An information theoretic framework for predictive channel reservation in VoIP over GPRS

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMSInt. J. Commun. Syst. 2006; 19:463–489Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/dac.785

An information theoretic framework for predictive channelreservation in VoIP over GPRS

Abhishek Roy1,n,y, Kalyan Basu2,z and Navrati Saxena3,}

1Wireless India Team, Conexant Systems India Inc., India2CReWMaN lab, Computer Science Department, The University of Texas at Arlington, U.S.A.

3Amity School of Computer Science, Amity University, India

SUMMARY

The wireless telecommunication industry is now slowly shifting its paradigm from traditional, circuit-switched, voice-alone domain to an integrated packet-switched architecture. This will give rise to a varietyof new applications, in the same infrastructure, in a cost-efficient way. However, due to the delay anddilemma behind new 3G mobile applications, it is important to use the legacy 2:5G access systems as muchas possible, to make the transition smooth. The recent industry-wide trend towards push-to-talk voiceservices in GPRS networks is a direct consequence of this emerging packet-switch services. In this paper,we propose a packet-switched-based architectural framework for efficient integrated push-to-talk voice (orVoIP) and data services in GPRS using low-bit-rate coding. The prime novelty and advantage of theframework lies in proposing new intelligent, advanced channel reservation techniques to reduce the voice-packet delay. Subsequent use of packet-classification and packet assembly scheme aids in reducing thepacket overhead and achieving the voice quality within ITU’s recommendations. The mutual effects of dataand voice packets over the entire system is analysed using suitable, two-stage performance modelling. It hasalso been shown that, for voice over IP (VoIP) services, our proposed framework results in more than 50%capacity gain over current GSM system using a silent-detection mechanism. Copyright # 2006 John Wiley& Sons, Ltd.

KEY WORDS: VoIP; GPRS; information theory; entropy; LZ-coding; Bayesian learning; on–off sources;imbedded Markov chain

1. INTRODUCTION

It has been a decade since the vision of real-time services emerged as the future of currentgeneration wireless networks. Rapid rise in wireless data services has already resulted in

Contract/grant sponsor: NORTEL Networks

Contract/grant sponsor: Computer Science Department of The University of Texas, Arlington

Received 1 March 2005Revised 1 October 2005

Accepted 1 October 2005Copyright # 2006 John Wiley & Sons, Ltd.

yE-mail: [email protected]

nCorrespondence to: A. Roy, C/o Dr. A. K. Saxena, 7/182 Swaroop Nagar, Kanpur 208 002, UP, India.

zE-mail: [email protected]}E-mail: [email protected]

migration from traditional circuit-switched telecommunication networks to packet-switchednetworks [1]. This will provide vast opportunities for new, cost-effective real-time services. Theproliferation of IETF (Internet engineering task force) standardized protocols and thespeculation behind the deployment of UMTS (Universal Mobile Telecommunications System)[2] is probably the first step towards this movement. It is expected that the evolving architectureswould bring in the virtues of packet-switched technologies on the already deployed, costlywireless telecommunication infrastructure. Recent statistics reveal that almost two-third oftoday’s mobile phones are of GSM (global system for mobile communications) standards,which support circuit-switched voice and short messaging services (SMS). But the everincreasing popularity of Internet usage, along with the voice traffic resulting from mobilephones, is exacerbating these existing circuit-switched telephony systems, by placing an immensetraffic burden. In order to reduce this burden, while improving the data capability in the circuit-switched GSM system, new GPRS (general packet radio services) standards [3, 4], with datarates at 128þ kbps; has been proposed. However, the benefit of packet-switched-basednetworks cannot be fully harnessed without successful integration of voice and data servicesunder the same networks. The recent upsurge of push-to-talk services in GPRS is a directconsequent of this realization. Emerging standards and protocols like H.323 [5] and SIP (sessioninitiation protocol) [6] for session establishment are catalyst for such voice–data integration [7].

In this paper first we have proposed a packet-switched architecture based on the subsystemsproposed in UMTS specification by 3GPP [8]. These subsystems are designed for push-to-talkvoice (or VoIP) and data services. It leverages the use of relevant, existing features of GPRSarchitecture and proposes new, intelligent, advanced channel reservation schemes to ensuresuitable QoS guarantee in voice over IP (VoIP) services. In order to get an optimal trade-offbetween voice-packet delay and overhead, we have used packet classification, low-bit-rate AMR(adaptive multi-rate) coding and packet-bundling mechanisms in the traffic plane. The push-to-talk voice packets are given higher priority over any in-coming data packets in the GPRSPCU (packet control unit). Our contribution is to show that the integration of these techniquestogether with the new intelligent PCU slot-allocation strategy, based on advanced channelreservation can meet the stringent quality of service requirements of voice packets. Performanceanalysis and simulation results identify that the voice-packet delay over the wireless links can bekept within 40–45 ms; thus resulting in an end-to-end delay within ITU’s recommendationð300 msÞ: Finally we demonstrate that for VoIP service, our proposed framework, leads toalmost 50% capacity gain over existing GSM radio spectrum, using silent-detection scheme.More precisely our contribution includes:

1. In order to resolve the conflicting requirements of voice and data traffic, we propose the useof a classifier to differentiate the two types of packets. The regular data packets will becarried over radio link protocol (RLP). The automatic repeat request (ARQ) is used tomaintain the integrity and low bit error rate (BER) required for data packets. However,RLP and ARQ increases the delay of the real-time push-to-talk voice packets. Hence, thevoice packets will bypass the RLP layer. Moreover, the real-time voice packets suffer fromhigh overhead. The addition of header at every layer of the network increases the overallheader-size, thereby putting increasing burden of the actual voice payload. In order toreduce this overhead, we have used the concept of packet bundling. Bundling combinesmultiple packets into a frame, thereby reducing the overall overhead associated with thevoice packets.

Copyright # 2006 John Wiley & Sons, Ltd. Int. J. Commun. Syst. 2006; 19:463–489

A. ROY, K. BASU AND N. SAXENA464

2. Upon receiving of a voice frame, a channel-allocation request is sent. The actual transferof the voice packets can begin only after receiving the response of this channel-allocationrequest. Actually, this channel is allocated in terms of PCU slots. A close look into thisscenario reveals that the inherent delay in the voice-packet transmission can bereduced by minimizing the delay between this request–response message. One majorcontribution of our work lies in the development of an early channel reservation schemeto achieve this reduction in delay. We have proposed two different predictive strategiesfor early channel reservation in GPRS VoIP services. While the proposed Bayesianpredictor operates on i.i.d. process, the optimal LZ-predictor operates on the higher-ordercontext models to minimize the overall channel-allocation delay. Although, thesetechniques have been used in the context of information theory and mobility manage-ment [9, 10] according to our knowledge we are the first to apply such informationtheoretic optimal algorithms in the context of channel reservation in real wireless standardslike GPRS.

3. In order to estimate the overall capacity of the system, a suitable two-stage modellingtechnique is developed. The overall dynamics of the delay associated with push-to-talkvoice and data packets are analysed using a two-dimensional imbedded Markov chain.Simulation results point out that using a suitable combination of early channel reservation,packet classification and packet-bundling, the voice-packet delay could be kept within theITU’s end-to-end delay budget. These experiments also demonstrate that using suitablesilent-detection scheme, the capacity of the overall system can be improved by almost 50%;while maintaining almost similar voice quality. Abridged versions [11, 12] of this work arealready published in IEEE International Conference on Global Mobile Congress, 2004 andInternational Teletraffic Conference (ITC) 2005.

The rest of the paper is organized as follows. Section 2 highlights the state-of-the-art works invoice–data communication over GPRS. The proposed architectural framework for emerging,packet-switched, VoIP services over GPRS is described in Section 3. Section 4 discusses thenewly proposed advanced GPRS channel reservation techniques for VoIP services. Performanceanalysis of overall packet delay and blocking is presented in Section 5. Subsequently, weperform simulation experiments and show the results of early channel reservation strategy,overall packet delay and blocking in Section 6.2. Section 7 concludes the paper with pointers tofuture researches.

2. BACKGROUND AND EXISTING WORKS

The concept of VoIP over wireless networks also simplifies its architectural complexity andreduces the voice transmission cost. Since the last few years, the wireless service providers aretrying to explore this concept to provide different integrated multimedia (voice, data and video)services at reasonable cost. Subsequently, research and development along this voice-dataintegrated services are also on the way. With uncertainties over the success of WCDMA indelivering advanced mobile data and multimedia services, mobile operators across the world areincreasingly faced with the dilemma of adopting 3G technologies [13]. Thus, 2:5G legacysystems, like GPRS remains as a viable, popular system for data communication over wireless.For example, NTT DoCoMo’s 3G FOMA [14] service has been adopted by just 2% of its


INFORMATION THEORETIC PREDICTIVE CHANNEL RESERVATION IN GPRS-VoIP 465

cellular customers, some three years after commercial launch. This contrasts sharply with therapid growth to 40 million subscribers of its 2:5G GPRS services [13]. Hence, research scientistsand service providers are looking into the possibilities of exploring integrated voice–dataservices over GPRS.

It has been already shown in [15] that the VoIP over GPRS has more capacity than circuitswitch voice, while guaranteeing acceptable QoS. The paper also demonstrates that with a singleradio frequency, the circuit switch system can provide 7 voice connections, while VoIP overGPRS can support 12–34 voice connections, using the same 13 kbps speech coder. A newnetwork element called VoIP Mobile Switching Center (VMSC) is introduced in [16] to replacestandard GSM-MSC. This work uses legacy H.323, GPRS, and GSM protocols and describesthe message flows for vGPRS registration, call origination, release and termination. Industry-wide emphasis over VoIP and push-to-talk services are also noticeable. For example, Sonim andTexas Instruments collaborate to create on-chip voice over GPRS solutions [17]. Similarly,Ericsson, Nokia and Siemens are jointly defining specifications for an open standard to speedthe adoption of direct-call push-to-talk service over GPRS [13]. However, multimediatransmission like voice needs to adhere specific QoS guarantee in terms of delay and blocking.Advanced channel reservation is a major step, which possesses the capability to reduce this delayand blocking. According to our knowledge, none of the existing works on GPRS has discussedon this channel reservation.

3. ARCHITECTURAL FRAMEWORK FOR VoIP IN GPRS

The proposed packet-switched architecture for integrated voice and data services operates intwo planes: the control plane and the traffic plane. Figure 1 shows the essential components ofthese two planes. The control plane is built over the IP multimedia subsystems (IMS) of theUMTS Release 5 architecture proposed by 3GPP [8]. The architectural and performance detailsof the control plane is out of the scope of this paper. We focus on the architectural components,intelligent algorithms, performance analysis and results in the traffic plane. Figure 2 delineatesthe protocol stack used in the architecture. The voice packets use the UDP and IP over wirelessstack of the mobile terminal (user equipment). On the other hand, the data applications arecarried over popular TCP/IP protocol stack. The control signal is propagated using SIP and

Control Plane

UMTSIP Multimedia

Subsystem

GPRS AccessNetwork

SIP based control

Proposed Framework

Voice Data

Traffic Plane

Figure 1. Architectural framework.



SDP signalling protocols. From the mobile terminal, the voice and data packets communicatewith the TCP/IP (or UDP) over wireline stack of the core networks using BSS/SGSN/GGSN ofGPRS.

Since we are considering integrated voice and data services, satisfying the QoS requirements,like, maximum voice-packet transmission delay and blocking are of prime concern. ITU’s E:750recommends an end-to-end voice-packet transmission delay should be kept within 300 ms [18].One major challenge behind the development of an integrated, packet-switched framework forwireless voice and data services is the conflicting requirements of these two types of packets.While, the data packets are delay tolerant and require low BER, the real-time voice packets arevery delay-sensitive and can tolerate BER. We have introduced the concept of packetclassification to separate the voice and data packets. The data packets are carried over the RLP[4] layer using ARQ, and the real-time, higher priority, voice packets are transmitted withoutany ARQ. For voice coding, we have chosen the fixed rate AMR-coding scheme [3, 4] for itshigh compression ratio and good voice quality. However, this AMR voice-coding introduceslarge overhead because, the low coding rates reduce the coded voice bits, but the associatedheader-size remains the same. Let r; f ; F and b; respectively, denote the rate, frame time, frame-size and bundling-factor in voice coding. If HIP; HUDP; HRTP and HGPRS represent the IP, UDP,RTP and GPRS headers, respectively, then the resulting GPRS packet size PGPRS and theassociated overhead for conversion of voice frames on UDP packets are given by PGPRS ¼F :bþ ðHIP þHUDP þHRTP þHGPRSÞ and overhead ¼ ½ðPGPRS � F :bÞ=PGPRS� � 100%: Inorder to reduce this huge overhead, we have used the concept of packet bundling [19]}groupinga number of voice packets together to make one UDP payload.

Control

TCP/RTP UDP

SIPServers

Appl

WAN

ControlDataVoiceVoiceDataAppl

SIP

SDP SDP

SIP

TCP/RTP UDP TCP/RTPUDP

IP IP IP

Wireline Stack Wireless Stack

Wireline Stack Wireless Stack

VoIP-enabled GPRS Mobile Terminal GGSN / SGSN/ BSSPacket Core Network

Figure 2. Protocol stack for the architecture.



Unfortunately, bundling results in additional delay. Table I shows the overhead penalty ofAMRC (adaptive multi-rate codes). As the delay reduces from 110 ms to only 20 ms; theoverhead increases from 95 to 400%: Thus, a suitable trade-off between delay and overheadneeds to be specified for voice transmission. The connection of a GPRS terminal includesnetwork components like wireless terminals, channels, access points, media gateways andde-jitter buffer. Considering the AMR coding with CS2 scheme [18] (inherent error correctionmechanisms), and assuming average values of the above-mentioned delay parameters as: 100 ms(including a 30 ms interleaving delay), 90, 85, 80, 100 ms; an estimation of total delay alongthese components is � 455 ms: Thus, even without considering Internet delay, this delay ismuch more than 300 ms (ITU’s specifications). Although it has been shown that Internet delayvaries between 100 ms to 1 s [20], emerging high-speed technologies like GMPLS [21] providesefficient traffic-engineering to reduce this Internet delay to close 20 ms; with suitable bufferallocation and management [22]. Thus, the potential areas of improvement now lies in voice-coding, bundling and GPRS channel-allocation strategies.

4. ADVANCED CHANNEL RESERVATION IN GPRS

Figure 3 provides the messaging diagram for coding, bundling and GPRS slot-allocation schemeduring voice-packet transmission. The GPRS channel module fragments the coded and bundledPDUs (packet data units) into a number of frames to transfer to BS (base station) using theGPRS-slots assigned by the TBF (temporary buffer flow)-allocation process [4]. At certaininstant, GPRS terminal has to send a TBF request to the BS by using the dynamic controlchannel as shown in Figure 3. The major delay at this point arises from TBF-response time,PCU queuing delay and delay in packet transmission. While the queuing delay and packettransmission time are dependent on the actual packet transfer mechanism, we propose a newadvance channel reservation (advance PCU request) to minimize the response delay. Intuitively,it is clear that the mobile needs to be cognizant of the time T it should send the TBF request sothat it receives TBF response as soon as the packets are ready for transmission, thereby reducingthe effective waiting time for TBF response becomes close to zero.

In order to be cognizant about the time, the mobile needs to be aware of the patternsassociated with the voice packets generated by the user. From a symbolic domain, we assumethe time instances of channel request generation is given by T1; T2; . . . ;Tn: The history of thisTBF request generation is thus a string, T1;T2; . . . ; of symbols (time instances). We argue thatthe current TBF request-generation time is merely a reflection of the history, which can belearned over time in an online fashion. Characterizing this TBF-request time instants as aprobabilistic sequence suggests that it can be formulated as a stochastic process T ¼ fTig;where the repetitive nature of identical patterns in the time instances adds stationarity as an

Table I. Overhead-delay trade-off for frame bundling.

Frames/bundle 1 2 3 4 5

Overhead (%) 400 200 150 120 95Delay (ms) 20 40 60 95 110



essential property, thereby leading into the following relation:

Pr½Ti ¼ ti� ¼ Pr½Tiþl ¼ ti� ð1Þ

for any shift l: A close look into this scenario reveals that the voice packets generated by theuser’s speech actually creates an uncertainty of time instant Ti: The more random the speechpattern, the more uncertain is the time instant Ti: The objective of the mobile is to reduce thisuncertainty to make an exact guess (prediction) of this time instant Ti: According to ourknowledge the concept of entropy in information theory provides the most fair measure for thisuncertainty.

4.1. Uncertainty and entropy

The traditional definition of entropy is given as follows.

Definition 1The entropy HðXÞ of a discrete random variable X with probability mass function pðxÞ isgiven by

HðXÞ ¼ �Xx

pðxÞlg pðxÞ ð2Þ

For any set of k discrete random variables fT1;T2; . . . ;Tkg; with the distribution given by:pðt1; t2; . . . ; tkÞ ¼ Pr½T1 ¼ t1;T2 ¼ t2; . . . ;Tk ¼ tk�; the joint entropy is given by

HðT1;T2; . . . ;TkÞ ¼Xki¼1

HðTijT1;T2; . . . ;Ti�1Þ

NIL

NUB

D

G

TBF Allocation

GPRS Frame Time

Response Time

TBF RequestTactual

Tp

Tres

Response Time

PCU : Packet Control Unit

Waiting Time

TerminalCodecUser

TalkBurst

Early TBF (Topt)

Voice Packet Transfer

Figure 3. Messaging diagram of voice-packet transmission over GPRS network



where

HðTi jT1Þ ¼ pðt1ÞHðTijT1 ¼ t1Þ ð3Þ

HðTi jT1Þ is termed as conditional entropy.

This motivates us to look for a smart predictive framework which minimizes the uncertaintyassociated with the system. While most predictive learning algorithms attempts to maximize theprediction success, they operates on different models based on different orders.

* In the order-0 model, all the variables are assumed to be independently identicallydistributed (i.i.d.). In other words, the variables (symbols) are assumed to have probabilityproportional to their relative frequencies.

* In order-1, the probability of a state is dependent only on its previous state and not on anyother state. This is indeed the popular Markov model.

* In higher-order (ith order) models the probability of a state depends on a finite number (i)of previous states. The additive terms in Equation (3) points out that the higher-ordermodels are more information rich than the lower-order models.

All we now need is a way to automatically arrive at the appropriate order dictated by the inputsequence. Fortunately, this is exactly what the class of LZ-78 text-compression algorithms [23]achieves. In order to investigate into the two opposite poles of the predictive learning scheme forchannel allocation, we have proposed to different predictive algorithms. While the Bayesianalgorithm [24] is pretty simple and treats the time instants as i.i.d. process, the Lempel–Ziv (LZ)algorithm [23] actually works on the appropriate order to make a more intelligent, efficientprediction, at the cost of more complexity.

4.2. Bayesian learning for channel reservation

We propose an online learning mechanism based on Bayesian aggregation technique [24] tomake the mobile knowledgeable enough to transmit the TBF request at exact time. The entiretime frame T is now divided into small time units T1;T2; . . . ;Tn of 1 ms: Each time unit Ti isassigned to a weight wi: Initially, all the weights are same, i.e. wi ¼ 1=n: The mobile nowestimates the future-TBF-request time by computing an weighted measure of these timeinstances. Formally, the predicted time Tp for TBF request is

Tp ¼Pn

i¼1 wi � TiPni¼1 wi

ð4Þ

The objective of the scheme is to send the TBF request early, so that the TBF response isobtained exactly when the coded voice packets get ready for transmission over the channels.According to this Bayesian channel reservation strategy, let, the mobile predicts to transmit theTBF request at time Tp and also estimates to receive the TBF response at time Tres: Also, let,Tactual represents the actual time when the coded voice packets gets ready for transmission.Thus, the objective is to make the system cognizant enough so that the TBF-response timecoincides with the time when voice packets get ready for transmission, i.e. Tres ¼ Tactual: At thiscondition the framework achieves optimal time for TBF request, i.e. Topt ¼ Tp: The most fairmeasure of the loss (error) associated with this process is entropic loss. This entropic loss li for



each time instant and its cumulative estimate in m iterations lci is given by

li ¼ �Toptln Ti � ð1� ToptÞlnð1� TiÞ

lci ¼Xmi¼1

li ð5Þ

At every iteration m; the weights associated with every time instant Ti are now updated as

wmþ1i ¼ wm

i e�li ð6Þ

This procedure is iterated until Tp � Tactual4E; where E is a predefined precision. The basicunderlying intuition behind this entire scheme is to probabilistically cluster the weights towardsthe optimal time instant Topt: The effects of time instants far from Topt are reduced byexponentially decreasing their associated weights. In the long run, the predicted time instant Tp

approaches Topt: The overall expected deviation form the optimality (lo) is given by thefollowing equation

lo4mini

lci þ lnðmÞ ð7Þ

Figure 4 provides a pseudo-code of this entire advanced TBF-request scheme.

4.3. Early channel reservation based on LZ-prediction

The optimal LZ text-compression schemes helps to reduce the cost of information acquisition byprocessing the symbols (time instants) in chunks. The entire sequence of sampled symbolswithheld since the last reporting is reported in an encoded form. Thus, the time-instant ‘t1t2t3:::’reaches the profile server as Cðo1Þ; Cðo2ÞCðo3Þ; where oi is the non-overlapping, distinctsegments of string ‘t1t2t3:::’ and CðoiÞ is the encoding for segment oi: For example, the inputstring ‘t1t2t4t4t5t5t2t3t3t1t1t2t4t4t5t5t2t1t1t2t4t4t5t5t2t1t1t2t4t4 . . .’ is parsed as distinct substrings(phrases): ‘t1; t2; t4; t4t5; t5; t2t3; t3; t1t1; t2t4; t4t5t5; t2t1; t1t2; t4t4; t5t5; t2t1t1; t2t4t4; . . . ;’. Such asymbolwise context model can be efficiently stored in a dictionary implemented as a search trie.Figure 5 shows these different phrases with their frequencies, where the frequency of every

Figure 4. Algorithm for advance TBF request using Bayesian learning strategy.



symbol is incremented for every prefix of every suffix of each phrase [9, 10]. The incrementalparsing accumulates larger and larger phrases in the dictionary, thereby accruing estimates ofhigher-order conditional probabilities and asymptotically outperforming any finite-orderMarkov model. Essentially, the algorithm approaches optimality for stationary ergodicstochastic process. Figures 6 and 7 show the encoder and decoder side of the predictivemechanism. The encoder resides at the GPRS mobile, while the decoder resides in the BS.

t (7)1 t (7)2

t (2)3

t (8)4t (6)5

t (2)1

t (1)2 t (2)1 t (2)4 t (1)

3t (2)

5t (2)4

t (2)5

t (1)1t (1)4 t (1)

5

Λ

Figure 5. Decoder trie for LZ-prediction-based channel reservation in the GPRS system.

Figure 6. Encoder at the mobile.

Figure 7. Decoder at the system.



Maintaining the inhabitants context in such a trie helps in efficient computation of theprobabilities of the different phrases (c). Following a PPM-style blending technique [21], ourprediction mechanism starts from the highest order of context (leaf of the trie) and escape tolower orders until order-0 (root) is reached. The probabilities of individual symbols (time-instants) are computed based on their relative weights on the phrase. The symbol (time instant)having maximum probability is chosen by the GPRS-PCU for sending the TBF request. Figure8 provides the pseudo-code of this probabilistic prediction scheme from the trie.

5. PERFORMANCE MODELLING AND ANALYSIS

As discussed earlier in Section 3, the PCU of the GPRS performs TBF allocation for packetstransmission. Sharing of multiple terminals by traffic sources leads to the voice-packet blocking,which is dependent on the number of active GPRS terminals in a cell and the intensity of thepackets generated. The burst-level blocking and delay analysis of packets is necessary toestimate the QoS effectiveness of voice and data packets. The contribution of this section lies indeveloping a new two-stage channel modelling to get an estimate of the end-to-end delay andblocking of the voice packets in GPRS. The effects data traffic over voice QoS is alsoinvestigated.

5.1. GPRS burst-level blocking

The channel blocking probabilities are determined by using the two-stage modelling technique.In the first stage, the probability of the number of active sources in the system is determinedusing a Markov process with quasi-random arrival. Figure 9 depicts this scenario, where everystate represents the number of active terminals in the cell.

Figure 8. Probabilistic LZ-78 prediction scheme for early channel reservation.



Let, nt; lt; 1=mt and P #kv ; respectively, represents the total number of GPRS terminals, arrivalrate of the free voice traffic source, mean service time of one source, and probability of #kv activevoice sources in the system. In the equilibrium state, for #kv terminals in the system, usingfamiliar Engset-loss model [25], we have

P #kv ¼ P0ltmt

� � #kv nt

#kv

!¼

nt

#kv

!g #kv

P #kv|¼0

nt

|

!g |

ð8Þ

where g ¼ lt=mt: The accepted voice sessions generate the voice bursts. The probabilities of thesimultaneous voice bursts of the active sources are estimated by considering the ‘on/off’ burstsource model with the individual sources as a Bernoulli process. Figure 10 depicts this ‘on–off’modelling, where lv and mv denote the transition probabilities between ‘on’ and ‘off’ states. If Rvdenotes the probability of being in ‘on’ state, then Rv ¼ mv=ðlv þ mvÞ: Hence, among i activeusers in the system, the probability that #kv voice sources will be in the ‘on’ state, at any instant, isgiven by

Pi #kv ¼i

#kv

!R #kvv ð1� RvÞ

i� #kv ð9Þ

There are 8 slots in the GPRS system with 200 kHz frequency channel in a cell. We assume thatvoice payload from one user can be mapped to one slot of this GPRS channel. The jointprobability that i active users in the system and #kv traffic bursts are generated is Pi �Pi #kv :Hence, the probabilities of blocking of the voice bursts is given by

Pr½B� ¼ 1�Xnti¼0

XMin½i;8�

#kv¼0

Pi �Pi #kv ð10Þ

0 1 2nt

µ 2µ 3µ nµ tt t t

ntλt (nt-1)λt (nt-2)λtλ t

Figure 9. State model for active voice sources.

OFFON

λv

µv

Figure 10. On–off source model for voice traffic.



5.2. Analysis of packet delay dynamics

The voice and data burst requests are received by the PCU from dynamic control channel tobuild the PCU queue, which is serviced in the discrete frame time. During each frame time, 8slots are allocated by the PCU. We assume that PCU is using FCFS scheduling discipline withvoice packets given higher priority over any in-bound data packets. Thus, a data packetis allocated a slot only when all the voice packets are serviced. This slot allocation iscommunicated by the PCU by using TBF allocation. The associated PDCHs (Packet DataChannels) of GPRS use slotted-ALOHA for packet transmission. In these cases, the behaviourof the system can be determined by looking at state of the queues only at the departure instantsof the packets, thus leading to an imbedded, discrete, Markov Process. Defining the state spaceof the burst requests in the system by tuples ðr; r0Þ 2 I (integer), where both r and r0 are countablyfinite (number of voice and data bursts cannot be more than the number of active ‘on–off’ voiceand data sources), then this Markov Process is a two-dimensional,discrete, imbedded Markovchain. Figure 11 demonstrates this, where the voice- and data-packet dynamics are represented,respectively, along the horizontal and the vertical axis. Initially, the system was in state ðr; r0Þ:During the previous frame duration min½m; r� voice bursts are served, where m representsnumber of times GPRS slots are allocated to a voice burst. During this frame time, new

kr’+1

k

r’k

k

0

k

0 r−m r−1 r r+1

r’+1 r’+1 r’+1 r’+1 r’+1

0 r−m r−1 r r+1

r’0

r’ r’ r’ r’r−m r−1 r r+1

0 r−m r−1 r r+1

0 0 0 0 0

k k k k k

data

tran

smis

sion

voice transmission

Figure 11. Imbedded Markov model of voice- and data-packet delay dynamics.



voice-bursts requests will arrive from ð #kv � rÞ active voice sources in the system. The PCUalgorithm can use different slot-allocation strategies for a voice burst, like allocating 1; 2 or 3slots per burst per frame. The arrival of the new voice bursts and the service of existing voice-bursts cause the horizontal transition from the current to the new states. Similarly, arrival ofdata packets from ð #kd � r0Þ active data sources and their service changes the states vertically.However, it should be noted that, since the data packets are given lower priority, service of datapackets happen only when excess GPRS slots are left after servicing the real-time voice packets.Hence, the data-packet service at any instance, is shown by dotted arrow in Figure 11. Thesedata packets are assumed to get one slot per packet.

Let Z denotes the number of slots allocated per voice burst. Then, as shown in Figure 11, thepossible destination states for horizontal transitions from state ðr; r0Þ are: Sv ¼ fðr�m; r0Þ;ðr�mþ 1; r0Þ; . . . ; ðr� 1; r0Þ; ðr; r0Þ; ðrþ 1; r0Þ; . . . ; ð #kv; r0Þg; where m ¼ 8=Z ¼ f8; 4; 2g: If Wv ¼½ #kv �maxð0; r�mÞ þ 1�; then the horizontal transition probability from state ðr; r0Þ to anyother state ðs; r0Þ is estimated as

}ðr;r0Þ;ðs;r0Þ ¼

Wv

s

!Rsvð1� RvÞ

Wv�s 8s 2Sv

0 otherwise

8>><>>: ð11Þ

Now, for data packets, two types of situations arise. If the number of voice bursts r58; thenðr0 � rÞ data packets are serviced, with possible destination states for vertical transitions beingSd ¼ fðr; r0 � 8Þ; ðr; r0 � 7Þ; . . . ; ðr; r0Þ; ðr; r0 þ 1Þ; . . . ; ðr; #kdÞg: On the other hand, if r58; thennone of the data packets can be serviced, thereby leading to S0d ¼ fðr; r

0Þ; ðr; r0 þ 1Þ; . . . ; ðr; #kd Þg:Hence, the vertical transition probability from state ðr; r0Þ to any other state ðr; s0Þ is given by

}ðr;r0Þ;ðr;s0Þ ¼

Wd

s0

!Rs0

d ð1� RdÞWd�s0 8r58

W0d

s0

!Rs0

d ð1� RdÞW0d�s

08s05r0

0 otherwise

8>>>>>>>><>>>>>>>>:

ð12Þ

where Wd ¼ ½ #kd �maxf0; r0 � ð8� rÞg þ 1� and W0d ¼ ½ #kd � r0 þ 1�: The steady-state probabilitiesfor a set of a total #k ¼ #kv þ #kd voice and data bursts are obtained by solving the equation:

%p #kðiÞ;ð |Þ ¼ %p #k

ðiÞ;ð |ÞP ð13Þ

where P represents the transition probability matrix of the imbedded Markov chain. Thisprocedure is repeated for all possible set of bursts B; such that 04jBj4 #k: As compared to thesession holding time and the inter-arrival time of the new session requests, the GPRS frameduration (burst service time) is very small. Hence, the regeneration points of the imbeddedMarkov chain occur as a micro-dynamics during the inter-arrival time and call holding time.Thus, we have considered the steady state probabilities to be independent of P #k (probability ofhaving #kv active voice and #kd active data sources), and the final, conditional steady-state



probabilities are obtained by solving the equation given below:

pðiÞ;ð |Þ ¼X‘¼ð #k¼ #kvþ #kd Þ

ð‘¼iþ|Þ¼1

P‘ � p‘ðiÞ;ð |Þ ð14Þ

If %tGPRS denotes the average GPRS frame-time, then the associated PCU-slot-allocation delayfor #kv voice and #kd data sources producing r voice and r0 data packets will be given as:

d #kvr ¼

r

Z� %tGPRS; d #kd

r0 ¼ r0 � %tGPRS ð15Þ

The average PCU-slot-allocation delay for all #kv voice and #kd sources are obtained as

d #kvavrg ¼

X#kvr¼0

d #kvr �

X#kd�r|¼0

pðrÞ;ð |Þ

!ð16Þ

d #kdavrg ¼

X#kdr0¼0

d #kdr0 �

X#kv�r0i¼0

pðiÞ;ðr0Þ

!ð17Þ

Assuming that advanced TBF request is sent in exact time to minimize the delay betweenbundling and TBF allocation, the average delay generated during coding and bundling of nvoice packets will be ðn� 1Þ �T: Hence, total average voice packet and data-packet delaywill be

dvtot ¼ ðn� 1ÞTþM

Z� 1

� �Tþ d #kv

avrg ð18Þ

ddtot ¼M

Z� 1

� �Tþ d #kd

avrg ð19Þ

where M and T ¼ 20 ms; respectively, represents number of PCU slots and AMR frame-time.

6. SIMULATION EXPERIMENTS AND RESULTS

In this Section, we first highlight the basic assumptions we have used for experiments of thepredictive channel reservation scheme in GPRS VoIP services. Subsequently, we provide a seriesof performance results to show the efficiency of the early channel reservation strategies. At theend, we show the overall blocking and delay in the network to support our two-stageperformance modelling scheme described in the previous section.

6.1. Assumptions and framework

The prime objective is to learn the profiles of the voice and data packets arriving from themobile users. Based on this profile, the advanced channel (PCU) reservation request is sent.The goal is to reduce the difference between the time when the channel response is obtained andthe voice–data packets are actually transmitted. The learning and channel reservation scheme iscontinued for 1 h ð3600 sÞ:



1. An event-driven simulation framework is developed to perform experiments for validatingour analysis and algorithm of early channel reservation.

2. We have used AMR Vocoder for encoding and decoding of voice signals into voicepackets. The coder operates GSM full-rate channel mode at a total bit rate of 22:8 kbps:The available bits are allocated to source coding ð11:85 kbpsÞ; channel coding andsignalling ð10:95 kbpsÞ: The data is collected in accordance with the GSM/AMRqualification test plan. All tests are performed using mean opinion of score (MOS) rate.

3. A voice call shows periods of activity and inactivity. Over the years, the duration of activityand inactivity are generally assumed to obey exponential distribution [26–30]. Subse-quently, the generation of voice bursts (a collection of packets) is assumed to be Poissondistributed [26–30]. Two different types of speech samples are considered: interactive andbusiness. The interactive speech samples have exponentially distributed ‘on’ and ‘off’period with mean lengths 1026 ms and 1171 ms; respectively [27]. On the other hand, thebusiness speech samples have exponentially distributed ‘on’ and ‘off’ period with mean1490 ms and 1722 ms; respectively [27]. Simulation results are evaluated for up to nt ¼ 400users with average traffic intensity b ¼ ½0:05; 0:1� erlang.

We now provide a series of simulation results of our newly proposed predictive channelreservation strategies for both interactive, business and mix speech samples.

6.2. Results of early channel reservation

The smartness, efficiency and intelligence of any predictive framework is generally measured byits predictive accuracy. This predictive accuracy is nothing but the over-reservation due toincorrect predictions. Figure 12 provides the dynamics of the predictive accuracy of ourproposed early channel reservation schemes. The results point out that LZ-78-based predictiveframework results in almost � 90% and � 97% accuracy for business and interactive speech.The system becomes cognizant of the user’s patterns within 200–300 s: After 500 s the system

0 500 1000 1500 2000 2500 3000 350010

20

30

40

50

60

70

80

90

100

Time (Seconds)

Pre

dict

ive

Acc

urac

y (%

)

Interactive Speech with LZ-78 CodingBusiness Speech with LZ-78 COdingInteractive Speech with Bayesian AlgorithmBusiness Speech with Bayesian Algorithm

Figure 12. Prediction accuracy for interactive and business speech.



becomes stabilized. This 200–500 s is actually the learning-stage of the system. Note that, inorder to make the system realistic the learning and prediction process are going online, withoutany prior training. Thus, the system takes some time � 200–500 s to attain stability. Theaccuracy of the Bayesian learning framework is relatively less (� 80–� 85%). Intuitively, thiscan be clear from the fact that the LZ-predictor actually acquires the knowledge of higher-ordercontext models and their associated dependencies, which the Bayesian predictor is ignorant of.To further investigate into the predictive framework, we have evaluated our algorithm over amixture of interactive and business speech, i.e. business speech followed by interactive speechand vice versa. Figure 13 demonstrates that the LZ-78-based predictive framework provides anaccuracy close to 87–90% for both interactive-business and business-interactive speech. TheBayesian prediction algorithm, on the other hand, provides a success rate of 75–80%.

Figure 14 shows the reduction in channel-allocation delay using our predictive frameworks.Without any early channel reservation scheme the channel-allocation delay would have been� 70 ms: However, the predictive channel reservation scheme based on LZ-prediction has thecapability to reduce the channel-allocation delay to only 15–20 ms for interactive and businessspeech, respectively. The corresponding delay in Bayesian predictive framework is around30–35 ms: The better predictive accuracy of the LZ-predictors results in lower channel-allocation delay than the Bayesian predictor. The results get stabilized quickly, in a span of10–15 min:We now look into the channel-allocation delay with mixed traffic (interactive speechfollowed by business speech and vice versa). Figure 15 shows that the channel-allocation delayfor both the mixed speech samples are initially around � 70 ms: The channel-allocation delaystarts reducing as the system acquires knowledge about the voice samples. The delay forinteractive speech followed by business speech takes a reverse-bell shaped curve. The minimumvalue ð� 20 msÞ of the curve specifies the minimum channel-allocation delay obtained by thescheme for interactive speech. The delay starts increasing again as the delay in the businessspeech is around 40 ms and is always higher than the delay in the interactive speech. However,the channel-allocation delay for the business speech followed by interactive speech continuously

0 500 1000 1500 2000 2500 3000 35000

10

20

30

40

50

60

70

80

90

100

Time (Second)

Pre

dict

ion

Acc

urac

y (%

)

Interactive-Business Speech with LZ-78 CodingBusiness-Interactive Speech with LZ-78 CodingInteractive-Business Speech with Bayesian AlgorithmBusiness-Interactive Speech with Bayesian Algorithm

Figure 13. Prediction accuracy of mixed speech.



reduces up to 20 ms: The Bayesian algorithm achieves a higher delay of almost 10 ms than theLZ-78 algorithm. Table II shows that the confidence interval of all these plots of Figures 14 and15 lie in the range 93–96%.

As discussed earlier, the major goal of the early channel reservation scheme is to reduce thetime difference between the channel-response time and the packet-transmission time. Ideally, ifthe packets are transmitted at exactly the same time, when the channel response is received, thenthe allocation delay is negligible. However, if the channel response arrives sufficiently before thepacket is actually ready for transmission, then the wireless channel is wasted and the packet is

0 500 1000 1500 2000 2500 3000 350010

20

30

40

50

60

70

Time (Second)

Ch

ann

el A

lloca

tio

n D

elay

(m

ili-s

eco

nd

s)

Interactive Speech with LZ-78 CodingBusiness Speech with LZ-78 CodingInteractive Speech with Bayesian AlgorithmBusiness Speech with Bayesian Algorithm

Figure 14. Channel-allocation delay for interactive and business speech.

0 500 1000 1500 2000 2500 3000 350010

20

30

40

50

60

70

80

Time (Second)

Cha

nnel

Allo

catio

n D

elay

(m

sec)


Figure 15. Channel-allocation delay for mix voice.



discarded or dropped. An intelligent predictive framework also attempts to reduce this packet-drop rate. Figure 16 demonstrates that the LZ-predictor starts with an initial packet-loss rate of� 7%: After around 10 min the packet-loss rate saturates at almost � 2% for both interactiveand business speech. The corresponding initial and final packet-loss rate for the Bayesianpredictor lies in the range 14 and 4%; respectively. We now look into the results of packet loss inmixed speech. Figure 17 points out that the LZ-78 algorithm reduces the packet loss of both themixed traffic from 10–12% to almost 1–2%. The Bayesian algorithm, on the other hand, resultsin a packet loss of 6–7%.

At this point of time, we are interested in investigating into the cost of the predictionalgorithms in terms of resource (computation and storage) efficiency. Both the Bayesian andLZ-predictor needs extra computational and storage overhead. Figure 18 depicts the extrastorage overhead associated with the mobile node for running this Bayesian and LZ-predictionstrategies. Although the storage requirement of LZ-predictor is more than that of the Bayesianpredictor, both the predictive frameworks require a modest amount (� 800 bytes) of extramemory in the mobile node.

0 500 1000 1500 2000 2500 3000 35000

2

4

6

8

10

12

14

Time (Second)

Per

cen

tag

e o

f P

acke

t L

oss

Interactive Speech with LZ-78 CodingBusiness Speech with LZ-78 CodingInteractive Speech with Bayesian AlgorithmBusiness Speech with Bayesian Algorithm

Figure 16. Packet loss in interactive and business speech.

Table II. Confidence intervals of channel-allocation delay.

Speech-type Confidence interval (%)

Interactive-LZ-coding 96.8Interactive with Bayesian learning 96.4Business with LZ-coding 95.6Business with Bayesian learning 93.1Interactive-business-mix with LZ-coding 94.1Interactive-business-mix with Bayesian learning 93.6Business-interactive-mix with LZ-coding 93.1Business-interactive-mix with Bayesian learning 93.7



6.3. Overall delay and blocking

We now look into the PCU-slot-allocation delay, overall delay over the wireless links and thetotal end-to-end delay. The PCU-slot-allocation delay is dependent on the different on the typeof slot-allocation strategy used. As discussed earlier, packets can be assigned 1, 2 or 4 GPRSslots. Figure 19 demonstrates that by allocating a single GPRS slot/user, PCU-slot-allocationdelay for 400 users is found to be 10.5 and 18 ms; respectively, for the two traffic intensities.Silent detection reduces this delay further to 3–4 ms:

In order to investigate the effect of different GPRS-slot-allocation strategies into the PCUdelay, the same procedure is repeated with 2 and 4 slots/burst. Analytical results in Figure 20

0 500 1000 1500 2000 2500 3000 35000

2

4

6

8

10

12

14

16

Time (Second)

Pac

ket L

oss

(%)


Figure 17. Packet loss in mixed speech.

0 500 1000 1500 2000 2500 3000 35000

100

200

300

400

500

600

700

800

900

1000

Time (Seconds)

Sto

rag

e R

equ

irem

ents

in t

he

Mo

bile

(B

ytes

)

LZ-78 Coding Scheme Bayesian Algorithm

Figure 18. Storage overhead associated with early channel reservation schemes.



demonstrates that the delay with 400 users for Z ¼ ½2; 4� is around 19; 20 ms and 10; 11 ms;respectively, for the two different traffic intensities. With the silent-detection scheme the delaycan be reduced to 11; 12 ms and 5; 6 ms; respectively.

The total voice-packet delay over the wireless links is now computed for two differentbundling strategies. In 2� 2 and 3� 2 AMR frame bundling, respectively, 2 and 3 AMR framesare grouped together to make an UDP payload and 2 GPRS slots are allocated to this payload.As shown in Figure 21, the average voice packet can be kept between 40–70 ms for therespective bundling strategies. Indeed, this delay includes the channel-allocation delay resulted

0 50 100 150 200 250 300 350 4000

2

4

6

8

10

12

14

16

Number of Users in the System

PC

U-D

elay

(m

sec)

with

1 s

lot /

bur

st

Traffic Intensity = 0.05Traffic Intensity = 0.1Traffic Intensity = 0.05 and Silent-DetectionTraffic Intensity = 0.1 and Silent Detection

Figure 19. PCU delay for voice traffic with 1 GPRS slot/burst.

0 50 100 150 200 250 300 350 4000

5

10

15

20

25


PC

U-D

elay

(m

sec)

with

Incr

easi

ng s

lots

/ bu

rst

4 slots/burst and traffic-intensity = 0.052 slots/burst and traffic-intensity = 0.054 slots/burst and traffic-intensity = 0.12 slots/burst and traffic-intensity = 0.14 slots/burst with silent-detection and traffic-intensity = 0.052 slots/burst with silent-detection and traffic-intensity = 0.054 slots/burst with silent-detection and traffic-intensity = 0.12 slots/burst with silent-detection and traffic-intensity = 0.1

Figure 20. Voice-packet delay with different slot allocations.



from the early channel reservation scheme (using LZ-predictor). Silent detection aids inreducing this delay even further. Clearly, the two bundling schemes result in a difference (� 20–25 msÞ in voice-packet delay over wireless links, and the 2� 2 frame-bundling gains constantlyover 3� 2:

As discussed earlier in Section 5, the data packets are always given lower priority than thereal-time voice packets. The data packets are transmitted over the GPRS channels, only when itis free from active voice sources. Intuitively, thus the data packets will acquire more delay thanthe real-time voice packets. Figure 22 points out that the delay for the data packets over thewireless links can be kept within 90–110 ms:

0 50 100 150 200 250 300 350 400

30

40

50

60

70

80

90

100

110


Ave

rage

Voi

ce-p

acke

t Del

ay (

mse

c)

3x2 Bundling and traffic-intensity = 0.053x2 Bundling and traffic-intensity = 0.12x2 Bundling and traffic-intensity = 0.052x2 Bundling and traffic-intensity = 0.1 3x2 Bundling with silent detection and traffic-intensity = 0.053x2 Bundling with silent detection and traffic-intensity = 0.12x2 Bundling with silent detection and traffic-intensity = 0.052x2 Bundling with silent detection and traffic-intensity = 0.1

Figure 21. Voice-packet delay with different bundling schemes.

0 50 100 150 200 250 300 350 400

80

85

90

95

100

105

110

115


Ave

rage

Del

ay (

mse

c)

0.05 traffic-intensity0.1 traffic-intensity

Figure 22. Data-packet delay dynamics.



Hence, from this delay model, we can conclude that additional overhead penalty of 2-framebundling is not causing significant voice-packet delay penalty to override the delay advantage of2-frame bundling over the 3-frame. Thus, the 2� 2 frame bundling is an optimal choice with adelay advantage of � 20 ms: Using this bundling scheme and assuming the delay componentsassociated with packet-interleaving, media gateway, de-jitter buffer and Internet as describedbefore in Section 3, the total end-to-end voice-packet delay for 300 users can be kept boundedby 300 ms (ITU’s recommendation) (Figure 23). Indeed, the delay associated with elastic datapackets are quite more, but these data packets are delay-tolerant. While the delay of the voicepackets can be kept within ITU’s delay budget, we now want to look into the blocking

0 50 100 150 200 250 300 350 400260

270

280

290

300

310

320

330

340

350

360

370


End

-to-

end

Del

ay (

mse

c)

Data traffic with intensity = 0.05Data traffic with intensity = 0.1Voice traffic with intensity = 0.05voice traffic with intensity = 0.1

ITU’s Recommendation

Figure 23. End-to-end delay for voice and data traffic.

0 50 100 150 200 250 300 350 4000

20

40

60

80

100

120

No of Users

Per

cent

age

of B

lock

ing

Traffic-intensity = 0.05Traffic-intensity = 0.1Traffic-intensity = 0.05 and Silent-detectionTraffic-intensity = 0.1 and Silent-detection

Figure 24. Voice burst blocking.



associated with our proposed strategy. The sharing of different terminals by voice and datasources leads to the packet blocking, which is dependent on the number of active GPRSterminals in a cell and the intensity of the packets generated. As shown in Figure 24, the schemecan support around 100 and 200 users, respectively, over 200 kHz bandwidth spectrum for thesetwo traffic intensities. Using silent detection, the mechanism has the potential of supportingaround 160 and 300 users, respectively, with low blocking probabilities, thus achieving acapacity growth of almost 50% over the current GSM spectrum.

6.4. Effects of data packets on voice QoS

At this point of time, we are interested in investigating into the effects of data packets into thevoice QoS (delay and blocking). Figure 25 shows the effects of both voice and data packets

Figure 25. Combined delay of voice and data packet.

Figure 26. Combined blocking of voice and data packets.



arrival on the overall voice-packet delay. The surface plot indicates that with low data arrivalrate (43 packets/s) the voice-packet delay can be kept considerably below the ITU’s budget of300 ms: However, with higher data arrival, the voice-packet delay starts increasing and it isalmost 300 ms for a voice arrival rate of 10 packets/s and data arrival rate of 5 packets/s.

Similarly, the arrival of data packets considerably affects the blocking of voice packets also(Figure 26). For a lower data arrival rate the voice-packet blocking is very low, almost 43%:However, for higher data arrival rate, the voice blocking is around 7%: Thus, we can concludethat even with sufficient data arrival rate, we can maintain a good voice QoS (blocking 47%and delay 4300 ms). This speaks for the efficiency of our newly proposed early channelreservation framework.

7. CONCLUSION

In this paper we have proposed an emerging architecture for packet-switched-based, integrated,VoIP services over GPRS networks. The architecture uses packet classification, coding andbundling scheme and proposes two different novel, advanced channel reservation schemes tooptimize the voice-packet delay and blocking. The channel reservation schemes have thecapability to reduce the channel-allocation delay with a reasonable memory consumption andpacket-loss rate. Performance modelling and experiments of these early channel reservationschemes, PCU slot-allocation strategies, bundling and packet classification in traffic plane pointsout the effectiveness of our framework to keep the delay of voice packets used in VoIP serviceswithin ITU’s specifications. The system is also capable of achieving more than 50% capacitygain over existing GSM systems, using silent detection. Our future interests lie in investigatinginto the effects of channel reservation scheme in voice over wireless LANs. We hope theseresearches should significantly contribute towards the deployment of VoIP services in nextgeneration, heterogenous wireless systems.

ACKNOWLEDGEMENTS

Initial part of the work was supported by NORTEL Networks, Richardson, Texas and Computer ScienceDepartment of The University of Texas, Arlington. Authors would also like to thank Dr Sajal K. Das forhis initial suggestions during the conference publications.

REFERENCES

1. Dixit S, Prasad R (eds). Wireless IP and Building the Mobile Internet. Artech House Publishers: Norwood, MA,2002.

2. Bos L, Leroy S. Toward an All-IP-based UMTS system architecture. IEEE Network 2001; 15(1):36–45.3. ETSI, GSM 03.60 Digital cellular telecommunications system (Phase 2þ): General Packet Radio Service, Service

Description Stage 1.4. ETSI, GSM 04.60 Digital cellular telecommunications system (Phase 2þ): General Packet Radio Service, Service

Description Stage 2.5. Packet based Multimedia Communication Systems. ITU Recommendation H.323, 1998.6. Handley M, Schulzrinne H, Schooler M, Rosenberg J. SIP: Session initiation protocol, RFC 2543. Internet

Engineering Task Force (IETF), 1999.7. Goode B. Voice over Internet protocol (VoIP). Proceedings of the IEEE 2002; 90(9):1495–1517.8. 3GPP TS 24.228: IP Multimedia (IM) Subsystem Stage 3.



9. Bhattacharya A, Das SK. LeZi-update: an information-theoretic approach for personal mobility tracking in PCSnetworks. ACM-Kluwer Wireless Networks (WINET) 2002; 8(2):121–137.

10. Misra A, Roy A, Das SK. An information-theoretic framework for optimal location tracking in multi-system 4G,wireless networks. Proceedings of the IEEE INFOCOM, Hong Kong, 2004.

11. Roy A, Basu K, Das SK. Advanced channel reservation for end-to-end delay education in low-bit-rate GPRS push-to-talk service. IEEE International Conference on Global Mobile Congress, Shanghai, China, 2004.

12. Roy A, Basu K, Das SK. An information theoretic framework for predictive channel reservation in GPRS push-to-talk service. International Teletraffic Conference (ITC), Beijing, China, 2005.

13. http://www.3g.co.uk14. http://www.nttdocomo.com/15. Shen Q. Performance of VoIP over GPRS. 17th International Conference on Advanced Information Networking and

Applications (AINA), Xi’an, China, 2003.16. Chang M-F, Lin Y-B, Pang A-C. vGPRS: a mechanism for voice over GPRS. ACM/Kluwer Wireless Networks 2003;

9(2):157–164.17. Brundahl K, Silverio R. Sonim and Texas Instruments Collaborate to Create Real-time Voice over GPRS Solutions

for Multiple Chipset Families. Sonim Newsroom. http://www.sonimtech.com/newsroom/18. Introduction to the E.750-series of recommendations on traffic engineering aspects of networks supporting mobile

and UPT services, ITU-T Recommendations, Revision 2, March 2000.19. Cisco Systems. Verifying customer QoS markings based on packet length. Beyond Basic IP 2003; 4(9):1–3.20. Paxson V. End to end Internet packet dynamics. Proceedings of the ACM International Conference on

Communications (SIGCOMM), Cannes, France, 1997; 139–152.21. Berger L. Generalized multi-protocol label switching (GMPLS) signaling resource reservation protocol-traffic

engineering (RSVP-TE) extensions. Request for Comments (RFC) 3473, January 2003.22. Supporting Differentiated Service Classes in Large IP Networks. Juniper White Papers. http://www.juniper.net/

solutions/literature/white papers/.23. Ziv J, Lempel A. Compression of individual sequences via variable-rate coding. IEEE Transactions on Information

Theory 1978; 24(5):530–536.24. Feder M, Freund Y, Mansour Y. Optimal universal learning and prediction of probabilistic concepts. Proceedings of

the International Symposium on Information Theory, Whistler, BC, Canada, 1995.25. Kleinrock L. Queuing Systems, Volume I: Theory. Wiley: New York, 1975.26. Agustn F, Ferrer D, Rosales CV. Performance analysis of cdma2000 MAC layer with voice/data integration. 14th

International Conference on Electronics, Communications, and Computers (CONIELECOMP 2004), Veracruz,Mexico, 16–18 February 2004.

27. Estepa R, Estepa A, Vozmediano JM. Background noise influence on VoIP traffic profile. Multimedia InteractiveProtocols and Systems (MIPS), Lectures Notes in Computer Science (LNCS) Interactive Multimedia and NextGeneration Networks. Springer: Berlin, 2004; 13–22.

28. Gusak O, Oliver N, Sohraby K. Efficient approximation of multiplexed voice traffic. 15th IASTED InternationalConference on Modelling and Simulation (MS 2004), Merinadel Rey, CA, U.S.A., 2004.

29. Sriram K, Whitt W. Characterizing superposition arrival processes in packet multiplexers for voice and data. IEEEJournal on Selected Areas on Communication (JSAC) 1986; 4(6):833–846.

30. Sahinoglu Z, Tekinay S. Traffic engineering for multimedia networks over xDSL access systems. NSF Industry/University Co-Operative Research for Next Generation Video, IAB Conference and Meeting Presentation, 2000.

AUTHORS’ BIOGRAPHIES

Abhishek Roy is currently a Senior Software Engineer in the Wireless India Team ofConexant Systems Inc. He has about 5 years of experience in different aspects ofR&D in wireless telecommunication. He received his MS degree (with Best MSResearch award) in 2002 from the University of Texas at Arlington (U.S.A) and BEdegree in 2000 from Jadavpur University, Calcutta, in Computer Science andEngineering. His research interests include mobility management, quality of services(QoS) provisioning in current and future generation wireless networks, voice over IP(VoIP) and mobile multicasting. He has published over 15 International conferencesand 3 International journals.



Kalyan Basu is the managing director of the UTA’s Center for Research in WirelessMobility and Networking Lab. He has more than 30 years of telecommunicationindustry experience, of which last 27 years is on telecommunication system research.Since 1980 he has worked at NORTEL Telecommunications Research Division, firstas manager of performance group at Bell Northern Research, Ottawa, Canada, andthen as senior manager of network planning and system analysis at BNR, RTP,North Carolina. He moved to BNR, Richardson, Texas in 1989 to set up the wirelesssystem engineering group and became the Director of Wireless System Engineering in1995: He left NORTEL in early 2001 to become the Vice President, SystemEngineering for Yotta Networks, Plano, Texas, and lead the architecture team todesign the Petabit Photonic Switch. He has received eight U.S. and internationalpatents and applied for another 10 patents on different wireless network

technologies. He has been involved in various IEEE/ACM sponsored international conferences and hascontributed as technical committee member, and has chaired various sessions. He has published more than50 papers at various international conferences during last 27 years. He obtained his BEE degree fromCalcutta University, India, and his MS from the Indian Institute of Technology, Delhi.

Dr Navrati Saxena is currently a lecturer and co-ordinator of the Next GenerationNetwork (NGN) technologies in the Amity School of Computer Science (ASCS) ofAmity University, India. She completed her PhD from Informatics and Telecommu-nication Department of the University of Trento, Italy. Prior to that she did herMCA and BSc from Agra and Kanpur Universities, respectively. Her prime researchinterests include data broadcasting and scheduling in wireless systems, VoIP inwireless, mobility management, multicasting and wireless sensor networks. She hasmore than 12 international conference and two international journal publications.



Documents

An information theoretic framework for predictive channel reservation in VoIP over GPRS