thesis1 plc

8/6/2019 Thesis1 PLC

1/87

Quality aspects of audio communication

Ian Marsh

A thesis submitted to KTH,the Royal Institute of Technology,

in partial fulfilment of the requirements forthe Licentiate of Technology degree.

May 2003

Laboratory for Communication NetworksDepartment of Microelectronics and Information Technology

KTH, Royal Institute of TechnologyStockholm, Sweden


2/87

TRITA-IMIT-LCN AVH 03:01ISSN 1651-4106ISRN KTH/IMIT/LCN/AVH-03/01SEc Ian Marsh May 2003

Printed by Universitetsservice US-AB 2003


3/87

Quality aspects of audio communication

Abstract

The Internet is increasingly being used to carry real-time voice traffic.Users of real-time voice services are sensitive to variable audio quality. Thequality of packet audio is largely determined by the mouth-to-ear delay andthe packet loss. The contribution of this thesis is to provide techniques toimprove the packet audio quality: dimensioning links specifically for packetvoice communication, modelling the packet audio arrival process at a re-ceiver, measuring connectivity quality in wide area networks, and reducingdelays in end systems.

The first study investigates how to allocate capacity to voice traffic in apurely packet switched network. We study an idealised case for VoIP ses-sions, where the voice traffic is separated from the data traffic. A Markovmodulated Poisson process model simulates the superposition of VoIP flowsinto a finite buffer. The model corresponds well with both packet level simu-lations and laboratory experiments. A second study looks at the interactionbetween voice and data traffic. We address the issue of how a constantrate VoIP stream is affected when multiplexed together with data traffic inrouter queues. We derive a Markov model which captures the effect of therandom delays experienced by packet audio data, plus the affect of silencesuppression at the sender and packet loss in the network.

Measurements made in 1999 and 2002 show that VoIP communication isfeasible between academic sites in Europe and the United States. However,we show that network connectivity on a global scale still does not providesufficient quality for satisfactory real-time voice communication. The datacollated as part of this study is one of the largest publicly available reposi-tories of VoIP data, containing over 18,000 sample sessions.

The end systems also contribute to the delay of interactive voice com-munication. Absorption of the variable delay, or jitter, is necessary in apacket switched network in order to replay voice samples smoothly with-out glitches. We have shown that by moving the buffer used to absorb the jitter into the operating system, significant time savings can be achieved.We have implemented a VoIP tool, Sicsophone, which shows very low delaycharacteristics.

Using the above techniques, we show that hundreds of milliseconds canbe saved in the delay budget of real-time voice communication, improvingthe audio quality considerably. The traffic models and measurement datapresented in this thesis, will also enable future research into quality aspectsof audio communication.

Keywords: Packetised voice, packetised audio, Voice over IP (VoIP),Quality of Service (QoS), speech quality, network measurements


4/87


5/87

5

Preface

This thesis proposes methods to improve the perceived quality of real-timevoice communication over the Internet. The cost of providing and runningvoice services using an IP infrastructure is considerably less than a tra-ditional public exchange system. Therefore, new types of operators sellingvoice services are emerging that use IP technology, allowing a broader choiceof operators for end users. Users however, would like to receive voice qualityakin to that provided by the traditional telephony network. Voice servicesusing IP in service today, focus on lower cost and less guarantees on servicequality. There is a common misconception about VoIP, which implies it of-fers lower quality, this does not necessarily need to be the case. The servicesoffered today simply use the Internet as a bearer for long haul links, and

live with the implications under the proviso it is cheaper than PBX-basedtelephony. The voice quality can be variable when using the IP infrastruc-ture, especially in peak hours, however it is possible to improve the qualitybut it is often not implemented. We1 are proposing ideas that will improvethe quality of voice communication at least initially, to that provided by thetraditional telephony network.

To achieve our goal of good quality audio communication, some changesmight be needed to the ubiquitous Internet. Providing strict quality guaran-tees has plagued researchers for many years and now industry is facing thesame challenges. The degradation of voice quality which can occur whenusing a multi-user packet switched network is the fundamental problem.

Unpredictable short term loads, lack of guarantees on network performance,lack of control over the end systems and stringent requirements on the voicequality make VoIP a challenging application to realise successfully on theInternet.

Our Quality of Service (QoS) research is orthogonal to the investigationsbeing carried out by the network community. These investigations focus onchanging the packet switching techniques to be more reliable, more timelyand more fair. This is especially the case for time sensitive traffic such asvoice. Protocols have been developed to signal routers and end systemsthat certain data types need to be treated differently, again in the case ofvoice traffic often at higher priority. The techniques presented in this thesis

do not rely on any ongoing research within the network community. Welook at allocating resources given the current conditions of the network oradapting to it, also by measuring the current state so that we can makedecisions based on these measurements rather than assuming the certainfunctionality will be available.

1The authors of the publications and I.


6/87


7/87

7

Content notes

The thesis contains five papers that address four distinct areas within qualityaspects of packet voice communication: dimensioning links for VoIP traffic,delay reduction at the end systems, the disruption of real-time voice streamsby traditional data traffic, and wide area measurements of VoIP quality.These four areas therefore include investigations at the network layer, in theoperating system and at the application.

We start with a short introduction and then provide some backgroundon the subject of this thesis. A problem definition is given, explaining whyand how we tackled each problem. The contribution of this thesis is givennext, showing exactly what has been achieved during the course of thisinvestigation. There is also a precis of the five included articles. Since all

of the works are co-authored, my contribution to each of the publications isstated. Finally we round off with some conclusions. The papers appear asthey were published.

Paper A Bengt Ahlgren, Anders Andersson, Olof Hagsand, and Ian Marsh.Dimensioning Links for IP Telephony. In Proceedings of the 2nd IP-TelephonyWorkshop, pages 14-24, New York, USA, April 2001.

Paper B Ingemar Kaj and Ian Marsh. Modelling the Arrival Process forPacket Audio. In Quality of Service in Multiservice IP Networks, pages35-49, Milan, Italy, February 2003.

Paper C Olof Hagsand, Ian Marsh and Kjell Hanson. Sicsophone: A Low-delay Internet Telephony Tool. To appear at the 29th Euromicro Conference,Belek, Turkey, September 2003.

Paper D Olof Hagsand, Kjell Hanson Ian Marsh. Measuring InternetTelephony Quality: Where are we today? In Proceedings of IEEE Globecom:Global Internet, pages 1838-1842, Rio De Janeiro, Brazil, December 1999.

Paper E Ian Marsh and Fengyi Li. Wide Area Measurements of VoIPQuality. To appear at Quality of Future Internet Services 2003, October,

2003, Stockholm, Sweden.


8/87


9/87

9

Acknowledgements

Writing this part of the thesis is actually enjoyable. First of all I must thankmy advisor Professor Gunnar Karlsson at KTH, and my manager Dr. BengtAhlgren at SICS. Without their co-operation and assistance this Licentiatewould not be completed. I am very thankful to Gunnar for his creative spiritduring this licentiate life, in particular I would like to thank him for ourfruitful discussions. To my manager Bengt, I am also very grateful, firstlyfor getting me started on this academic path, and secondly for his continualencouragement during the degree, particularly in the last stages Hur gardet med licen? 2 I was frequently asked. The people at SICS are a fantasticwell of information. This includes all the people with whom I have had coffeeroom discussions, in particular the members of the CNA group. I would liketo extend my special gratitude to Laura the Estonian Feeney and HerrDoctor Engineer Thiemo Voigt for their careful reading and comments onthe text I loosely referred to as English. This also extends to Dr. AdrianBullock who, at least, has the same notion of spelling. Dr. Olof Hagsand,who I have tracked from SICS to Dynarc and now at KTH, has set a greatexample of how to work effectively. He is one of the rare people who can getthings done both quickly and with high quality. Finally in the SICS gang,I would like to thank Bjorn 200,000 volts Gronvall who has helped almosteveryone at SICS, not least myself. This extends from FreeBSD installationquestions to getting ID cards (yes plural). Outside of Stockholm, I would liketo thank Professor Ingemar Kaj at Uppsala University; his excellent course,

book and support has been a source of inspiration during my research andis reflected in this thesis.

The other half of my working life revolves around KTH University.Most of the people there have become friends rather than working colleagues.Two weddings and a Christmas holiday in their home countries only goes toillustrate this point. Our United Nations style lunch time gatherings arealways something I look forward to. The innumerable humoristic momentskept me sane during the past three years of pseudo-student life. ParticularlyI would like to say Dudey Wudey to Iyad Diad Al-Khatib, I will neverforget our numerous classic moments, unfortunately not many of them areprintable in a licentiate thesis. Being in an academic environment allows

one to advise; with two Chinese masters students finished and a couplemore on the horizon, I would like to say (to you) it is a pleasure to beinvolved in your education. In particular the latest and greatest FengyiLi, whose effort is also evident in this thesis, you should be awarded firstprice! Not too far from China, newly married Evgueni Dude Ossipov fromSiberia, is always a welcoming site including a firm handshake. His presenceautomatically provokes a rye smile from me (not the English though :-). On

2How are things going with the licentiate work?


10/87

10

the subject of English, I want to extend my deepest gratitude to Professora

Nil Neely Wheely Tarim for her many last minute proof readings of mydrafts. Even after working all night, she still has the time and energy tocorrect my carelessness, her accent might be American but her English is just perfect:-)

There is more to life than work and study but not much more! OutsideKTH and SICS circles I would like to say orange juice please to the Englishgang in the Loft. I must extend my thanks to the staff and students at thelocal gym, called World Class (the WC). Performing (almost) mindless exer-cise works wonders in alleviating stress, creating ideas and preparing oneselffor the next days of rigorous research. Alphabetically Christer, Elisabeth,Eric, Eshan, Maria, Mia, Petra, Sabine and Sarah(Z), thanks! Of course Ihave to thank my beloved mother and Ray for their never ending support,this time you have been spared the manuscripts!


11/87

Contents

Preface 5

Content Notes 7

Acknowledgements 9

1 Introduction 13

2 Background 13

3 Problem definition 14

4 Contribution of this thesis 14

5 Quality aspects of audio communication: A 30 year perspec-

tive 14

5.1 A decade of research: 1973 - 1983 . . . . . . . . . . . . . . . . 155.2 Emergence of Internet applications: 1990 - 1995 . . . . . . . . 165.3 Times of measurement: 1996 - present . . . . . . . . . . . . . 17

6 Summary of the individual papers and their contributions 18

6.1 Paper A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186.2 Paper B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206.3 Paper C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

6.4 Paper D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236.5 Paper E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

7 Conclusions 27

References 29

Paper A: Dimensioning Links for IP Telephony 35

Paper B: Modelling the Arrival Process for Packet Audio 47

Paper C: Sicsophone: A Low-delay Internet Telephony Tool 63

Paper D: Measuring Internet Telephony Quality: Where are

we today? 73

Paper E: Wide Area Measurements of Voice over IP Quality 81


12/87


13/87

1 Introduction

The success of the Internet has been phenomenal. Since the introduction ofthe Web, the benefit of a world wide data network has been truly realised.There are many reasons for the Webs success: a well designed protocolsuite, large colour displays, the Mosaic browser, interesting sites (even then)plus no viable alternative. It is argued that the telephone companies alsodesigned and deployed a global network, but were very conservative by onlyallowing voice to be carried. The telephone network could have been thestart of the Internet. Now the telephone network can be considered part ofthe Internet as it partly carries data traffic via modems. The question posedin this thesis is: Can the opposite be applied? Can the Internet be used

to carry the worlds voice traffic? The savings would be enormous, if onenetwork could carry all the voice and data of the worlds users. People wouldretain their computers and phones in the homes and offices, but outside theseareas the data and voice would be carried along the same communicationlines, using the IP protocol.

2 Background

In a packet switched network, the voice is sampled, packetised, transmit-

ted, received, de-packetised and replayed. This sequence is not problematicwhen using a packet switched network for voice per se. A functioning, wellprovisioned packet switched network delivers voice data reliably and withlow jitter. We have extensive results that show this to be the case (PaperE). The major problem faced by real-time traffic on the Internet is the un-predictable competing load. If a network is neither totally reliable nor verypredictable, as on an IP network, delay, jitter and packet loss will affect theaudio quality.

Quoting the definition of the IP protocol (RFC 791) The internet pro-tocol does not provide a reliable communication facility. During high loadsituations data may be discarded to keep the network operational, and the

IP specification is not violated by doing so. In the Internet, reliability is nor-mally addressed at the transport layer using TCP or an application specificsolution. Neither of these are perfectly suitable for real-time communica-tion, mainly due to the delay requirements of interactive voice. This clearlyhas implication for the worlds telephony traffic if IP is to be its bearer.Within the IP protocol, there is functionality defined to allow time-sensitivepackets to be transmitted with higher precedence in the case of high load.However this has, until now, not been widely deployed and is not expectedto be for some time to come.


14/87

14

3 Problem definition

We are now ready to define the problem: How to achieve acceptable qualityvoice communication on an unreliable, unpredictable packet switched net-work? The requirements of the voice communication can also be defined.The network and application should not delay the voice samples above a wellknown number of milliseconds. The network ideally should also deliver a suf-ficient number of packets in order for the voice stream to be reconstructedwith acceptable quality. Realising these two requirements effectively definesthe problem we address in this thesis.

4 Contribution of this thesis

Taken as a whole, the contribution of this thesis can be summarised as tech-niques to improve the quality of real-time voice communication. Moreovershould all the ideas presented in this thesis be implemented, they would re-sult in improved perceived quality. We have identified and addressed issueswithin this research topic that have a direct significance on the improve-ment of real-time voice services on the Internet. We propose solutions toreal-world problems concerning packet audio communication. The thesislooks at quality aspects of audio communication at different layers (in the

ISO sense) and from a theoretical and practical perspective. Since the out-comes of real-time voice research are implementable, we also consider thepractical results of this work a valuable contribution.

5 Quality aspects of audio communication: A 30

year perspective

The problem addressed in this thesis is by no means new. Researchers haveinvestigated real-time voice communication on packet switched networks for

over 30 years now. This section covers chronologically the most significantfindings related to our work. It is important to emphasise this point, as thereare probably hundreds (but not many thousands) of articles which have somerelevance to real-time voice on packet switched networks. Therefore we in-clude only the ones which are relevant and widely cited. We also includesurvey articles and theses for the interested reader; they usually give com-prehensive lists of relevant publications. Finally we should state there are anumber of textbooks about Voice over IP and Internet Telephony, howevernone of them adequately cover the quality aspects in sufficient detail.


15/87

5 Quality aspects of audio communication: A 30 year perspective 15

5.1 A decade of research: 1973 - 1983

Advances in low data rate coders [1] and the deployment of a distributed(viable) packet switched network led to early findings on real-time voicebeing published. In 1974 William Naylor published A status report on thereal-time speech transmission work [2] and James Forgie published Speechcommunications in packet-switched networks in the Journal of AcousticSociety of America in 1976 [3]. In 1977, Naylor published his PhD thesistitled Stream traffic communication in packet switched networks [4].

Another researcher who published his early experiences with voice onpacket switched networks was Danny Cohen. In 1977, he suggested thatthe packetisation algorithm and data rate should be varied according tothe network load [5]. Interestingly, this adaptive approach of reacting to

the network load has been popular in recent years. Cohen also states thatthe time spent at the receiver (called waiting period in his paper) shouldbe a function of the network performance. Tuning the size of the playoutbuffer in voice systems to the network load has occupied researchers formany years. Indeed, Cohen states that the parameters in a real-time voicecommunication system are heavily dependent on the network performanceand a systematic method of predicting it must be developed.

In 1979 John Gruber looked at the issue of Variable delays in a sharednetwork environment handling voice traffic [6]. His vision was a packet andcircuit switched hybrid network called Transparent message switching forhandling both voice and data traffic. The ideas were novel (and preliminary):

The basic entities processed are messages rather than calls. The messagesdo belong to an established call, however they may be completed or blockedat the network periphery. Voice messages are given priority where delaysare being exceeded, however where loss is being experienced, voice packetsmay be discarded initially. However, some loss in voice communicationis tolerable. Gruber suggests that channel contention can be resolved bybuffering messages at the edges only. Once sufficient capacity is obtained,the network behaves as a circuit, switching the messages. Messages areswitched on the fly, thus eliminating the need for the whole message toarrive. The rest of the paper explores the benefits of using this techniquefor voice traffic. The paper includes 76 references, encompassing nearly all

of the early work on real-time voice on packet switched networks. The ideaof having a circuit switched core for the Internet has gained popularityrecently with schemes such as Multi Protocol Label Switching (MPLS).

In 1980 Giulio Barberis and Daniele Pazzalia published their seminalAnalysis and Optimal Design of a Packet-Voice Receiver [7]. They con-clude that in order to obtain the optimal voice reconstruction, an accurateestimation of the delay suffered in the communication network is required.Their conclusions echo those by Cohen, however they suggest meeting thisrequirement by using a synchronisation algorithm that reduces the gap be-


16/87

16

tween the transmitter and receiver clocks to zero.

In December 1983 Warren Montgomery published Techniques for PacketVoice Synchronization in an IEEE JSAC special edition on Packet SwitchedVoice and Data Communication [8]. He considers the local and wide areanetwork situations separately. Round trip estimates are sufficient for thelocal area case, while more sophisticated methods are needed for the widearea case. He suggests that the addition of timing information and incor-porating extra delay at the receiver should be sufficient to yield satisfactoryvoice quality for the wide area case. This is the approach taken by mostmodern real-time packet voice applications. It is effective, simple and cheapto implement.

The period from 1984 to 1990 was relatively inactive as far as real-timevoice over packet switched networks was concerned. Two notable excep-tions include Prabandham Gopals Analysis of Playout Strategies for VoiceTransmission Using Packet Switching Techniques [9] and Mehmet Alis Re-assembly Buffer Requirements in a Packet Voice Network [10].

5.2 Emergence of Internet applications: 1990 - 1995

In the early nineties, Domenico Ferraris group at UC Berkeley produced anumber of significant publications about the effect of jitter and delay on real-time communication applications as part of the TENET suite [11][12]. Theirwork proposed a distributed mechanism for controlling the delay jitter in apacket-switching network. They argued that if the advantages are sufficient

to justify the higher costs of the distributed jitter control mechanism, thenimplementing it is worthwhile. Although no such scheme was deployed, theirwork is still referenced.

Research on IP multicast was actively being carried out in the earlynineties. Multicast was to be the vehicle on which multimedia sessions wereto be transmitted over the Internet. Indeed many people listened to theearly Mbone transmissions [13]. An array of real-time applications were pro-duced, notably VIC, VAT and wb (whiteboard) from the Network ResearchGroup at LBL [14]. Other tools surfaced such as Nevot [15], Freephone [16]and RAT [17]. These works led to a standardised synchronisation protocol,RTP, the real-time protocol for use with real-time media flows. The au-

thors of the standard were those of the above mentioned applications. Oneof them was Van Jacobson, who gave a Sigcomm tutorial in London 1994entitled Multimedia conferencing on the Internet [18]. In this presentationhe suggested using a simple synchronisation protocol to restore the originaltiming information at the receiver and a small adaptable buffer to absorbdelay variations. This influential presentation moulded the approach takenby researchers in real-time voice for many years.

Henning Schulzrinnes 1993 PhD thesis Reducing and characterizingpacket loss for high-speed computer networks with real-time services looked


17/87

5 Quality aspects of audio communication: A 30 year perspective 17

at congestion control, scheduling, and loss correlation of real-time traffic [19].

Schulzrinne highlighted the practical importance of scheduling packet audioin the context of the DARTnet project.

During 1993-1996 Jean Bolot produced a series of papers that reportedand characterised the loss and delay behaviour of packet audio on the In-ternet [20, 21, 22]. They were theoretical works supported by experimentalevidence, advocating the use of techniques such as redundancy protectionagainst packet loss. Their publication also had an effect on the researchcommunity, by highlighting the need to conduct theoretical, but applicableresearch on the Internets behaviour. Comparisons of voice playout algo-rithms have been made by Ramachandran Ramjee et. al. in 1994 [23]. Thiswork was extended to include performance bounds of the algorithms, withSue Moon as the primary investigator [24]; it was also later published in theMultimedia Systems journal in January 1998 [25].

In 1997, Nicolas Maxemchuk and Shau-Ping Lo measured the loss anddelay variation for intra-state, inter-state and international links [26]. Twoimportant, but unsurprising conclusions, were that the quality depends onthe number of hops and the time of day. In 1999 Dong Lin also concludedthat even calls within the USA could suffer large jitter spikes [27]. Herresults on packet loss also agree with those above in [22], which is interestingas the latter measurements were taken some four years later. We did notobserve such effects in our measurements, which were conducted over purelyacademic networks.

One piece of work which is worthy of note is Christian Sieckmeyers

master thesis entitled Evaluation of adaptive playout algorithms for packetaudio done at TU Berlin in 1995 [28]. This is a comprehensive evaluationof jitter buffer playout algorithms using C++ implementations. Because itis written in German, this thesis has not received the attention it deserves.

5.3 Times of measurement: 1996 - present

Recently, measurements have been very much in vogue. Three significanttheses presenting VoIP measurements were published at the turn of the cen-tury. First, Dong Lins master thesis Real-time voice transmissions over theInternet 1999 undertaken at the University of Illinois in 1999 investigated

the use of interleaving and reconstruction for improved speech fidelity [27].Second, Henning Sannecks PhD work Packet Loss Recovery and Control forVoice Transmission over the Internet in 2000 looks at intra-flow hop-by-hopschemes where VoIP flows are repaired within the network itself [29]. Therouters look into individual flows and attempt to interpolate any missing au-dio packets. Third, Sue Moons Measurement and Analysis of End-to-EndDelay and Loss in the Internet in 2000 [30] looked at correcting the sys-tematic errors introduced by clock skew between the sender and the receiver[31]. Her findings on the playout delay incurred by a buffer at the receiver


18/87

18

are relevant to our work. In particular, she and her co-authors computed

the upper and lower bounds for the delay given a number of losses. Theyalso showed that these bounds are tight. A new spike detection algorithmwhich takes into account the sudden delay peaks of audio transmissions ispresented in [25].

Robert Cole and Joshua Rosenbluth published Voice Over IP Perfor-mance Monitoring in 2001 which advocates using the ITUs E-model withsimplifications for use on packet switched networks [32]. In 2001 MansourKaram and Fouad Tobagi published Analysis of the Delay and Jitter ofVoice Traffic Over the Internet [33]. They looked at voice delay and theeffect of network parameters on the delay caused by voice traffic, assumingit uses separate queues. They state the importance of bandwidth to reducethe delay percentile incurred by voice. Additionally and unsurprisingly, fornetworks over 10Mbits per second, the transmission delay becomes negligiblewithin the end-to-end delay.

Catherine Boutremans recent (December 2002) PhD thesis Delay As-pects in Internet Telephony presents an adaptive error control scheme thatis delay aware [34]. The end-to-end delay is considered when choosing whichparameters should be used for forward error correction coding. She extendsthis idea to include a playout buffer implementation that takes into accountthe choice of the FEC scheme. Boutremans states that link and router fail-ures are the dominant sources of degradation for VoIP sessions, despite IProute protection.

6 Summary of the individual papers and their con-

tributions

6.1 Paper A

Bengt Ahlgren, Anders Andersson, Olof Hagsand, and Ian Marsh. Dimen-sioning links for IP telephony. In Proceedings of the 2nd IP-Telephony Work-shop, pages 14-24, New York, USA, April 2001.

Summary: Currently many telephony providers use IP technology as a

bearer for voice data. One method to achieve good quality voice communi-cation is to dimension the network, as is done in the traditional telephonesystem. We propose that the capacity allocation for voice traffic on IPnetworks can be performed as in traditional telephony or ATM voice net-works. We argue that the research and models derived from the traditionaltelephony and ATM fields are suitable for todays IP networks.

Data network operators do not necessarily know how to dimension theirnetworks for voice traffic. A naive approach could be to allocate the numberof calls based on the capacity of the link and the maximum speech coder


19/87


rate. This will yield good quality, but will under-utilise the link due to the

lost capacity by not using the statistical multiplexing characteristics of thevoice calls. The other alternative is to allow as many calls onto the networkas possible, however in busy periods this may result in unpredictable orpoor quality. We can phrase one possible formulation of the problem asHow many calls can an operator allow over a link (or portion thereof) witha packet loss rate under 1%?.

We propose a model based on the Markov modulated Poisson process(MMPP) which calculates packet loss probabilities for a set of super-po-sitioned voice input sources. We assume the talk and silence periods areexponentially distributed. Because of the exponentially distributed inter-arrival times during a talkspurt (a series of voice packets), the emission ofpackets in a talk period can be regarded as a Poisson process with a givenintensity. The superposition of Poisson processes is also a Poisson process.We can therefore simply add the intensities of the sources that are currentlyin a talkspurt and obtain a new Poisson process for the superposition. Wecan use a two state birth-death process to describe the packet generation;one state represents the idle periods and the other state the talkspurts. Thisarrival process is fed into a */D/1/K queue. It is a single FIFO server withdeterministic service times and a buffer size K-1. The size of the buffer isvariable. This solution is sufficient to tackle the dimensioning problem givena number of possible known and unknown quantities.

Our simulations and laboratory measurements are in good agreementwith the Markov model chosen. This shows that many of the earlier efforts

on network dimensioning actually match the real environment they wereproposed for. A second contribution of this work is the inclusion of bothpacket level simulations and laboratory measurements to verify the Markovmodel. As far as we know of, no other researchers have presented andcompared results from three different environments.

Contribution of this work: The contribution of this work is a planningtool with which to dimension networks for voice traffic. We have establishedrelationships amongst certain parameters of a packet voice network; namelythe speech coding, the capacity of the voice network, the number of users,

the buffer sizes, the acceptable packet loss in order dimension the networkfor voice communication.

We propose an engineering solution to the problem of achieving goodquality using theoretical work as a basis. Many engineers unfortunately donot use results from the theoreticians and we attempted in this work to provethe algorithms based on theoretical results are correct and are implementablein a real network. This work is a little unusual in that we compared a model,simulation and a laboratory testbed, which raises interesting questions initself. What are the differences between the three approaches, and how do


20/87

20

these differences manifest themselves in the results? As a simple example,

the time needed to transmit a packet is not included in the model, but itis included in the simulation environment. In the testbed milieu, packetsmight be sent in a bursty manner due to system behaviour. These effects aredifficult to account for in the model and the simulation. The work broughtdifferences such as these to our attention, and though it is not possible todo anything about them specifically, it is useful to know of their existence.

My Contribution: The original idea to perform such a study was mine.Within the project I supervised a masters student, Anders Andersson, whoimplemented the MMPP model in Matlab and corresponding simulationscripts in ns-2. I implemented most parts of the testbed environment andthe traffic generator that was used to simulate the superposed telephonyflows. I wrote the majority of the paper.

6.2 Paper B

Ingemar Kaj and Ian Marsh. Modelling the Arrival Process for PacketAudio. In Quality of Service in Multiservice IP Networks, pages 35-49,Milan, Italy, February 2003.

Summary: In this work, we model the arrival process of audio packetsthat have passed through a number of routers. The first objective wasto gain an insight into the processes that determine this behaviour: Does

any theory exist that can explain the distribution of the arriving audiopackets? The second objective was to use a model of the arrival process forthe generation of artificial audio streams, thus resembling those that havetraversed the Internet. A model is clearly superior to static trace files as itcan produce variable temporal relationships between packets of a flow whichis not possible with a recorded session. To evaluate our work, we comparethe probability density functions of the gathered and generated data.

Normally packets are sent with constant time spacing from a sender,and due to buffering at intermediate routers, arrive at the receiver withnon-constant time spacing. We separate the queueing delay caused by ourown packets (in front of us) from the delays induced by cross-traffic present

in the buffers of routers. This is because the contribution of the delay fromour packets is known and solvable. The delay contribution of the cross-traffichowever is more difficult and is the problem we address in this study. It iscomplicated by the fact that most VoIP tools implement silence suppres-sion and packets may be discarded. To make the problem more tractable,we assume that the waiting times in the routers buffers are exponentiallydistributed.

The solution, based on Markov theory, attempts to model the delay vari-ation of audio packets sent at a constant rate. The packets are assumed to be


21/87


subjected independently (of each other) to delays when traversing the net-

work. The waiting time in intermediary buffers is, as stated, assumed to beexponentially distributed. The observed delay of the packets can be shownto be Markovian. The use of Markov theory allows silence suppression andloss to be incorporated into the model as they are considered independentfrom the delay variation introduced by the network.

Given a certain distribution of the delays in the network (e.g. exponen-tial, Gaussian) it is possible to create a distribution of the arrivals whichmimics the arrival process of real data streams.

Contribution of this work: The main contribution of this work is in-sight into the problem of how packet audio streams become distorted whentraversing a network like the Internet. A Markov model that does not useany transforms is developed from first principles. The model is extensible,and therefore allows us to include both silence suppression at the sender andpacket loss during the transmission. A simple method to estimate packet lossbased on observed interarrival times is also given, independent of whethersilence suppression is activated or not.

My Contribution: The idea came about from Ingemar Kajs course Sto-chastic Traffic Modelling, with whom it was jointly conceived. My contribu-tion was the original wide-area measurements of several VoIP flows. Usingthese measurements I made some suggestions to the possible processes actingupon the streams. I also wrote several tools to process the data. I co-wrote

the paper.

6.3 Paper C

Olof Hagsand, Ian Marsh and Kjell Hanson. Sicsophone: A Low-delay In-ternet Telephony Tool. To appear at the 29th Euromicro Conference, Belek,Turkey, September 2003.

Summary: Users of interactive VoIP applications demand low latencyconversations. Replaying packetised audio requires that sufficient numberof packets are available to the application in order to avoid audible glitches.

The standard method to solve this problem is to introduce a small inter-mediary buffer between the decoded voice and the audio hardware. Thiscreates a dam and hence a temporary reservoir of packets for immediateplayout. We describe a Voice over IP system, called Sicsophone, that cou-ples the low level features of audio hardware with a jitter buffer playoutalgorithm. Using the sound card directly eliminates unnecessary bufferingas well as giving us fine control over timers needed by soft real-time applica-tions such as VoIP. A soft real-time application is one which does not havestrict deadlines, but nevertheless requires low latency to obtain acceptable


22/87

22

performance. The delay in real-time voice communication is an example of

such a requirement.Constructing a fully functional low delay VoIP system is non-trivial.

There is a clear tradeoff of speed against flexibility; for example differentaudio formats make the system more difficult to optimise. PCM voice codingis considered the fast path in Sicsophone. PCM encoded speech samples aredelivered to the loudspeaker the quickest. GSM is supported, but requiresextra buffers which have not been placed on the optimised path.

A standard statistical-based approach for inserting packets directly intoaudio buffers is used in conjunction with low level control of the audio hard-ware. The buffers in the sound card memory act as the playout buffersrather than using memory in the application. This saves valuable time

when copying data to the application, as later it must be copied back to theaudio hardware after de-jittering. The application calculates the length ofthe current playout buffer, because simple statistics need to be calculated.However adjustments to the buffer length are done in the silent periods bythe operating system. Late arrivals are detected in the hardware itself byusing two pointers, one which writes the incoming data and one which reads(and hence replays) it immediately afterwards. If the read pointer passesthe write pointer, the next incoming data is simply not written to the (cir-cular) buffer. In addition we developed a scheme for inhibiting unnecessarycorrections in the playout buffer size. Adjusting the buffer size in the veryshort term only induces unwanted instability. This is especially undesirable

in our approach as we are dealing with the hardware directly.Reducing unwanted changes to low level buffers on the sound card main-

tains good performance of the system. We found this combination of lowlevel access to the audio hardware plus a relatively simple technique of adjusting the playout buffer gave excellent delay characteristics for many testcases. We performed live tests with the implemented software for compari-son with off-the-shelf VoIP tools and found the system to exhibit much lowerdelay characteristics.

Contribution of this work: The contribution of this work is a consid-erable reduction in the delay incurred by VoIP end systems. This delay isan important factor in determining the perceived quality. Since this work isan engineering solution, it is rarely seen in the research community. Peoplehave looked at optimising and reducing jitter buffer sizes, but do not realisetheir ideas in real systems. Small theoretical improvements can be almostnegligible in a real system. A key artifact of this work is Sicsophone, a fullyfunctional VoIP application. It also has been used for the measurementwork, with some modifications.


23/87


My Contribution: I wrote the RTCP part of Sicsophone. I also made

some changes to the tool for measurement work. I wrote the conference pa-per. I also performed comparisons between the playout delay of Sicsophoneand the optimal playout delay.

6.4 Paper D

Olof Hagsand, Kjell Hanson and Ian Marsh. Measuring Internet TelephonyQuality: Where are we today? In Proceedings of IEEE Globecom: GlobalInternet, pages 1838-1842, Rio De Janeiro, Brazil, December 1999.

Summary: Users of Internet telephony applications demand good qualityaudio playback. This quality depends on the instantaneous network condi-tions and the time of day. In this paper, we describe a scheme for measuringnetwork connectivity and motivate the development of a new metric, asym-metry, for judging quality. Work such as this gives useful feedback to usersand operators of IP telephony networks and important information for de-velopers of Voice over IP applications.

This paper outlines a scheme to send a pre-recorded telephone call, inPCM format, between a central site (Stockholm) and four satellite sites.The call was simplex, it was transmitted in one direction only, i.e. either toor from Stockholm. The call probes the links and routers of the interveningconnections, giving an estimation of the quality at the receiver. Using thesetechniques we measured the quality of the intervening links. Our tests in-cluded a wide range of geographically distributed sites. We concluded thatat four of the five sites we had access to, it was feasible to run successfulVoIP applications according to the ITU-T G.114 (delay specific) standard,which states the end-to-end delay should not exceed 150 ms.

Contribution of this work: In 1999 we reported QoS measurements oflinks between remote five sites and a central site. As far as we are aware thejitter and asymmetry results were relatively new within the VoIP measure-ment community. A further contribution of this work was updating the

available VoIP traces for research. The same ones were being used in manydifferent works and were becoming out of date. Three of the sites involvedin this study have been used in the work in Paper B and for comparisonof the VoIP quality in Paper E. This work can be seen as precursor to thework described next.

My Contribution: Although my name appears last on the (alphabetic)list, the idea, work and text of the paper were mine.


24/87

24

6.5 Paper E

Ian Marsh and Fengyi Li. Wide Area Measurements of VoIP Quality. To ap-pear at Quality of Future Internet Services 2003, October, 2003, Stockholm,Sweden.

Summary: In Paper D from 1999, we reported experiments to quantifythe quality of Internet telephony. In this work we improved our measurementmethodology, included more hosts, probed more sessions and compared thequality of the links that remained unchanged over the past years.

Once again, a pre-recorded telephone call was sent between nine sitesavailable to us, however this time the sites were connected as a full mesh

allowing us, in theory, to measure the quality of 72 different Internet paths.In practice, some of the combinations were not usable due to certain portsbeing blocked, thus preventing the audio to be sent to some sites. There werefour such cases. Bi-directional sessions were scheduled on an hourly basisbetween any two given end systems. Calls were transferred only once perhour due to load considerations on remote machines. This time nine siteswere carefully chosen with large variations in hops, geographic distances,time zones and connectivity to obtain a better diversification of distributedsites. One limitation of the sites was they were all located at academicinstitutions, which are typically associated with well provisioned networks.

In order to gather more comprehensive measurement data, we includedfour new test sites, and automated the process of sending and measuring thetest files. Other extensions included hourly bi-directional conversations ona 24 hour basis. In contrast to the first set of experiments, where we set upcalls between a central site and the satellites, this time we used a full-meshscheme so each co-operating site could send and receive to all of the others.Only PCM coding was used and call signalling was not included, we simplystarted sending a UDP voice stream to an awaiting receiver, thus assumingthe signalling has been established between the two communicating parties.

Contribution of this work: The contribution of this work is a reporton the quality of Voice over IP in 2002. We defined the quality as one-way

delay, loss and jitter. With a large undertaking, we have gathered more than24,000 sample sessions from nine globally distributed sites. for three sites,we have been able to compare the quality from 1999.

One further contribution is by combining the results of this work andPaper C: we can estimate the mouth-to-ear delay of a VoIP system. WithoutGPS and other specialised equipment this is a difficult quantity to measure.Paper C accounts for the delay incurred by the end systems and paper E thedelay by the network. Seen as one result these two results are an estimateof the total mouth-to-ear delay of a wide area VoIP system.


25/87


My Contribution: The idea to improve on the measurements from 1999

was mine. I advised a Masters student, Fengyi Li, to perform the measure-ments using Sicsophone with my modifications. I wrote a tool to processthe session files and we jointly wrote the paper (based on the Lis masterthesis).


26/87


27/87

7 Conclusions 27

7 Conclusions

In the first phase of my doctoral studies I have investigated selected topicswithin real-time audio communication. I have suggested techniques to im-prove the quality of packet audio: dimensioning links specifically for packetvoice communication, modelling the packet audio arrival process at a re-ceiver, measuring connectivity quality in wide area networks, and reducingdelays in end systems. The common theme in this work is an engineering ap-proach to real-world issues. We have also tried to solve them independentlyof current quality of service research agendas.

In the first study, we looked at the scenario where an operator uses alink (or portion thereof) exclusively for voice over IP. Taking advantage ofthe statistical properties of the call properties, such as silence periods, al-

lows much higher utilisation of the link capacity. We used existing Markovtheory to ascertain how many calls can be allocated to the link accordingto a specific quality. Given parameters we can define an appropriate opera-tional range. Through modelling the scenario, simulation and a laboratoryimplementation we are able to conclude the model and approach are valid.

In the second study, we investigated the effect that bulk TCP data hason a single audio stream. This extends the previous work to include theinteraction between the two traffic types. Small audio packets multiplexedwith large data packets in the queues of routers can distort the originaltiming of the speakers voice pattern. By identifying independence betweenthe delay experienced by each packet in the buffer and the observed network

delay we can show that the observed delay of the packets is Markovian. Thisallows us to construct a Markov model in which we are able to model thearrival process of packet audio streams. The interarrival histograms for themodel and the gathered data are similar, confirming that the model producesan arrival process similar to those observed. The knowledge gained allows usto generate representative and reproducible packet audio streams, which canbe used to test jitter buffer playout algorithms. The alternative is extensivefield testing, which although representative, is not very flexible. We haveused data from our measurement effort to construct the model, so realisticdata has been incorporated.

Measurement work has been an important part of this thesis. We con-

ducted two separate studies to report for the quality of VoIP on the Internet.We implemented a real-time VoIP tool called Sicsophone (Paper C). Thistool was modified to enable us to measure the delay, loss and jitter betweenglobally distributed sites. Our findings show that the VoIP quality is accept-able for communication between academic sites in Europe and the UnitedStates, providing the end systems do not add excessive delay to the audiostreams. The measurements have also assisted us in gaining some insightinto the loss and delay processes at work on the Internet. We have gatheredover 18,000 traces and made them publicly available. This can be seen as a


28/87

28

contribution of this work in its own right.

Finally, we have looked more closely at the contribution of the end sys-tems to the mouth-to-ear delay. The end systems are an important, andoften overlooked, part of a real-time voice communication system. Theyare also one component of a VoIP system that can be finely tuned. Weshow that the delay incurred by the end points can be reduced from 100sof milliseconds to 10s of milliseconds per host. This is achieved by movingthe jitter buffer to the memory of the sound card, resulting in less copyingof the data and direct access to the sound samples. Because humans areparticularly sensitive to delays over 175 milliseconds, the reduction in delayis critical to achieving good quality audio communication.

We believe that a combination of research and engineering solutions canyield significant improvements for the quality of real-time voice services.We have investigated four areas from different perspectives, with a par-ticular focus on delay. Using the techniques we have proposed, we showthat hundreds of milliseconds can be saved in the delay budget of real-timevoice communication, improving the audio quality considerably. We havealso gained valuable insight into more fundamental issues through modellingand measurements of real time voice systems. In future work we will look atimplementing more efficient jitter reduction techniques. We will also investi-gate the bound on the lowest possible delay that can be attained in a packetaudio system. This work will be complemented by further measurements toascertain the contribution of the (variable) queueing delay within the totalmouth-to-ear delay budget.


29/87

References

[1] D. T. Magill, Adaptive speech compression for packet communicationsystems, in Conference record of the IEEE National Telecommunica-tions Conference, pp. 29D1 29D5, 1973.

[2] W. E. Naylor, A status report on the real-time speech transmissionwork at UCLA. NSC Note 52, Dec 1974.

[3] J. Forgie, Speech communications in packet-switched networks, Jour-nal of the Acoustic Society of America, vol. 59, no. 1, 1976.

[4] W. E. Naylor, Stream traffic communication in packet switched net-works. PhD thesis, UCLA, 1977.

[5] D. Cohen, Issues in transnet packetized voice communications, inProceedings of the Fifth Data Communications Symposium, (Snowbird,Utah), pp. 610 613, ACM, IEEE, Sept. 1977.

[6] J. G. Gruber, Delay related issues in integrated voice and data net-works a review and some experimental work, in 6th Data Commu-nications Symposium (ACM Sigcomm Computer Communication Re-view), (Pacific Grove, California), pp. 166180, ACM/IEEE, Nov. 1979.

[7] G. Barberis and D. Pazzaglia, Analysis and optimal design of a packet-voice receiver, IEEE Transactions on Communications, vol. COM-28,pp. 217227, Feb. 1980.

[8] W. A. Montgomery, Techniques for packet voice synchronization,IEEE Journal on Selected Areas in Communications, vol. SAC-1,pp. 10221028, Dec. 1983.

[9] P. M. Gopal, J. W. Wong, and J. C. Majithia, Analysis of playoutstrategies for voice transmission using packet switching techniques,Performance Evaluation, vol. 4, pp. 1118, Feb. 1984.

[10] M. K. M. Ali, C. M. Woodside, and J. F. Hayes, Re-assembly bufferrequirements in a packet voice network, Computer Networks and ISDN

Systems, vol. 15, no. 2, pp. 109120, 1988.

[11] D. C. Verma, H. Zhang, and D. Ferrari, Delay jitter control for real-time communication in a packet switching network, Tech. Rep. TR-91-007, University of California, Berkeley, CA, 1991.

[12] D. Ferrari and D. C. Verma, A scheme for real-time channel estab-lishment in wide-area networks, IEEE Journal on Selected Areas inCommunications, vol. 8, no. 3, pp. 368379, 1990.


30/87

30

[13] S. L. Casner and S. E. Deering, First IETF Internet audiocast, ACM

Computer Communication Review, vol. 22, pp. 9297, July 1992.[14] V. Jacobson and S. McCanne, vat - LBNL audio conferencing tool,

July 1992. Available at http://www-nrg.ee.lbl.gov/vat/.

[15] H. Schulzrinne, Voice communication across the Internet: A networkvoice terminal, Technical Report TR 92-50, Dept. of Computer Sci-ence, University of Massachusetts, Amherst, Massachusetts, July 1992.

[16] http://www-sop.inria.fr/rodeo/fphone/, 1999.

[17] V. Hardman, A. Sasse, M. Handley, and A. Watson, Reliable audio foruse over the Internet, in Proc. of INET95, (Honolulu, Hawaii), June

1995.

[18] V. Jacobson, Multimedia conferencing on the Internet, in SIGCOMMSymposium on Communications Architectures and Protocols, (London,England), Aug. 1994. Tutorial slides.

[19] H. Schulzrinne, Reducing and characterizing packet loss for high-speedcomputer networks with real-time services. PhD thesis, University ofMassachusetts, Amherst, Massachusetts, May 1993.

[20] J. C. Bolot, Characterizing end-to-end packet delay and loss in theInternet, Journal of High Speed Networks, vol. 2, no. 3, pp. 305323,1993.

[21] J. C. Bolot, End-to-end packet delay and loss behavior in the In-ternet, in SIGCOMM Symposium on Communications Architecturesand Protocols (D. Sidhu, ed.), (San Francisco, California), pp. 289298,ACM, Sept. 1993. also in Computer Communication Review 23 (4),Oct. 1992.

[22] J. C. Bolot, H. Crepin, and A. Garcia, Analysis of audio packet loss inthe Internet, in Proc. International Workshop on Network and Operat-ing System Support for Digital Audio and Video (NOSSDAV), LectureNotes in Computer Science, (Durham, New Hampshire), pp. 163174,Springer, Apr. 1995.

[23] R. Ramjee, J. Kurose, D. Towsley, and H. Schulzrinne, Adaptive play-out mechanisms for packetized audio applications in wide-area net-works, in Proceedings of the Conference on Computer Communications(IEEE Infocom), (Toronto, Canada), pp. 680688, IEEE Computer So-ciety Press, Los Alamitos, California, June 1994.

[24] S. B. Moon, J. Kurose, and D. Towsley, Packet audio playout delayadjustment algorithms: performance bounds and algorithms, research


31/87

report, Department of Computer Science, University of Massachusetts

at Amherst, Amherst, Massachusetts, Aug. 1995.[25] S. Moon, J. F. Kurose, and D. F. Towsley, Packet audio playout de-

lay adjustment: performance bounds and algorithms, Multimedia Sys-tems, vol. 5, pp. 1728, Jan. 1998.

[26] N. F. Maxemchuk and S.-P. Lo, Measurement and interpretation ofvoice traffic on the Internet, in Conference Record of the InternationalConference on Communications (ICC), (Montreal, Canada), June 1997.

[27] D. Lin, Real-time voice transmissions over the Internet, master thesis,University of Illinois, Urbana-Champaign, 1999.http://manip.crhc.uiuc.edu/Wah/papers/TM16/TM16.pdf.

[28] C. Sieckmeyer, Bewertung von adaptiven Ausspielalgorithmen furpaketvermittelte Audiodaten (Evaluation of adaptive playout algo-rithms for packet audio - in German), Studienarbeit, Dept. of Electri-cal Engineering, TU Berlin, Berlin, Germany, Oct. 1995.

[29] H. Sanneck, Packet Loss Recovery and Control for Voice Transmissionover the Internet. PhD thesis, Technical University of Berlin, Oct. 2000.

[30] S. Moon, Measurement and Analysis of End-to-End Delay and Loss inthe Internet. PhD thesis, University of Massachusetts, 2000.

[31] S. Moon, P. Skelly, and D. Towsley, Estimation and removal of clockskew from network delay measurements, in Proceedings of the Con-ference on Computer Communications (IEEE Infocom), (New York),Mar. 1999.

[32] G. Cole, Robert and H. Rosenbluth, Joshua, Voice over IP Perfor-mance Monitoring, ccr, vol. 31, pp. 924, Apr 2001.

[33] M. Karam and F. Tobagi, Analysis of the delay and jitter of voicetraffic over the Internet, in Proceedings of the Conference on ComputerCommunications (IEEE Infocom), (Anchorage, Alaska), Apr. 2001.

[34] C. Boutremans, Delay Aspects in Internet Telephony. PhD thesis,EPFL, Dec 2002. 2715.


32/87

Paper A

Bengt Ahlgren, Anders Andersson, Olof Hagsand, and Ian Marsh. Dimen-sioning Links for IP Telephony. In Proceedings of the 2nd IP-TelephonyWorkshop, pages 14-24, New York, USA, April 2001.


33/87


34/87

IPTEL2001 35

Dimensioning Links for IP TelephonyBengt Ahlgren, Anders Andersson, Olof Hagsand and Ian Marsh

SICS

CNA Laboratory

Sweden

{bengta, olof, andersa, ianm}@sics.se

Abstract

Packet loss is an important parameter for dimensioning

network links or traffic classes carrying IP telephony traf-

fic. We present a model based on the Markov modulated

Poisson process (MMPP) which calculates packet loss prob-

abilities for a set of superpositioned voice input sources and

the specified link properties. We do not introduce another

new model to the community, rather try and verify one of

the existing models via extensive simulation and a real world

implementation. A plethora of excellent research on queu-ing theory is still in the domain of ATM researchers and we

attempt to highlight its validity to the IP Telephony commu-

nity.

Packet level simulations show very good correspondence

with the predictions of the model. Our main contribution is

the verification of the MMPP model with measurements in

a laboratory environment. The loss rates predicted by the

model are in general close to the measured loss rates and the

loss rates obtained with simulation. The general conclusion

is that the MMPP-based model is a tool well suited for di-

mensioning links carrying packetized voice in a system with

limited buffer space.

KeywordsLink Dimensioning, Markov Process, IP Tele-

phony, MMPP/D/1/K

I. INTRODUCTION

Voice applications, such as telephony, have been used

on the best effort service provided by the Internet for quite

some time. Currently many telephone operators have ad-

vanced plans to use IP technology as a bearer also for the

regular telephone service. This, however, requires that the

IP network can provide service guarantees.

Quality of Service (QoS) issues are being addressed bymany forums, committees and researchers. Research on IP

QoS has concentrated on the issues of classifying, schedul-

ing and admission of packets into a network. Less has been

done on how to dimension an IP network carrying real time

traffic.

This paper focuses on dimensioning IP network links

intended to carry packetized telephony or voice calls. It is

feasible that existing carriers would like to allocate a por-

tion of their bandwidth for this service and through mech-

anisms like differentiated services [11] provide superior

Node 1a0

a1

a3

Buffer

1.536Mbits/secSources (60-80)sink

Voice Node 0

Fig. 1. Problem: dimensioning a link for voice sources over IP.

service for this kind of data and subsequently levy highercharges.

Our approach is to look at work done in both the ATM

and traditional telephony communities as well as to use

tools and simulators from the IP community to verify these

ideas in an environment relevant for the Internet today. We

have seen very little work which has taken this approach.

The research community is divided into one of the two

camps (but is changing as ATM and telephony people are

more engaged in Internet research now).

Figure 1 illustrates the problem scenario we are address-

ing. A number of packet voice sources are multiplexedonto a link. The link has a limited amount of buffering

which sometimes will result in the loss of packets with the

obvious consequences on sound quality. With a link of a

given bandwidth and a number of voice sources, what kind

of quality could be expected if we ran 60 sources? What if

we increased to 80can we still expect adequate quality?

How will we affect the system by changing the amount of

buffering in the router?

We present a mathematical model based on a Markov

modulated Poisson process (MMPP) which can predict the

packet loss probability. We first verify the model using theNS packet level simulator. The main contribution of this

paper is the verification of the MMPP model with mea-

surements in a lab network. These experiments show a

very good correspondence between the loss rate predicted

by the model and the loss rate measured in the lab.

The rest of the paper is organized as follows. After

summarizing relevant related work in the next section, we

present the MMPP-based mathematical model and the rea-

soning leading to this model in Section III. Section IV

describes the parameters we used in the experiments. Sec-


35/87

IPTEL2001 36

tions V and VI describe the NS simulations and the labo-

ratory experiments, respectively. The experimental results

are presented and discussed in Section VII and the paper

is concluded with Section VIII.

II . RELATED WORK

Link dimensioning for voice has been a research topic

for several decades in both academia and the telecommuni-

cations industry. Starting a little more than ten years back,

the research focus has been on link dimensioning for ATM

networks. Most of the results in the domain of ATM net-

works are also applicable in the domain of IP networks,

since both are packet switching systems. The majority of

the results from previous research is theoretical or results

from simulations. Our research also has results from mea-

surements of a real system.

Several approaches have been suggested in the litera-

ture to solve the problem of dimensioning links in packet

switched networks. Anick, Mitra and Sondhi [2] study amultiplexer with infinite buffer with a stochastic fluid flow

model but it is shown by Zheng [14] that this model only

works for a multiplexer under heavy load. Tucker [15]

studies a multiplexer with finite buffer using the fluid

flow model, but it does not fit the model well for small

buffers. Heffes and Lucantoni [7] uses a two-state Markov

modulated Poisson process (MMPP) quite successfully to

estimate the delay in a multiplexer with infinite buffer

size. They suggest that the same approach for calculat-

ing the parameters of the MMPP can be used for a mul-

tiplexer with finite buffer size, but Nagarajan, Kurose andTowsley [10] show that this does not work in the case of

finite buffer size. Instead, they develop a different method

for finding the parameters of the MMPP. Baiocchi et

al. [4] approximate the arrival process with a two-state

MMPP and suggest a method called asymptotic matching

for the calculation of the parameters of the MMPP. This

approach is used by Andersson [1] together with a proce-

dure to calculate the loss probabilities developed by Baioc-

chi, Melazzi and Roveri [3] to study a multiplexer loaded

with a superposition of voice sources.

III . MATHEMATICAL MODEL

In this section we develop a mathematical model for di-

mensioning a link carrying voice traffic. We start with the

arrival process of a single IP telephony source and pro-

ceed with the superposition of independent identically dis-

tributed sources. The sources are then multiplexed on a

bottleneck link through a queue of limited size. A more

detailed description of this model can be found in previous

work by one of the authors [1]. The model is based on a

model developed by Baiocchi, Melazzi and Roveri [3].

T

ON OFF

t

Packet

size

T T

Fig. 2. Characteristics of a single source.

A. Single source properties

Most standard voice encodings have a fixed bit rate and

a fixed packetization delay. They are thus producing a

stream of fixed size packets. This packet stream is however

only produced during talk-spurtsthe voice coder sends

no packets during silence periods.

The behavior of a single source is easily modeled by a

simple on-off model (Figure 2). During talk-spurts (ON-

periods), the model produces a stream of fixed size packetswith fixed inter-arrival times T. Note that the first packetis produced one packet time after the start of an on-period.

This is the result of the packetizationthe voice coder

has to collect voice samples before it can produce the first

packet.

The number of packets in a talk-spurt, denoted with the

stochastic variable Nb, is assumed to be geometrically dis-tributed on the positive integers with mean n. This meansthat we can never have zero packets in a talk-spurt. This

variant of the geometric distribution is sometimes called

first success distribution (see for instance Gut [6, page258]), and has the probability function:

P(Nb = k) = qpk1, k = 1, 2, 3, . . . (1)

where q represents the probability that a packet is thelast one in a talk-spurt. This means that p = n1

n. This

fact implies that the ON-periods have a expected value of

= nT, where n is the expected value of the number ofpackets in a talk-spurt.

We assume that the OFF-periods are exponentially dis-

tributed with mean , which is well documented and dis-cussed by Sriram and Whitt [13]. A voice source may be

viewed as a two state birth-death process with birth rate and death rate . The OFF state represents the idle peri-ods and the ON state represents the talk-spurts. While in a

talk-spurt, packets are generated with a rate of 1T

packets

per second.

B. Approximating the single source

We have chosen to approximate the above model using

exponentially distributed inter-arrival times with mean T


36/87

IPTEL2001 37

Exp(1/T)

ON OFF

t

Packet

size

Exp(1/T)

Fig. 3. A single source approximated with exponentially dis-

tributed inter-arrivals.

instead of fixed inter-arrival times. The purpose of the ap-

proximation is to simplify the modelling of many sources.

We let Exp( 1T

) denote the stochastic variablewhich describes the inter-arrivals during talk-spurts, and

Nb be the geometrically distributed stochastic variablewith the probability function stated in Equation 1 with

mean n describing the number of packets in a talk-spurt.Moreover and Nb are assumed to be independent. It can

be easily seen that the ON-periods (denoted U) are expo-nentially distributed and that the mean length of a talk-

spurt is the same as in the deterministic inter-arrival case

(nT). Figure 3 illustrates the behaviour of a single sourcewith exponentially distributed inter-arrivals.

As in the previous section the OFF-periods are assumed

to be exponentially distributed with mean . Because ofthe exponentially distributed inter-arrival times during a

talk-spurt, the emission of packets during an ON-period

can be regarded as a Poisson process with intensity T. Wecan use the two state birth-death process to describe the

packet generation with one state representing the idle pe-riods and the other state representing the talk-spurts where

packets are generated as a Poisson process with inten-

sity T.

C. The superposition of independent voice sources

The superposition of voice sources can be viewed as a

birth-death process where the states represent the number

of sources that are currently in the ON-state. Here state

i represents that i sources are active in a talk-spurt. Werefer to the birth-death process as the phase process J(t).The birth rate is given by the mean of the exponentially

distributed idle periods, and we denote the mean as 1

. The

death rate is determined by the mean of duration of the

talk-spurts and is denoted 1

. The probability pon that asource is on is given by:

pon =

+ .

D. Markov modulated Poisson process

The Markov modulated Poisson process (MMPP) is a

widely used tool for analysis of tele-traffic models (see,

Poisson rates

N-1 N10 . . . . .

N (N-1) 2

N2 (N-1)

T NTN-1 T

Fig. 4. Superposition ofN voice sources with exponentiallydistributed inter-arrivals.

100

101

102

103

104

0

2

4

6

8

10

12

14

16

18

20

IndexofDisp

ersionforIntervals(IDI)

# consecutive intervals (k) in log 10

N=1

N=10

N=60

N=130

Poisson

Fig. 5. k-interval squared coefficient of variation curves forsuperposition ofN voice sources.

e.g., Heffes and Lucantoni [7]). It describes the superpo-sition of sources of the type described in Section III-B.

When the phase process is in state i, i sources are on. Themodel graph of the MMPP is shown in Figure 4.

The superposition of Poisson processes is also a Poisson

process. We can therefore simply add the intensities of the

sources that are currently in a talk-spurt and receive a new

Poisson process for the superposition.

To validate the accuracy of approximating with a MMPP

process, we calculated the index of dispersion of intervals

(IDI) using a formula from Sriram and Whitt [13]. The

IDI, also called the squared coefficient of variation, givesus some measure of how similar the traffic is in terms of

burstiness. A value of 1 shows the traffic is as bursty as

Poisson traffic, whereas a value as 18 is the burstiness of a

single voice source. The high value accounts for the fact

that the source is indeed bursty. The time period under

which one observes this behaviour is very important.

Figure 5 shows c2kN, the IDI, versus k for k between 1and 2000 and the number of sources, N, equal to 1, 10, 60and 130. As a reference we have added the value ofc2kNfor a Poisson process. Data was obtained from simulations


37/87

IPTEL2001 38

using a Matlab program. The solid line shows the c2kNfor sources with deterministic inter-arrival times between

packets during a talk-spurt, and the dashed lines show the

c2kN for sources with exponentially distributed inter-arrivaltimes, i.e., the MMPP approximation.

We see in the figure that the two descriptions of a sin-

gle source behave in a similar way when they are super-

positioned. The figure also shows that the superpositioned

arrival process behaves as a Poisson process if we look at

it for a short instant of time but it is much burstier if we

study it over a longer period of time.

E. The multiplexer: MMPP/D/1/K queue

The arrival process described by the MMPP model is fed

into a simple D/1/K queue. It is deterministic, has a sin-

gle FIFO server and a buffer size (waiting room) which we

vary. This kind of model is described in detail by Baioc-

chi et al. [3], [4]. We use their method and formulas for

calculating the loss probability.

IV. PARAMETER VALUES

We used the following parameters to run the MMPP

model, simulations and lab experiments:

32 kb/s ADPCM voice encoding with 16 ms packet inter-

arrival time, which results in 64 bytes of voice payload per

packet

A protocol header overhead consisting of 12 bytes for

RTP, 8 bytes UDP and 20 bytes IP. We do not include any

link layer headers. The resulting total packet size is 104

bytes, and the resulting bit rate is 52 kb/s.

The number of successive packets in one talk-spurt is

geometrically distributed on the positive integers with a

mean of 22, which results in a mean talk-spurt length of

352 ms. The idle time between two successive bursts is

exponentially distributed with a mean of 650 ms. The re-

sulting average fraction of time a source is in a talk-spurt

is 0.351.

The bottleneck is a T1 link with a bandwidth of

1.536 Mb/s.

These values coincide with Sriram and Whitt [13] as well

as previous work done by Zheng [14] whilst at SICS and

Andersson [1], except that we in this paper include proto-

col header overhead for the RTP/UDP/IP protocol stack.

Figure 6 shows loss curves computed with the MMPP

model for a sample set of buffer sizes. The next steps are

to compare these loss probabilities from the model with

results from NS simulations and measurements from a lab

network.

1e-05

0.0001

0.001

0.01

0.1

1

0 10 20 30 40 50 60 70 80 90 100

Loss

probability

Number of buffers

Mathematical MMPP model

60 sources

65 sources

70 sources

75 sources

80 sources

Fig. 6. Loss probabilities computed with the MMPP model.

Sources (N) Load ()

29 34.5 %

60 71.4 %

80 95.3 %

84 98 %

TABLE I

NETWORK LOAD FOR A NUMBER OF SOURCES.

A. Load

We use between 60 and 80 sources to load the link. To

define a load that is independent of the link bandwidth the

load factor, or , is used in the literature:

Load() =N Pon Ratepeak

C

where N is number of sources, C is the link capacity, Ponis the probability that the source is on and Ratepeak speaks

for itself. Table I shows loads for different numbers of

sources.

We decided to run between 60 and 80 sources as 84

sources is where the mean bandwidth of the sources equals

the bandwidth of the link. The peak allocation is as low as

29 sources (100 % utilisation whenPon = 1

) so taking ad-

vantage of the probability that a source is off yields much

higher link utilisation.

B. Buffer size

We have chosen to simulate a multiplexer with an output

link capacity of 1.536 Mb/s and buffer sizes ranging from

2 to 100 packets. With this choice of parameters we in-

troduce a maximum queueing delay of 54 ms in the buffer.

According to ITU recommendation G.114 [8] a delay of 0-

150 ms acceptable for telephony, between 150 and 400 ms


38/87

IPTEL2001 39

set cbr($i) [new Agent/CBR/UDP]

set exp($i) [new Traffic/Expoo]

$exp($i) set packet-size 104

$exp($i) set burst-time 0.352s

$exp($i) set idle-time 0.65s

$exp($i) set rate 52K

$cbr($i) attach-traffic $exp($i)

Fig. 7. Tcl code fragment defining a source NS-2.

can also be acceptable, but over 400 ms is not. The total

acceptable delay must be divided into a delay budget for

each node in the path between the sender and receiver. If

the path has 15 hops, and half of the delay budget can be

allocated to queueing delay, then we get 13.3 ms per hop.

This translates to approximately 24 buffers per hop. For

higher bandwidth links, the queueing delay per buffered

packet decreases inversely proportional to the bandwidth.

V. NS SIMULATION

We used ns-2 [5], a packet level simulator to verify

the MMPP model. Figure 1 shows the topology used in

the simulations and Figure 7 the Tcl code that is used to

start agents. They are constant rate sources, denoted by

CBR/UDP. Traffic/Expoo generates traffic based on an

exponential on/off distribution with the parameters speci-

fied in the next four lines. Each CBR source $i$ uses a

different random number seed, hence the sources will start

independently of each other.

The simulation should run long enough for the system

to reach steady state, ideally the system should be run for

an infinite amount of time, however this is not practical

due to time and resource constraints. A reasonable trade-

off is to use a simulated time of 1000 seconds in both the

simulation and the lab experiments. 1000 seconds with an

interval of 16 ms generates 22000 packets per source and

1.32 million packets for 60 sources or 1.76 million for 80

sources.

V I . LAB NETWORK MEASUREMENTS

A. Topology

Figure 8 shows the experimental setup. A single

machine acts as a traffic generator and emulates several

IP Telephony calls multiplexed together. The traffic is

then sent on a shared 100 Mb/s Ethernet and received by

two hosts: (1) a machine configured as a router; (2) a sink

machine for measurement purposes. An outgoing link of

the router is connected to the sink. In this configuration the

traffic is emitted by the generator, passes through the router

and is received by the sink. Since the sink can observe the

1.536Mbits/s

Hub

dummynetfxp1

fxp2 fxp2

fxp1

SICS net

fxp0

Traffic generator Router

queue

60-80 sourcesSink

100Mbits/s

Fig. 8. Topology for Laboratory. The outgoing interface of the

router is also connected to the sink.

packets before it enters the router, it can directly compare

latency and loss of each individual packet. The outgoing

link of the router is constrained to 1.536 Mb/s using Dum-

mynet [12] which is explained in the next section. All the

machines in the experiment were running FreeBSD 3.4.

B. Dummynet

Dummynet is a link emulator which allows arbitrary

bandwidths and latencies to be specified. It is often used

for emulating a slower link than what is physically avail-

able. Buffer sizes can be set for a given link and loss rates

set to emulate the effect of lossy links. It is possible to

create the illusion for TCP/UDP and IP that the link is like

a WAN rather than a LAN. We are primarily interested

in the lower bandwidth and configurable queue sizes. We

modified the output functionality slightly to enable simpler

calculation of the total number of packets received as well

as the drop rate. Recording the total number of packets re-

ceived gives us an additional check if the traffic generator

or any system component lost/dropped packets during theexperiment.

The total number of sent packets remained the same for

a given source count and can be checked with the output

of the traffic generator. It is trivial with a script to divide

the loss by the total number of packets to obtain the loss

rate.

C. Packet capture

To verify the loss rate we gathered the packets on the

sink machine via a program that we developed1 using the

Berkeley Packet Filter [9] . Figure 8 shows that the outputof the generator is attached directly to the sink machine

as well as the outgoing link of the router. This enables us

to capture all the packets and the ones not dropped by the

router. A simple difference between the two should ver-

ify the loss rate reported by Dummynet. Our bpf program

captures packets with a specific destination and port, and

prints the time of arrival, RTP src and seq fields.

1Not tcpdump. We wrote out our own kernel filter to extract the pack-

ets we wanted as well as a user space program to output headers from

2 interfaces simultaneously.


39/87

IPTEL2001 40

#define INVERSE_M ((double) 4.6566128200e-10)

/* little number */

int calc_length(double burstlen) {

double rand, logvalue;

rand = INVERSE_M * random();

logvalue = burstlen * -log(rand);

return ((int)(logvalue + 0.5));

}

Fig. 9. C code to randomize a burst length

D. Traffic generator

The idea of the traffic generator is to create a sequence

of packets that resemble many individual IP telephony

calls multiplexed together. Furthermore, it should perform

this job as accurately as possible with each packet emerg-

ing with a given deadline.

D.1 Trace file generation and playback

In order to be able to repeat experiments, we first pre-

calculate the sending times of the packets and generate

trace files. These files are then fed into the traffic generator

which sends packets according to the trace. The trace files

also allow us to test our setup to see if packets were be-

ing generated at the right times (such as inter-arrival times

and sequence). The files are generated on a per source ba-

sis. The average length of a burst is calculated as shown in

Equation 2.

burst length = rand

Poninterval

(2)

The C-code for the rand function is shown in Figure 9.

Using the logarithm of the random variable generates

burst lengths which are exponentially distributed.

The same calculation is applied for the idle (with Poff)period. The result is (reading vertically for each source) an

exponentially distributed series of ON and OFF sequences

with a mean ON of 0.351 seconds, OFF of 0.65 seconds

which results in a burst length of 22 packets. An example

of a trace file2 with ten sources is shown in Figure 10.

The file shows for each time step (in this case 16 ms)which of the 10 sources are on or off. In the example,

sources 1, 3, 5, 7 and 9 sends packets in the first time step.

The traces of one source can be followed by reading a col-

umn downwards. Source 2, for example, sends no packet

in the first timestep, but then sends a packet in each of the

succeeding steps.

If there are n sources, each timestep is further subdi-vided into n sub steps. Each sub step defines the sending

2Actually it is converted into a binary format for more compact rep-

resentation

source

0 1 2 3 4 5 6 7 8 9

time

0 0 1 0 1 0 1 0 1 0 1

1 0 1 1 0 1 1 0 1 1 1

2 1 1 1 1 1 1 0 1 0 1

3 1 0 1 0 1 0 1 0 0 1

4 1 0 1 0 1 0 1 0 0 1

5 0 1 1 0 1 1 0 1 1 0

Fig. 10. Traffic generator trace file.

5

16ms

0 1 2 3 4 5 6 7 8 9

0

0 1 2 3

1

source interval

timestep

3 9 11 27

Fig. 11. Traffic generator sending times

interval for each source. For example, with ten sources and

a time step of 16 ms starting at t, source 0 sends its packetwithin [t, t + 1.6]; source 1 sends within [t + 1.6, t + 3.2],etc. If a source does not send its packets within its interval,

it is said to miss its deadline. Packets that miss their dead-

line are recorded by the generator and printed when the

run has completed as well as the largest value by which a

packet was delayed.

So for the trace file above, the first steps of a packet se-

quence is shown in Figure 11. The sending of each packet

is depicted as a horizontal interval, corresponding to theentering and leaving of the send system call, respectively.

In the picture, the packets of source 5 and 7 missed their

deadlines. The actual sending time on the link can be mea-

sured by an external mechanism, such as the packet cap-

ture program described previously.

D.2 Traffic generator verification

As a simple test for a trace file of 220000 packets we

obtained values 36.9 % for the on time, 63.1 % for the off

time by simply counting the ones and zeros in one column

of the file. The mean number of packets in a burst equalled22.5. Using the trace files turned out to be more useful than

we first expected, despite the performance gains of replay-

ing pre-calculated files they also allowed us to test the per-

formance of our traffic generator (setting all the sources

on), cross check parameters as just stated as well as gener-

ating special sequences for analysing queue behaviour.

D.3 Traffic generator verification

We calculated the index of dispersion of intervals, or

IDI (see Section III-D), also for the lab traffic generator.


40/87

IPTEL2001 41

0

2

4

6

8

10

12

1 10 100 1000

Indexofdispersionforintervals(IDI)

# consecutive interval (k) in log 10

IDIs for lab and simulation"idi_lab_75.txt""idi_sim_75.txt""idi_poisson.txt"

Fig. 12. IDI curves for superposition of 75 sources

In Figure 12

we can see that the simulation and lab traffic generator

produce similar types of traffic. The larger the observa-

tion time the more skewed the traffic is. One voice sourceis equal to about 18.1 also a value is given for a Poisson

sources. The graphs show the result of a trace which was

10000 simulated seconds, resulting in 17.3 million packets

for the lab and 16.3 for the simulation.

The traffic generator was also tested to ensure it (and the

machine on which we run on) was capable of outputting

packets as close to their deadlines as possible

VII. RESULTS

In this section we present and dis

thesis1 plc

Documents