introduces co-channel interference in the low-interference traffic pattern,and heavy contention in the high-interference traffic patternbecause the caller, intermediate, and callee nodes are with-in the transmission range of their counterparts in the otherflow.

Using a 150 ms delay budget and no header compres-sion, we plot the DVQ and MOS for grid topologies inFig. 10 as a function of the number of calls per flow. Unlikethe linear topologies, there is some oscillation in the DVQ(and hence MOS) in the grid topologies. Therefore, whenwe tabulate the results in Table 6, we find the number ofcalls supported by the first MOS value below 3, and thenfind the number of calls supported for last MOS valueabove 3. This gives us a range on the number of calls

supported. Using this method, our adaptive protocol sup-ports from [0–10] calls per flow in the low-interferencetraffic pattern and from [0–5] calls per flow in the high-interference traffic pattern with fair listening quality.

We compare the performance of grid topologies and theline-variant topologies with two and three-hop paths asboth of these topologies have two competing flows. Thenumber of calls per flow supported in each topology iscomparable; see Tables 5 and 6.

The final static scenarios that we consider are the irreg-ular-grid topologies. Using a delay budget of 150 ms andno header compression, we present the number of callssupported per flow in a low-interference traffic pattern(two flows), and in a high-interference traffic pattern (fourflows) in Fig. 11. The variance of the results is high becausein the irregular-grid topologies the caller–callee pairs areselected at random. Table 6 shows that from [0–3] callsper flow are supported in the low-interference traffic

Table 6Number of calls supported per flow for grid topologies with at least fairMOS (i.e., MOS P 3.0) for a 150 ms delay budget and no ROHC. In the gridtopology, the low interference (LI) and high interference (HI) trafficpatterns each have two flows. In the irregular-grid topology, the LI trafficpattern has two flows while the HI traffic pattern has four flows.

Number of calls per flow

LI Pattern HI Pattern

GridAdaptive protocol [0–10] [0–5]Non-adaptive G.711 (30 ms) [0–3] [0–2]Non-adaptive G.729 (30 ms) [0–4] [0–3]Non-adaptive G.723 (30 ms) [0–4] [0–3]Non-adaptive GSM-FR 6.10 (20 ms) [0–3] [0–2]

Irregular-gridAdaptive protocol [0–3] 0Non-adaptive G.711 (30 ms) [0–1] 0Non-adaptive G.729 (30 ms) 0 [0–1]Non-adaptive G.723 (30 ms) 0 0Non-adaptive GSM-FR 6.10 (20 ms) [0–1] [0–1]

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 773

pattern, but no calls of fair listening quality are supportedin the high-interference traffic pattern.

Our adaptive protocol supports roughly twice the num-ber of calls for each interference pattern in grid topologiescompared to any of the non-adaptive protocols. The sameis true for irregular-grid topologies, but only for the low-interference pattern. For the high interference pattern,the adaptive protocol does not support any calls withMOS P 3 while the G.729 and GSM-FR 6.10 occasionallysupport one call.

5.3. Results for mobile scenarios

Table 7 tabulates the number of calls per flow sup-ported by the adaptive protocol for the event, march, and

0

1

2

3

4

5

2 4 6

MO

S

Numbe

0

0.2

0.4

0.6

0.8

1

DVQ

Fig. 11. DVQ and MOS as a function of number of calls per flow in irregular-grid

pursuit applications using the nomadic, column, and pursuitmobility models, respectively. In these mobile scenarios,the node separation is very small (610 m) compared tothe node separation in the static topologies (P150 m). Asa result, the signal power is very strong and the flows areable to tolerate more interference and are consequentlyable to support a higher number of calls per flow in theadaptive protocol. Even though the presence of mobility af-fects performance, since the nodes are moving as a groupand are relatively close to each other, high performanceis achieved. The results depend on the traffic pattern (re-flected by large error bars in each of the figures). The adap-tive protocol supports at least five times more calls whencompared to any non-adaptive approach.

6. Performance bounds

To gain an understanding of how the performance ofour protocol compares to an upper bound, we quantifythe theoretical maximum number of concurrent calls thatcan be supported on a single-hop IEEE 802.11b accesspoint (AP) for the compression rates and packet sizes wehave used in our simulations. We assume that the trafficis saturated and that no time is wasted in contention.

The transmission of a voice packet over an IEEE 802.11bnetwork triggers the following steps. RTP, UDP, and IPheaders totalling 40 bytes are added to the voice packet.As well, a 6 byte LLC sub-network access protocol (SNAP)header is included to reflect the transported network-layerprotocol [39]. A 24 byte MAC header is required, togetherwith a 4 byte Frame Check Sequence (FCS) calculated overthe entire frame. The channel is sensed to see if it is clearfor a distributed inter-frame space (DIFS) duration. If so, aphysical layer convergence protocol (PLCP) preamble is

8 10 12 14

r of Calls

Low-interference Traffic PatternHigh-interference Traffic Pattern

topologies using the adaptive protocol (150 ms delay budget, no ROHC).

Table 7Number of calls supported per flow for mobile scenarios with at least fairMOS (i.e., MOS P 3.0) for a 150 ms delay budget and no ROHC. The event,march, and pursuit applications use the nomadic, column, and pursuitmobility models, respectively. Two, four, and eight concurrent flows areconsidered in the event and march applications, while up to threeconcurrent flows are considered for the pursuit application.

Number of calls per flow

2-Flows 4-Flows 8-Flows

Event applicationAdaptive protocol 46 21 9Non-adaptive G.711 (30 ms) 5 2 1Non-adaptive G.729 (30 ms) 8 4 2Non-adaptive G.723 (30 ms) 8 4 2Non-adaptive GSM-FR 6.10 (20 ms) 5 2 1

March applicationAdaptive protocol 46 21 8Non-adaptive G.711 (30 ms) 5 2 1Non-adaptive G.729 (30 ms) 8 4 2Non-adaptive G.723 (30 ms) 8 4 2Non-adaptive GSM-FR 6.10 (20 ms) 5 2 1

1-Flow 2-Flows 3-Flows

Pursuit applicationAdaptive protocol 96 46 29Non-adaptive G.711 (30 ms) 10 5 3Non-adaptive G.729 (30 ms) 15 8 5Non-adaptive G.723 (30 ms) 16 8 5Non-adaptive GSM-FR 6.10 (20 ms) 10 5 3

Table 8Default parameter values per frame sent by IEEE 802.11b DCF.

Parameter Value

Distributed Inter-Frame Space (DIFS) 50 lsShort Inter-Frame Space (SIFS) 10 lsRTP/UDP/IP headers 40 bytesLLC/MAC headers 34 bytesPayload Codec dependentLong PLCP (preamble and header), 192 bits 192 lsFrame Check Sequence (FCS) 4 bytesAcknowledgement (ACK) at 2 Mbps 14 bytesSlotTime 20 lsCWmin, CWmax 32 slots, 1024 slots

0

50

100

150

200

250

32241612853

Cal

ls

Compression Rate (Kbps)

Upper bound at 2MbpsUpper bound at 11Mbps

Fig. 12. Bounds on the number of calls supported as a function ofcompression rate.

774 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

added. The short frame format requires 72 bits of the PLCPpreamble to be transmitted at a required rate of 1 Mbpsand 48 bits of the PLCP header to be transmitted at arequired rate of 2 Mbps. The frame is then transmitted bythe caller at the IEEE 802.11 data rate in use (one of 2,5.5, or 11 Mbps). After waiting a short inter-frame space(SIFS) duration, the callee creates a 14 byte acknowledg-ment (ACK) frame, and adds a PLCP preamble and headerto be transmitted at the required rates of 1 and 2 Mbps,respectively. The callee transmits an ACK at the IEEE802.11b data rate.

Since IEEE 802.11b supports three transmission rates,the time needed to transmit a packet depends on the rateused. However, regardless of the data rate in use by theadaptive protocol, some fields are transmitted at a fixedrate as specified by the standard [40]. The default parame-ter values for IEEE 80211b DCF are shown in Table 8.

The packet transmission time (PTT), in ls, of a voicepacket is calculated as:

PTT ¼ DIFSþ SIFS þ 2� ðPLCP Preamble þ PLCP HeaderÞ

þ ðRTP=UDP=IP=LLC=MAC Headersþ Payloadþ ACKÞ � 8data rate

The number of packets per a voice call (PPVC) per secondis equal to:

PPVC ¼ Compression Rate ðbpsÞPayload� 8

� �� 2:

The multiplication by two is to account for the bidirec-tional nature of a call. Given the equations for PTT andPPVC, the maximum number of concurrent calls that aresupported is given by:

Maximum Number of Calls ¼ 106

PTT� PPVC

$ %:

Fig. 12 uses these equations to plot the maximum num-ber of calls supported as a function of the compression ratefor data rates of 2 Mbps and 11 Mbps with the minimumand maximum payload, respectively. Since the analysis isdone for a single-hop IEEE 802.11b access point, it boundsthe results for the single-hop line topology most closely.The adaptive protocol supports 64 calls in this case, whichlies between the two bounds. The analysis does not takeinto account that the adaptive protocol varies the modula-tion, compression, and packet size, over the call lifetimeand is therefore only a loose bound on performance.

7. Related work

We now overview work on voice over IP in ad hocnetworks, wireless local areas networks (WLANs), and alsoin related wireless mesh networks (WMNs). It is difficult tocompare directly the voice capacity of our protocol tothose discussed since the details of the experimentationare not fully known. Nevertheless, we attempt a compari-son and also indicate avenues of future research.

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 775

7.1. VoIP over ad hoc networks and wireless LANs

Specialized protocols that focus on voice support overad hoc networks have been proposed. Wang et al. [1] pro-pose the combined use of multicasting and multiplexing ofmultiple voice packets into one packet as a way of reducingthe per-packet overhead. As a result, the protocol shows anincrease in the network capacity and a decrease in thedelay experienced by voice calls. Priority queuing isemployed as a way of preventing competing TCP trafficfrom starving voice traffic of resources. The analysis for or-dinary VoIP capacity for ETSI Global System for Mobile(GSM) communications 06.10 Full Rate (FR) speech coder[37], ITU G.711, G.729, and G.723 voice codecs using IEEE802.11b DCF access scheme at 11 Mbps shows voice capac-ities similar to our experimental and analytical results forthe non-adaptive protocol. The small difference betweenour results may be due to the use of a packet-loss rate be-low 1% compared to our 10%. The voice capacity of themulticast scheme, which improves the ordinary VoIPcapacity by close to 100%, is less than that achieved byour adaptive protocol.

A modification of IEEE 802.11 is proposed in Dong et al.[41] in which the cyclic redundancy codes are computedonly over those parts of the voice frame that have a highimpact on the perceived quality rather than over the entireframe. In this way, less bandwidth is wasted in retransmis-sion and less delay is introduced. In [42], the use of newspeech coding techniques for supporting voice over adhoc networks is proposed. One such technique is multipledescription coding. It involves creating more than one bitstream from the source signal. Each independent streamrepresents a coarse description of the transmitted signal.If more than one description is received, a refined signalis reconstructed. Another technique is scalable speech cod-ing, which consists of sending a base stream at a minimumrate and one or more enhancement streams. Our workcomputes the FCS over the entire frame and does not makeuse of these speech coding techniques.

Obeidat and Syrotiuk [43] study the performance ofadaptive voice communications over multi-hop wirelessnetworks; this work extends that work significantly. Inparticular, a statistically designed experiment is used toquantify significant factors and their interactions on voicequality. This motivated the integration of end-to-end adap-tation. In addition, the use of real audio traces allows theevaluation of audio quality metrics. We also consider morecomplex topologies and scenarios integrating mobility instudying the protocol to better understand how it performsin situations more representative of battlefield and emer-gency scenarios.

Fasolo et al. [44] present a cloud of nodes that commu-nicate with one gateway by means of multi-hop ad hocconnections to study the effect of multi-rate on voicecapacity. They assessed their analysis through ns-2 simu-lations using IEEE 802.11b DCF access scheme at 11 Mbpsand ETSI GSM 06.60 Enhanced Full Rate (EFR) voice codec[36]. Their results for a delay budget of 100 ms and lessthan 1% loss probability show a maximum of 6 and 3concurrent voice connections for single-hop and multi-hop scenarios, respectively. Our adaptive protocol achieves

higher voice capacity perhaps due to differences in the de-lay budget and loss probability. Moreover, our evaluationconsiders more extensive multi-hop and mobile scenarios.

A number of works consider voice capacity of WLANs.Adaptive modulation and adaptive compression have beenapplied separately in VoIP-based wireless and wired net-works [45–48]. Supporting packet voice over IEEE 802.11has been investigated for both the DCF and PCF, howeverthe performance is poor [49,50].

Garg and Kappes [51] analyze the number of simulta-neous VoIP calls a single AP running the IEEE 802.11bDCF can support. Their experimentation uses an ITUG.711 a-Law codec with 10 ms of voice data. At 11 Mbps,6 calls are supported by the AP with acceptable quality.An analytical model is developed for three standard codecs(ITU G.711 a-Law [33], G.723 [35], and G.729 [34]) consid-ering DCF compliance and data transmission rates of theAP varying from 1 Mbps to 11 Mbps to validate the exper-imental results. Our model in Section 6 is similar and re-ports essentially the same number of VoIP callssupported for these standard codecs.

Hole and Tobagi [52] quantify the capacity of a wirelessLAN using IEEE 802.11b at 11 Mbps carrying VoIP callsusing analysis and simulation. The analytic upper boundmatches the simulation results when channel quality isgood. The capacity of the network is found to be highlydependent on the delay constraints of the carried voice. Gi-ven a delay budget constraint and non-ideal channel condi-tions they offer a means to select the voice data packet size(in ms) for the ITU G.711 and G.729 codecs. Our work on thenon-adaptive protocols shows close results for the VoIPcalls supported for the same standard codecs. We agree thatthe combined effects of delay and packet loss must be takeninto consideration on the quality of the voice, hence we gobeyond fixed codec attributes and offer a protocol thatopportunistically adapts modulation, compression, andpacket size to maximize call capacity and quality.

Along the same lines of research, Anjum et al. [53] inves-tigate the capacity of wireless LANs for VoIP traffic and as aresult suggest the use of controlled back-off and priorityqueuing at the AP when voice and data traffic co-exist.

7.2. VoIP over wireless mesh networks

The advantage of using multiple radios on voice capac-ity has been investigated in [54,55]. Kim et al. [54] proposea model to accurately infer network capacity of VoIP callsin multi-channel multi-radio (MCMR) WMNs. This is neededsince accurate connection admission and control dependon accurate estimation of call capacity. Coordination ofradios and channels is accomplished using the hybrid mul-ti-channel protocol (HMCP). The model is validated throughboth test bed measurements and ns-2 simulations, accu-rately estimating capacity to within 6% of actual measure-ments and simulations. With speech compressed at 8 Kbps,up to 80 calls can be supported over a 5-hop line topology.

Bayer et al. [55] investigate the feasibility of VoIP overWMNs through measurements from a designed test bed.The use of dual radios is shown to provide significantlybetter performance than single radios. However, suchimprovements are seen only for large packet sizes. As a

776 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

result, a hop-to-hop aggregation algorithm is proposed.Packets are held at intermediate nodes until there areenough packets to make a preset minimum size. However,the holding of packets is done as long as their delay has notreached a certain threshold. The network simulator is usedto investigate the performance of the aggregation algo-rithm over an 802.11a with a basic rate of 6 Mbps, a datarate of 24 Mbps and a node separation of 45 m. Withspeech encoded using G.729a with voice activity detection,results show that around 350 calls can be supported with aMOS of 3.5.

While the use of multiple radios and the proper assign-ment of channels can result in an increase in networkcapacity, it requires the use of such configurations withcorresponding changes in the protocol stack. Our focus inthis work is on the more common single-radio endsystems.

Kamoun et al. [56] propose a packet scheduling algo-rithm that takes into account wireless channel conditions,class of service of data carried, and whether a connection isnew or handoff. A handoff occurs as the source of an ongo-ing multimedia session moves from the range of one wire-less mesh router to that of another. The scheme favourshandoff calls over new calls, and realtime traffic overnon-realtime traffic. The algorithm successfully limits thedelay of realtime traffic to 135 ms. The rate at whichspeech is compressed, and the protocol overhead is consid-ered, but the mobility pattern considered is not describedmaking it is hard to relate to their results. We plan to aug-ment our work with scheduling and drop policies that takeinto account the nature of voice and possibly packet size.The channel-aware nature of this scheduling algorithmmakes it a particularly good candidate as it gives a short-term prediction of network conditions.

El-Hennawey et al. [57] study the performance of VoIPover a WMN running IEEE 802.11e for QoS provisioning.Both call quality and throughput are quantified. Using astatic line topology, results show that over a single-hopup to 8 calls are supported, over 2-hops up to 6 calls, over3-hops up to 4 calls, and over 4-hops up to 2 calls. A callis considered supported if it meets a MOS of 3.1. Fairnessis also quantified to determine whether the network treatscalls with identical QoS requirements fairly. Results showthat a high degree of fairness is exhibited. Another aspectthat is quantified is whether non-overlapping backgroundtraffic has an effect on call quality. A 3-hop call is separatedfrom background traffic by 2-hops, 1-hop, and no-hops.The results show that the smaller the separation, the high-er the impact on quality. The study does not consider theeffect of mobility or frame bursting. While we do notinvestigate fairness or separation of background traffic,comparison with their results for line topologies reflectsthat our protocol shows superior performance.

Siddique and Kamruzzaman [58] estimate the VoIP callcapacity of a single-hop WMN using analytic modelling.Network capacity is modelled as a maximization problemgoverned by quality constraints involving network param-eters. The model can be expanded to multi-hop networksand to other types of realtime traffic. The main contribu-tion is in the detailed modelling of delay and loss sourcesto capture impairment factors contributing to quality

compromise. The model is solved numerically and its re-sults are verified by simulations using ns-2. The resultsshow that increasing the number of voice frames per pack-et results in an overall increase in network capacity butonly to a certain degree beyond which packetization delayresults in call quality degradation. In addition, lower datarate coders, those more aggressive in compressing speech,result in a higher capacity, even though the coder’s impair-ment factor can affect such a trend. The results also dem-onstrate the effect of increasing the data rate from11 Mbps to 54 Mbps. The increase in network capacity isnot matched by a comparable increase in call capacity.Further, higher data rate coders such as G.711 result in rel-atively higher gains in capacity than higher compressioncoders such as G.729a. This is because G.711 generateslarger packets with less per-packet overhead. Lastly,employment of RTS/CTS is found to negatively affect thenumber of calls supported. Simulation results are solelyof one-hop network with no mobility and are similar orinferior to the results of our protocol.

Kulkarni and Devetsikiotis [59] propose a cross-layerdesign for increasing the VoIP call capacity of a WMN.The study identifies parameters deemed crucial acrossthree layers, MAC data rate, routing approach, and voicepacketization interval. Four different MAC data rates areconsidered as provided by the IEEE 802.11b standard.Two routing approaches are investigated: hop-count andlink-rate aware routing. Using G.711 for encoding speech,ten different packetization rates and corresponding packetsizes are considered. Simulations in ns-2 are used to gen-erate responses to variations of the parameters. An n-facto-rial analysis and linear regression fitting are used to derivealgebraic equations for the call capacity. Fitting equationsare found using the SAS GLM procedure. In plotting thesefunctions, parameter-combinations that provide the high-est capacity are found. As for the goodness of fit, an analysisof variance (ANOVA) R2 greater than 70 is considered anindicator of acceptable call quality. Results show thepositive effect of using link rate-aware routing. Packetiza-tion has an effect on capacity but only to a certain degreebeyond which it becomes negligible. We only considerhop-count as a link metric in our routing protocol. How-ever, link-rate aware routing is shown to give a substantialimprovement and appears to be worthwhile to consider.

Packet aggregation is proposed by many studies as away of mitigating the per-packet overhead of inherentlysmall voice packets [60–62]. Hasegawa et al. [60] proposethe use of bidirectional packet aggregation and networkcoding for the support of VoIP over WMN. The proposedprotocol is implemented in a test bed and is also verifiedthrough simulations. Using a line topology, bidirectionaltraffic is aggregated then network-coded using an XORoperation. Aggregation opportunities are increased by hav-ing intermediate routers hold packets for a time periodequal to their queuing delay share of the total delaybudget. With node separation of 100 m and a number ofhops varying between 2 and 7, the protocol is shown tosupport around 23 calls of speech compressed usingG.711 over a 7-hop connection. A call is consideredsupported if its network delay is limited to 150 ms andits loss rate is within 5%.

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 777

Kurien et al. [61] propose a dynamic approach to packetaggregation to increase VoIP call capacity in WMNs. Aggre-gation is performed only on packets going to the same nexthop. The optimal aggregation size is chosen based on thesignal-to-noise and interference ratio (SNIR) of the outgoinglink. Knowledge of the receiving MAC of the SNIR is used tocompute bit error rate (BER) for the employed modulationtechnique. BER is then used to compute the frame error rate(FER). The algorithm then chooses an aggregation size thatlimits FER to less than 0.1%. This value is chosen so that theend-to-end error rate is small. Nodes maintain a queue foreach outgoing link. Aggregation takes place whenever aqueue grows past certain threshold or when oldest packethas crossed certain delay threshold. Performance is inves-tigated using the network simulator ns-2 and is comparedagainst non-dynamic aggregation and plain 802.11. Using astatic line topology, the approach is shown to have supe-rior performance in terms of all network parameters andin call capacity. However, it is not obvious what is consid-ered acceptable call quality.

Kim and Hong [62] propose a scheme integrating packetaggregation and header compression to limit overhead andmaximize VoIP capacity of a WMN. Aggregation takes placeboth end-to-end and hop-to-hop with the first contributingto the end-to-end delay and the latter working within theMAC delay. End-to-end aggregation is applied intra-flow,to packets coming from the same flow, while hop-to-hopaggregation is applied between flows. Since end-to-endaggregation is applied intra-flow, the scheme is augmentedwith header elimination of the second to the last packets ofan aggregated packet. Simulations using the network simu-lator show that using G.729a speech, the scheme can resultin supporting more than 10 calls over 4 to 8 hops of a linetopology. While our results using the same coder show a callcapacity of 10 calls over one hop, the use of aggregation andheader elimination enables their scheme to support 7 timesas much (a line topology provides an ideal scenario foraggregation). Many studies reach to the same conclusionregarding the merit of aggregation and we intend to incorpo-rate it into our future work.

Aggregation-aware routing is investigated in [63,64].Liwlompaisan and Phonphoem [63] propose a routingscheme that combines packet aggregation, multi-pathrouting, utilization awareness, and event-triggered rerout-ing. A link that can be part of many paths allows for higherchances of aggregation, and hence is more attractive inroute discovery. This, however, may result in hot spot rout-ing behaviour. As a result, the saturated utilization is takeninto account in the cost so that routes go around suchspots. As an additional measure to limit the hot spot effect,backward traffic is sent on a path different from forwardtraffic. Also, an intermediate hot spot node sensing highmedium utilization may request certain source nodes toreroute their traffic. Simulations are conducted using thenetwork simulator ns-3 of an 802.11a WMN with speechencoded at a rate of 64 Kbps. Quality constraints are300 ms of delay budget and loss rate of 10%. The resultsshow an increase in the number of supported calls overlonger paths (4–9 hops). The delay behaviour is not im-proved but is not aggravated in comparison with similarprotocols.

Along the same lines, Ramprashad et al. [64] use a the-oretical framework to investigate the joint effect of routingand admission with packet aggregation, bursting, and rateadaptation of multiple packets in a single transmissionopportunity on VoIP call capacity of a multi-hop 802.11network. Analytic results are verified through simulationof a 2-hop scenario using the ns-2 network simulator.Results show that the analytical framework provides atight upper-bound when compared with simulation. Inthe presence of channel errors, around 22 calls can be sup-ported. As for rate adaptation, the results show that jointoptimization of other factors is only of interest in a2–3 dB SNR region between rate switches. Our cross-layerframework does not include the routing layer. Incorporat-ing more layers involves a tradeoff between performanceand protocol complexity.

8. Conclusions and future work

Adaptation and cross-layer design are two approachesto address the challenges of supporting voice over ad hocnetworks. We identified the factors of compression, modu-lation, and packet size to adapt based on the QoS require-ments of voice. Our resulting opportunistic protocolcombines adaptation on two time scales: hop-by-hop andend-to-end. The performance of our protocol was evalu-ated through simulations in static and mobile scenarios,carrying real-time audio traffic using both quantitative(DVQ) and qualitative (MOS) audio metrics.

Our work may be extended in several ways. The protocolmay be combined with a multi-path diversity approachwhere multiple paths are used between a caller–callee pair.Different paths may carry voice packetized, compressed,and modulated differently to optimize network perfor-mance and call quality. In general, QoS-aware routing,which takes interference of the flows into account, ratherthan following the shortest hop-count path may be useful.

The use of forward error correction (FEC) is another ave-nue of work. Even though the use of FEC introduces extraoverhead, it can curb the rate of lost and late packets. Anode can decide whether to use no compression and expe-rience a high loss rate or consider aggressive compressionwhile applying FEC.

The impact of traffic heterogeneity, where voice, data,and video are supported concurrently, is another importantstudy. Unlike real-time applications which are particularabout delay but more resilient to losses, data applicationsare bandwidth-greedy, delay-elastic, and intolerant to loss.Employing special measures, such as the use of priorityqueuing, may be needed to ensure appropriate supportfor voice applications.

Finally, experiments using human subjects to obtainMOS results in battlefield or emergency situations wouldbe useful for future work on supporting voice in these typesof scenarios.

Acknowledgments

We are grateful to area editor, and to the anonymousreferees, whose thoughtful comments led us to signifi-cantly improve the results of our paper.

778 S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779

References

[1] W. Wang, S.C. Liew, V.O.K. Li, Solutions to performance problems inVoIP over a 802.11 wireless LAN, IEEE Transactions on VehicularTechnology 54 (1) (2005) 366–384.

[2] R.Y.W. Lam, V.C.M. Leung, H.C.B. Chan, Polling-based protocols forpacket voice transport over IEEE 802.11 wireless local area networks,IEEE Wireless Communications Magazine 13 (1) (2006) 22–29.

[3] S.B. Lee, A.T. Campbell, INSIGNIA: An IP-based quality of serviceframework for mobile ad hoc networks, Journal of Parallel andDistributed Computing 60 (4) (2000) 374–406.

[4] A. Goldsmith, S.B. Wicker, Design challenges for energy-constrainedad hoc wireless networks, IEEE Wireless Communications Magazine9 (4) (2002) 8–27.

[5] A. Goldsmith, Wireless Communications, Cambridge UniversityPress, 2005.

[6] B. Raman, P. Bhagwat, S. Seshan, Arguments for cross-layeroptimizations in bluetooth scatternets, in: Proceedings of the IEEE2001 Symposium on Applications and the Internet, 2001, pp. 176–184.

[7] V. Kawadia, P.R. Kumar, A cautionary perspective on cross layerdesign, IEEE Wireless Communications Magazine 12 (1) (2005) 3–11.

[8] V. Srivastava, M. Motani, Cross-layer design: a survey and the roadahead, IEEE Communications Magazine 43 (12) (2005) 112–119.

[9] ITU-T Recommendation G.114, One-way transmission time,International Telecommunication Union, Geneva, 1996.

[10] K. Sriram, M.H. Sherif, Voice packetization and compression inbroadband ATM networks, IEEE Journal on Selected Areas inCommunications 9 (3) (1991) 294–304.

[11] D. Miras, A survey of network QoS needs of advanced internetapplications, Internet2 QoS Working Group, Working Document,December 2002.

[12] D.E. McDysan, D. Spohn, ATM Theory and Applications, McGraw-Hill,1999.

[13] C. Perkins, O. Hodson, V. Hardman, A survey of packet loss recoverytechniques for streaming audio, in: IEEE Network Magazine, 1998,pp. 40–48.

[14] W. Stallings, High Speed Networks: TCP/IP and ATM DesignPrinciples, Prentice Hall, 1998.

[15] C. Fraleigh, F. Tobagi, C. Diot, Provisioning IP backbone networks tosupport latency sensitive traffic, in: Proceedings of the 22nd AnnualJoint Conference of the IEEE Computer and CommunicationsSocieties (Infocom’03), 2003, pp. 375–385.

[16] S.A. Obeidat, Cross-layer opportunistic adaptation for voice overwireless ad hoc networks, Ph.D. Thesis, Arizona State University,Tempe, AZ, May 2008.

[17] B. Sadeghi, V. Kanodia, A. Sabharwal, E. Knightly, Opportunisticmedia access for multirate ad hoc networks, in: Proceedings of the8th Annual International ACM Conference on Mobile Computing andNetworking (MobiCom’02), 2002, pp. 24–35.

[18] T. Chen, M. Kazantzidis, M. Gerla, I. Slain, Experiments on QoSadaptation for improving end user speech perception over multihopwireless networks, in: Proceedings of the IEEE InternationalConference on Communications (ICC’99), 1999, pp. 708–715.

[19] The Network Simulator — ns-2. <http://www.isi.edu/nsnam/ns>.[20] S. Singh, P.A. Acharya, U. Madhow, E.M. Belding-Royer, Sticky CSMA/

CA: Implicit synchronization and real-time QoS in mesh networks,Ad Hoc Networks 5 (6) (2007) 744–768.

[21] T. Camp, J. Boleng, V. Davies, A survey of mobility models for ad hocnetwork research, Wireless Communications and Mobile Computing2 (5) (2002) 483–502 (special issue on Mobile Ad hoc Networking:Research, Trends and Applications).

[22] M.M. El Saoud, MANET reference configurations and evaluation ofservice location protocol for MANET, Masters Thesis, CarletonUniversity, 2005.

[23] The Monarch Group, The Monarch Project: Wireless and MobilityExtensions to ns-2. <http://www.monarch.cs.rice.edu/cmu-ns.html>.

[24] R. Punnoose, P. Nikitin, D. Stancil, Efficient simulation of Riceanfading within a packet simulator, in: Proceedings of the IEEEVehicular Technology Conference (VTC’00), 2000, pp. 764–767.

[25] Speex: A Free Codec for Free Speech. <http://www.speex.org>.[26] C. Bormann, C. Burmeister, M. Degermark, H. Fukushima, H. Hannu,

L. Jonsson, R. Hakenberg, T. Koren, K. Le, Z. Liu, A. Martensson, A.Miyazaki, K. Svanbro, T. Wiebke, T. Yoshimura, H. Zheng, Robustheader compression (ROHC): Framework and four profiles: RTP,UDP, ESP, and uncompressed, IETF RFC 3095, 2001.

[27] S. Rein, F. Fitzek, M. Reisslein, Voice quality evaluation for wirelesspacket voice: A tutorial and performance results for ROHC, IEEEWireless Communications Magazine 12 (2005) 60–76.

[28] P. Seeling, M. Reisslein, F.H.P. Fitzek, S. Hendrata, Video qualityevaluation for wireless transmission with robust headercompression, in: Proceedings of the 4th International Conferenceon Information, Communications and Signal Processing (ICICS’03),vol. 3, 2003, pp. 1346–1350.

[29] R. Jain, S. Munir, J. Iyer, Performance of VBR voice over ATM: Effect ofscheduling and drop policies, ATM Forum/97-0608, 1997.

[30] ITU-T Recommendation P.800, Methods for subjective determinationof transmission quality, International Telecommunication Union,Geneva, 1996.

[31] ITU-T Recommendation P.862, Perceptual evaluation of speechquality (PESQ): an objective method for end-to-end speech qualityassessment of narrow-band telephone networks and speech codecs,International Telecommunication Union, Geneva, 2001.

[32] ITU-T Recommendation G.728, Implementors guide for ITU-Trecommendation G.728: coding of speech at 16 kbits/sec usinglow-delay code excited linear prediction, InternationalTelecommunication Union, Geneva, 1992.

[33] ITU-T Recommendation G.711.1, Wideband embedded extension forG.711 pulse code modulation, International TelecommunicationUnion, Telecommunication Standardization Sector, 2008.

[34] ITU-T Recommendation G.729, Coding of speech at 8 kbit/s usingconjugate-structure algebraic-code-excited linear prediction (CS-ACELP), International Telecommunication Union, Telecommu-nication Standardization Sector, 2007.

[35] ITU-T Recommendation G.723.1, Dual rate speech coder formultimedia communications transmitting at 5.3 and 6.3 kbit/s,International Telecommunication Union, Telecommunication Stan-dardization Sector, 2006.

[36] ETSI European Telecommunications Standards Institute, Enhancedfull rate (EFR) speech transcoding (GSM 06.60 version 8.0.1), ETSIDigital Cellular Telecommunications System (Phase2+), 1999.

[37] ETSI European Telecommunications Standards Institute, Full ratespeech, transcoding (GSM 06.10 version 8.2.0), ETSI Digital CellularTelecommunications System (Phase2+), 2005–2006.

[38] ITU-T Recommendation P.830, Subjective performance assessmentof telephone-band and wideband digital codecs, InternationalTelecommunication Union, Telecommunication StandardizationSector, 1996.

[39] M. Gast, 802.11 Wireless Networks: The Definitive Guide, second ed.,O’Reilly Media, Inc., 2005.

[40] IEEE standard 802.11: W-LAN medium access control and physicallayer specifications, December 1999.

[41] H. Dong, I.D. Chakares, C.-H. Lin, A. Gersho, E. Belding-Royer, U.Madhow, J.D. Gibson, Selective bit-error checking at the MAC layerfor voice over mobile ad hoc networks with IEEE 802.11, in:Proceedings of the IEEE Wireless Communications and NetworkingConference (WCNC’04), 2004, pp. 1240–1245.

[42] H. Dong, I.D. Chakares, C.-H. Lin, A. Gersho, E. Belding-Royer, U.Madhow, J.D. Gibson, Speech coding for mobile ad hoc networks, in:Proceedings of the Asilomar Conference on Signals, Systems, andComputers (ACSSC’03), vol. 1, 2003, pp. 280–284.

[43] S.A. Obeidat, V.R. Syrotiuk, An opportunistic cross-layer architecturefor voice in multi-hop wireless LANs, International Journal ofCommunications Systems 22 (4) (2009) 419–439.

[44] E. Fasolo, F. Maguolo, A. Zanella, M. Zorzi, S. Ruffino, P. Stupar, VoIPcommunications in wireless ad-hoc network with gateways, in:Proceedings of the 12th IEEE Symposium on Computers andCommunications (ISCC’07), 2007, pp. 69–74.

[45] A. Barberis, C. Casetti, J.C. De Martin, M. Meo, A simulation study ofadaptive voice communications on IP networks, in: Proceedings of theInternational Symposium on Performance Evaluation of Computerand Telecommunication Systems (SPECTS’00), 2000, pp. 531–542.

[46] S. Shenker, Fundamental design issues for the future interent, IEEEJournalon Selected Areas in Communications 13 (7) (1995) 1176–1188.

[47] K. Balachandran, S.R. Kadaba, S. Nanda, Channel quality estimationand rate adaptation for cellular mobile radio, IEEE Journal onSelected Areas in Communications 17 (7) (1999) 1244–1256.

[48] T. Ue, S. Sampei, N. Morinaga, K. Hamaguchi, Symbol rate andmodulation level-controlled adaptive modulation/TDMA/TDDsystem for high-bit-rate wireless data transmission, IEEETransactions on Vehicular Technology 47 (4) (1999) 1134–1147.

[49] M.A. Visser, M. El Zarki, Voice and data transmission over an 802.11wireless network, in: Proceedings of the IEEE InternationalSymposium on Personal, Indoor, and Mobile RadioCommunications (PIMRC’95), 1995, pp. 648–652.

[50] E. Ziouva, T. Antonakopoulos, CBR packetized voice transmission inIEEE 802.11 networks, in: Proceedings of the IEEE Symposium onComputers and Communications, 2001, pp. 392–398.

S.A. Obeidat et al. / Computer Networks 56 (2012) 762–779 779

[51] S. Garg, M. Kappes, Can I add a VoIP call? in: Proceedings of the IEEEInternational Conference on Communications (ICC’03), vol. 2, 2003,pp. 779–783.

[52] D. Hole, F.A. Tobagi, Capacity of an IEEE 802.11b wireless LANsupporting VoIP, in: Proceedings of IEEE International Conference onCommunications (ICC’04), vol. 1, 2004, pp. 196–201.

[53] F. Anjum, M. Elaoud, D. Famolari, A. Ghosh, R. Vaidyanathan, A.Dutta, P. Agrawal, T. Kodama, Y. Katsube, Voice performance inWLAN networks - an experimental study, IEEE GlobalTelecommunications Conference (Globecom’03) 6 (2003) 3504–3508.

[54] S. Kim, M. Ji, J. Ma, Voice call capacity model for hybrid multi-channel protocol over multi-hop multi-channel multi-radio wirelessmesh networks, in: Proceedings of the International Conference onAdvanced Communication Technology (ICACT’11), 2011, pp. 1239–1244.

[55] N. Bayer, M.C. de Castro, P. Dely, A. Kassler, Y. Koucheryavy, P.Mitoraj, D. Staehle, VoIP service performance optimization in pre-IEEE 802.11s wireless mesh networks, in: Proceedings of the IEEEInternational Conference on Circuits and Systems for MultimediaWireless Communications (ICCSC’08), 2008, pp. 75–79.

[56] W. Mansouri, F. Zarai, K. Mnif, L. Kamoun, New scheduling algorithmfor wireless mesh, in: Proceedings of the International Conference onMultimedia Computing and Systems (ICMCS’11), 2011, pp. 391–396.

[57] D.V. Geyn, H. Hassanein, M.S. El-Hennawey, Voice call quality using802.11e on a wireless mesh network, in: Proceedings of the IEEE34th Conference on Local Computer Networks (LCN’09), 2009, pp.792–799.

[58] M.A. Siddique, J. Kamruzzaman, VoIP call capacity over wirelessmesh networks, in: Proceedings of the IEEE Global CommunicationsConference (Globecom’08), 2008, pp. 601–611.

[59] V. Kulkarni, M. Devetsikiotis, Cross-layer response surfacemethodology applied to wireless mesh network VoIP call capacity,in: Proceedings of the 41st Annual Simulation Symposium, 2008, pp.15–22.

[60] J. Hasegawa, H. Yomo, Y. Kondo, P. Davis, R. Suzuki, S. Obana, K.Sakakibara, Bidirectional packet aggregation and coding for VoIPtransmission in wireless multi-hop networks, in: Proceedings of theInternational Conference on Communications (ICC’09), 2009, pp.113–118.

[61] J.M. Okech, Y. Hamam, A. Kurien, A cross-layer adaptation for VoIPover infrastructure mesh network, in: Proceedings of the 3rdInternational Conference on Broadband Communications,Information Technology and Biomedical Applications(BroadCom’08), 2008, pp. 97–102.

[62] K. Kim, S. Hong, VoMESH: voice over wireless mesh networks, in:Proceedings of the IEEE Wireless Communications and NetworkingConference (WCNC’06), 2006, pp. 193–198.

[63] W. Liwlompaisan, A. Phonphoem, Call capacity improvementtechniques for VoIP over wireless mesh networks, in: Proceedingsof the International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and InformationTechnology (ECTI-CON’09), 2009, pp. 902–904.

[64] S.A. Ramprashad, D. Li, U.C. Kozat, C. Pepin, An analysis of jointaggregation, bursting, routing, and rate adaptation for increasing

VoIP capacity in multi-hop 802.11 networks, IEEE Transactions onWireless Communications 7 (8) (2008) 3128–3139.

Suhaib A. Obeidat completed his Ph.D atArizona State University in 2008. Currently, heis an Assistant Professor of Computer Scienceat Bennett College, Greensboro, NC. Hisresearch interests include wireless and mobilead hoc networks, and adaptive and cross-layermultimedia communications over wireless.

Abraham N. Aldaco earned his M.S. in Com-puter Science and B.S. in Systems and Elec-tronics from Monterrey Institute ofTechnology (ITESM), Mexico, in 2000 and1988, respectively. Currently he is a graduatestudent pursuing a Ph.D. in Computer Scienceat the School of Computing, Informatics andDecision Systems Engineering, Arizona StateUniversity, Tempe, AZ. His current researchinterest include mobile ad hoc networks,cognitive radio networks, and wireless sensornetworks.

Violet R. Syrotiuk earned her Ph.D. in Com-puter Science from the University of Waterloo(Canada). She is currently an Associate Pro-fessor of Computer Science and Engineering inthe School of Computing, Informatics, andDecision Systems Engineering at Arizona StateUniversity. Dr. Syrotiuk’s research has beensupported by grants from NSF, DSTO (Aus-tralia), ONR, DoD, and contracts with LANL,Raytheon Co., General Dynamics, and ATCCorp. She serves on the editorial boards ofComputer Networks, Computer Communica-

tions, and the International Journal of Communication Systems and on thetechnical program and organizing committees of several major confer-ences including IEEE Infocom, ACM Mobicom, and ACM MSWiM, among

many others. Her research interests include MAC and higher layer pro-tocols for multi-hop wireless networks.