5 Next-Generation Telecommunications Networks: Core Network part. 49 ...... Heinanen J., Baker F., Weiss W., Wroclawski J., âAssured Forwarding PHB Group,â ...
Tampereen teknillinen korkeakoulu Julkaisuja Tampere University of Technology Publications
Yevgeni Koucheryavy
Multimedia Traffic Delivery over NextGeneration Telecommunications Networks
Tampere 2004
v
ABSTRACT Third-generation (3G) systems have created the basis for multimedia services with Quality of Service (QoS) support on the move. It is expected that 3G architecture and related QoS provision approaches will evolve to face the challenges of future Next-Generation Networks (NGN). QoS assurance frameworks developed for fixed networks bring a huge experience to many areas. Often QoS in fixed networks is not assured for each particular traffic flow, but rather is guaranteed for the aggregated flow of a given class. This approach does not seem valid for future scenarios where plenty of applications and different user profiles will require a finer characterization of QoS levels. Additionally, within NGN framework calllevel performance indicators such as call blocking probability and handover failure rate, typical of Second-generation (2G) systems, should be integrated with packet-level QoS metrics dealing with packet error probability, packet delay, and so on. Therefore, inherent characteristics of mobile users, like their mobility and traffic demands, as well as typical features of mobile systems, like unstable nature of the air interface, have to be addressed before the required quality and transparency of user services will be achieved by NGN systems. Finally, current 3G systems are characterized by lack of energy efficiency considerations, end-to-end seamless transport, reconfigurability, and application scalability. These limitations have to be overcome in future NGN systems. The first part of this research work introduces the NGN system design and deals with cross-layer design, both for performance optimization and network efficiency. The proper NGN system design becomes a very important task, especially accounting the provision of effective solutions for ubiquitous access to the Internet and the QoS guarantees for different traffic classes. It presumes that the system must be optimized to improve the performance perceived by users and the efficient use of network resources, which is an important prerogative for making revenue for operators and for allowing the realization of services at affordable costs. Based on the results of the NGN system design, the cross-layer black-box framework is developed. This framework allows qualitative and quantitative evaluation of QoS expectations of user applications running over the wireless interface. The second part of this research work is dedicated to practical realization of the results from the first part. Taking into account that the QoS assurance problem falls into radio access network QoS and core network QoS, this problem is addressed from two different perspectives. At first, practical evaluation of the available off-the-shelf wireless technologies under different conditions is performed. The dedicated testbed supporting WLAN and GPRS access networks is built for these purposes in accordance with the cross-layer black-box framework developed in the first part of the thesis. Then the core network is considered. Based on the Differentiated Services Internet architecture model, analytical studies of the NGN core network comprised of a sequence of QoS-enabled domains is performed. Two analytical models reflecting transmission services are proposed and evaluated. The models are characterized by Differentiated Services ingress node parameters and bounded delay within the network nodes along the path of the behavior aggregate and allow predicting the QoS degradation which is experienced by both the behavior aggregate and single microflows.
Preface
The research work presented in this thesis was carried out during the years 2001 – 2004 at the Institute of Communications Engineering, Tampere University of Technology, Tampere, Finland. I would like to express my deepest gratitude to Prof. Jarmo Harju, my thesis advisor, for his great support and allowing me the time to undertake this project. I also would like to thank Sari Kinnari and Tarja Er¨alaukko, secretaries of the laboratory, for their invaluable help with practical matters. Many thanks go to Dmitri Moltchanov, Andrey Krendzel and Roman Dunaytsev for outstanding technical and non-technical discussions. Last, but no means least, I would like to say a simple and heartfelt thank you to my wife Maria and my children Ivan and Egor, to my dearest parents and sister, to all my nearest relatives for understanding and patient forbearance throughout. YEVGENI KOUCHERYAVY Tampere, August 25, 2004
vii
Contents
Preface
vii
List of Publications
xiii
List of Symbols and Acronyms
xv
1 Introduction 2 Quality of Service for Next-Generation Networks 2.1 Teletraffic and Quality of Service Assurance in the Internet . . . . . . . 2.1.1 Teletraffic and its properties . . . . . . . . . . . . . . . . . . . 2.1.2 Factors Affecting Quality of Service . . . . . . . . . . . . . . . 2.2 Differences between wireline and wireless networks . . . . . . . . . . 2.3 Wireless QoS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Basic ideas of QoS provision . . . . . . . . . . . . . . . . . . . 2.3.2 Multimedia and wireless . . . . . . . . . . . . . . . . . . . . . 2.4 Networking after 2G: a services-driven tendency . . . . . . . . . . . . 2.4.1 3G Network Architecture . . . . . . . . . . . . . . . . . . . . . 2.4.2 NGN Network Architecture . . . . . . . . . . . . . . . . . . . 2.4.3 Reconfigurable Access Networks . . . . . . . . . . . . . . . . . 2.4.4 Always Best Connected concept . . . . . . . . . . . . . . . . . 2.5 NGN System design for performance optimization and network efficiency 2.6 End-to-end QoS of NGN networks . . . . . . . . . . . . . . . . . . . 2.6.1 Service differentiation . . . . . . . . . . . . . . . . . . . . . . 2.6.2 Connection-oriented resource reservation . . . . . . . . . . . . 2.6.3 Integrated approach . . . . . . . . . . . . . . . . . . . . . . . 2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
5 6 7 7 9 10 10 11 11 12 12 13 13 13 16 17 17 18 20 ix
x
CONTENTS
3 Cross-Layer Design for NGN environment 3.1 End-to-end service configuration in NGN 3.2 Components of cross-layer model . . . . 3.2.1 Mobility models . . . . . . . . . 3.2.2 Teletraffic models . . . . . . . . 3.2.3 Integrated models . . . . . . . . 3.2.4 Models of wireless channel . . . . 3.3 Cross-layer black-box framework . . . . 3.4 Summary . . . . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
23 24 26 26 26 26 27 29 30
4 Next-Generation Telecommunications Networks: Access Network part 4.1 Wireless Access Peculiarities . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Multimedia traffic treatment in current evolutions of wireless access networks 4.2.1 Services under consideration . . . . . . . . . . . . . . . . . . . . . 4.2.2 Current WLAN implementation . . . . . . . . . . . . . . . . . . . 4.2.3 Current GPRS implementation . . . . . . . . . . . . . . . . . . . . 4.3 Practical study: testbed configuration . . . . . . . . . . . . . . . . . . . . 4.4 Tools and Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 On video streaming service implementation and testing . . . . . . . 4.4.2 On mp3-based services implementation and testing . . . . . . . . . 4.5 Test scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 WLAN end-to-end performance testing . . . . . . . . . . . . . . . 4.5.2 GPRS end-to-end performance testing . . . . . . . . . . . . . . . . 4.5.3 Notes on end-to-end performance testing for different access technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.4 Video streaming service . . . . . . . . . . . . . . . . . . . . . . . 4.5.5 mp3-based services . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43 43 44 46
5 Next-Generation Telecommunications Networks: Core Network part 5.1 QoS support on base of DiffServ . . . . . . . . . . . . . . . . . . 5.2 Real-time service . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.2 Real-time service configuration . . . . . . . . . . . . . . . . 5.2.3 Source configuration . . . . . . . . . . . . . . . . . . . . . 5.2.4 Modeling of real-time service traffic . . . . . . . . . . . . . 5.3 Real-time service traffic models . . . . . . . . . . . . . . . . . . . 5.3.1 Stochastic and deterministic traffic modeling . . . . . . . . . 5.3.2 Stochastic real-time service traffic model . . . . . . . . . . . 5.3.3 Deterministic real-time service traffic model . . . . . . . . .
49 50 50 50 51 52 52 53 53 54 54
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
33 33 34 35 36 37 39 40 40 41 42 42 42
xi
CONTENTS
5.4 AF PHB transmission service . . . . . . . . . . . . . 5.4.1 Token bucket parameters . . . . . . . . . . . . 5.4.2 Estimation of AF class queue parameters . . . . 5.4.3 Per-source QoS degradation . . . . . . . . . . 5.5 EF transmission service . . . . . . . . . . . . . . . . 5.5.1 Definition of the Service . . . . . . . . . . . . 5.5.2 Token Bucket Parameters Estimation . . . . . . 5.5.3 Token bucket parameters violation . . . . . . . 5.5.4 EF queue parameters estimation . . . . . . . . 5.6 Real-time service traffic treatment within interior nodes 5.7 Summary . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
55 56 57 60 61 61 62 63 64 66 67
6 Conclusions
69
7 Summary of Publications 7.1 Overview of publications . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Author’s Contribution to the Publications . . . . . . . . . . . . . . . . . .
71 71 73
References
75
Publications
87
List of Publications
[P1] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Analytical Estimation of EF PHB Service Parameters for Aggregated MPEG Traffic,” in Proc. of the 16th Nordic Teletraffic Seminar, NTS’02, Espoo, Finland, August 2002, pp. 279 – 290. [P2] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”An Analytical Evaluation of VoD Traffic Treatment within the EF-enabled DiffServ Ingress and Interior Nodes,” in Proc. of the IEEE International Conference on Telecommunications, ICT’03, Papeete, Tahiti, French Polynesia, February 2003. [P3] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”A Top-Down Approach to VoD Traffic Transmission Over DiffServ Domain Using the AF PHB Class,” in Proc. of the IEEE International Conference on Communications, ICC’03, Anchorage, Alaska, USA, May 2003. [P4] Y. Koucheryavy and D. Moltchanov, ”Notes on Quality of Service and Performance Evaluation 4G All-IP networks,” in Proc. of the 1st International Workshop on Wireless, Mobile and Always Best Connected, ANWIRE’03, University of Strathclyde, Glasgow, UK, April 2003, CD ISBN 0-9545660-0-9. [P5] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Performance Evaluation of Live Video Streaming Service in 802.11b WLAN Environment Under Different Load Conditions,” in Proc. of the ACM International Workshop on Multimedia Interactive Protocols and Systems, MIPS’03, LNCS 2899, Napoli, Italy, November 2003, pp. 30 – 41. [P6] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Impact of Mobility on Entertainment Services’ Performance in Heterogeneous Wireless Environment,” in Proc. of the Australian Telecommunications, Networks and Applications Conference, ATNAC’03, Melbourne, Australia, December 2003, CD ISBN 0-646-42229-4. [P7] Y. Koucheryavy, D. Moltchanov, J. Harju, and G. Giambene, ”Cross-Layer BlackBox Approach to Performance Evaluation of Next Generation Mobile Networks,” in Proc. of the IEEE International Conference on Next Generation Teletraffic and Wireless/Wired Advanced Networking, NEW2AN’04, St.-Petersburg, Russia, February 2004, pp. 266 – 272, ISBN 952-15-1132-X. xiii
xiv
List of Publications
[P8] Y. Koucheryavy, D. Moltchanov, G. Giambene, and J. Harju, ”Teletraffic Requirements and System Aspects for Future Mobile Communications,” in Proc. of the 8th IASTED International Conference on Internet & Multimedia Systems & Applications, IMSA’04, Kauai, Hawaii, USA, August 2004.
List of Symbols and Acronyms
2G 3G 3GPP 4G ABC ACF ADSL AF AL AN ARQ ATM BB BER BMAP BR BS CAC CBR CDMA CN CR CSCN
Second Generation System Third Generation System 3rd Generation Partnership Project Forth Generation System Always Best Connected Autocorrelation Function Asymmetric Digital Subscriber Line Assured Forwarding Application Layer Access Network Automatic Repeat Request Asynchronous Transfer Mode Bandwidth Broker Bit Error Rate Batch Markovian Arrival Process Border Router Bearer Service Connection Admission Control Constant Bit Rate Code Division Multiple Access Core Network Core Router Circuit Switched Core Network xv
xvi
List of Symbols and Acronyms
D-BMAP DL D-MAP D-MMBP DoS DSCP EDGE EF ER FCFS FEC HLR HMM HTTP QM QoS QoSD GERAN GGSN GPRS IMS IP IPv6 ITU LAN LOS LR MAC MANET MAP NGN OS PCM PDF PDP
Discrete Batch Markovian Arrival Process Data Link layer Discrete Markovian Arrival Process Discrete Modulated Bernoulli Process Denial-of-Service Differentiated Services Code Point Enhanced Data Rates for Global Evolution Expedited Forwarding Edge Router First Come First Served Forward Error Correction Home Location Register Hidden Markov Model Hypertext Transfer Protocol Queue Management Quality of Service Quality of Service Domain GSM-EDGE Radio Access Network Gateway GPRS Support Node General Packet Radio Service IP-based Multimedia Services Internet Protocol Internet Protocol, version 6 International Telecommunication Union Local Area Network Line Of Sight Local Router Medium Access Control Mobile Ad-hoc NETworks Markovian Arrival Process Next Generation Network Operating System Pulse Code Modulation Probability Distribution Function Packet Data Protocol
List of Symbols and Acronyms
PDN PHB PHY PSCN RAN R&D RED RLP RNC RTSP SGSN SIP SLA SNR TC TCA TCP TL UDP UMTS USIM UTRAN VBR VPN WAN WAP WCDMA WLAN
Packet Data Network Per Hop Behavior Physical layer Packet Switched Core Network Radia Access Network Research and Development Random Early Detection Radio Link Protocol Radio Network Controller Real-Time Streaming Protocol Serving GPRS Support Node Session Initiation Protocol Service Level Agreement Signal-to-Noise Ratio Traffic Category Traffic Conditioning Agreement Transmission Control Protocol Transport Layer User Datagram Protocol Universal Mobile Telecommunications System Universal Subscriber Identity Module Universal Terrestrial Access Network Variable Bit Rate Virtual Private Network Wide Area Network Wireless Application Protocol Wideband Code Division Multiple Access Wireless Local Area Network
xvii
Chapter
1
Introduction
N
EW telecommunication services and telematic applications are the strong drivers of progress and they pose new requirements to network design and construction issues. To implement new services within a new-generation mobile network, a set of novel network architectures, protocols, and traffic-related mechanisms have to be invented. Nowadays, when third-generation (3G) networks are coming into play, it is a fair question to ask what a fourth-generation network (4G), in other words Next-Generation Networks (NGN), might or should be. The answer is not clear so far, but the trend is to regard NGN as a new paradigm for telecommunications networks, mobile and fixed, where network and service characteristics of reconfigurability, interworking, adaptable services, and interoperability are met. It is highly anticipated that the most important, value-added, and revenue-expected new services for the mobile network will be Internet access and Internet Protocol (IP) multimedia applications. As shown in a number of R&D projects worldwide, on the basis of a plain fixed Internet access it is possible to implement a set of brand-new applications for which certain Quality of Service (QoS) requirements need to be provided. No doubt, QoS for Internet applications will be in heavy demand by end-users. Indeed Internet QoS goes mobile is why, with respect to the existing experience on the implementation of QoS within fixed networks, design, development, and implementation of QoS-capable mobile networks are of paramount importance. NGN systems should be flexible enough to support all those services of today’s fixed networks without any limitations, and to provide them in an Always Best Connected (ABC) access strategy, so that the users are given with the most suitable connection for a given environment, satisfying as much as possible the QoS requirements of the service and the user at a price which is right for both parties. Such features should be realized through multi-access networks and should be transparent to the user. The main objective of the research being conducted by the research community is to develop and to investigate new approaches, techniques, methods, models, strategies, and tools for the analysis, design, control, and evaluation of future advanced NGN systems supporting user mobility, multimedia applications, and interworking. Special attention is given to QoS and related aspects in both access networks and core networks in the presence of 1
2
INTRODUCTION
mixed multimedia traffic. To accomplish these tasks, new analytical tools, software implementations, and prototypes have to be developed. In connection with the issues mentioned, the specific topics of research can be formulated as follows [1]: • Traffic engineering, congestion control, and parameters estimation in packetswitched mobile and wireless environment • QoS for future packet-switched mobile and wireless networks • Network planning and architectures Traffic engineering encompasses the application of scientific principles and technology to the measurement, modelling, characterisation, and control of multimedia multiclass traffic and the application of such knowledge and techniques in order to achieve specific performance objectives, including the planning of network capacity under QoS guarantee and the efficient, reliable transfer of information. The need to allocate and balance resources among different traffic classes to accomplish the best use of network resources is a crucial traffic engineering problem. As a matter of fact, traffic engineering and QoS issues need to be jointly considered. The major objective of traffic engineering is to improve network performance while maintaining the QoS requirements through the optimization of network resources, with the main focus of the optimisation being the minimization of the over-utilisation of capacity in certain parts of the network while other capacity is available/under-utilised in the same network. Multiprotocol Label Switching (MPLS) concept, if expanded and translated into mobile scenarios, may provide one route towards traffic engineering and QoS solutions. The optimization objective depends on the specific goal of network operators, which may include minimizing congestion, minimizing packet loss/delay, or minimizing the blocking probability. Network management and control can be considered very complex, and, thus, will require robust, possibly intelligent, control methodologies to obtain satisfactory (if feasible, optimal) performance. The development of efficient and effective management and control techniques may include issues regarding resource management, congestion control, connection admission control, and active queue management. A further aim is the evolution of techniques for anomality and event analysis based on QoS and traffic monitoring in multimedia mobile networks such as detection of specific sources for events, malformed packets, security violations, and attacks with special consideration of mobile networking, as for instance vulnerabilities occurring at the translation point between the wireless protocols and the wireline (fixed) protocols. It should be emphasized that traffic patterns generated by IP multimedia services are quite different from traditional Poisson models used for circuit-switched voice traffic. As a result, the network parameters can be underestimated if inadequate traffic models and analytical approaches are adopted. Therefore, within NGN traffic engineering problems, a particular problem is that of the performance analysis of network elements taking into account the self-similar nature of multi-service traffic. Hence, it is necessary to derive, for instance, the upper and lower bounds of a service provision rate for 3G and in systems beyond 3G. Among the latter one could already count wireless LAN (WLAN) and mobile ad-hoc network clouds as access networks, so relevant considerations for traffic engineering should already extend to include WLANs and Mobile Ad-hoc NETworks (MANETs).
3
QoS provisioning in wireless environment involves mechanisms, algorithms, and schemes at various layers of the OSI Reference Model; in particular, physical layer, Medium Access Control (MAC) layer, IP layer and transport layer. The basic idea to be investigated is that QoS support requires joint collaboration among all these layers; how it may be done, new definitions of cross-layer protocols, and so forth, are important areas for research. The investigated techniques will be able to manage multimedia traffic with different characteristics in terms of burstiness, QoS, load, etc. Also regarding this topic, a huge number of issues for investigation exist, some of which are statistical traffic models for NGN, mobility and location awareness, dynamic resource allocation mechanisms, and adaptive MAC protocols depending on traffic load and channel propagation conditions and based on QoS requirements, mobile Virtual Private Networks (VPN), security, handoff techniques, etc. Novel and improved analytical and simulation models will be instrumental for the definition of correct techniques for the design and planning of multi-service wireless IP networks with QoS guarantees. The principal contribution of the first part of the thesis is to study the area of NGN system design and to develop a cross-layer abstract model for NGN. To perform the proper NGN system design, a number of important issues must be addressed, namely: network architecture; access network peculiarities; QoS assurance at different layers; etc. Developing of cross-layer abstract model based on the NGN system design principles allows to define QoS cross-layer framework for NGN systems. The principal contribution of the second part of the thesis is to apply the QoS cross-layer framework developed in the first part. The QoS assurance task comprises access network (AN) QoS assurance and core network (CN) QoS assurance. The cross-layer framework can be implemented as a practical testbed supporting different access networks. Evaluation of QoS expectations that a user application may experience running over the wireless channel is very sophisticated task which involves a number of interdependent stochastic factors. This practical testbed is a complex translator of parameters at the input of the system to corresponding output characteristics. Therefore, it allows practical evaluation of QoS expectations. The CN QoS assurance problem can be considered independently. Properly parametrized NGN CN would not heavily contribute to QoS degradation. Analytical techniques are the good choice in this case, they allow obtaining parameters either analytically or via simulations. The rest of the thesis is organized as follows. Chapter 2 deals with QoS for NGN systems and the NGN system design. In Chapter 3 the cross-layer framework is developed. In Chapters 4 and 5, the main research work of the thesis is described. Chapter 6 includes some conclusions of the thesis and Chapter 7 introduces the publications of the thesis.
Chapter
2
Quality of Service for Next-Generation Networks
T
HE idea to provide certain applications with different types of networking services was conceived since the beginning of the Internet. Bandwidth limitations and buffer overload are two of major factors limiting the assured timely delivery of packets end-to-end. The introduction of a wireless part in the physical (PHY) layer Internet infrastructure poses new specific challenges not only in bottom layers, but also in upper ones. However, one of the first major challenges is to address the quality of service (QoS) question in the end-to-end IP scenario. This is an inherent problem for many service types even in fixed networks. Wireless and mobility add their own QoS problems on top of this inherent IP flaw. Below, the main focus is dedicated to new ways through which IP-based NGN networks may provide QoS guarantees, thus requiring the definition of appropriate network architectures and QoS frameworks as well as traffic and mobility issues. Current multimedia applications like audio, video, and data are characterized by much higher bit rates compared to those produced by compressed speech information or even by high-quality pulse code modulation (PCM) codecs, which have been used for a long period of time in both fixed and mobile networks. These new services require the development of new methods of traffic theory, optimization, and design. For example, it is necessary to develop a theory of multi-dimensional distributed queues with priority service for integrated traffic in multi-point radio channels with multiple accesses. Among other things, special attention should be paid to traffic modeling – it provides the starting point in theoretical analyses of QoS experienced by application. 3G systems have created the basis for multimedia traffic with QoS support on the move. Therefore, 3G system architecture and related QoS provision approach should evolve to face such challenges of future systems. QoS is not assured for each particular traffic flow, but rather is guaranteed for the aggregated flow of a given class. This approach does not seem to be valid for future scenarios where plenty of applications and different user profiles will require a finer characterization of QoS levels. Additionally, call level performance indicators such as call blocking probability and handover failure rate, typical of 2G systems, 5
6
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
should be integrated with packet-level QoS metrics dealing with packet error, packet delay, and so on. Therefore, inherent characteristics of mobile users, like their mobility and traffic demands, as well as typical features of mobile systems, like unstable nature of the air interface, must be addressed before the required quality and transparency of user services can be achieved by mobile systems. Finally, current 3G systems are characterized by lack of energy efficiency considerations, end-to-end seamless transport, reconfigurability, and application scalability. These limitations have to be overcome in future NGN systems.
2.1 TELETRAFFIC AND QUALITY OF SERVICE ASSURANCE IN THE INTERNET The quality of communication services comprises several aspects such as the availability, reliability, and dependability of the services. These aspects relate to the expectation of users that services are always provided when needed. Other aspects relate to the quality of established communication sessions and are connected with delay, errors, and loss in the data transfer. For instance, a user of an application such as IP telephony has the expectation that an established call allows a conversation to take place, but normally the conversation does not break down just because of a high delay making interaction difficult, or errors that make speech unintelligible, or losses which interrupt the conversation. The session’s quality depends both on the design and performance of the system running the application as well as on the network. Therefore, some QoS degradation components introduced by the network can be smoothed away by the application’s logic and mechanisms. Network quality depends on the traffic load of the links in the network. This in turn depends on the amount and characteristics of the traffic entering the network, as well as on the change in characteristics due to multiplexing and the routing of the traffic. This apparently simple problem becomes very complex when a network is large geographically, administratively, and in terms of the population of attached devices. Indeed, traffic in the Internet is extremely complex, resulting from a very large and everlastingly changing range of applications. Traffic management consists in characterizing, measuring, classifying, and serving this traffic. A vast amount of work has been performed over the years on the various aspects of IP traffic management. Architectures like Integrated Services (IntServ, [2]) and Differentiated Services (DiffServ, [3]) have been proposed which would allow the network to offer a set of QoS guarantees to different types of applications. The traffic generated by various applications has been widely studied [4–6] and the statistical properties of individual flows [7–9] and traffic aggregates are beginning to be fairly well understood [11, 12, 67]. Despite these significant advances, it remains true to say that much uncertainty surrounds the future evolution of the Internet. One can argue that an alternative service model is required and that new network architectures and additional traffic controls are required. Research continues on characterizing traffic and on developing analytical modeling techniques that can enable performance evaluation and effective network design [13].
TELETRAFFIC AND QUALITY OF SERVICE ASSURANCE IN THE INTERNET
7
2.1.1 Teletraffic and its properties Traffic in the Internet results from the uncoordinated actions of a very large population of users and must be described in statistical terms. It is important to be able describe this traffic somewhat briefly in a manner which is useful for network engineering. The relative traffic proportions of TCP and UDP have varied quite little in recent years and tend to be the same throughout the Internet. More than 90% of data is in TCP connections [14, 15]. New streaming applications running over UDP are certainly becoming popular, but this extra UDP traffic is compensated by the increase in regular data transfers using TCP. For traffic engineering purposes it is not necessary to identify all the different applications comprising Internet traffic. It is generally sufficient to distinguish just three fundamentally different types of traffic: elastic traffic, streaming traffic, and control traffic. Elastic traffic corresponds to the transfer of data over TCP and is named so because the rate of transfer can vary in response to changes in network load. Streaming traffic results from audio and video applications which generate flows of packets having a target rate which must be preserved by limiting packet delay and loss. Control traffic is formed by a variety of signaling and network control protocols. While the efficient handling of control traffic is clearly important for the correct operation of the network, its relatively small volume makes it a kind of minor consideration for traffic management purposes. The traffic can be described in terms of the characteristics of a number of objects, including packets, bursts, flows, sessions, and connections. Traffic characterization is most convenient at flow level [16]. A flow is defined as the unidirectional sequence of packets relating to one instance of application (often referred to as a microflow). Packet-level characteristics of elastic flows are mainly induced by the transport protocol and its interactions with the network. On the other hand, streaming flows have a target rate that must be preserved as the flow traverses the network. The arrival process of flows in a backbone link typically results from the superposition of a large number of independent sessions and has complex correlation behavior. The size of elastic flows is extremely variable and might have a so-called heavy-tailed distribution: most documents are small (a few kilobytes), but the small number which are very big tend to contribute a high volume of traffic [17]. The precise distribution clearly depends on the types of applications. For example, e-mail has very different characteristics to mp3 files downloading. A number of studies on application traffic volumes exist [18– 20]. 2.1.2 Factors Affecting Quality of Service QoS is a widely used term today. However, the meaning of the term is not so well defined. It can mean different things in different contexts. Indeed, the Internet’s best effort delivery approach currently in use is not capable of efficiently supporting the different requirements of the various traffic types. The fact that video, audio, data, and multimedia applications have significantly different requirements led to the idea that a differentiated services approach could assure a differentiated quality of service.
8
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
Here and later on, a relatively simple approach is employes to build the QoS concept. Let us consider only three layers of the so-called QoS model, namely application, network, and data link. All these layers can be seen to have a different definition. At the application layer, QoS reflects user satisfaction. Basically, this depends on the perceived service level, user expectations, and cost. Service level is merely a reflection of the personal impression of the users, and consequently the best way to estimate it is testing conducted with a selected set of persons. The test results are able to show that for example 95% of persons were satisfied with the quality of a certain service. The challenge lies in the fact that applications and people vary and results are influenced by many parameters, which in turn are relatively difficult to define. Nowadays there are two most commonly used mechanisms for maintaining user satisfaction with multimedia services. • Adaptation. For example, if streaming video service is jeopardized due to the performance of lower layers, the client software can request a lower bandwidth coded video from the server. • Buffering. For example, typically lower layers contribute randomized delay factor to the delivery of stream packets, i.e. a destination-side application deals with jitter that can easily deny service just because of too high delay of a particular packet. Implementation of buffer on the destination side allows a certain number of packets to be kept between the network and the user’s application. The term QoS refers to the capability of a network to provide better service for selected network traffic over various technologies, including Frame Relay, Asynchronous Transfer Mode (ATM), Ethernet and IEEE 802.11 networks, SONET, and IP-routed networks that may use any or all of these underlying technologies. The primary goal of QoS is to provide priority, including dedicated bandwidth, controlled jitter and latency (required by some real-time and interactive traffic), and improved loss characteristics. Another important QoS issue is making sure that providing priority for one or more flows does not cause other flows to fail. QoS technologies provide the elemental building blocks that will be used for future business applications on campuses, wide area networks (WANs), and service provider networks (domains). QoS can be described qualitatively (relative) or quantitatively (absolute). Relative QoS definitions relate the treatment received by a class of packets to some other class of packets, while absolute definitions provide metrics such as delay or loss, either as bounds or as statistical indications. An example of absolute bounds are statements such as no more then 10% of packets will be dropped or no packets will experience a delay more of than 100 ms. A set of such statements along with guarantees about reliability are often called a service level agreement (SLA). As long as the sum of the bandwidths of the ingress links exceeds the minimum capacity of a network, QoS can be offered only in one of two ways: either by predicting the traffic and engineering the network to make violations of the committed QoS sufficiently unlikely or by restricting the total amount of traffic competing for the same resources. In many cases the network capacity is effectively partitioned by packet prioritization, so that higher priority traffic is largely unaffected by lower priority traffic. Applications differ in their QoS requirements. While data applications can recover from packet loss via retransmission, loss above 5% generally leads to very poor effec-
DIFFERENCES BETWEEN WIRELINE AND WIRELESS NETWORKS
9
tive throughput. Data applications such as file transfer are not delay sensitive, although human patience imposes lower throughput bounds on applications such as web browsing. Continuous media applications such as streaming audio and video generally require a fixed bandwidth. The diversity of applications running over current best effort networks, makes these networks inadequate. 2.1.2.1 Packet loss UDP cannot provide a guarantee that packets will be delivered at all, so in order. Packets will be dropped under peak loads and during periods of congestion. Regarding real-time multimedia, due to the time sensitivity of such transmissions, the normal TCP-based retransmission schemes are not appropriate. Approaches used to compensate for packet loss include interpolation of information by replaying the last packet and sending redundant information. Packet losses greater than 10 % are generally intolerable, unless the encoding scheme provides extraordinary robustness. 2.1.2.2 Jitter IP networks cannot guarantee the delivery time of data packets (or their order); the data could arrive at very inconsistent rates. The variation in inter-packet arrival rate is known as jitter and is introduced by variable transmission delays over the network. Removing jitter to allow a steady stream requires collecting packets and storing them long enough to permit the slowest packets to arrive in time to be played in the correct sequence. The jitter buffer is used to remove the packet delay variation that each packet encounters transiting the network. Each jitter buffer contributes to the end-to-end delay. 2.1.2.3 Latency Latency is just the time delay of a packet delivery. One-way latency and round trip latency are distinguished. The lower the latency, the more users are satisfied with the perceived QoS.
2.2 DIFFERENCES BETWEEN WIRELINE AND WIRELESS NETWORKS Both wireline and wireless networks support different service types. Because of limited frequency allocations wireless networks have peculiarities that distinguish them from conventional wireline networks [21]. • The wireless channel varies over time and space. It has short memory due to multipath. These variations are caused by either motion of the wireless device or changes in the surrounding environment. This causes a burst of errors to occur during which packets cannot be successfully transmitted over the link. Small-scale channel variations due to fading are such that states of different channels asynchronously switch from good to bad within a few milliseconds and vise versa. Furthermore, very strong forward error correction (FEC) codes cause very slow rates and hence cannot be used to eliminate errors because this technique leads to reduced spectral efficiency. • In addition to small-scale channel variations, there are also spatiotemporal variations on a much greater timescale. Large-scale channel variation means that the average channel state condition depends on user location and interference levels. Thus, due to small-scale and large-scale changes in the channel, some users may inherently
10
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
demand more channel access time than others based on their location or mobile velocity, even if their data rate requirement is the same as that of other users.
2.3 WIRELESS QOS 2.3.1 Basic ideas of QoS provision QoS has become an important issue both in fixed and in wireless networks, mostly due to the requirements of today’s multimedia applications. In wireless networks, QoS requires even more attention than in fixed networks. This is because of the different nature of the physical medium – the wireless channel is significantly more error prone than a fixed one. Introduction of wireless access to the network brings at least the following new technical challenges: • Quality of the wireless channel is typically different for different users and changes randomly with time on both slow and fast time scales. • Wireless bandwidth is usually a scarce resource that needs to be used efficiently. One cannot overprovision the wireless link). • An excessive amount of interference and higher error rates are typical. • Mobility complicates resource allocation. The data transfer in a wireless access network should be reliable, which means that the error rate should be as low as possible. There are two different error control methods that are used for providing reliability. They are FEC and Automatic Repeat Request (ARQ). FEC is based on redundant coding information that is added to transferred data. The redundant information is then used at the receiver side to detect and correct transmission errors. The maximum number of errors that can be detected or corrected depends on the coding mechanism. Basically, the more redundant information is added, the more errors can be tolerated. ARQ means that the receiver can detect erroneous packets and request them to be retransmitted. For detecting the erroneous packets, some sort of FEC coding which does not correct errors is used. There are two basic retransmission request mechanisms: go-backN and selective repeat. In go-back-N, every packet that has been sent after an erroneous packet has to be retransmitted. In selective repeat, only the erroneous packets are retransmitted because the receiver can keep correctly received packets in its buffer instead of discarding all the packets that come after an erroneous packet. The selective repeat ARQ scheme is commonly agreed to have better performance, but go-back-N has the advantage of simpler implementation. By combining ARQ and FEC error control schemes, a hybrid ARQ mechanism can be designed. In a hybrid ARQ scheme, FEC is used to reduce the frequency of retransmission. Efficient error control for a certain frequency, mobile velocity, and range are always a trade-off between data efficiency and throughput. In the context of wireless networking, delay and jitter are mostly generated by queuing, subsequent transmission over the wireless access network, and ARQ retransmissions. Typically, queuing occurs on a mobile terminal’s side or at a base station (access point) where
NETWORKING AFTER 2G: A SERVICES-DRIVEN TENDENCY
11
packets are waiting to gain access to wireless link. In packet-based wireless networks, the queuing delay is one of the central QoS issues. The wireless medium access mechanisms and so-called schedulers (for the packets with different priority) are of paramount importance. Unreliability of behavior in terms of signal spreading and the shared nature of wireless transmission impose ARQ to add an unpredictable component to the queuing delay. 2.3.2 Multimedia and wireless In order to work properly and satisfy end-user expectations, multimedia applications impose requirements on some communication parameters, such as bandwidth, drop rate, delay, and jitter. Exceeding these requirements in wireless environments is very challenging. It either decreases considerably the communication quality or even ruins it completely. Using multimedia in a mobile environment causes a critical situation due to the properties of radio links, such as interference, fading, etc. These parameters are also more variable than in a wired or fixed environment, which calls for adaptive protocols or for robust control algorithms, but this last point is not optimal when the channel conditions are good. To deal with these problems, many wireless communication standards have been defined. Some of them enhance the QoS of the whole system, others differentiate between the priorities of each mobile host, offering them different quality of service parameters, e.g. different bandwidth, delay, etc.
2.4 NETWORKING AFTER 2G: A SERVICES-DRIVEN TENDENCY Mobile communication services originated in voice telephony. However, if every person in the country owns a mobile terminal, one neither expects the number of subscribers to exceed the population nor hope for an increase in traffic merely through voice telephony. Therefore, for mobile communications in the 21st century, three strategic objectives exist: implementation of multimedia services, ubiquitous services, and global services. Multimedia services are expected to diversify services and increase the volume of traffic by migrating traditional voice-oriented services over to services centering on text data, image, and other non-voice services. Ubiquitous services aim to expand the object of communications services, which have been limited to humans so far, to everything and anything. In principle, a wireless terminal may be attached to anything convenient; wireless chip may be attached to a box being delivered via snail mail. As such, if everything or anything (other than humans) that moves becomes an object of mobile communications, the number of mobile terminals of extremely diverse types will increase dramatically and lead to much greater traffic. In a society with advanced multimedia and ubiquitous communications services, computers and communication equipment will be in all forms centering on mobile networks, not only between people, but also between person and machine (computer), and between a machine and another machine. Under these circumstances, a huge increase in non-voice traffic will be expected in person-person, person-machine, and machine-machine communications.
12
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
2.4.1 3G Network Architecture The 3G network architecture known in Europe as Universal Mobile Telecommunication System (UMTS) [22] is an evolution of the GSM network. The network mainly consists of two parts (subnetworks): a UMTS terrestrial radio access network (UTRAN) and core network (CN). Further, the CN is logically and physically divided into a circuit switched CN (CSCN) and packet switched CN (PSCN). There are a number of UMTS releases of standards, each of which covers different RAN and CN transmission technologies. For example, Release 3 (3GPP R99) [23] of the UMTS system supports WCDMA access technology at the air interface and ATM transport technology, whereas Release 4/5 (3GPP R00) [24] supports two RAN technologies. Since the main objective of Release 4/5 is to use the same CN for the two RANs both available types of RAN GERAN (GSM-EDGE RAN) and UTRAN have to be connected to the same CN. 2.4.2 NGN Network Architecture NGN networks are not as well defined as 3G ones. The basic features which are assumed to be implemented in these systems are given below. The scope of an IP-based mobile NGN system focuses on both core network and access networks. In NGN systems, PSCN and CSCN are assumed to be replaced by single IP-based packet switched network (PSN). Therefore, the whole end-to-end path between mobile terminal and service access point is assumed to be IP-based, all mobile terminals are to be fully IP capable, and all services are IP-based. A clear separation between radio access network (RAN) and CN in 3G networks has already led the network to multi-access environment. The main objectives of the RAN are to provide an access technology at the air interface and to hide all access specific features from the CN. Therefore, the RAN part comprises all functions that enable a user to access services. Because of that, the CN has little impact on the introduction of new RAN and can evolve independently of it. In order to enable seamless IP-based services for users in hot spot areas and on the move, it is necessary to study a system architecture that combines broadband wireless access technologies (HIPERLAN/2, IEEE 802.11) with RANs (UTRAN, GERAN), since WLANs were primarily developed to match requirements of non real-time services like file transfers, remote access to LAN, etc., while UTRAN and GERAN are mainly targeted on high quality voice communications. The main reason for considering such a heterogeneous environment is that it is almost impossible to define a RAN that combines all the advantages of the different types of RANs. In addition to multiaccess environment, NGN systems will have a layered network infrastructure with at least two hierarchical levels. Such architecture provides a high flexibility for current monoservice cellular network and it is expected that it will be able to perform well in a multiservice IP environment. In accordance with the proposed layered network infrastructure, there will be cells of different sizes (picocells, microcells, macrocells) each of which serve users with different mobility patterns. Layers with picocells or microcells are able to provide a high capacity with high bandwidth in hot-spot areas, i.e. serve slow mobility users with relatively high traffic demands. Macrocells serve users with wide-scale mobility patterns, e.g. driving users. Different RANs can be considered as different levels of the network hierarchy.
NGN SYSTEM DESIGN FOR PERFORMANCE OPTIMIZATION AND NETWORK EFFICIENCY
13
2.4.3 Reconfigurable Access Networks Reconfigurable access networks, also known as ad-hoc networks, are a valuable topic of ongoing research. These networks do not have a fixed configuration and all stations are assumed to be nomadic. Wireless links between them must be configured on demand. Within the NGN framework ad-hoc networks can be considered as a special structure of RAN. Ad-hoc configuration of RAN must be able to provide plenty of benefits to service providers, especially in areas of low population. However, both the concept of ad-hoc networking and basic protocols for these networks are still under heavy development [25] and, therefore, one can expect that at least the first evolution of NGN system will not employ such a type of access network. 2.4.4 Always Best Connected concept Always Best Connected (ABC) [26] is a concept that allows users of NGN system to choose the most suitable RAN which is the best fit for its applications from the QoS point of view. ABC should be implemented as an inherent feature of future NGN systems. ABC relates directly to vertical (intersystem) handover, which is one of the fundamental issues of the NGN framework. Indeed, by proper implementation of vertical handover it is possible to achieve seamless service performance over multi-access AN in NGN environment.
2.5 NGN SYSTEM DESIGN FOR PERFORMANCE OPTIMIZATION AND NETWORK EFFICIENCY As mobile wireless communications evolve from supporting circuit-switched, voice-centric systems to future NGN systems, the reliable and efficient support of heterogeneous multimedia services becomes important [27]. In this scenario, system design becomes an important task, especially from the point of view of the provision of effective solutions for ubiquitous access to Internet and the QoS guarantee for different traffic classes. The system must be optimized to improve the performance perceived by users and the efficient use of network resources, which is an important prerogative for making revenue for operators and for allowing the realization of services at affordable costs. The NGN system should be achieved as an evolution of 3G ones. Hence, it is important to consider basic aspects of the QoS mechanism implemented in 3G networks. 3G systems are based on the end-to-end classical layered QoS architecture, based on the Bearer Service (BS) concept (see Fig.2.1). This is a flexible way of designing a kind of bit pipe at a certain network level, between given network entities, with certain QoS attributes and capacity. Each bearer service relies on the QoS-enabled services of lower layers [28, 33]. The QoS functions in such architecture can be summarized as follows. • End-to-end QoS management and support. • Scheduling and dynamic resource allocation (air interface, RAN level). • Transmission adaptivity and diversity to counteract the dynamics of the radio channel (air interface, RAN level).
14
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
USIM
ME
Node B, RNC
SGSN
GGSN
UE
End-to-end service traffic classification Local BS
3G network BS
Radio access BS
radio resource management
Radio BS
Iu BS
Physical layer
Physical BS
external IP QoS External BS
CN BS Backbone serv.
List of acronyms: ME = Mobile Equipment GGSN = Gateway GPRS Support Node RNC = Radio Network Controller SGSN = Serving GPRS Support Node UE = User Terminal (= ME+USIM) UMTS = Universal Mobile Telecommunications System (a 3G system) USIM = Universal Subscriber Identity Module
Fig. 2.1 Bearer Services architecture both at different 3G interfaces and different layers for the packet-switched domain of the core network.
• Connection Admission Control CAC (air interface, RAN level). As was mentioned above, future 3G releases are oriented towards packet-switched traffic, IP, and high-speed access. In particular, an innovative feature of 3G realization in UMTS – Release 4/5 [24] is represented by IP-based Multimedia Services (IMS), with an IP core network (IPv6 based) and the handling of multimedia services using session initiation protocol (SIP) signaling and the bearers offered by the packet-switched domain. However, 3G systems still leave some unsolved problems for NGN. The substantial aspects are outlined below. • Bandwidth restrictions limit the number of operators and constrain the data rate available per user. • Energy consumption poses significant constraints in the design of terminals and calls for innovative concepts to be adopted for system design (especially the air interface protocol stack). • System dynamics is continuously changing in terms of traffic patterns and loads, user locations, network congestion, and radio channel conditions. • Lack of an end-to-end seamless transport mechanism spanning different wireless and wired networks.
NGN SYSTEM DESIGN FOR PERFORMANCE OPTIMIZATION AND NETWORK EFFICIENCY
15
• Coexistence of heterogeneous services needing different QoS levels. • Lack of air interface reconfigurability for mobile terminals which would allow access to different wireless systems based on distinct technologies. • Lack of scalability in applications and services depending on the access networks, etc. • Difficulties in roaming across distinct service networks (i.e., RANs). According to the above 3G limitations, mostly related to RAN, the system aspects that need improvements in the migration from 3G to NGN can be outlined as follows. • A new improved radio interface (physical layer, antenna system). • Support of different RANs, e.g. through the Software Defined Radio technology. • Vertical (intersystem) handover support between RANs. • Interoperation of different networks with joint and optimized resource management: novel inter-system handover protocols and introduction of a resource broker for a centralized resource control with management of QoS on a session basis. • Optimization of scheduling techniques for multimedia traffic under diverse constraints, QoS requirements and objectives. • IPv6-based CN; IPv6-based macro mobility control for QoS support; use of IP micromobility protocols for an efficient management of intra-domain mobility. • Novel design of the air interface protocol stack: (i) a cross-layer approach is needed to allow the exchange of information also between non-adjacent protocol layers; (ii) introduction of a middleware for the support of different access devices with heterogeneous characteristics. Therefore, the NGN system should be based on new optimization paradigms in accordance with two different and complementary approaches, as outlined below [37]. Vertical Approach. NGN system optimization calls for a vertical design of the air interface protocol stack. Such a cross-layer approach requires interfaces between different platforms across the layers, which exchange control information beyond the standard OSI structure. Cross-layer interfaces can be within, between, or beyond adjacent abstraction layers. Although interfaces between adjacent layers are in general preferable, there can be the need for efficient and direct interaction between non-adjacent layers; in general, a layer should be aware of the other layers of the protocol stack. In fact, IP layer and above ones often need direct interfaces to DL layer, e.g. for handover support. Another example concerns transmission parameters (e.g. transmission mode, channel coding and DL layer retransmissions) that must be related to application characteristics (e.g. type of information, source coding, etc.), network characteristics, user preferences and context of use. Moreover, DL layer should be aware of higher layer (IP and transport) behaviors in order to make decisions on traffic management. The air interface protocol architecture with
16
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
OSI layers Applications and services
7
Middleware Transport
4
IP
3
Link layer
2
Physical layer
1
Fig. 2.2 Possible cross-layer interactions for the air interface in NGN systems.
the proposed interactions among the different layers and with a middleware is shown in Fig.2.2. Horizontal Approach. In the NGN framework, different wireless technologies will operate to allow the best access to users, depending on their locations, mobility characteristics, applications, user profile, etc, i.e. in accordance with the ABC paradigm. Therefore, it is necessary that the use of the resources in the different RANs be globally coordinated by means of a resource brokerage function. Such intelligence is centralized and allocates sessions to RANs or switches them from one to another when some conditions are meet.
2.6 END-TO-END QOS OF NGN NETWORKS A major weakness of current IP networks is the lack of a QoS guarantees provision. Indeed, only the best effort service is supported. Hence inadequate QoS levels may be obtained by real-time services. This problem will definitely become worse when multimedia services are extended to the air interface. Currently, the challenge is to add IP QoS in the reservation mechanisms of radio network technologies and to propose a consistent end-to-end resource reservation model. In the past, IETF proposed two QoS frameworks for IP networks. These are IntServ [2], based on the connection-oriented resource reservation principle, and DiffServ [3], based on the service differentiation principle. IntServ, based on CAC procedures, can provide deterministic QoS guarantees and requires a signaling protocol in order to inform network elements about the necessary resource reservation. On the other hand, DiffServ employs a different approach. In order to distinguish between packets with different QoS requirements, the DiffServ specification defines packets marking procedures. It provides probabilistic QoS guarantees to aggregated traffic flows and uses CAC algorithm, which is based on Service Level Agreement (SLA) between subscribers and service providers or between two service providers.
END-TO-END QOS OF NGN NETWORKS
17
Since interworking between the NGN system and the public Internet is expected, QoS provision should be based on the frameworks available for IP-based fixed networks. These are the following. • Service differentiation. • Connection-oriented resource reservation. • Integration of service differentiation and connection-oriented resource reservation. 2.6.1 Service differentiation Due to high scalability, service differentiation can nowadays be seen as the most obvious approach for QoS provisioning in IP-based networks. IETF DiffServ working group has standardized two Per Hop Behavior (PHB) groups. • Assured Forwarding (AF PHB, [35]). • Expedited Forwarding (EF PHB, [36]). AF PHB has been designed for a range of applications which require different QoS guarantees. There are four classes of PHB identification codes within the AF PHB group. Within each class there are three distinct DiffServ CodePoints (DSCP) with different packet drop precedence. EF PHB is targeted to applications which require strict guarantees of endto-end delay and should not suffer from packet losses. One of the major advantages of the service differentiation approach is the unification of service classes within a certain domain. The availability of SLAs between domains makes it possible to implement service differentiation on a network basis. Due to the limited number of service classes, the mapping between user application requirements and network services can be easily implemented. 2.6.2 Connection-oriented resource reservation All network elements in IntServ architecture have to be aware of a certain connection establishment protocol and should maintain per-flow state for each connection. Since the number of users within the coverage area of a cell is statistically limited, a connectionoriented resource reservation can be successfully implemented on a RAN level. Implementation of connection-oriented resource reservation within the RAN can make it possible to achieve a deep granularity of network services. Indeed, in this case the set of possible services is not limited to a predefined number of classes, but can be extended to fit specific requirements of a wide range of applications. Since the parameters of resource reservation at each node along the path of a connection establishment request can be arbitrarily defined in accordance with application requirements, the source can request an arbitrary amount of forwarding resources. Such an approach requires the implementation of the resource reservation protocol user’s terminals, which results in additional software requirements.
18
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
2.6.3 Integrated approach One of the promising approaches to maintain end-to-end QoS guarantees is to incorporate connection-oriented resource reservation within the RAN and service differentiation within both the NGN CN and the Internet routers. Indeed, the service differentiation framework operates on traffic aggregates and no per-flow state has to be maintained in the core routers, while the connection-oriented resource reservation approach, when used with a specific connection establishment protocol, can provide deterministic QoS guarantees for applications. The integration of both approaches is a promising research challenge since it provides the following advantages for user applications. • Proper granularity of user QoS requirements. • Deterministic QoS guarantees within the RAN. • High scalability of network routers within the CN. • Seamless interoperation between NGN networks and the public Internet. However, there are several obvious drawbacks of the integrated approach. • Implementation of a connection establishment protocol in both mobile terminals and RAN routers. • QoS mapping procedures between RAN and NGN CN. A simple configuration of the proposed integrated approach is shown in Fig.2.3 where RAN and CN consist of a number of IP-capable routers. In particular, there are four types of IP routers. • Local Routers (LR). • Edge Routers (ER). • Core Routers (CR). • Border Routers (BR). LR is directly connected to radio access hardware. It must be responsible for the following functionalities. • Relaying messages of the IntServ connection establishment protocol. • IntServ resource reservation. ER should allow both IntServ and DiffServ functions, which can be summarized as follows. • Aggregating, policing, and shaping incoming traffic. • Mapping between IntServ QoS specifications and DiffServ service classes. • DiffServ class-based resource reservation.
19
END-TO-END QOS OF NGN NETWORKS
LR
4G All-IP CN
BR
with public internet
ER
4G All-IP RAN
CR LR
SLA
BR
SLA with other 4G All-IP
Fig. 2.3 Foreseen architecture for future QoS-enabled NGN networks.
• Encapsulating and tunneling messages of the IntServ connection establishment protocol. BR should serve as a gateway router between two domains and must allow the following functionalities. • DiffServ resource reservation capabilities. • Tunneling messages of the IntServ connection establishment protocol. Additionally BR may perform policing and shaping functions to align incoming traffic with the SLA with a neighboring domain. Depending on the type of neighboring domain two types of BRs are distinguished: BRs interworking with public Internet and BRs interworking with other NGN systems. Both types of BR may actually have the same functionalities. CR is just a simple DiffServ interior network node. The only requirement that should be added is tunneling messages of the IntServ connection establishment protocol. A mobile terminal should be responsible for end-to-end QoS negotiation mechanisms using one of the existing connection establishment protocols. When it is not possible to maintain a desired QoS, a renegotiation process may be performed. IP QoS mechanisms have been specified with wired networks in mind. Whereas IntServ and DiffServ have already received considerable attention from researchers, only few studies on supporting a wireless IntServ or DiffServ have been published [30–34]. In fixed networks, packet losses primarily occur due to buffer overflow within the routers, while the error rate of the transmission medium is small (less than 10E-9). Therefore, packet losses caused by errors at the PHY layer can be neglected. In contrast, in dealing with NGN systems, packet losses, caused by bit errors in the wireless transmission medium must be accounted. Therefore, the weak chain in NGN end-to-end QoS architecture is the air interface between the mobile terminal and the first fixed point. To maintain the best possible QoS at the air interface, novel methods of FEC, ARQ, and hybrid techniques should be developed and properly parameterized. Therefore, extension of QoS frameworks
20
QUALITY OF SERVICE FOR NEXT-GENERATION NETWORKS
Unreliable medium
Mobility
Teletraffic
Fig. 2.4 Components that influence QoS at the air interface.
to the air interface introduces problems that should be solved before the required QoS of user services can be achieved.
2.7 SUMMARY The analysis of packet-level QoS at the air interface is a very challenging and complicated task due to a number of factors. Wireless links have highly time-varying behaviors due to shadowing, multipath fading, and other effects. Besides, the nomadic behavior of users introduces another uncertainty. It can cause a handover procedure between RANs built on technologies dedicated to different coverage areas. Hence, a new path in the fixed part of the RAN and possibly in the CN has to be established. This new path may have different delay, loss, and bandwidth characteristics. The movement of the user may also influence its activity, which, in turn, influences call and packet-level characteristics of the generated traffic. Therefore, teletraffic specific issues like the use of several simultaneous Variable Bit Rate (VBR) applications must be accounted. Components that influence QoS at the air interface can be represented by the interaction of mobility, teletraffic characteristics and wireless medium, as shown in Fig.2.4. A mechanism responsible for QoS support in NGN systems should predict the future state of the network based on both user mobility and its traffic parameters. Such a mechanism has to incorporate a CAC algorithm which denies or allows the admission of new calls on the basis of the network state. Practical investigation of RAN behavior under different traffic loads and wireless channel conditions is of paramount importance. Special bias must be given to multimedia realtime traffic. Monitoring of a real network can be done using specific measurement tools. In those cases when implementation of a network does not exist, adequate traffic models emulating user behavior should be adopted for planning and optimization purposes. Therefore, it is crucial to develop traffic models of various types of sources which are assumed to be used in NGN networks.
SUMMARY
21
End-to-end QoS investigation is of paramount importance as well. Indeed, every QoS dependent service can be implemented smoothly only in end-to-end manner. QoS can easily be denied if fixed CN is not well provisioned, even if RAN provides adequate treatment of that service’s traffic. Thus, the QoS assurance problem falls into RAN QoS assurance and CN QoS assurance. Indeed, those problems can be treated independently, but delivering of end-to-end QoS is possible only when solutions for every problem are combined.
Chapter
3
Cross-Layer Design for NGN environment
I
N addition to providing broadband wireless access to the Internet, an IP-based NGN system has to satisfy requirements of QoS-aware applications. Indeed, this is an inherent problem for many services even in fixed networks. Wireless and mobility add their own QoS problems on top of this inherent IP flaw. Characteristics of wireless channels, such as high and correlated bit error rate and limited bandwidth of wireless channels, have to be addressed before wireless Internet services can be commercially deployed. To facilitate development of these networks, novel methods of teletraffic theory, wireless channel and mobility modeling, optimization, and design must be developed. To develop tools for NGN analysis, it is important to define whether performance evaluation methods available to date can be successfully applied to forthcoming NGN systems. Within such a framework it is necessary to consider both current traffic modeling and wireless channel modeling approaches available in literature. Indeed, surveys of literature have shown that due to the circuit-switched technology utilized by 2G systems, traffic models designed for conventional 2G mobile systems are primarily dealt with capturing session-level parameters of traffic sources, while networklayer traffic demands were assumed to be constant during the whole duration of a session [38–41]. Therefore, those traffic models are not appropriate when mobile systems operate over IP protocol as an end-to-end transport technology. Regarding wireless channel modeling techniques it is noticeable that most studies performed so far have been devoted to performance evaluation of data link (DL) layer protocols, while there have been no studies targeted on IP layer performance evaluation. Therefore, those wireless channel models developed so far have to be extended to the network layer. Indeed, parameters of DL layer error concealment protocols depend on both user’s traffic demands and wireless channel characteristics. Therefore, the model of user’s traffic demands and the wireless channel model should be merged rather than considered independently.
23
24
CROSS-LAYER DESIGN FOR NGN ENVIRONMENT
FFAP
Radio AN
SLA
SAP Fixed network
mobile user Fixed AN FFAP
SLA
fixed user FFAP - First Fixed Access Point; AN - Access Network; SLA - Service Level Agreement; SAP - Service Access Point. Fig. 3.1 End-to-end service architecture in NGN All-IP network.
3.1 END-TO-END SERVICE CONFIGURATION IN NGN Let us consider an end-to-end client-server service configuration in NGN as shown in Fig. 3.1. The NGN framework assumes a number of different access networks (AN). However, only two ANs are shown in the figure. These are fixed AN and radio access network (RAN). In accordance with All-IP architecture, RAN should take on a certain configuration and consists of a number of IP-capable routers, part of them are actually combined with radio access equipment. These nodes are first routers in uplink direction when mobile terminals access the network. In what follows, they are called first fixed access points (FFAP). According to the architecture under consideration, a service that a user wishes to access is located in the fixed part of the IP-based network. In this case, packets generated by a client are routed via air interface to FFAP, then through the gateway node between RAN and the fixed IP network, and via a number of routers located in the fixed part of the IP network. To evaluate QoS expectations in the considered service configuration a user’s application may experience, all those packet losses and delays that may occur at each point up to the destination and in the backward direction must be accounted. It is well-known that packet losses may occur due to both bit errors in transmission medium and network nodes’ congestion situations. Given the current evolution of fixed transmission medium, it is well understood that packet losses caused by bit errors at the physical (PHY) layer can be neglected. For these reasons, it is always claimed that packet losses in fixed networks occur due to buffer overflows caused by congestion of network nodes. Therefore, to evaluate the performance of user applications in fixed networks, it is sufficient to evaluate those packet losses caused by buffer overflows. In contrast, in dealing with wireless network, packet losses caused by bit errors in wireless transmission medium cannot be neglected. Moreover, it is well-known that these errors may contribute a lot to end-to-end performance degradation of a packetized user application. Indeed, they are completely different in nature compared to what is accounted in
END-TO-END SERVICE CONFIGURATION IN NGN
SRP
IP router
FFAP
SAP
Client
Server
TCP IP RLP PHY
TCP IP RLP PHY
wireless channel
SRP SAP RLP PHY
-
25
IP
IP
IP
ATM
ATM
ATM
fiber optic
fiber optic
server request point; service access point; radio link protocol; physical layer. Fig. 3.2 End-to-end protocol model of NGN All-IP network.
fixed networks so far. For these reasons, the wireless interface between mobile terminal and FFAP becomes an unexplored point in future wireless end-to-end IP QoS architecture. To determine what kind of performance degradation should be expected at the wireless interface, let us consider an end-to-end protocol model of service architecture considered in the previous subsection, as shown in Fig. 3.2. Since the error rate of fixed transmission medium is very low, it is possible to assume that the delay caused by PHY and DL layers is constant and equal to the sum of the propagation time over the fixed transmission medium and the time taken to decode and packetize these bits at the PHY and DL layers before sending them to IP layer. At the wireless interface, error concealment procedures may significantly increase the delay of transmission by applying various error correction techniques like forward error correction (FEC), automatic repeat request (ARQ), or combinations of them. Additionally, those DL layer protocols may allow some errors to propagate up to IP layer, which in turn may cause retransmissions through underlying layers. During long periods of worse quality of wireless channel, both errors and delays can be very high and therefore may significantly affect the performance of packetized user applications. Therefore, in dealing with end-to-end performance evaluation of NGN, in addition to IP level peculiarities at the wireless interface, behavior of underlying layers must be accounted. Therefore, an extension of QoS guarantees to the wireless interface introduces problems that must be solved before the required QoS of user services can be achieved. These new challenges require development of new cross-layer integrated methods of performance evaluation at the wireless interface. There are a lot of publications related to performance evaluation of fixed networks (see [42–48] among others). On the contrary, there is only little work devoted to packet-level performance evaluation of packetized user applications running over wireless channels. To date, the only comprehensive study is presented in [49], where authors consider the convergence of wireless channel modeling and queuing analysis.
26
CROSS-LAYER DESIGN FOR NGN ENVIRONMENT
3.2 COMPONENTS OF CROSS-LAYER MODEL 3.2.1 Mobility models There are a number of different approaches towards realistic mobility models have been developed in recent decade [50–52, 69] etc. Relatively comprehensive survey on mobility modeling is presented in [57,58]. In accordance to [68] mobility models can be categorized into two classes: • microscopic mobility models describe mobility behavior of individuals; they are often based on analytical descriptions of user movements. There are numerous models including the random walk model [59], the random Gauss-Markov model [58, 60], the random waypoint model [50], and the reference point group model [52]. These models are not able to reflect macroscopic behavior and could exhibit undesired properties if applied improperly [53]. • macroscopic mobility models [54–56, 62–64] describe aggregated effects of mobility; they are often based on statistical data collection. Some of these models address mobility together with teletraffic demand. However, due to their nature, these models have a simplified representation of the teletraffic part, while the mobility part receives the main attention. 3.2.2 Teletraffic models NGN will support a huge number of applications like in current IP networks. It is known that each application can be described by its own traffic characteristics and these characteristics should be measured from real traffic traces and then used to parameterize a model. It is recognized from traffic modeling in fixed networks that most applications have to be characterized using different traffic models. In those cases when a general traffic model is used, parameterization must be done depending on the traffic characteristics of a certain application. The recent trend is to use general models like Markovian arrival process (MAP) or batch MAP (BMAP) in order to represent the teletraffic part in NGN [65, 66]. QoS characterization in NGN All-IP networks is not limited to call level parameters, like in 2G systems. All NGN calls, including voice ones, will be IP based and require different, often variable rates. Therefore, dealing with performance evaluation of NGN, in addition to call level QoS parameters it is important to consider IP level QoS ones as well. Due to these facts, users traffic models which take into account both call-level and packet-level behavior of the single user are needed. An example of correspondence between call level and packet level traffic demands for voice source is shown in [71]. 3.2.3 Integrated models Based on the above-mentioned considerations, there is an increasing need to develop new integrated traffic models incorporating both mobility and teletraffic parts. These models have to be further applied to investigate the QoS level provided for each application using analytic or simulation techniques.
COMPONENTS OF CROSS-LAYER MODEL
27
To date, only a limited number of studies [67–71] which simultaneously consider user’s mobility behavior and its traffic demands are available . In [67], the authors highlight the importance of the correct modeling of user mobility when considering traffic in a mobile system. The authors have shown that traffic load could affect mobility models accuracy, this means that the verification of mobility models must be undertaken with great care. In [68], the authors developed a novel approach towards a mobility model based on statistical data collection augmented with traffic predictions to form a workload model for wireless metropolitan area networks. The proposed model employs macroscopic mobility model, traffic demands are parameterized by measured statistical data. The authors have failed to show that the model is adequate. In [70], the authors developed an integrated model based on general assumptions that user’s mobility and user’s teletraffic demands can be represented as a combination of two MAPs, one of which describes user’s mobility while the other one specifies the teletraffic characteristics of applications. Despite obvious advantages, the model is complex, and therefore its application in performance evaluation is limited. In [71], the authors developed an integrated cross-layer traffic model for NGN. It consists of two different parts: a mobility model and a teletraffic one. The mobility behavior of the user is captured by a Markov chain with a finite state space, while the teletraffic characteristics at the call-level are represented by a Markov chain with two states. The authors have shown how to extend the proposed model to capture packet-level characteristics of constant bit rate (CBR) and variable bit rate (VBR) traffic types. Because of complexity the main application area of the model is in simulation studies. 3.2.4 Models of wireless channel In wireless networks, the effective transmission bandwidth is highly variable, depending on many factors such as noise, distance, multiple path propagation, interference, etc. The fluctuation of radio signal can be divided into multipath fading (or fast fading) and shadowing (or slow fading) [21]. • Multipath fading is caused by multiple path propagation of the wave between transmitter and receiver. It characterizes the interference among multiple versions of transmitted signal arrived at receiver. Due to unknown number of possible paths between transmitter and receiver, multipath fading is a complex stochastic process that depends on the mobility of the user. • Shadowing is caused by the loss of the line of sight (LOS) between transmitter and receiver due to shadowing of the propagating wave by large obstacles, it describes the attenuation with irregular terrains, in other words. The received signal power varies in accordance with alternating interruptions and release of LOS. Size and location of obstacles are random, and therefore the occurrence and duration of interruptions are also random. The movement of the user between shadowers is a mobility-dependent process. Due to the multipath propagation in urban areas, the signal amplitude is Rayleighdistributed, if there is no LOS. For the case of an additional LOS we expect a Rice-
28
CROSS-LAYER DESIGN FOR NGN ENVIRONMENT
distribution. This distribution superposes the log-normal distribution that is caused by the slow fading. Early work in wireless channel modeling dedicated to stochastic modeling of channel dynamics at PHY layer, measured by signal strength and bit error rate (BER) [72, 73]. These models cannot be directly used to evaluate high layer network performance [49]. The models capturing fast fading characteristics of wireless channels are based on Markov chains where the state space of the models is designed to capture a histogram of relative frequencies of received signal-to-noise ratio (SNR) level [74–77]. These models are also able to capture autocorrelation properties found in empirical SNR traces. Typically, such model is constructed by partitioning the range of the received SNR into a set of non-overlapping intervals. Each interval is represented by a nominal BER, which represents a certain channel quality. Several studies have addressed the wireless channel modeling through appropriate partitioning of the received SNR [78–84]. Moreover, it has been demonstrated that hidden Markov models (HMM) are capable to model accurately fast fading characteristics of wireless channels [86–88]. In particular, in [87] the authors present results for the commonly considered Rayleigh fading channel. The models capturing slow fading are represented much weaker – to date only few studies are available [85]. To make modeling feasible, the first objective must be in identifying what are the important statistics of fading channel to packet-level network performance. In [49] the authors investigated the impact of multipath and shadowing fading channel dynamics on the packet-level data queueing performance. Based on [89] the authors find that the quality of the wireless channel can be measured by the variation of the received SNR experienced by a mobile terminal. It is a complex mobility-dependent stochastic process resulting in all fading components, each of which significantly influences the performance of the wireless channel. The abovementioned models primarily capture only one fading component which is dominant for a given type of wireless channel. However, it rarely happens that the performance of a wireless channel is affected by only one fading component. Given the movement of a user, one should expect the existence of all fading components, where the influence of each component cannot be neglected. Moreover, the abovementioned models do not account IP layer at which user’s traffic demands and QoS requirements are usually represented. Indeed, those models give only basic ideas of how SNR changes over time depending on fading characteristics of wireless channels and they do not clarify how many IP packets are lost or retransmitted in every particular case. As was mentioned above, for some types of wireless channels, BER of the wireless channels is a complex but deterministic function of the received SNR. Therefore, provisioning of IP layer QoS for wireless channels is difficult due to high and correlated BER caused by changes in received SNR. To overcome these problems, a radio link protocol (RLP) with a suitable error control mechanism has to be used at the DL layer [90, 91]. The choice of error concealment strategy and its parameters depends on both wireless channel characteristics and the nature of the traffic source [92]. Therefore, an adequate and ideal wireless channel model at IP layer can be represented as a complex function of the following components: the user’s mobility behavior, the user’s
CROSS-LAYER BLACK-BOX FRAMEWORK
29
application type, SNR, the coding scheme used, the PHY layer, and error correction techniques at the DL layer.
3.3 CROSS-LAYER BLACK-BOX FRAMEWORK In context of packet-switched mobile data communications in NGN, classic performance metrics which have been traditionally applied to conventional 2G systems and plain fixed Internet no longer hold. More elaborate and sophisticated ways to define QoS are being considered. Among the many parameters used for this purpose are the following: • at IP layer – throughput, packet loss, delay, delay jitter; • at DL layer – modulation technique, error concealment procedure, etc. • at PHY layer – coding schemes, etc. Moreover, mobility of the user affects both teletraffic characteristics [P4] and the quality of the wireless channel [P6]. The wireless channel model and the user’s traffic model are complex functions of a number of environmental characteristics like user preferences including current mobility pattern, DL layer functions and procedures, application type, landscape, etc. In this context, evaluation of QoS expectations that a user application may experience running over the wireless channel is very sophisticated task which involves a number of interdependent stochastic factors. Therefore, both the quality of the wireless channel and the user’s traffic demands should not be considered independently of each other, but treated together. One among other possible ways to evaluate QoS expectations is to integrate the wireless channel model with the teletraffic model. This integration has to be performed at the IP layer. However, it leads to high complexity of the resulting model. Indeed, the resulting model in addition to cross-layer structure of each counterpart, should reflect all interdependence (correlation) properties between different parts of two models. Another way to deal with complicated model is to employ a so-called black-box framework – a strategy for investigating a complex object without knowledge or assumptions about its internal make-up, structure or parts [93]. The framework aims at either a formal description of the transformation rules linking inputs and outputs or the construction of a model exhibiting a behavior that approximates what is observable from the outside of the black-box. In accordance with this framework, if a set of parameters is given at the input of the system, the black-box translates it into set of parameters at the output. For example, from the mathematical point of view, neural networks are black-box models [94]. The cross-layer model accounting QoS parameters of PHY, DL and IP layers and based on black-box framework can be parameterized by the following inputs: a set of possible applications with corresponding session-level and packet-level characteristics; a set of various mobility behaviors of a user; a set of PHY layer’s coding schemes, and a set of DL layer error correction methods with appropriate sets of their parameters; etc. In [P7] the authors proposed elegant approach where black-box model substituted with real testbed. A testbed is supposed to be fed with certain parameters as specified above, and then appropriate performance measures can be collected at the output of the system. These
CROSS-LAYER DESIGN FOR NGN ENVIRONMENT
QoS performance evaluation metrics
AL TL IP DL PHY
+
MOBILITY
cross-layer set of parameters
black-box
Perceived QoS
30
+
AL TL IP DL PHY
Fig. 3.3 Cross-layer black-box framework.
performance measures include IP layer performance parameters as well as other higher level performance measures, like the perceived quality obtained by the user. A testbed is seen as a complex translator of parameters at the input of the system to corresponding output characteristics. Therefore, one can tune certain input parameters like PHY layer coding schemes or DL layer error correction methods to optimize them to match the desired output. The fourth chapter of this thesis is dedicated to the implementation of the cross-layer black-box framework. The testbed has been built and fed with a set of possible parameters like different traffic patterns and mobility behaviors. At the output of the testbed, IP layer performance degradation via perceived quality has been evaluated.
3.4 SUMMARY The evolution of mobile communications from current 3G to future NGN requires a novel design of the system architecture in order to integrate different access technologies with a globally optimized approach. Suitable techniques need to be developed in order both to coordinate the resources in different RANs (horizontal approach) and to redesign the air interface protocol architecture, allowing cross-layer interactions among protocols at different levels (vertical approach). A mandatory task is to improve the network efficiency and to increase the degree of satisfaction of users (QoS expectations). New teletraffic methods must be developed to adequately predict QoS expectations that a particular application may experience at various levels of wireless channel congestion. Special attention should be paid to joint traffic and wireless channel modeling issues. Traffic models developed for wired networks cannot be effectively used in wireless environment since both mobility of users and unreliability of transmission medium have to be taken into account. Those models developed for 2G networks have a simplified structure of the teletraffic part. Therefore, an adequate model for the NGN environment should explicitly take into account both mobility of users and teletraffic characteristics of various applications at per-user granularity. Currently available wireless channel models are primarily concerned with stationary characteristics of the wireless channel behavior. However, mobility of the user, an inherent property of mobile networks, restricts the application of such models. To facilitate the per-
SUMMARY
31
formance evaluation of QoS at the air interface, the wireless channel models should either depend on the mobility behavior of the user or explicitly integrate the traffic model. The cross-layer black-box framework for qualitative and quantitative evaluation of QoS expectations of user’s applications running over the wireless interface has been proposed. The major advantage of that framework is avoidance of complexity of cross-layer traffic and wireless channel modeling methods at the IP layer. The practical applicability of that framework is shown in the fourth chapter of the thesis.
Chapter
4
Next-Generation Telecommunications Networks: Access Network part
T
HE radio access network is one of the weakest points in the NGN framework. NGNs are supposed to be equipped with numerous RANs and therefore wireless interface is of paramount importance. Moreover, taking into account that wireless channels are affected by a number of factors such as multipath fading and shadowing mainly induced by mobility, it is wise to state that much effort must be undertaken in the area of QoS assurance in the presence of wireless interface. On the other hand, NGN is being designed to make new multimedia services available to mobile users with certain QoS requirements. Multimedia traffic makes wireless interface a great bottleneck. It is normally heavy and QoS demanding. Currently, only a handful of off-the-shelf RAN technologies are available – WLAN and GPRS. In this chapter, implementation and testing of several multimedia real-time services over those RANs is considered. The testbed has been built in accordance with the crosslayer black-box framework developed in the third chapter of this thesis. The testbed has been fed with a set of possible parameters, like different traffic patterns and mobility behaviors. At the output of the testbed, IP layer performance degradation via perceived quality has been evaluated.
4.1 WIRELESS ACCESS PECULIARITIES Being different from wired networks, wireless networks typically have time-varying and nonstationary channels. Therefore, the quality of the wireless channel varies, which can be measured by the variation of the signal-to-noise ratio (SNR) or the bit-error rate (BER) [89]. Such variations result in time-varying of the available transmission bandwidth at the DL layer, which also leads to time-varying delay of arrival packets at the application 33
34
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
layer, especially when automatic repeat request (ARQ) mechanisms are employed at the DL layer. Since the buffer size at the DL layer is finite, the time-varying channel service rate can induce buffer overflow due to the bit rate mismatch between the transmitting packet and the channel service rate. Normally in streaming services, due to variation in the arrival time of packets, some packets may become useless during playback if their arrival time exceeds a certain threshold.
4.2 MULTIMEDIA TRAFFIC TREATMENT IN CURRENT EVOLUTIONS OF WIRELESS ACCESS NETWORKS New multimedia services attracted by NGNs, such as downloading and streaming, are striving hard towards commercial market. Both the limited QoS support and scarce bandwidth of wireless environment restrain their wide deployment in NGN. Currently, at least two wireless access networks which can be used in NGN systems are available. These are WLAN IEEE 802.11b and GPRS. In addition to multi-access environment, it is becoming clear that NGN systems will have a layered infrastructure with at least two hierarchical levels. In accordance with layered network infrastructure, there should be cells of different sizes (picocells, microcells, macrocells) each of which serves users in areas with different population densities. Layers with picocells or microcells are able to provide a high capacity with high bandwidth in hot-spot areas. They can serve slow mobility users with high traffic demands. It is assumed that in NGN systems this role will be assigned to WLANs. Therefore, one can state that WLANs and NGN RANs are not competitors, but complement each other to allow coverage in areas with different population densities. Investigation of existing wireless access networks behavior under different signal conditions, mobility patterns, and load conditions is of paramount importance. Perceived QoS as well as QoS parameters suffer very much if signal strength is low and not stable. FEC and ARQ schemes can easily modify the traffic profile delivered to the destination of a wireless channel compared to one issued by a source. Mobility of the station is one strong reason among others why signal strength varies from excellent to poor. The user of the station may have different mobility patterns like walking, driving, etc. A well-known parameter of QoS, i.e. blocking probability, describes how neighboring stations affect the performance of the station under investigation, indeed in paper [P5] the authors have shown how WLAN segment could be overloaded just by single misbehaving station. Recently only a few studies have addressed performance of WLAN and GPRS under such circumstances. In [95] performance of real-time streaming service over WLAN has been investigated under different SNR conditions. Authors in [96] considered performance of synthetic TCP load in GPRS network, while performance of real services was not taken into account. In [97] the author considered performance of real-time service in 802.11b WLAN environment under different ranges of SNR and competing load conditions. Authors in [98] considered the multimedia streaming service over 802.11b WLAN. They defined a number of SNR ranges and evaluated the perceived QoS provided to the user. SNR varies heavily during the whole duration of a call or session due to the station’s mobility. That is why it is necessary to consider the mobility behavior of the station explicitly. Additionally, as has been stateed above that NGN systems introduce an additional notion
MULTIMEDIA TRAFFIC TREATMENT IN CURRENT EVOLUTIONS OF WIRELESS ACCESS NETWORKS
– the Always Best Connected (ABC) concept [26]. ABC should allow users to choose the most suitable RAN at any instant of time during the duration of a call. Particularly, this feature is claimed to be very attractive for users with complex mobility patterns. Technical implementation of ABC is to be based on intersystem (vertical) handover, which should be implemented in a seamless way between any types of access networks [99, 100]. In the study presented below, a series of tests for both wireless access networks has been performed. The obtained parameters have been compared to fixed local area network (LAN) IEEE Ethernet 802.3 access, which is currently the de-facto standard in multimedia networking. Moreover, special attention is given to WLAN performance evaluation under heavy load. To fulfill the defined objectives, the testbed, which includes three access networks, namely WLAN, fixed Ethernet, and GPRS, has been put-up. Attention has been dedicated to tracing of real traffic, but not synthetic one. Indeed, it is highly anticipated [101] that testing of real implementations can bring better understanding and new knowledge on the area. 4.2.1 Services under consideration Multimedia applications are continuously growing in popularity. The availability of highspeed access networks is the primary reason behind that. Today, it is strategically important to support these services over wireless networks. Basically, multimedia traffic consists of one or more media streams and can be characterized by strict delay requirements while can tolerate some losses. It is supposed that applications emerging from Internet will become capable of defining the required QoS level soon. However, currently multimedia traffic within the Internet is treated similarly to ordinary best effort data traffic, which does not pose strict delay bounds. Therefore, it is crucial to predict the QoS degradation that may be experienced by multimedia applications over wireless access networks. Below the following entertainment multimedia services are considered: • mp3 file transfer • mp3-based Internet Live Radio • live video streaming From the user point of view, these services behave quite similarly and can be described by two phases: a prefetching phase and playing phase. While in prefetching phase, an application stores data and then moves to playing phase. When an application is in prefetching phase, it uses all available bandwidth to prefetch data. However, during playing back, the streaming player continues downloading at the maximum available bandwidth while the service restricts itself to a certain target rate. From the traffic transmission point of view, mp3 file transfer service has essential characteristics such as relatively short time of connection/flow (transmission phase) duration. Indeed, traffic load simulators allow us to obtain some type of stationary traffic processes. However, taking into account the average sizes of mp3 files, MPEG-based compression techniques and the unpredictable nature of music sources, one can assert that this application is intended to produce traffic which cannot be stochastically described by stationary processes. Even in those cases when, based on certain available information about codecs
35
36
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
and music sources, one can expect stationary behavior of mp3 traffic, it is still not possible to prove statistically because of the relatively short duration of mp3 sessions. The second type of service under consideration is the real-time streaming service. Internet Live Radio stations are widespread nowadays in the global network and QoS perceived by the end-user varies on average from good up to excellent. For example, some stations are already available in 128 Kbps stereo format and can be streamed continuously for several days without a break in connection via the public Internet. The last service under consideration is the live video streaming service, which belongs to real-time services, like the previous one. It is widespread nowadays in the Internet and the QoS perceived by end-users varies on average from good up to excellent. For example, some live streaming videos are already available in 300 Kbps and can be streamed continuously without a break in service via well-provisioned parts of the public Internet. It is supposed that live video streaming service will be very challenging and demanding in NGN systems. 4.2.2 Current WLAN implementation 4.2.2.1 802.11 family Considering the increasing popularity of wireless stations, WLAN becomes an important NGN solution offering new growth opportunity. Many European, American, and Asian providers started public nationwide WLAN service around 2001 – 2002. The service is currently based mostly on the 802.11b standard [102], and it will evolve to 802.11a [103] or 802.11g [104] for enhanced services. 3G CDMA service enables relatively low data rates, but supports mobility on a wide scale. The position of WLAN is located between the two, and service enables easy and high-speed Internet access through a notebook PC or PDA with restricted mobility in hotspot areas. Besides public WLAN service, it is remarkable that home WLAN service is very promising, because so-called post-industrial society presumes high penetration of home wired high-speed Internet access. With growing numbers of homes having PC and other networking devices, the home WLAN market is emerging as a strong potential market combined with existing wired access. Using access points integrated in Asymmetric Digital Subscriber Line (ADSL) modems, home WLAN provides a wireless home network that enables multi-terminal connection while using one wired access line of ADSL. Current WLAN service has several drawbacks such as short coverage, a limited number of deployed hotspots, etc. It also has to overcome problems related to ease of use, security, mobility, and network management. However, with advances in wireless communications technology along with integration with other fixed and mobile technologies, the wireless broadband Internet market is believed to be another major market in the long term. The 802.11x [105] specifications are wireless standards that specify an over-the-air interface between a wireless client and a base station (access point), as well as among wireless clients. The 802.11 specifications address both the physical and medium access control (MAC) layers and are targeted to resolve compatibility issues between manufacturers of WLAN equipment. Two basic operating modes are defined: infrastructure and ad-hoc. Most dedicated hardware provides a basic service set that builds the wireless infrastructure. It allows clients to roam between access points. The ad-hoc mode allows individual stations to participate in peer-to-peer communication without an access point.
MULTIMEDIA TRAFFIC TREATMENT IN CURRENT EVOLUTIONS OF WIRELESS ACCESS NETWORKS
4.2.2.2 On QoS implementation over WLAN 802.11 family Quality of service is not yet supported in WLAN 802.11x, but a great amount of work is going on in different directions. There are many examples [106, 107] of the implementation of different QoS methods and approaches over the 802.11b standard. These methods and approaches can be used to modify other 802.11x networks as well, but IEEE recently has approved the 802.11e standard [109, 110] including QoS support. The QoS support in 802.11e still presents several open issues like the following. • traffic categories (TC) assignment: the standard does not define how the TCs are assigned. They can be assigned directly by the application or by another entity in the network. The introduction of a middleware component may be required. If the network is already QoS-enabled (i.e. DiffServ), the proper TC can be assigned by mapping the upper layer QoS parameter. • Content Based Policies: within the same flow, each packet may have a different level of importance with respect to the perceived quality. • Hybrid networks: the deployment of 802.11e will most probably require a hardware substitution or a firmware update, depending on the manufacturer. During initial phases of deployment, the stations that use different technologies like 802.11b and 802.11e may well be in the same network (being served by an 802.11e-enabled Access Point), and such situations may strengthen unfairness between 802.11b and 802.11e clients. The transition from plain WLAN802.11b to QoS-enabled 802.11e could be very expensive and WLAN hot spots which are already deployed will not be substituted by new technology very fast. Therefore, at the current stage one can expect 802.11e implementation and start to build testbeds and hotspots as soon as the equipment is available, but 802.11b commercial hotspots have been launched and traffic is running over these networks. Implementation of traffic-related tests with respect to perceived QoS is of paramount importance, mainly due to demanded contribution to the following reasons, at least • Network planning problems solutions. • Understanding of WLAN behavior under diverse multimedia traffic loads. • Contribution to understanding of DoS attacks on wireless interface. 4.2.3 Current GPRS implementation 4.2.3.1 Prerequisites Data exchange of GPRS phones is unsymmetrical. Data uploading (e.g. sending an e-mail message) is slower than downloading (e.g. reading an e-mail message). The higher number indicates the phone’s ability to employ channels for downloading, while the smaller one indicates the ability to employ channels for uploading. Thus, a GPRS (3+1) phone can send data by one channel and receive it by three channels. However, downloading data by three channels does not mean that the data transmission rate increases three times. 40 Kbps and other data transmission rates are distributed between all employed channels. Initially, GPRS supports up to 8 channels on each bearer
37
38
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
frequency. In order to utilize the channel’s capacity better, four different coding methods are used (CS-1 9.05, CS-2 13.4, CS-3 15.6, and CS-4 21.4 Kbps per one channel). Currently, most commercial networks support only two coding schemes (CS-1 and CS-2) and therefore theoretical GPRS rate cannot be achieved. It may be the case that all channels will never be used for GPRS data transfers since GSM voice calls employ the same channels. The other reason is that it would require greater computing efficiency (which results in greater power consumption), which phones lack. 4.2.3.2 On QoS implementation over GPRS GPRS architecture is comprised of a set of GPRS Supporting Nodes (GSNs), which are nodes equipped with the GPRS protocol stack. Two types of GSNs are defined, Supporting GSN (SGSN) and Gateway GSN (GGSN). The latter acts as a logical interface to external Packet Data Networks (PDNs) and consequently to global Internet, maintaining, also, routing information used to tunnel IP datagrams to the correct SGSN, while the former is responsible for servicing the mobile stations currently located within its service area. In GPRS, a specific QoS profile, a part of the so-called packet data protocol (PDP) Context Profile, is assigned to every subscriber on his or her attachment to the network. This profile contains information like: • traffic precedence class • delay class • reliability class • peak throughput • mean throughput class Traffic precedence class may be of high, normal and low priority. Four delay classes and five reliability classes are defined. Nine peak throughput (i.e. 8, 16, 32, 64, 128, 256, 512, 1024, 2048 kbps) and 19 mean throughput classes (from best effort up to 111 kbps) are defined. The GPRS QoS profile can either be requested by the mobile user during the PDP context activation phase or, if no profile is requested, a default one assigned to the user on his or her subscription to the network, is being activated. This default QoS profile is defined in the Home Location Register (HLR) along with other information relative to the subscriber. SGSN is the entity responsible for fulfilling the QoS profile request of the mobile station. After the activation of a certain profile, no modification by the mobile station is allowed as long as the current one is active. If the mobile station wishes to modify its QoS profile, it should have to detach from the network and then attach again, specifying a new QoS profile. SGSN, on the other hand, can modify an active QoS profile at any time, depending on the network load and the resources available. Although a QoS profile is defined, this profile does not differentiate treatment of data flows within the core GPRS network (that is between the SGSN and the GGSN). All flows within the network are treated uniformly (best effort IP) and the QoS parameters mentioned before are used to allocate resources outside the core network. Thus, after the activation of a certain QoS profile, the mobile station is responsible for shaping its traffic to the negotiated QoS and the GGSN is responsible for restricting data flows to the mobile station based
PRACTICAL STUDY: TESTBED CONFIGURATION
39
on the QoS profile. This implementation can be considered as a network with guaranteed bandwidth on the access interface, but with no provision for allocating resources within the GPRS core network.
4.3 PRACTICAL STUDY: TESTBED CONFIGURATION The testbed was built to test and compare performance of the different access networks performance for different workloads and mobility patterns under three multimedia services. The three access networks are WLAN, GPRS, and fixed Ethernet LAN. Testbed configuration is presented in Fig. 4.1. Several computers equipped with different operating systems (OS) and different access network devices were employed. WLAN tests were carried out on a base of running implementation of 802.11b WLAN in the campus area of Tampere University of Technology. To enable fixed Ethernet access, 100 Mbps Ethernet LAN of the Institute of Communication Engineering was employed. Both WLAN and LAN networks are connected to the public Internet via broker-gw edge router. For carrying out tests over the GPRS network, a connection to the GPRS network of a commercial Finnish GSM operator was employed.
iperf-server mp3-live
mp3-source
helix-server
LAN Public Internet
broker-gw
WLAN
GGSN mobila
. . .
GPRS
SGSN
GPRS RAN real-client
. . . iperf-client
iperf-client
Fig. 4.1 Testbed configuration.
Let us explain the testbed. To emulate the mp3 file server mp3-source, a desktop PC PIII under Win2000 OS connected to 100 Mbps Ethernet LAN was used. To ensure fixed station performance against OS-specific issues, the tests were validated with different fixed stations, which were based on different PC PIII with Linux OS RedHat 7.2. This is be-
40
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
cause the QoS perceived by the end-user may vary significantly from one OS to another. Different users use different devices to access Internet services. To evaluate the performance of live mp3 streaming service, one of the stable Internet Live Radio was chosen. The name of this machine is mp3-live. It is clear that paths between the end-user and both mp3 file server and the Internet Live Radio station are different. To overcome this obstacle, traffic tests of each destination were performed over relatively long time periods. It appears that both paths are stable and therefore can be used in the testbed. The mobila station was used as a sink for both mp3 services’ traffic, or for traffic from mp3-source and mp3-live, in other words. The mobila was Mac PowerBook G4 under Jaguar v.10.2 OS. It was equipped with WLAN 802.11b (interior) and Bluetooth facilities (used to connect the computer with a GPRS adapter, namely GPRS phone Nokia 7650). To ensure this mobile station performance against OS-specific issues, the tests were validated against a different mobile station which was based on IBM ThinkPad PIII laptop under Win2000 OS and equipped with an exterior Cisco’s Aironet 350 802.11b WLAN card and the same Bluetooth adapter. To organize video streaming service over the testbed, some more stations were set up. The mobile station called real-client, IBM ThinkPad PIII laptop under Win2000 OS equipped with Cisco’s Aironet 350 802.11b WLAN card, was used as a sink for video streaming service from helix-server. To ensure mobile station performance against OS-specific issues, the tests were validated against different mobile stations. The fixed station called helix-server was desktop PIV under Win2000 OS connected to 100 Mbps Ethernet LAN. To generate competing traffic on WLAN interface, the well-known iperf client-server utility [111] was employed. To maintain the iperf server, desktop PC PIII under Linux OS connected to 100 Mbps Ethernet LAN was chosen. The name of this station is iperf-server. A set of mobile stations iperf-client were used to organize heavy background connections over WLAN to the iperf-server station. The access point (AP) was Avaya’s ORiNOCO range extender. To hide implementationspecific issues, all tests were carried out with only one WLAN access point.
4.4 TOOLS AND ENVIRONMENT 4.4.1 On video streaming service implementation and testing In the testbed, the commercial implementation of client-server streaming service was used. The combination of RealNetworks Helix server and RealNetworks RealOne player was chosen. To perform actual streaming, Helix server employs slta utility, which is part of the distribution. Free distribution of Helix server publicly available at [112] was used. Helix server has the ability to stream a lot of well-known medias, including both proprietary and standard ones. Real media streaming format was chosen because of the following reasons. At first, free distribution of Helix server allows all server-side capabilities only when real media format is used. Nowadays, the real media format is very popular in the Internet because of the relatively good quality of low bit rate videos. Additionally, when real media format is used, free distribution of Helix server makes it possible to
TOOLS AND ENVIRONMENT
41
serve clients with different bandwidth capabilities. Moreover, the bandwidth at which the client is served can also be changed dynamically during the connection lifetime. In order to achieve that, the video should be coded at different target rates, each of which is specific for a certain bandwidth capability of a client. The video streaming service operates as follows. The server continuously listens to specific ports for connection requests. When the request arrives, the server sets up a real-time streaming protocol (RTSP) connection, adds the client to a connection pool, and then begins streaming at the rate which is the most appropriate for the requesting client. However, if the bandwidth capability of the client changes, the server can adapt connection by increasing or decreasing the target rate of the video. To evaluate the performance of the live streaming service, the fragment of high-motion pre-recorded video was chosen. Using the slta utility, it is possible to carry out tests continuously as long as required. To encode video to real media format, Helix Producer v9.0 was used. The resolution of video was set to 240×352 pixels, while the target rates were chosen to be 56 Kbps, 150 Kbps, and 350 Kbps. Note that paths between both helix-server and iperf-server and corresponding clients are stable, and pass only one router, broker-gw. The broker-gw router is not considered as bottleneck of the testbed configuration since all tests were carried out when both LAN and WLAN were in unloaded conditions. Additionally, both the Helix server and iperf-server were located within the same Ethernet segment. Such a condition cannot also be considered the bottleneck, since the bandwidth of fixed LAN is substantially higher than that of 802.11b WLAN. Therefore, the only bottleneck in the testbed is the WLAN. The connection was assumed to have failed after 30 seconds of unsuccessful attempts and the player was not allowed to prefetch data. In [P5] the authors proposed to assign WLAN channel quality classes on the basis of measured SNR levels. The ranges of SNR and corresponding user-friendly channel conditions were chosen as follows (note that other partitioning of SNR is also possible): ≤ 10dB (very bad); 10 – 20dB (poor); 20 – 30dB (fair); 30 – 40dB (good); ≥ 40 (excellent). Additionally to different SNR ranges, tests were carried out under different competing traffic volumes. Both the UDP and TCP competing traffic patterns were generated by iperf-sever and sink to iperf-client stations. To capture traffic and obtain statistics, an Ethereal software package [113] in conjunction with post processing by own Perl scripts was used. 4.4.2 On mp3-based services implementation and testing As stated above, both of the mp3-based services were implemented by means of server – client interaction. Servers were connected to the public Internet via fixed access, while clients were connected via wireless accesses. In each service scenario, a client was responsible for posing content requests to a server. In the testbed environment, both mp3-based services under different user mobility patterns were evaluated. The term mobility pattern is used to denote global behavior of a user. Three global mobility patterns were defined. These are: fixed user, walking user, and driving user. Since WLANs are assumed to be accessible only in relatively small hot-spot
42
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
areas, they can serve both fixed and walking users. The GPRS access network serves all three types of users. LAN is only accessible for a fixed user.
4.5 TEST SCENARIOS At first, it is necessary to explore performance characteristics of WLAN, GPRS, and Ethernet under unloaded conditions. The aim is to obtain networks’ service parameters for traffic of each access network. Several advanced UNIX-based utilities [114] were used. These utilities estimate service parameters on transport layer. 4.5.1 WLAN end-to-end performance testing The end-to-end performance parameters of the path between iperf-server and iperf-client were obtained using the iperf utility. Performance parameters of iperf-server – iperf-client and helix-server – real-client paths are similar since iperf-client and real-client were on the same WLAN access point, while iperf-server and helix-server were on the same Ethernet LAN segment. To obtain stable characteristics of WLAN, it is important to measure them on a wide time scale. To obtain statistically tractable values, 5 periods of 120 minutes testing in each period were performed. The following parameters were of particular interest: maximum throughput of WLAN, end-to-end round trip time (RTT), and jitter. A statistics summary is presented in [P5], Table 2. 4.5.2 GPRS end-to-end performance testing To obtain end-to-end performance parameters of the GPRS network, the route from mobila to mp3-source was used. The values for several important parameters like maximum throughput, RTT, loss probability, and jitter are presented in [P6], Table II. These parameters are measured between nomadic station mobila and edge router. Using the RTT parameter measured between nomadic station and edge router, it is possible to characterize any particular access network. In the testbed, in order to characterize the GPRS access network, GGSN was used as an edge router. As one can see in [P6], Table II, the differences between end-to-end RTT and RTT between end station and edge router are relatively small. End-to-end RTT gives us an estimation of joint access network and core network parameters. However, since the GPRS core network is built on a base of high-speed wired technology for which the values of delays and probability of loss are much smaller than in the GPRS access network, it is wise to neglect it. To obtain stable characteristics of GPRS, it is important to measure them on a wide time scale. To obtain statistically tractable values, 5 periods of 120 minutes testing for each period were performed.
TEST SCENARIOS
43
4.5.3 Notes on end-to-end performance testing for different access technologies Estimation of loss probability for the transport layer of WLAN and GPRS is a non-trivial task because both wireless access technologies use extensive error recovery algorithms. Low WLAN maximum throughput is explained by the fact that the tests were conducted under realistic conditions; moreover, the throughput was measured at a transport layer while 11 Mbps of 802.11b can be achieved only at a physical one. 4.5.4 Video streaming service In the testbed environment, the performance of live video streaming service in WLAN environment under different SNR levels and different competing traffic loads was evaluated. Traffic was generated by a user of a real-client station by posing requests on live video streaming from helix-server. 4.5.4.1 Without competing traffic For the worst SNR case of less than 10 dB, TCP’s 3-way handshake procedure was successful and mean throughput was measured to be 0.22 Mbps, but the RTSP connection was not established. This stems from the fact that under this SNR condition the quality of the air interface becomes very unstable and therefore there were many of bandwidth renegotiations performed by RTSP. After the server had failed to establish a connection in 30 seconds, the connection request was rejected by the client. For the case when SNR was in 10 – 20 dB range, the RTSP connection was successfully established and it did not take substantially more time compared to the 20 – 30 dB range. The perceived QoS for all SNR classes between 10 – 40 dB and higher was good. Therefore, the most critical point in live video streaming service is the RTSP connection establishment phase. Indeed, there were cases when the SNR was fluctuating a lot while the RTSP connection was being set up. It caused a lot of errors at the wireless channel, which resulted in frequent bandwidth negotiations. Therefore, sometimes the client’s connection establishment timer had expired before the connection was set up. The summary of statistics under different SNR ranges and in unloaded network condition is presented in [P5], Table 3. 4.5.4.2 With competing traffic Then, the same tests given the different load conditions of the WLAN were performed. The following cases of service performance under different background loads (or competing traffic in other words) were tested • four UDP streams with a target bandwidth of 2 Mbps each • four UDP streams with a target bandwidth of 1 Mbps each • four TCP streams with an initial window size of 60 Kb each The results are presented in [P5], Table 4. The competing traffic adds additional (to that given by SNR) fluctuations to the bandwidth available at the path between client and server and induces the bottleneck at the wireless interface of WLAN.
44
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
It was found that the RTSP connections were not established under all considered types of competing traffic when the SNRs were in ≤ 10 dB, 10 – 20 dB, and 20 – 30 dB ranges. The reason was that TCP was not able to complete 3-way handshake procedure, which is necessary to establish the RTSP connection. The live video streaming service did not work for all SNR ranges when the network was loaded by four UDP traffic sources, each of which was targeted on 2 Mbps. The client permanently failed to perform TCP’s 3-way handshake. When the target rate of each competing source was decreased to 1 Mbps, Helix server and RealOne player completed 3-way handshake and established the RTSP connection for SNR of more than 30 dB. For the case when SNR was 30 – 40 dB, RTSP establishment phase took a much longer time compared to unloaded network conditions and the perceived quality was unacceptably bad – long pauses without picture and even without sound appeared, and most frames were corrupted. Despite that, the service was operated without a break. The duration of 3-way handshake procedure when SNR was more than 40 dB and the network was loaded by four 1 Mbps UDP source, was much greater than that of the network in unloaded conditions, as well as for the time duration of the RTSP connection establishment phase. However, the perceived quality of live video streaming service was good and can be roughly compared to the perceived quality in the network in unloaded conditions. Almost similar observations were made when the network was under the load of four competing TCP connections, each of which had 60 Kb initial window size, and SNR was fluctuating in the 30 – 40 dB range. However, one can note that the duration of the RTSP connection establishment phase was smaller compared to the previous case. It is explained by fairness between competing TCP connections whose bandwidth changes dynamically depending on losses. On the other hand, a UDP connection tries to get as much as it needs because the target rate is strictly defined. The quality of the picture for this case was sometimes slightly deteriorated, i.e. some frames were blurry and truncated. However, the picture quality was acceptable. While having different background loads on wireless interface, some attempts to use not only real-time streaming multimedia application, but plain Internet services like WWW as well, were made. It was found that even WWW services can easily get denial of service if the wireless interface is overloaded. This happens mostly because of the TCP timer expiring. This is true especially for the case with heavy UDP competing traffic. The throughput graphs for each considered case are given in [P5]. 4.5.5 mp3-based services 4.5.5.1 mp3 file transfer The next type of service under consideration in the testbed was mp3 file transfer. Note that if one examines an mp3 downloading service user’s behavior, it can be seen as sudden transfers with relatively long periods of silence. To be fair in user behavior estimation, it is necessary to take into account the traffic produced by the user while surfing mp3 repository (it could be HTTP- or WAP-based service). However, such traffic is beyond the scope of this study. In the testbed, this type of traffic was generated by mobila station by posing requests on mp3 file transfers from a mp3-source server located in a fixed part of the public Internet. To download files, the FTP over TCP service was employed. The summary of statistics obtained for LAN, WLAN, and GPRS is presented in [P6], Table III.
TEST SCENARIOS
45
Brief summaries of results obtained for every access network type are as follows. • The perceived QoS of mp3 file transfer in LAN environment was excellent. Despite such performance, it should be noted that during the test a relatively long collision occurred. However, it did not affect the perceived QoS significantly. Because of that, the perceived QoS obtained with LAN access can serve as a reference for wireless access technologies. • The perceived QoS of mp3 file transfer in WLAN for users with fixed location was very good and comparable to LAN results. The throughput is significantly less with WLAN since both the theoretical and practical bandwidths of 802.11b WLAN are considerably lower. Additionally, one can notice that several collisions on WLAN occurred. However, it also did not affect the perceived QoS dramatically. The perceived QoS did not change dramatically in WLAN environment with a walking user compared to the same environment with a fixed user. • The perceived QoS of mp3 file transfer in GPRS environment was quite bad for every mobility pattern compared to both LAN and WLAN environment. It took approximately 820 seconds to download the entire mp3 file of 3 MB size for a fixed user and there was an approximately 720 seconds long prefetching phase (actually, such long prefetching time is due to the software implementation and obviously it can be significantly reduced). The throughput was constant at approximately 4900 bytes per second with every mobility pattern under consideration. However, the goodput varied substantially. Additionally, the other phenomenon with GPRS wireless access was observed. Given the different mobility patterns of mobile users, the perceived QoS almost did not change. The RTT values are spread uniformly over a constant interval for all mobility patterns, which allows us to assume that they were mostly contributed by specific implementation of the GPRS network. Moreover, the deviation of throughput increased insignificantly when more mobility was added to the behavior of the user. All correspondent Stevens TCP graphs, RTT graphs, and throughput graphs for every considered case are located in the Appendix of [P6]. 4.5.5.2 mp3-based Internet Live Radio Most current low-quality (low target rate) Internet Live Radio stations are using TCP protocol instead of the proposed UDP/RTP/RTCP set of protocols [P6]. This noticeable case can be explained by the results obtained in paper [115] dedicated to VoIP performance testing throughout several years. Among other conclusions drawn by the authors, one should note essential fact that the overall quality of the fixed public Internet network is becoming significantly better every year. The lack of session control protocol for UDP is also one of the reasons behind the use of TCP. To obtain results, one of the stable Internet Live Radio, mp3-live server, was chosen. Because of GPRS bandwidth limitations, a 24 Kbits mp3 stream was taken. The stability of Internet Live Radio was evaluated by survival tests, i.e. how many hours the service is provided without break. For all three available levels of quality, 24 Kbps, 56 Kbps, and 128 Kbps, the chosen station gave very strong results – no break in service within 24 hours.
46
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: ACCESS NETWORK PART
The prefetching phase allows receiver software to buffer a certain amount of streaming data to smooth end-to-end packet delay variation (jitter). For Internet Live Radio tests, the application level buffer size was set to 240 Kb, which is equal to 10 seconds of playback. Three-way handshake procedure performed by every TCP connection may introduce additional unnecessary delay in a service running over a low-speed network. On average this procedure takes quite a short time, not exceeding one second even for GPRS and obviously it can be neglected. For certain packets the PUSH flag in the TCP segment is set. This attempts to ensure timely delivery of a whole TCP segment to the application. The summary of statistics for a fixed user under controlled transfer is presented in [P6], Table IV. Brief summaries on the results obtained for every access network type are as follows. • The perceived QoS of Internet Live Radio in LAN environment was excellent. The prefetching phase took the least time and then the service operated steadily. Due to the high available bandwidth, there was no re-prefetching during the whole time of connection. • The perceived QoS of Internet Live Radio in WLAN environment for both a fixed and walking user was excellent. The prefetching phase took a bit more than in the fixed LAN case and then the service operated steadily. Due to the relatively high available bandwidth, there was no re-prefetching during the whole time of connection for a fixed user, but it happened once for a walking user. The perceived QoS was not affected dramatically by the mobility, but low SNR in several shadowed areas could be the reason for perceived QoS degradation. • The perceived QoS of Internet Live Radio in GPRS environment was good. One break in service was observed. It was the case of a walking user in an underground parking area. Moreover, the perceived QoS was not affected by the mobility of the user at all. However, one can note that the variability of throughput increases with increase in velocity of the GPRS user. At the same time the RTT values for all three cases stay intact and are uniformly distributed over the same constant interval. Let us also highlight the considered driving profile: during the first half of the test interim the user drove over suburb streets at 50 km/h and during the second half along a highway at 120 km/h. All correspondent Stevens TCP graphs, RTT graphs, and throughput graphs for all explained cases are located in the Appendix of [P6].
4.6 SUMMARY Based on the results obtained in [P5], [P6], the experience gained from testbed put up, and the tests carried out, the following important conclusions are drawn. • It is already possible to add some low-bandwidth multimedia services like low-rate Internet Live Radio for nomadic users even in slow GPRS networks. The only drawback of the current evolution of mobile systems that restricts implementation of current multimedia service available in the Internet is the lack of bandwidth on wireless
SUMMARY
47
interface. All other issues like the absence of specific wireless-oriented TCP implementations are not of vital importance. Therefore, one can expect that deployment of WCDMA technology at the air interface of the next stage of 3G systems’ evolution will immediately open the possibilities for wide use of different multimedia services currently available on the Internet. Enhanced Data Rates for Global Evolution (EDGE), a 3G radio technology standardized by ITU and 3GPP [116, 117], built on the existing GSM/GPRS network, is expected contribute significantly to the development of wireless multimedia services and will act as catalyst for further integration of IP-based services and mobile networks. • It is not possible to support multimedia services like live video streaming on a commercial basis over the current evolution of IEEE 802.11b WLAN. Despite a common belief bandwidth-greedy live streaming services are not ready for wide implementation in hot-spot areas where both high traffic volume and relatively weak signal strength (less than 30 dB) may deny the service easily. • Measurements carried out have shown that live video streaming service cannot be easily implemented over wireless medium with reliable QoS. The use of TCP in the connection establishment phase may easily deny the service even the network in unloaded conditions. The live video streaming service performs well in the presence of any type of considered competing traffic only when excellent channel conditions (greater than 40 dB) are met. However, the perceived quality becomes unacceptable in the presence of any type of considered competing traffic with an SNR of 30 – 40 dB. A high volume of competing traffic easily denies service when the SNR is under 30 dB. Frequent bandwidth fluctuations, caused by both SNR and competing traffic loads, stimulated numerous bandwidth negotiations, and therefore, the connection establishment’s timer often expires before the RTSP connection is set up. However, these bandwidth fluctuations do not actually indicate that the streaming service fails due to scarceness of the bandwidth, since the actual streaming is performed over a bandwidth-greedy UDP protocol. • Denial-of-Service (DoS) attacks nowadays are a very popular method of malicious violation in Internet. Internet goes mobile and DoS attacks can be easily introduced within wired interfaces [118]. The results presented above (in part of WLAN performance under competing traffic) allow claiming that using periodic heavy load spikes in a wireless network can easily deny TCP-based services and seriously affect UDPbased services’ perceived QoS. The results obtained during thesis preparation have initiated interest in the area and stimulate research currently being undertaken [119].
Chapter
5
Next-Generation Telecommunications Networks: Core Network part
T
HE Quality of Service (QoS) assurance task comprises access network (AN) QoS assurance and core network (CN) QoS assurance. The task of QoS assurance has been considered independently in the previous chapter. This chapter is dedicated to the task of QoS assurance in CN. Indeed, NGN presumes independent treatment of AN and CN [22]. As already discussed in chapter 2, Differentiated Services (DiffServ, [3]) architecture paradigms can be used in NGN systems for building core network transmission facilities. Indeed, in many recognized projects [120–122] the use of many DiffServ principles for the implementation of NGN QoS architectures can be found. It is assumed that wireless accesses must maintain QoS on their own, therefore those studies relate to the provision of QoS at the backbone interdomain level. In this chapter, based on the DiffServ model, two end-to-end transmission services for real-time service traffic delivery are developed. In order to satisfy the demands of the realtime nature of the service, it is proposed to use assured forwarding (AF) [35] and expedited forwarding (EF) [36] per-hop behavior (PHB) groups. For both transmission services the specific traffic profile parameters which can be adequately tuned to fit real-time service traffic requirements are provided. Using these parameters and AF PHB or EF PHB class, it is shown how to construct well-defined transmission services that are suitable for real-time service traffic delivery. Using the QoS parameters which should be provided for real-time service traffic, one can evaluate the required capacity not only within the DiffServ ingress node, but within the interior nodes as well. The transmission services are designed in such a way that they are characterized by the parameters of DiffServ ingress and interior nodes, bound losses and delays along the path of behavior aggregate and allow the prediction of the QoS degradation which can be experienced with both behavior aggregate and single microflows. 49
50
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
5.1 QOS SUPPORT ON BASE OF DIFFSERV The DiffServ approach is based on the service differentiation principle. It aggregates flows with the same QoS requirements and assigns to them the same treatment at the routers along the path of aggregated traffic flow. The service differentiation is achieved by implementing several simple PHB in the routers. DiffServ provides probabilistic guarantees for aggregated traffic and uses a sort of static connection admission control (CAC) algorithm which is based on service level agreement (SLA). The IETF DiffServ working group has standardized two PHB groups. These are AF PHB and EF PHB. AF PHB is designed for a range of applications which need different QoS guarantees. There are four classes of PHB identification codes within the AF PHB group. Within each class there are three distinct DiffServ codepoints (DSCP) with different packet drop precedence. EF PHB is targeted on applications which require strict guarantees of end-to-end delay and should not suffer from packet losses. Because of the ability of AF PHB and EF PHB to provide relative and strict priority in service respectively together with adequate bandwidth reservation, in order to handle such type of traffic within the DiffServ domains appropriately two approaches are considered: real-time service over AF PHB and real-time service over EF PHB. QoS parameters which are provided by the designed transmission service are characterized by the parameters of DiffServ ingress and interior nodes and in particular by the buffer capacity and link share which are assigned to a single AF PHB class or whole EF PHB group. The transmission algorithm which is based on AF PHB allows the loss of packets and can provide the necessary delay that satisfies the needs of real-time nature of the real-time service traffic. In the case of EF PHB, there are no packet losses and delay is bounded. Both services provide a clear way to tune queuing parameters which directly affect losses and delays within the DiffServ nodes.
5.2 REAL-TIME SERVICE 5.2.1 Definition Real-time service is expected to be one of the very popular applications in broadband multiservice networks. From the traffic transmission point of view, real-time service can be classified as one of the most delay- and loss-intolerant client-server application. Traffic generated by this sort of applications needs strict QoS guarantees in terms of bandwidth, losses, delays, and jitter. In order to satisfy the demands of the real-time nature of real-time service traffic, an appropriate network transmission facility should be provided. Currently, Internet does not provide any QoS guarantees for its applications. It provides the best-effort service only, which does not satisfy the needs of real-time applications. Modern compression algorithms mostly operate in variable bit rate (VBR). VBR mode allows very high compression ratios while the quality of decoded picture remains almost the same. It is known that the real-time service traffic should have the necessary amount of reserved bandwidth and/or some type of priority in service. Since EF PHB is used to construct transmission services with constant end-to-end bandwidth, it can be inappropriate
REAL-TIME SERVICE
51
Real-time service user ISP N
1
ISP 1 SLA QoSD
...
QoSD
M
SLA
Real-time service user
Fig. 5.1 Real-time service configuration in NGN environment.
from the network dimensioning point of view to transmit VBR traffic over a CBR channel. However, in several cases, this is almost the only way to transmit compressed video over Internet with strict QoS guarantees. It is also known that loss-free transmission is preferable for real-time service traffic. However, keeping in mind the high peak rate and drastic short-term rate changes found in real-time service sequences it may not be wise to provide a loss-free transmission in all cases. Moreover, the development of effective FEC algorithms and the evolution of codecs’ native destination-based error concealment techniques allow to reduce requirements on packet losses. At the same time delay requirements remain very strict. For those cases where there is no need to provide strict QoS guarantees, AF PHB service is proposed. 5.2.2 Real-time service configuration Assume that there are N neighboring Internet domains making up a chain. Consider that these domains belong to different large internet service provides (ISP). Both here and after, let us under ISP assume an NGN operator, or any independent provider running own DiffServ domain is considered. Let us follow Mobydick’s approach [120] and call such a domain a QoS Domain (QoSD). Further, assume that every QoSD has its own DiffServ implementation in accordance with IETF specifications [3] and that SLA is established between neighboring QoSDs. Such a configuration is depicted in Fig. 5.1 and it is may be the case in the NGN framework. Let us call this scenario the remote real-time service configuration. Both here and after, under real-time service such services as live video / audio streaming, VoIP, etc. are assumed. Therefore, in this configuration the QoSD may not have their own real-time service source or dedicated connection to real-time service content provider. Moreover, the nearest real-time service source can be located far from the home domain of the real-time service user. However, the desired QoS guarantees must be provided between the real-time service entities. It is also assumed that there are several users in QoSD N who wish to use real-time service. However, only QoSD 1 has a dedicated connection to a real-time service content
52
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
provider. In this case, real-time service (and, therefore, respective QoS guarantees) should be provided across all N domains which belong to different QoSDs. 5.2.3 Source configuration A real-time service source possesses information on traffic issued by it by means of requested user service parameters. That is why the source can adapt a generated traffic profile to an available QoS class offered by the network on the basis of wireless interface technology and available resources on a fixed part of the network. Above it is assumed that wireless interface peculiarities are not considered in this chapter. Therefore, real-time service traffic generated by a source station is delivered over wireless interface with a certain QoS and assumed to be treated by a fixed part of the NGN system, namely the backbone or CN. Below only traffic treatment within the NGN system CN case is considered. Let us introduce several assumptions about the real-time service source configuration. Assume that the length of all IP packets which belong to different microflows within the multiplexed real-time service traffic stream is constant and that its value R lies between acceptable bounds. Thus, the real-time service IP packets’ service time is a constant value and equal to the time it takes to transmit one packet on the outgoing link. This is true at least for a DiffServ ingress node. Let us also assume that domains do not implement IP packet segmentation procedures along the path between the real-time service source and the end-user equipment. In this case the service time for each real-time service IP packet is constant in every DiffServ network node along the path to the destination. Below let us assume that the real-time service source performs a host marking procedure and that all real-time service packets which enter the network already have an appropriate DSCP codepoint. 5.2.4 Modeling of real-time service traffic Commonly used models like the Poisson and Bernoulli process, in spite of their computational tractability, often do not incorporate important characteristics of real traffic, especially multimedia traffic. A discrete time batch Markovian arrival process D-BMAP originally has been defined in [123] and is the discrete time version of the Markovian arrival process MAP defined in [124], which was originally called N-Process by M. Neuts in [125]. Let us define the D-BMAP process. Consider a discrete time Markov chain with transition matrix D. Suppose that at time k this chain is in some state i, 1 ≤ i ≤ m. At the next time instant k + 1 there occurs a transition to another or possibly the same state and a batch arrival may or may not occur. With probability dij (0), 1 ≤ i ≤ m, there is a transition to state j without an arrival, and with probability dij (n), 1 ≤ i ≤ m, n ≥ 1, there is a transition to state j with a batch arrival of size n. Therefore, ∞ X m X n=0 j=1
dij (n) = 1
(5.1)
REAL-TIME SERVICE TRAFFIC MODELS
53
The matrix D(0) with elements dij (0) governs transitions that correspond to no arrivals, where the matrices D(n) with elements dij (n), n ≥ 1 govern transitions that correspond to arrivals of batchesP of size n. ∞ The matrix D = n=0 D(n) is the transition matrix of the underlying Markov chain. Let ~π be a stationary probability vector of this Markov process, i.e. ~π D = ~π , ~π~e = 1
(5.2)
where ~e is a column vector of 1’s. D-BMAP can be employed to model various traffic sources [126], in particular on the basis of D-BMAP superposition it is possible to model such VBR sources like streaming audio and video [127]. Indeed, a number of well known arrival processes can be obtained as special case of D-BMAP. • The discrete time Markovian arrival process D-MAP [128]. • The Bernoulli arrival process [129]. • The discrete time Markov modulated Bernoulli process D-MMBP [130, 131].
5.3 REAL-TIME SERVICE TRAFFIC MODELS 5.3.1 Stochastic and deterministic traffic modeling It is well known that in order to analyze a network performance, models of different types of traffic should be introduced. Let us consider two approaches to traffic modeling. In accordance with first the approach, a traffic source or multiplexed traffic from a certain network element is modeled by a stationary stochastic process. Since such models should capture most relevant statistical characteristics of real traffic sources, the availability of representative statistical data as well as efficient methods of statistical analysis and stochastic modeling are of paramount importance. Moreover, in most cases the decision regarding whether the designed model is a good predictor of traffic source is often taken after a statistical comparison of modeled data with statistical data. The described approach is very popular and mostly used in traffic modeling. It is called the stochastic approach to traffic modeling. However, there are some cases when the statistical data of traffic sources are unavailable or those data cannot be represented by some type of stationary stochastic process, or when even the nature of traffic is unknown. In the presence of such uncertainty one can make an assumption that the traffic is unknown but satisfies certain regularity constraints [42]. Note that in most networks such an assumption holds – traffic entering the network is constrained by a certain regulator, e.g. token bucket policer. In most cases these constraints are represented by a simple deterministic model which describes the model of unknown traffic. Using these simple traffic models, it is possible to derive bounds of performance parameters within the wide variety of network elements [42]. It should be noted also that the stochastic models can be used to define the deterministic models of network traffic to which the traffic is constrained. This is called the deterministic approach to traffic modeling.
54
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
Below both approaches are used. The task can be defined as one of modeling the multiplexed real-time service traffic at the entrance to the network. Since the traffic source is known and statistical data are available [132], first the stochastic approach is employed. The deterministic model is based on the stochastic model. 5.3.2 Stochastic real-time service traffic model The packet stream at the network interface between the real-time service source and DiffServ ingress network node consists of multiplexed traffic from a number of real-time service traffic sources – the real-time service traffic behavior aggregate. Further, let us model both the traffic from a single real-time service source and the multiplexed traffic from a number of real-time service sources by the discrete-time batch Markovian arrival process (D-BMAP). It was shown [133, 134] that the D-BMAP process matches well the probability distribution function (PDF) of the single real-time service traffic source and its autocorrelation function (ACF). In order to define the PDF of the number of arrivals of the D-BMAP process, an empirical histogram of relative frequencies can be used. Since the ACF of the D-BMAP process obeys the geometrical sum distribution, it produces a good approximation of the ACF of empirical data [133,134]. These considerations give us an assurance that these statistical characteristics of the D-BMAP model match their empirical counterparts well. It has been determined that the ACF of multiplexed traffic from a number of real-time service sources obeys a near geometrical sum distribution [134]. Thus, the D-BMAP process can also model such type of traffic. The construction of Markov modulated processes from the empirical data involves the inverse eigenvalue problem. It is known that a general solution of the inverse eigenvalue problem does not exist. However, to date several papers have addressed the solution of this problem when some limitations on the form of eigenvalues are set [133, 134]. 5.3.3 Deterministic real-time service traffic model In order to construct the deterministic traffic model, let us assume that the stochastic behavior of the real-time service traffic behavior aggregate is unknown but constrained to the burstiness regulator (r, b) as proposed by R. Cruz [42], where b is the length of the regulator’s queue and r is the constant output rate. The behavior of the (r, b) regulator is similar to the token bucket mechanism. One can use the stochastic traffic model or empirical traffic pattern to compute the parameters of the model. Since both stochastic models of both a single real-time service traffic source and multiplexed traffic from a number of real-time service traffic sources have been defined, one can define their deterministic (r, b) models.
AF PHB TRANSMISSION SERVICE
b0
DiffServ ingress node
...
Real-time service traffic
G/D/1/K0 r0 tb(r0,b0)
Network interface
55
1
B0
... K0-1 2
Fig. 5.2 Model of behavior aggregate’s treatment within the DiffServ ingress node.
5.4 AF PHB TRANSMISSION SERVICE To assure QoS parameters, real-time service traffic transmission service should have the necessary amount of reserved bandwidth and/or some type of priority in service. Let us assign to real-time service traffic a whole AF PHB class since each AF PHB class has a predefined minimum amount of forwarding resources such as bandwidth and buffer space at each node along the path of the AF behavior aggregate. Because of the predefined amount of forwarding resources and taking into account the distributed control of network nodes, one can assume that the process of AF traffic treatment is independent of other traffic treatment within the network nodes. Let us consider the DiffServ ingress node with AF PHB implementation. In accordance with specification [3] such a node must perform two major functions: conditioning of the behavior aggregate and configuring of the node so that the behavior aggregate has a minimum departure rate and a certain amount of buffer space. The queuing model of the DiffServ ingress node which serves the real-time service traffic behavior aggregate is shown in Fig. 5.2, where ’1’ is the traffic conditioning block, ’2’ is the AF class queue, r0 and b0 are the parameters of the traffic conditioning block, (K0 − 1) is the length of the buffer, and B0 is the outgoing link share which is assigned to the AF PHB class as the minimum departure rate. The DiffServ interior network nodes should perform only the queuing functions. The queuing model of the DiffServ interior node is presented in Fig. 5.3, where (Ki − 1) is the length of the buffer and Bi is the outgoing link share which is assigned to the AF PHB class as the minimum departure rate. Let us then consider the transmission path of the real-time service traffic behavior aggregate through the DiffServ domain. After leaving the real-time service source, the real-time service traffic behavior aggregate enters the DiffServ ingress node and then passes through a number (say N ) of DiffServ interior nodes. After that it leaves the DiffServ domain. Let us exclude the DiffServ egress node from the path of the real-time service traffic behavior aggregate because its functions can be identical to those performed by interior nodes.
56
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
DiffServ interior node G/D/1/Ki ...
Bi
Ki-1
Fig. 5.3 Model of behavior aggregate’s treatment within the DiffServ interior node.
In considering real-time service traffic behavior aggregate transmission over DiffServ domain, it is necessary to outline that the packet losses may occur within each network node and the remarking of packets may occur within the traffic conditioning mechanism. In order to compute the part of the SLA which QoS parameters expectation consist of and which should be provided to the user as a valuable part of the SLA the configuration parameters of all network nodes should be estimated. Each DiffServ interior node which serves the real-time service traffic behavior aggregate is characterized by two parameters (Ki , Bi ), i ∈ {1, 2, . . . , N }, while the DiffServ ingress node is described by two pairs (r0 , b0 ) and (K0 , B0 ). Thus, appropriate values of (r0 , b0 ), (K0 , B0 ), and (Ki , Bi ), i ∈ {1, 2, . . . , N } should be provided. These configuration parameters define the QoS expectations which are experienced by the real-time service traffic behavior aggregate. Note that for a case where loss of packets may occur in every network node, it is almost impossible to predict the performance parameters of the real-time service. Thus, in order to define the required capacity for the AF class which serves the real-time service traffic behavior aggregate, let us first analyze the treatment of the real-time service traffic behavior aggregate within the DiffServ ingress node. 5.4.1 Token bucket parameters The traffic conditioning functions implemented within the DiffServ ingress nodes are based on traffic profiles. Traffic profiles specify the temporal properties of traffic stream selected by a classifier [3]. One of the most popular traffic profiles used in IP-based networks is based on a token bucket policing mechanism. In general, a token bucket can be used for two purposes. First, it shapes incoming traffic up to some horizon. Shaping bounds of the simple token bucket are given by two parameters: bucket depth b and token rate r. A token bucket can also serve as a traffic marker – the DSCP of packets which do not conform to the token bucket specification (r, b) can be changed. The token bucket traffic profile defines rules for determining whether a particular packet is in-profile or out-of-profile. Out-of-profile packets are those packets which arrive when insufficient tokens are available in the bucket. Different actions may be applied to in-profile and out-of-profile packets.
AF PHB TRANSMISSION SERVICE
57
The bucket depth and token rate are often given in bytes. This is because the packet lengths can vary. When all packets within the real-time service traffic behavior aggregate have the same length, one can measure the bucket depth and token rate in packets without the loss of generality. Assume that the real-time service traffic behavior aggregate after entering the DiffServ ingress network node is policed and shaped via a simple token bucket mechanism. Token buckets allow packet remarking in accordance whith negotiated traffic profiles, which is highly undesirable for real-time service traffic transmission over an AF PHB service. Let us consider the token bucket which allows packet remarking. In this case a certain part of arriving packets will be remarked to other DSCPs. The number of packets which should be remarked depends on the token bucket parameters and on the arrival process. The DSCPs to which these packets are remarked may represent lower drop precedence within the AF PHB class or even the DSCP corresponding to the best effort service. Since these remarked packets may belong to different microflows which produce behavior aggregate, in the case of congestion these microflows can experience unpredictable delays, reordering of packets, and even losses which are caused by traffic conditioning functions. 5.4.2 Estimation of AF class queue parameters In order to characterize the developed transmission service, one should define the AF class queue parameters (B0 , K0 ) which should be assigned to the real-time service traffic behavior aggregate within the DiffServ ingress network node. Those parameters depend on the desired QoS parameters, which in turn are expressed in terms of losses and delays. The number of losses can be given by the mean number of lost packets or by the PDF of lost packets. The delay requirements can be expressed in terms of mean delay, PDF of the delay or delay bound. In most cases it is sufficient to fulfill the requirements which are given by the mean number of losses and delay bound. In order to introduce the model of the AF PHB class queue, it is necessary to consider the queue management (QM) algorithm. The AF PHB specification [35] proposes to use of some sort of active QM algorithm within each class of AF PHB. An example of such an algorithm is random early detection (RED, [135]) and its extensions to the case of several drop precedence levels. RED-like algorithms monitor the instantaneous congestion level and compute the congestion level in order to determine when packets should be discarded. These algorithms work well in the presence of TCP connections which use a congestion avoidance mechanism [136,137]. It can be demonstrated that in the presence of UDP flows the efficiency of algorithm significantly decreases [138]. Since the real-time service traffic transmission does not implement congestion avoidance mechanisms, let us use a simple DropTail QM. This algorithm makes use of first come first served (FCFS) queuing discipline. Since the packet lengths within the real-time service traffic behavior aggregate are constant and the AF PHB class has a fixed minimum amount of bandwidth and buffer space, the AF PHB class queue can be modeled as a discrete-time queuing system G/D/1/K0 , where (K0 − 1) is the length of the buffer and X denotes the arrival process from the traffic conditioner. Since the traffic conditioning functions do not affect the time structure of the real-time service traffic behavior aggregate, the arrival process to the queue is the D-BMAP
58
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
B0
... Real-time service traffic behavoir aggregate
Server (K0-1) waiting positions
Fig. 5.4 AF PHB class queue.
process which models the real-time service traffic behavior aggregate. Finally, the queuing system can be represented as D-BMAP/D/1/K0 (Fig. 5.4). The time in the D-BMAP/D/1/K0 system is slotted and the slot duration time is given by ∆ = R8/B0 , where R is the length of a packet in bytes and B0 is the outgoing link share which is assigned to the AF PHB class as the minimum departure rate measured in Mbps. Consider the system at the end of an arbitrary time slot. For such a system the following equation holds [139]: S [Q] (n + 1) = max 0, S [Q] (n) − 1 + min W [A] (n + 1), K − S [Q] (n) , (5.3) where W [A] (n + 1) denotes the number of arrivals from the D-BMAP process which models the real-time service traffic behavior aggregate in the (n + 1)th slot and S [Q] (n) ∈ {1, 2, . . . , K0 } is the state of the system. The complete description of the system requires a two dimensional Markov chain {S [A] (n), S [Q] (n), n = 0, 1, . . . } embedded at the moments of packet departures, where S [A] ∈ {1, 2, . . . , M } is the state of the arrival process in time slot n. One can use the iteration algorithm [139] to compute the steady state probabilities that k packets are in the system and the arrival process is in the state j: xkj = lim P r{S [Q] (n) = k, S [A] (n) = j}, n→∞
(5.4)
where k ∈ {0, 1, .., K0 } and j ∈ {1, 2, .., M [A] }. The probability that v, v = 0, 1, . . . packets are lost in the slot in the D-BMAP/D/1/K0 system is given by the next expression [140]: fL (v) =
K X
x~k D(K − k + v)~e, v = 1, 2, .. ,
k=0
fL (v) = 1 −
∞ X K X
x~k D(K − k + i)~e, v = 0,
(5.5)
i=1 k=0
where M is the number of states of modulating Markov chain of the D-BMAP process, xki are the elements of stationary distribution, D(= k), k = 0, 1, . . . are the matrix elements denoting the transition of the modulating Markov chain from state to state with k arrivals.
59
AF PHB TRANSMISSION SERVICE
Analyzing (5.6), one can note that losses depend on the arrival process, the capacity of the system K0 , and the service time duration ∆ (outgoing link share B0 ). Thus, the next functional dependency takes place: P r{L = v} = f (K0 , B0 , D − BM AP ).
(5.6)
Therefore, using the given arrival process and varying B0 and K0 parameters, one can choose the appropriate regions of the AF class queue pair (K0 , B0 ). Those regions fulfill the requirement on losses. In addition to losses it is also necessary to fulfill the requirements on delay. However, there must be areas within the chosen queue parameters which allow a very high delay of certain packets. Such behavior corresponds to very high values of K0 and low minimum departure rates B0 . Thus, in order to fulfill the requirements on delay and obtain the appropriate values of (K0 , B0 ), one should take into account the delay experienced by packets within the D-BMAP/D/1/K0 queuing system. The probability that the packet will wait in the D-BMAP/D/1/K0 system during w time slots is given by [140]: fQ (w) =
∞ X m=w
fQ (w) =
w
x~0 D(m)
X 1 ~e + + m
[A] [A] ∞ M X XM X
m=K i=1 j=1
∞ X
x~k D(m)
k=1 m=w−k+1
xoi dij (m)
1 , w = K, m
1 ~e, w = 1, 2, .., K − 1, m (5.7)
where r is the size of an arriving batch. Using the (K0 , B0 ) parameters which fulfill requirements on packet losses (computed at the previous step), it is possible to choose the appropriate (K0 , B0 ) pairs. These pairs will not allow the delay of any packet during more than w time slots. Note that all of these computations can be performed before the beginning of SLA establishing procedure. Moreover, in the case when the QoSD uses some sort of dynamic resource allocation through bandwidth brokers (BB) or some other devices, traffic profiles and AF PHB class queue parameters can be computed by real-time service source software dynamically. The given parameters enable us to compute those SLA parameters which can be negotiated with users. These parameters should include less sophisticated ones compared to those mentioned before. They can include mean delay and mean losses and probably the quantiles of their PDFs and/or other user friendly parameters. These parameters can be estimated using (5.6) and (5.7). It is important to mention that there are some cases when one is given the real-time service traffic behavior aggregate and AF class queue parameters (K0 , B0 ) and one has to compute the QoS parameters. It may be the case when the real-time service should be implemented within a limited resource environment. Note that this problem is the inverse of the one which has been treated. It is proposed to solve that as follows. Using the expressions (4) and (5), the PDF of the number of lost packets and the probability of at least one packet loss are estimated. Then, applying (7), the PDF of the delay of packets is estimated.
60
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
The extension of the latter task states that the number of sources composing the real-time service traffic behavior aggregate which can be successfully delivered to their destinations given the AF class queue parameters (K0 , B0 ) and desired QoS guarantees should be evaluated. This task is similar to the latter one except for the arrival process dimension. 5.4.3 Per-source QoS degradation Since the real-time service is the microflow-oriented one, the behavior aggregate’s performance parameters do not provide a full description of the service. From this point of view, it is crucial to analyze per-source QoS degradation which can be caused by the AF PHB class queue. In order to evaluate per-source QoS degradation, the tagged real-time service traffic source which is multiplexed with (N − 1) background real-time service traffic sources is considered. These N sources together compose the real-time service traffic behavior aggregate. Both the tagged real-time service source and multiplexed traffic from (N − 1) sources are modeled by D-BMAP processes, and therefore it is possible to define the multiplexed traffic from N sources as a superposition of two D-BMAP processes. In this case the following D-BMAP[T AG] +D-BMAP[BACK] /D/1/K0 queuing system takes place. It is equivalent to the D-BMAP[SU P ] /D/1/K0 queuing system where the D-BMAP[SU P ] arrival process is the superposition of tagged and background arrival processes. Superscripts [T AG] [BACK] , and [SU P ] are used to distinguish between parameters which belong to three types of defined arrival processes. Note that from the practical point of view the number of states of a superposed process should not exceed several tens. The probability of the loss of v T , v = 0, 1, . . . packets in the slot by the tagged DBMAP process within the D-BMAP[T ] +D-BMAP[B] /D/1/K0 queuing system is given by the expression [140]: [T ] fL
v
[T ]
=
K X
[T ] i+v X
[T ] i+v X
[T ] M MB XX
1×
i=0 k[T ] =v [T ] k[B] =i−k[T ] +v [T ] l=1 h=1
×
[T ] M X
[T ]
dlj
×
[B] M X
k [T ]
[B]
dhj
k [B] ×
j=1
j=1 [T ]
(K−i−v
[T ]
Ckv[T ] C(R−k[T ] ) [T ]
k CR
)
, v [T ] = 1, 2, .. .
(5.8)
where Cab is the combination of a elements from b possible choices. P∞ [T ] Note that fL (0) = P r L[T ] = 0 = 1 − k=1 P r L[T ] = k [T ] = 1 − P∞ [T ] [T ] k . Both the first and second moment of the number of lost packets in the k=1 fL slot of the tagged process can be obtained immediately from their probability distribution function (5.8).
EF TRANSMISSION SERVICE
61
The probability that an arbitrary packet of a tagged source will suffer a w, w = 1, 2, . . . , K0 slots delay is given by [140]: [T ] fV (w[T ] )
=
[T ] w X
∞ X
∞ X
1×
m=1 k[T ] =m k[B] =w[T ] −m
× Ω(k [T ] , k [B] , 0) +
[T ] [T ] w −i+1 Xw X
i=1
m=1
k [T ]
∞ X
1 + + k [B] ∞ X
1×
k[T ] =m k[B] =w[T ] −m
1 , w[T ] = 1, 2, .., K − 1, + k [B] ∞ ∞ K X X X [T ] 1× fV (w[T ] ) = × Ω(k [T ] , k [B] , i)
k [T ]
m=1 k[T ] =m k[B] =K−m
× Ω(k [T ] , k [B] , 0)
1 , w[T ] = K. k [T ] + k [B]
(5.9)
It is important to note that in practice computation of (5.8) and (5.9) requires effective realization of computation of stationary distributions of two dimension Markov chains having their transmission probability matrices in the Hessenberg form [141, 142]. The presented analytical evaluation allows to compute parameters values for ingress AF node. It could be used by QoSD providers to assure certain QoS level for AF behavior aggregate.
5.5 EF TRANSMISSION SERVICE 5.5.1 Definition of the Service To satisfy the demands of the real-time traffic nature of the service under consideration, such a type of traffic should have the necessary amount of reserved bandwidth and/or some type of priority in service. EF PHB has a predefined minimum amount of forwarding resources, such as a well-defined minimum departure rate (bandwidth) and a predefined amount of buffer space at each node along the path of the EF behavior aggregate. Moreover, the process of EF traffic treatment is independent of other traffic treatment within the network nodes [36]. In accordance with the DiffServ specification [3], interior network nodes do not implement any traffic conditioning functions. All these functions must be performed by ingress network nodes. Let us consider the DiffServ ingress node which implements EF PHB in addition to the best effort service. In accordance with the EF PHB specification [36], such a node must perform two major functions: conditioning of the EF behavior aggregate via policing and/or shaping, and configuring of the node so that the EF behavior aggregate has a well-defined minimum departure rate. The functional blocks of the DiffServ ingress node
62
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
... Real-time service traffic behavoir aggregate
(b0-1) waiting positions
Server r0 Served Real-time service traffic behavoir aggregate
Fig. 5.5 The token bucket equivalent queuing system.
which serves the EF behavior aggregate consist of two blocks: a traffic conditioning block and an EF queue at the output port of the DiffServ ingress node. 5.5.2 Token Bucket Parameters Estimation Assume that the real-time service traffic behavior aggregate after entering the DiffServ ingress node is policed and shaped via a simple token bucket mechanism. In this case, since all packets within the real-time service traffic behavior aggregate have the same length, the bucket depth and token rate are measured in packets. The token bucket can be represented by an equivalent queuing system with a single server (Fig. 5.5). The service rate in such a queuing system equals the token rate r, and b is to the capacity of the system. The probability of loss in this queuing system corresponds to the probability of remarking. The remarked packets will not enter the EF queue, which is highly undesirable for real-time service traffic transmission over EF PHB. Consider the token bucket with parameters (r, b) which allows packet remarking at the DiffServ ingress node. In this case a certain part of the arriving packets is remarked to other DSCP. A number of packets which should be remarked depends on (r, b) parameters of the token bucket and on the arrival process as well. The DSCP to which these packets are remarked may represent AF PHB classes or best-effort service. In both cases, microflows can experience unpredictable delays, reordering of packets, and even losses which are caused by traffic conditioning functions. Therefore, the token bucket mechanism must not allow the packet losses (remarking). In this scenario it is important to assure that the traffic conforms SLAs established between QoSD providers. To provide loss-free transmission of real-time service traffic through the DiffServ domain, such (r0 , b0 ) should be provided that the probability of packet remarking is equal to zero. This probability can be derived from the analysis of simple queuing system DBMAP/D/1/b0 , where D-BMAP models the arrival process of the real-time service traffic behavior aggregate and b0 is the capacity of the system measured in packets. The probability of the loss of v, v = 1, 2, . . . packets during the slot in D-BMAP/D/1/b0 system is given by (5.6). Analyzing the expression (5.6), it is important to note that the packet losses depend on the arrival process and the capacity of the system bucket depth b0 . Note that the D-BMAP arrival process depends on the service time duration, which in turn depends on the outgoing link rate token rate r0 .
EF TRANSMISSION SERVICE
63
Thus, using the given arrival process and varying b0 and ∆ parameters, one can choose the appropriate regions of token bucket parameters (r0 , b0 ). Recall that these regions do not allow the loss of the packets of the real-time service traffic behavior aggregate. The services which are based on EF PHB must guarantee a bounded delay in addition to low losses. However, there must be areas within the chosen token bucket parameters which allow a very high delay of certain packets. Such behavior corresponds to high values of b0 , which in turn correspond to low token rates r0 and depend on their values. Therefore, in order to bound the delay of packets within the token bucket mechanism and obtain the appropriate (r0 , b0 ) values, one should take into account the delay experienced by packets within the D-BMAP/D/1/b0 queuing system. The probability that the packet will wait in the D-BMAP/D/1/b0 system during w, w = 1, 2, . . . , K time slots is given by (5.7). Using the (r0 , b0 ) parameters which do not allow packet losses computed at the previous step, it is possible to choose the appropriate (r0 , b0 ) pairs. These pairs should not allow the delay of any packet during more than wmax time slots within the token bucket mechanism and the delay bound within the token bucket is bounded by Vtbmax = w∆. All of these computations can be performed before the SLA establishing phase. Moreover, in a case where the QoSD uses some sort of dynamic resource allocation through bandwidth brokers or some other devices, traffic profiles can be computed by real-time service source software dynamically. 5.5.3 Token bucket parameters violation Consider a misbehaving real-time service source. Assume that there will be a time instant when the real-time service source will not follow the predefined traffic profile without permission from the network. In this case it is important to predict the QoS degradation which can be experienced by the real-time service traffic behavior aggregate within the token bucket mechanism. In addition, it is crucial to analyze the per-source QoS degradation which is caused by token bucket parameter violation. Let us consider these tasks consecutively. Consider the loss of packets of the real-time service traffic behavior aggregate caused by token bucket profile violation. Assume that the real-time service traffic behavior aggregate feeds the traffic conditioner at the DiffServ ingress node. Let (rold , bold ) be the token bucket profile which the real-time service source should follow. Therefore, the network uses token bucket parameters (rold , bold ) which do not allow the remarking of packets in accordance with established SLA. At some time instant the real-time service source no longer follows this traffic profile and packet remarking occurs within the token bucket mechanism. Since the packet remarking corresponds to packet losses within the equivalent D-BMAP/D/1/b0 queuing system, one should consider the inverse task to that considered previously. Let us solve it as follows: given the token bucket parameters (rold , bold ), estimate the PDF and mean of losses using the expression (5.8). In order to evaluate the per-source QoS degradation, consider the tagged real-time service traffic source which is multiplexed with (N − 1) background real-time service traffic sources. These N sources compose the real-time service traffic behavior aggregate. Let us model both the tagged real-time service source and multiplexed traffic from (N − 1) sources by D-BMAP processes. In this way it is possible to define the multiplexed traf-
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
Real-time service traffic
b0
DiffServ ingress node
...
64
G/D/1/K0 r0 tb(r0,b0)
Network interface
1
B0
... K0-1 2
Fig. 5.6 Functional blocks of DiffServ ingress node serving real-time service traffic behavior aggregate.
fic from N sources as a superposition of two D-BMAP processes. In this case the DBMAP[T ] +D-BMAP[B] /D/1/b0 queuing system takes place, which is equivalent to the DBMAP[SU P ] /D/1/b0 queuing system where the arrival process is the superposition of tagged and background arrival processes. Superscripts [T ] , [B] , and [SU P ] are used to distinguish between those parameters which belong to tagged, background, and superposed arrival processes. From the practical point of view, the number of states of a superposed process should not exceed several tens. The probability of the loss of v packets during the slot in the tagged D-BMAP process within the D-BMAP[T ] +D-BMAP[B] /D/1/b0 queuing system is given by (5.8). The maximum delay experienced by packets of the real-time service traffic behavior aggregate is equal to the token depth multiplied by the time interval to transmit one packet at tb the token rate Vmax = b∆. Using (5.8), it is possible to obtain the PDF of the waiting time of packets within the queuing system. Moreover, the probability that the arbitrary packet of a tagged source will suffer a w, w = 1, 2, . . . , K slots delay in the queuing system is given by (5.9). In practice, it is difficult to obtain the values which are given by expressions (5.8) and (5.9). However, it may be preferable to carry out such a type of preliminary evaluation. Finally, consider the case when the real-time service source violates the traffic profile (rold , bold ) but the queuing system does not suffer from packet losses. Note that it may take place when the initial token bucket parameters are higher than necessary. In this case the real-time service traffic behavior aggregate may still not suffer from packet losses while the bounded delay may increase substantially. The new upper delay bound of the real-time service traffic behavior aggregate can be derived using [42]. 5.5.4 EF queue parameters estimation After the policing and shaping procedures the real-time service traffic behavior aggregate enters the EF queue at the output port of the DiffServ ingress node. Since the EF traffic behavior aggregate must have a well-defined minimum departure rate that is independent of the dynamic state of the node, one can consider the EF queue within the DiffServ ingress node in isolation from the other queues within that node. The functional blocks of the DiffServ ingress node are shown in Fig. 5.6, where the blocks correspond to those depicted in Fig. 5.7.
EF TRANSMISSION SERVICE G/D/1/K Ain (t)=(rt+b) Real-time service traffic behavoir aggregate
65
Server
...
(K0 -1) waiting positions
B0
Aout (t) Served Real-time service traffic behavoir aggregate
Fig. 5.7 EF PHB queue model.
The services which are based on EF PHB should not allow packet losses and the endto-end delay should be bounded. Both requirements can be fulfilled only if the delay of the packets is bounded and the packet losses are controlled within each network node along the path of EF traffic. In [P2] it has been shown that the appropriate choice of token bucket parameters bounds the delay within the token bucket mechanism. Thus, in order to bound the delay within the DiffServ ingress node to a certain value Vmax , in addition to bounding the tb , it is necessary to bound the delay within delay within the token bucket mechanism Vmax q the EF PHB queue Vmax . As long as the lengths of all packets are equal and EF PHB has a minimum amount of guaranteed departure rate, the EF PHB queuing system can be represented by discrete-time system G/D/1/K, where K is the capacity of the system and G is the arrival process from the token bucket. To define the arrival process to the EF queue at the DiffServ ingress node, one should consider the general features of the output process from the D-BMAP/D/1/b0 queuing system. It is known that the output process from the D-BMAP/D/1/b0 queuing system constitutes a special type of D-BMAP – D-MAP process. Therefore, the buffer at the output port of the DiffServ ingress node can be modeled by the D-MAP/D/1/K0 queuing system. In order to define the parameters of the EF PHB queue (B0 , K0 ), let us employ the arrival curve approach [42]. Indeed, the real-time service traffic behavior aggregate’s arrival function Rin (t), t ≥ 0 after leaving the token bucket system is constrained to deterministic model (arrival curve) Ain (t) ∼ (r0 , b0 ), t ≥ 0 as follows: Rin (t) ≤ Ain (t) = (r0 t + b0 ), t ≥ 0, where r0 is the token rate and b0 is the bucket depth. Let S(t) be the service curve of the EF PHB queue. By definition, the service curve of the EF PHB queue is given by B0 t, r ≥ 0, where B0 is the minimum departure rate of the EF PHB at the DiffServ ingress node. Let Aout (t), t ≥ 0 be the arrival curve at the egress of the EF PHB queue (Fig. 5.7). Therefore, from the above definition it follows that Aout (t) = min(B0 t, Ain (t)), t ≥ 0. The upper bound of the number of packets within the EF PHB buffer is given by: Kmax = max(Ain (t) − B0 t). t≥0
(5.10)
It is important to note that the maximum delay is equal to the time needed to serve the maximum burst of packets stored in the EF PHB queue. Note that the maximum delay q Vmax that a packet can experience waiting in the EF PHB queue is bounded by: q Vmax =
Kmax . B0
(5.11)
66
NEXT-GENERATION TELECOMMUNICATIONS NETWORKS: CORE NETWORK PART
Using (5.10) it is possible to calculate the region of minimum departure rate which should be assigned to the EF PHB within the DiffServ ingress node and guarantees no losses. Note that certain values in this region allow a very high maximum queuing delay of packets. Therefore, one should choose those which guarantee a delay of packets no more than in accordance with (5.11).
5.6 REAL-TIME SERVICE TRAFFIC TREATMENT WITHIN INTERIOR NODES There is an opportunity to decrease requirements on minimum departure rates within the DiffServ interior nodes. Indeed, the real-time service traffic behavior aggregate become constrained after leaving the AF PHB class queue. One can use the results obtained in [42] to compute the necessary departure rates and corresponding buffer spaces within DiffServ interior nodes given the end-to-end delay bound it is necessary to provide. Obviously, the transmission service will become more sophisticated. However, such a case can be seen as a worthwhile option of the proposed service. Consider the transmission path of the real-time service traffic behavior aggregate through the DiffServ domain. After leaving the real-time service source, the real-time service traffic behavior aggregate enters the DiffServ ingress node and then passes through a number (say N ) of DiffServ interior nodes. After that it leaves the DiffServ domain. The DiffServ egress node is excluded from the path because its functions can be identical to those performed by interior nodes. Considering the real-time service traffic behavior aggregate transmission path over the DiffServ domain, one should note that packet delays and packet losses can occur within each network node. In order to compute that part of the traffic conditioning agreement (TCA) which consists of QoS parameters expectations and which should be provided to the user as a valuable part of the SLA the configuration parameters of all network nodes should be estimated. Each DiffServ node which serves the real-time service traffic behavior aggregate is characterized by two parameters (Ki , Bi ), i = 1, 2, . . . , N , while the DiffServ ingress node is described by two pairs (r0 , b0 ) and (K0 , B0 ). These configuration parameters define the QoS parameters which are experienced by the real-time service traffic behavior aggregate. Note that in the case when delay and loss of packets can occur in every network node, it is almost impossible to predict the performance parameters of the real-time service. Two solutions of this problem are proposed. In accordance with the first solution, let us assign the minimum departure rate within the DiffServ interior nodes as high as required by the traffic which leaves the DiffServ ingress node. It is possible to compute these rates because, after leaving the AF, the queue real-time service traffic behavior aggregate is constrained to certain queue parameters (K0 , B0 ) and satisfies the deterministic model of the real-time service traffic behavior aggregate. Thus, within the DiffServ interior nodes, let us simply assign Bi ≥ B0 , i = 1, 2, . . . , N . By doing so, it is possible to bound the delay to zero and keep zero losses within every DiffServ interior node. Therefore, QoS degradation can occur within the DiffServ ingress node only. These QoS degradations depend on DiffServ ingress node parameters (K0 , B0 ) and implicitly determine the quality of decoded video.
SUMMARY
67
In the context of the second approach it is important to note that there is an opportunity to decrease the requirements on minimum departure rates within the DiffServ interior nodes. Indeed, the real-time service traffic behavior aggregate becomes constrained after leaving the AF PHB class queue. Given the end-to-end delay bound, one can use the results obtained in [42] to compute the necessary departure rates and corresponding buffer spaces within the DiffServ interior nodes. Obviously, the transmission service will become more sophisticated since one should fulfill the requirements on delay bounds within each DiffServ interior node.
5.7 SUMMARY In order to define the transmission services, a lot of peculiarities of the real-time service were taken into account. These are real-time service configuration, real-time service source configuration, traffic generation, and modeling issues. After that, implementation-ready transmission services within the DiffServ IP networks were developed. It is proposed to use the special traffic profile parameters which adequately fit real-time service traffic requirements. Using these parameters and the AF PHB class or EF PHB, it is shown how to construct well-defined transmission services that are suitable for real-time service traffic delivery. Using the QoS guarantees which should be provided the real-time service traffic behavior aggregate, one can evaluate the required capacity not only within the DiffServ ingress nodes but also within the interior nodes. Moreover, the designed transmission services are designed in such a way that they are characterized by DiffServ ingress node parameters and bound delay within the network nodes along the path of the behavior aggregate and allow us to predict the QoS degradation which is experienced by both the behavior aggregate and single microflows. It is also shown how to compute these valuable parameters. It is allowed to use more than one real-time service source within a designed transmission service. The only requirement is that the additional real-time service traffic behavior aggregates should satisfy the deterministic model (r, b). In this case one can use those results provided in [42] to compute the minimum departure rate Bi , ∀i ∈ {1, 2, . . . , N } within the DiffServ interior nodes. It should also be noted that the presented analysis gives worst-case results because some additional bandwidth may be allocated to the AF PHB class in accordance with fair sharing algorithm which can be implemented by DiffServ nodes along the path of the real-time service traffic behavior aggregate.
Chapter
6
Conclusions
T
HE fundamental concept of different telecommunications networks’ convergence in Next-Generation Network (NGN) is being developed by the worldwide communications research community and also by vendors and operators/providers. This concept is rather complicated and its implementation cannot be delivered just from scratch. NGN being developed as successor of Second- and Third-generation (2G and 3G) systems, but obviously it is a much more sophisticated and flexible system. Introducing of multimedia services within NGN framework pose new requirements to network system design, which includes, among others, the Quality of Service (QoS) assurance problem. This problem has stayed hot for more than 10 years already and it is one of the most evasive and confusing topics in networking today especially in the design of NGN. Indeed, it is highly anticipated that the most important, value-added and revenueexpected new services for NGN will be Internet access and IP multimedia applications. As shown in a number of recent R&D projects world-wide, on the basis of a plain fixed Internet access it is possible to implement a set of brand-new applications for which certain QoS requirements need to be provided. Undoubtedly, QoS for Internet applications will be heavily demanded by the NGN end-users. Internet QoS going mobile is why, with respect to the existing experience on the implementation of QoS within fixed networks, the design, development and implementation of QoS-capable mobile networks is of paramount importance. This thesis consists of two interconnected research parts. The first research part of the thesis has been concentrated on building of the cross-layer framework based on the NGN system design. Three important components of the QoS assurance problem in NGN systems are defined: teletraffic, mobility, and wireless channel. It was proved that the proper NGN system design should account all of these components, otherwise the desired QoS levels for different services can not be achieved. The original contribution in this part also includes a new approach for QoS evaluation of services – a cross-layer black-box framework was developed and presented in detail. This frameworks allows performing theoretical and practical qualitative and quantitative evaluation of QoS expectations of user’s applications running over complicated NGN systems. The major advantage of this frame69
70
CONCLUSIONS
work is avoidance of complexity of cross-layer nature of teletraffic, mobility, and wireless channel models. The second research part of the thesis has been devoted to practical applying of theoretical results obtained in the first part. At first, cross-layer framework was implemented as a testbed for practical QoS evaluation for existing multimedia services running over existing wireless technologies. The testbed is fed with a set of different parameters like traffic patterns and mobility behaviors. At the output of the testbed (or cross-layer framework), IP layer performance degradation via perceived QoS was evaluated. Several important conclusions were drawn and future research directions were defined. Moreover, based on the fact that the QoS assurance problem comprises access network (AN) and core network (CN) QoS assurance, an elegant analytical method of QoS assurance for real-time service operating over Differentiated Services-based (DiffServ) NGN CN was proposed. The both available possibilities, assured forwarding (AF) and expedited forwarding (EF) perhop behavior (PHB), of real-time traffic delivery over DiffServ domain were considered. Therefore, two transmission services were developed, real-time service over AF PHB and real-time service over EF PHB. The both transmission services are based on estimation of queue parameters and accounting of token bucket parameters. The obtained formulas allow computing values for the transmission services and therefore help tuning network nodes’ parameters to assure certain QoS levels.
Chapter
7
Summary of Publications
T
HE publications can be divided into three categories. Publications of the first category, [P1] – [P3], are devoted to theoretical study of real-time multimedia handling peculiarities related to a fixed part of the Next-Generation Networks (NGN) system built on a base of Differentiated Services architecture. Publications of the second category, [P5] and [P6], deal with practical performance evaluation of real-time multimedia traffic delivery over current evolution of wireless access networks such as WLAN IEEE 802.11b and GPRS. The last category of publications, [P4] , [P7], and [P8], is concentrated on teletraffic requirements, performance evaluation approaches, and quality of service (QoS) assurance in NGN.
7.1 OVERVIEW OF PUBLICATIONS [P1] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Analytical estimation of EF PHB service parameters for aggregated MPEG traffic,” in Proc. the 16th Nordic Teletraffic Seminar, NTS’02, Espoo, Finland, August 2002, pp. 279 – 290. Description. This publication concentrated on analytical construction of Video-onDemand (VoD) transmission service based on expedited forwarding per-hop behavior (EF PHB) of Differentiated Services (DiffServ) architecture. To construct such a transmission service, several conditioning functions based on traffic profiles must be implemented and properly parametrized. In this paper the method of EF PHB traffic profiles built on token bucket mechanism calculation is proposed. Moreover, an analytical approach for worstcase delay evaluation of EF traffic treatment within the DiffServ ingress node is developed. Finally, a VoD end-to-end theoretical model was elaborated. [P2] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”An Analytical Evaluation of VoD Traffic Treatment within the EF-enabled DiffServ Ingress and Interior Nodes,” in Proc. IEEE 2003 International Conference on Telecommunications, ICT’03, Papeete, Tahiti, French Polynesia, February 2003. Description. This paper is an extended version of [P1]. The VoD end-to-end theoretical 71
72
SUMMARY OF PUBLICATIONS
model considered in [P1] has been modeled analytically under certain circumstances. Analytical expressions of behavior aggregate and per-source QoS degradation resulting from token bucket parameter violation are obtained. [P3] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”A Top-Down Approach to VoD Traffic Transmission Over DiffServ Domain Using the AF PHB Class,” in Proc. IEEE International Conference on Communications, ICC’03, Anchorage, Alaska, USA, May 2003. Description. Within this paper the framework from [P1] was considered, but another PHB type has been used – assured forwarding (AF PHB), based on which an analytical transmission service has been developed. On the basis of this analytical model, it is possible to parametrize DiffServ’s ingress and interior nodes using QoS parameters normally defined for VoD service by service level specification (SLS) – transmission service is fully characterized by the parameters of the DiffServ ingress node; it allows losses and delays to be bounded along the path of the behavior aggregate. [P4] Y. Koucheryavy and D. Moltchanov, ”Notes on Quality of Service and Performance Evaluation of 4G All-IP Networks,” in Proc. ANWIRE’s 1st International Workshop on Wireless, Mobile and Always Best Connected, University of Strathclyde, Glasgow, UK, April, 2003, CD ISBN 0-9545660-0-9. Description. The paper proposes a framework of NGN network design aspects mainly focusing on QoS and teletraffic issues. A breakthrough proposition on integrated traffic models has been made – an adequate load model in wireless environment should take into account both mobility of users and teletraffic characteristics of upper layer applications. [P5] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Performance evaluation of live video streaming service in 802.11b WLAN environment under different load conditions,” in Proc. ACM International Workshop on Multimedia Interactive Protocols and Systems, MIPS’03, LNCS 2899, Napoli, Italy, Nov. 2003, pp. 30 – 41. Description. This paper brings a practical flavor to the thesis. It is devoted to live video streaming service evaluation over IEEE 802.11b wireless LAN (WLAN), which could be used as a part of the layered infrastructure of NGN systems to provide coverage in highly populated areas. Experiments have been performed under different signal-to-noise ratio (SNR) classes and different competing TCP and UDP traffic loads. SNR classes have been defined based on perceived QoS and proper granularity. To obtain results, a testbed comprising several PCs with dedicated software has been designed. [P6] Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Impact of Mobility on Entertainment Services’ Performance in Heterogeneous Wireless Environment,” in Proc. of Australian Telecommunications, Networks and Applications Conference, ATNAC’03, Melbourne, Australia, Dec. 2003, CD ISBN 0-646-42229-4. Description. Within this paper another practical investigation of up-to-date wireless networks performance such as WLAN 802.11b and GPRS against real-time multimedia services and different mobility patterns has been carried out. Special attention has been paid to stability of wireless access networks. Numerous parameters have been obtained based on TCP connections capturing. Hence, round trip time (RTT), throughput and Stevens graphs for different cases over different access networks are presented in the Appendix.
AUTHOR’S CONTRIBUTION TO THE PUBLICATIONS
73
[P7] Y. Koucheryavy, D. Moltchanov, J. Harju, and G. Giambene ”Cross-Layer Black-Box Approach to Performance Evaluation of Next Generation Mobile Networks,” in Proc. IEEE International Conference on Next Generation Teletraffic and Wireless/Wired Advanced Networking, NEW2AN’04, St.-Petersburg, Russia, Feb. 2004, pp. 266 – 272, ISBN 952-151132-X. Description. This paper is a sequel of [P4]. Based on the proposed earlier framework a unified cross-layer black-box approach for performance evaluation of NGN networks is elaborated. [P8] Y. Koucheryavy, D. Moltchanov, G. Giambene, and J. Harju, ”Teletraffic Requirements and System Aspects for Future Mobile Communications,” in Proc. of the 8th IASTED International Conference on Internet & Multimedia Systems & Applications, IMSA’04, Kauai, Hawaii, USA, August 2004. Description. The paper primarily deals with the QoS assurance problem in NGN systems. It delivers the state-of-the-art knowledge in telecommunication networks design areas. System design, advanced resource management schemes, and vertical optimization of employed protocol stack as well as horizontal integration of different systems are the challenges considered throughout the paper.
7.2 AUTHOR’S CONTRIBUTION TO THE PUBLICATIONS The Author is the primary author of all original papers [P1] – [P8]. All the research devoted to the thesis topic has been carried out by the Author in collaboration with Dmitri Moltchanov, naturally supported and guided by Prof. J. Harju. Although, quite many ideas were born in mutual discussions, the Author contributed to the papers significantly. The idea of possible decomposition of end-to-end QoS assurance for NGN systems as well as practical and theoretical approaches employed in [P1] – [P3], [P5], and [P6] originally were proposed by the Author. The Author participated heavily in practical testbed design and services evaluation, the results of which form the basis for [P5] and [P6]. Theoretical models of AF and EF PHB services in [P1] – [P3] were developed by the Author. The framework on teletraffic requirements, performance evaluation approaches, and QoS assurance in NGN systems [P4], [P7], and [P8] were proposed together with co-authors and developed by the Author.
References
[1] Koucheryavy, Y.; et. al; COST 290 Memorandum of Understanding, Brussels, ESF, March 2003. [2] Braden, R.; Clark, D.; Shenker, S.; Integrated Services in the Internet Architecture: an Overview, IETF, RFC 1633, June 1994. [3] Blake, S.; Black, D.; Davies, E.; Wang, Z.; Weiss, W.; An Architecture for Differentiated Services, IETF, RFC 2475, 1998. [4] Roberts, J.; Traffic theory and the Internet, Communications Magazine, IEEE, Volume: 39, Issue : 1, Jan. 2001, pp. 94 – 99. [5] Koucheryavy, Y.; Krendzel, A.; Analysis of Web traffic and users behavior modeling during busy hour, 6-th EUNICE Open European Summer School, the Netherlands, pp. 9 – 12, September 2000. [6] Guangwei Bai; Traffic modeling for multimedia communication systems, Communication Technology Proceedings, 2000. WCC – ICCT 2000. International Conference on,Volume: 1,21-25 Aug. 2000, pp. 50–53. [7] Pandit, K.; Schmitt J.; Steinmetz, R.; Aggregation of Heterogeneous Real-Time Flows with Statistical Guarantees, Procceedings of 2002 International Symposium on Performance Evaluation of Computer and Telecommunication Systems SPECTS’02, San Diego, USA, pp. 57 – 64. [8] Siripongwutikorn, P.; Banerjee, S.; Per-flow delay performance in traffic aggregates, Proc. GLOBECOM’02, IEEE ,Volume: 3, Nov. 2002, pp. 2634 – 2638. [9] Tsuru, M.; Takine, T.; Oie, Y.; Inferring traffic flow characteristics from aggregatedflow measurement, Proc. SAINT’02, Feb. 2002, pp. 256 – 261. [10] Basgeet, D.R.; Irvine, J.; Munro, A.; Dugenie, P.; Kaleshi, D.; Lazaro, O.; Impact of mobility on aggregate traffic in mobile multimedia system, Wireless Personal Multimedia Communications, 2002. The 5th International Symposium on,Volume: 2 ,2730 Oct. 2002, pp. 333 – 337. 75
76
REFERENCES
[11] Delcoigne, F.; Proutire, A.; Rgni, G.; Roberts, J.; Modelling integration of streaming and data traffic, 15th ITC Specialist Seminar on IP traffic, Wrzburg, July 2002. [12] Hui-Kai Su; Cheng-Shong Wu; Kim-Joan Chen; Session classification for traffic aggregation, IEEE International Conference on Communications ICC’04,Volume: 2 ,June 2004, pp. 1243 – 1247. [13] Smirnov, M.; Biersack, E.; Blondia, C.; Bonaventure, O.; Casals, O.; Karlsson, G.; Pavlou, G.; Quoitin, B.; Roberts, J.; Stavrakakis, I.; Stiller, B.; Trimintzios, P.; Van Mieghem, P. (Eds.); Quality of Future Internet Services, COST Action 263 Final Report. Lecture Notes in Computer Science 2856 Springer, 2003. [14] http://www.cs.columbia.edu/ hgs/internet/traffic.html [15] Fomenkov M.; Keys, K.; Moore, D.; claffy, kc.; Longitudinal study of Internet traffic from 1998-2003, CAIDA Tech Report, 2004. [16] Puqi Perry Tang; Tai, T.-Y.C.; Network traffic characterization using token bucket model, Proc. INFOCOM’99, Volume: 1 ,March 1999, pp. 51 – 62. [17] Brownlee, N.; claffy, kc.; Internet stream size distributions, ACM SIGMETRICS Performance Evaluation Review, June 2002, Volume: 30, Issue 1, pp. 282 – 283. [18] Dunaytsev, R.; Krendzel, A.; Koucheryavy, Y.; Harju, J.; Estimation of Web traffic generated by users in home networks, 8th IASTED Internet and Multimedia Systems and Applications, IMSA 2004, August 16 – 18, Kauai, Hawaii, USA. [19] Kamei, S.; Kimura, T.; Practicable network design for handling growth in the volume of peer-to-peer traffic, Proc. IEEE PACRIM’03,Volume: 2,Aug. 2003. [20] Caberera, J.B.D.; Ravichandran, B.; Mehra, R.K.; Statistical traffic modeling for network intrusion detection, Modeling, Analysis and Simulation of Computer and Telecommunication Systems, 2000. Proceedings. 8th International Symposium on,Sept. 2000. [21] Rappaport, T.; Wireless Communications: principles and practice, Communications engineering and emerging technologies, Prentice Hall, 2nd edition, 2002. [22] Kaaranen, H.; Ahtiainen, A.; Laitinen, L.; Naghian, S.; Niemi, V.; UMTS networks. Architecture, mobility and services, John Wiley & Sons, 2001. [23] ETSI TS 123 060 V3.5.0 (2000-10); Technical specification digital cellular telecommunications system (Phase 2+) (GSM); Universal Mobile Telecommunications System (UMTS); General Packet Radio Service (GPRS); Service description; Stage 2, 3GPP TS 23.060 version 3.5.0, Release 1999. [24] 3GPP TS 23.002 V5.5.0 (2002-01); 3rd Generation Partnership Project (3GPP); technical specification group services systems aspects; network architecture (Release 5). [25] Chalmtac, I.; Conti, M.; Liu, Jenifer J.-N.; Mobile ad hoc networking: imperatives and challenges, Elsevier Ad Hoc Networks Journal, vol. 1, 2003, pp. 13 – 64.
REFERENCES
77
[26] O’Droma, M.; Ganchev, I.; Morabito, G.; Narcisi, R.; Passas, N.; Paskalis, S.; Friderikos, V.; Jahan, A.; Tsontsis, E., Bader, C.; Rotou, J.; Chaouchi, H.; Always Best Connected enabled 4G wireless world, IST Mobile and Wireless Communications Summit, 2003. [27] Moustafa, M.; Habib, I.; Naghshineh, M.; Guizani, M.; QoS-enabled broadband mobile access to wireline networks, Communication Magazine, IEEE, 40(4):50 – 56, April 2002. [28] 3GPP TR 23.107 (2003); Quality of service (QoS) concept and architecture. [29] 3GPP TS 23.002 V5.5.0 (2002-01); 3rd Generation Partnership Project (3GPP); technical specification group services systems aspects; network architecture (Release 5). [30] Mahadevan, I.; Sivalingam, K.M.; Quality of Service architectures for wireless networks: IntServ and DiffServ models, Parallel Architectures, Algorithms, and Networks, 1999. (I-SPAN ’99) Proceedings. Fourth InternationalSymposium on,23-25 June 1999, pp. 420 – 425. [31] Moon, B.; Aghvami, H.; RSVP extensions for real-time services in wireless mobile networks, Communications Magazine, IEEE ,Volume: 39 ,Issue: 12 ,Dec. 2001, pp. 52 – 59. [32] Yu Cheng; Weihua Zhuang; DiffServ resource allocation for fast handoff in wireless mobile Internet, Communications Magazine, IEEE ,Volume: 40 ,Issue: 5 ,May 2002, pp. 130 – 136. [33] Maniatis, S.I.; Nikolouzou, E.G.; Venieris, I.S.; QoS issues in the converged 3G wireless and wired networks, Communications Magazine, IEEE , Volume: 40 , Issue: 8 , Aug. 2002, pp. 44 – 53. [34] Grilo, A.; Nunes, M.; Sergio, G.; Ciulli, N.; Integration of IP mobility and QoS for heterogeneous wireless access in MOICANE, Personal, Indoor and Mobile Radio Communications, 2003. PIMRC 2003. 14th IEEE Proceedings on ,Volume: 1 ,7-10 Sept. 2003, pp. 470 – 475. [35] Weiss, W.; Heinanen, J.; Baker, F.; Wroclawski J.; Assured Forwarding PHB group, RFC 2597, IETF, 1999. [36] Davie, B.; Charny, A.; Bennett, J.; Benson, K.; Le Boudec, J.; Courtney, W.; Davari, S.; Firoiu, V.; Stiliadis, D.; An Expedited Forwarding PHB (Per- Hop Behavior), RFC 3246, IETF, 2002. [37] 3GPP TR 25.881; Technical specification group radio access network; improvement of RRM across RNS and RNS/BSS (release 5). [38] Fang, Y.; Chlamtac, I.; Teletraffic analysis and mobility modeling of PCS networks, IEEE Transactions on Communication, 47(7):1062 – 1072, July 1999.
78
REFERENCES
[39] Fitzpatrick, P.; Lee, C.; Warfield, B.; Teletraffic performance of mobile radio networks with hierarchial cells and overflow, IEEE JSAC, 15(8):1549 – 1557, October 1997. [40] Orlik, P.; Rappaport, S.; A model for teletraffic performance and channel holding time charachterization in wireless cellular communication with general session and dwell time distributions, IEEE JSAC, 16(5):788 – 803, June 1998. [41] Orlik, P.; Rappaport, S.; Traffic performance and mobility modeling of cellular communications with mixed platforms and highly variable mobilities, Proc. of the IEEE, 86(7):1464 – 1479, July 1998. [42] Cruz, R.; Quality of service guarantees in virtual circuit switched networks, IEEE Journal on Selected Areas of Communications, 13:1048 – 1056, August 1995. [43] Heindl, A.; Decomposition of general queuing networks with MMPP inputs and customers losses, Performance Evaluation, 51:117 – 136, 2003. [44] Heindl, A.; Telek, M.; Output models of MAP/PH/1(K) queues for an efficient network decomposition, Performance Evaluation, 49:321 – 339, 2002. [45] Li, S.-Q.; Hwang, C.-L.; On the convergence of traffic measurement and queuing analysis: a statistical-matching and queuing (SMAQ) tool, IEEE/ACM Transactions on Networking, 5:95 – 110, December 1997. [46] Kim, Y.; Li, S.-Q.; Timescale of interest in traffic measurement for link bandwidth allocation design, In Proc. IEEE Infocom, pages 738 – 748, San Francisco, USA, March 1996. [47] Feng, H.-W.; Chang, J.-F.; Connection-wise end-toend performance analysis of queuing networks with MMPP inputs, Performance Evaluation, 43:39 – 62, 2001. [48] Feng, H.-W.; Chang, J.-F.; Connection-wise end-to-end performance analysis of queuing networks with MMPP inputs, IEICI Trans. on Comm., E83-B(3):659 – 671, March 2000. [49] Kim, Y.-Y.; Li, S.-Q.; Capturing important statistics of a feding/shadowing channel for network performance analysis, IEEE JSAC., 17(5):888 – 901, May 1999. [50] Broch, J.; Maltz, D.A.; Johnson, D.B.; Hu, Y.-C.; Jetcheva, J.; A performance comparison of multi-hop wireless ad hoc network routing protocols, in Proceedings of IEEE/ACM Mobicom’98, pp. 85 – 97, Oct. 1998. [51] Zonoozi, M.M.; Dassanayake, D.; User mobility modeling and characterization of mobility patterns, IEEE Journals on Selected Areas in Communication,15(7):1239 – 1252, 1997. [52] Hong, X.; Gerla, M.; Pei, G.; Chiang, C.-C.; A group mobility model for ad hoc wireless networks, Proc. ACM/IEEE MSWiM’99, pp. 53 – 60, Aug. 1999. [53] Yoon, J.; Liu, M.; Noble, B.; Random waypoint considered harmful, INFOCOM 2003, IEEE ,Volume: 2 ,30 March-3 April 2003, pp. 1312 – 1321.
REFERENCES
79
[54] Lam, D.; Cox, D.M.; Wilson, J.; Teletraffic modeling for personal communication services, Communications Magazine, IEEE, 35(2): 79 – 87, 1997. [55] Scourias, J.; Kunz, T.; Activity-based mobility modeling: realistic evaluation of location management schemes for cellular networks, Wireless Communications and Mobile Computing (WCMC), pp. 296 – 300, Sep.1999. [56] Rocha, M.; Mateus, G.; da Silva, S.; QoS and simulation models in mobile communication networks, Proc. ACM International Workshop on Modeling Analysis and Simulation of Wireless and Mobile Systems MSWiM 2000, pp. 119 – 122, Aug. 2000. [57] Bettstetter, C.; Mobility modeling in wireless networks: categorization, smooth movement, and border effects, ACM SIGMOBILE Mobile Computing and Communications Review, 5(3): 55 – 66, 2001. [58] Camp, T.; Boleng J.; Davies, V.; A survey of mobility models for ad hoc network research, Wireless Communications and Mobile Computing (WCMC): Special Issue on Mobile Ad Hoc Networking: Research, Trends and Applications, 2(5): 483 – 502, 2002. [59] Akyildiz, I.F.; Yi-Bing Lin; Wei-Ru Lai; Rong-Jaye Chen; A new random walk model for PCS networks, Selected Areas in Communications, IEEE Journal on ,Volume: 18 ,Issue: 7 ,July 2000, pp. 1254 – 1260 [60] Shukla, D.; Mobility models in ad hoc networks, Master’s Thesis, KReSIT-ITT, Bombay, November 2001. [61] Madhavapeddy, S.; Basu K.; Roberts, A.; Adaptive paging algorithms for cellular systems, WINLAB April (1995), pp. 976 – 980. [62] Murakami, K.; Haase, O.; Shin, J.; LaPorta, T.F.; Mobility Management Alternatives for Migration to Mobile Internet Session-Based Services, Selected Areas in Communications, IEEE Journal on ,Volume: 22 ,Issue: 5 ,June 2004, pp. 818 – 833. [63] Pagtzis, T.; Kirstein, P.; Hailes, S.; Afifi, H.; Proactive seamless mobility management for future IP radio access networks, Elsevier Computer Communications, Volume 26, Issue 17, 1 November 2003, pp. 1975 – 1989. [64] Tseng, Y.-C.; Chen, L.-W.; Yang, M.-H.; Wu, J.J.; A Stop-or-Move Mobility model for PCS networks and its location-tracking strategies, Computer Communications, Volume 26, Issue 12, 21 July 2003, pp.1288 – 1301. [65] Klemm, A.; Lindemann, M.; Lohmann, M.; Traffic modeling of IP networks using the batch markovian arrival process, Technical report, Department of Computer Science, University of Dortmund, 2001. [66] Klemm, A.; Lindemann, M.; Lohmann, M.; Traffic modeling and characterization for UMTS networks, Technical report, Department of Computer Science, University of Dortmund, 2002.
80
REFERENCES
[67] Basgeet, D.R.; Irvine, J.; Munro, A.; Barton, M.H.; Importance of accurate mobility modeling in teletraffic analysis of the mobile environment, Vehicular Technology Conference, 2003. VTC 2003-Spring, Volume: 3 ,22-25 April 2003, pp. 1836 – 1840. [68] Hollick, M.; Krop, T.; Schmitt, J.; Huth, H.-P.; Steinmetz, R.; Modeling mobility and workload for wireless metropolitan area networks, Computer Communications, Volume 27, Issue 8, May 2004, pp. 751 – 761. [69] Lee, K.-D.; Kim, S.; Modeling variable user mobility with stochastic correlation concept, Computer Networks, Volume 38, Issue 5, 5 April 2002, pp. 603 – 612. [70] Antunes, N.; Pacheco, A.; Rocha, R.; An integrated traffic model for multimedia wireless networks. Computer Networks, 38:2541, 2002. [71] Moltchanov, D.; Koucheryavy, Y.; Harju, J.; Multi-Application Packetized Traffic Model of Single Source for Next Generation IP-based Mobile Systems”, NEW2AN 2004, St.Petersburg, Russia, February 2004, pp.273 – 281. [72] Tsai, S.; Markov characterization of the HF channel, IEEE Trans. Commun. Technol., vol. 17, pp. 24 – 32, Feb. 1969. [73] Gans, M. J.; A power-spectral theory of propagation in the mobile radio environment, IEEE Trans. Veh. Technol., vol. 21, pp. 2738, Feb. 1972. [74] Steffan, H.; Adaptive generative radio channel models, Proc. 5th IEEE Int. Symp. Personal, Indoor and Mobile Radio Communications, vol.1, pp. 268 – 273, 1994. [75] Bischl, H.; Lutz, E.; Packet error rate in the non-interleaved Rayleigh channel, IEEE Trans. Commun., vol.43, pp. 1375 – 1382, Feb./Mar./Apr. 1995. [76] Zorzi, M.; Rao, R. R.; Milstein, L. B.; Error statistics in data transmission over fading channels, IEEE Trans. Commun., vol.46, pp. 1468 – 1477, Nov. 1998. [77] Hassan, M.; Krunz, M.M.; Matta, I.; Markov-based channel characterization for tractable performance analysis in wireless packet networks, IEEE Trans. Wireless Comm., 3(3): 821 – 831, May 2004. [78] Vucetic, B.; An adaptive coding scheme for time-varying channels, IEEE Trans. Commun., vol. 39, pp. 653 – 663, May 1991. [79] Rice, M.; Wicker, S. B.; Adaptive error control for slowly-varying channels, IEEE Trans.Commun., vol. 42, pp. 917 – 925, 1994. [80] Wang, H. S.; Moayeri, N.; Finite-state Markov channel – A useful model for radio communication channels, IEEE Trans. Veh. Technol., vol. 44, pp. 163 – 171, Nov. 1995. [81] Zorzi, M.; Rao, R. R.; Milstein, L. B.; On the accuracy of a firstorder Markov model for data transmission on fading channels, in Proc. IEEE ICUPC95, pp. 211 – 215. [82] Zorzi, M.; Rao, R. R.; On the statistics of block errors in bursty channels, IEEE Trans. Commun., vol. 45, pp. 660 – 667, June 1997.
REFERENCES
81
[83] Zhang, Q.; Kassam, S. A.; Finite-state Markov model for Rayleigh fading channels, IEEE Trans. Commun., vol. 47, pp. 1688 – 1692, Nov. 1999. [84] Guan, Y. L.; Turner, L. F.; Generalized FSMC model for radio channels with correlated fading, IEEE Proc. Commun., vol. 146, pp. 133 – 137, Apr. 1999. [85] Huo, D.; Simulating slow fading by means of one dimensional stochastical process, in Proc. IEEE VTC96, pp. 620 – 622. [86] Turin, W.; Jana, R.; Martin, C.; Winters, J.; Modeling wireless channel fading, Proc. Vehicular Technology Conference, VTC’01, vol.3, Oct. 2001, pp. 1740 – 1744. [87] Turin, W.; Zorzi, M.; Performance analysis of delay-constrained communications over slow Rayleigh fading channels, Wireless Communications, IEEE Transactions on ,Volume: 1 ,Issue: 4 ,Oct. 2002, pp. 801 – 807. [88] Zorzi, M.; Mazzini, G.; Tralli, V.; Giovanardi, A.; Some results on the error statistics over a WCDMA air interface, Proc. MMT2000 Florida, Dec. 2000. [89] Sklar, B.; Digital Communications Fundamentals & Applications, Englewood Cliffs, NJ: Prentice-Hall, 1988. [90] van Nobelen, R.; Seshadri, N.; Whitehead, J.; Timiri, S.; An adaptive radio link protocol with enhanced data rates for GSM evolution, Personal Communications, IEEE,Volume: 6 ,Issue: 1 ,Feb. 1999, 54 – 64. [91] Lin, H.; Das, S.K.; Performance study of TCP/RLP/MAC in next generation CDMA systems, Personal, Indoor and Mobile Radio Communications, 2003. PIMRC 2003. 14th IEEE Proceedings on ,Volume: 1 ,7-10 Sept. 2003, pp. 648 – 652. [92] Lin, H.; Das, S.K.; ARLP: An Adaptive Link Layer Protocol to Improve TCP Performance over Wireless Fading Channels,” to appear in Wiley Journal on Wireless Communication and Mobile Computing. [93] Zadeh, L.; On the Identification Problem, Circuit Theory, IRE Transactions on ,Volume: 3 ,Issue: 4 ,Dec 1956, pp. 277 – 281. [94] Takagi, H.; Suzuki, N.; Koda, T.; Kojima, Y.; Neural networks designed on approximate reasoning architecture and their applications, Neural Networks, IEEE Transactions on ,Volume: 3 ,Issue: 5 ,Sept. 1992, pp. 752 – 760. [95] Kuang, T.; Williamson, C.; RealMedia Streaming Performance on an IEEE 802.11b Wireless LAN, Proc. IASTED Wireless and Optical Communications (WOC) Conference, Banff, AB, Canada, pp. 306 – 311, July 2002. [96] Gurtov, A.; Passoja, M.; Aalto, O.; Raitola, M.; Multi-layer protocol tracing in a GPRS network, Proc. VTC 2002-Fall. 2002 IEEE 56th , Volume: 3 , Sept. 2002, pp. 1612 – 1616. [97] Sagar, A,; Performance evaluation of emerging multimedia services over wireless LAN, Master’s Thesis, Tampere University of Technology, Finland, Dec. 2003.
82
REFERENCES
[98] Kuang, T.; Williamson, C.; A Measurement Study of RealMedia Audio/Video Streaming Traffic, Proc. SPIE ITCOM, Boston, MA, pp. 68 – 79, July 2002. [99] Fitzek, F.; Kopsel, A.; Krishnam, M.; Reisslein, M.; Providing application-level QoS in 3G/4G wireless systems: A comprehensive framework based on multi-rate CDMA, IEEE Wireless Communications, 9(2):42 – 47, April 2002. [100] Leijonhufvud, G.; Multi access networks and always best connected (ABC), Proc. MMC workshop, Berlin, Germany, November 2001. [101] National science foundation workshop on network research testbeds. Final report. Available at: http://gaia.cs.umass.edu, NSF, 2002. [102] IEEE standard for information technology - telecommunications and information exchange between systems - local and metropolitan area networks - specific requirement. Part 11: wireless LAN medium access control (MAC) and Physical layer (PHY) specifications. Amendment 2: higher-speed physical lyaer (PHY) extension in the 2.4 GHz band - corrigendum 1; IEEE Std 802.11b-1999/Cor 1-2001 ,7 Nov. 2001. [103] Information technology- telecommunications and information exchange between systems- local and metropolitan area networks- specific requirements, part 11: wireless LAN medium access control (MAC) and physical layer (PHY) specifications amendment 1: high-speed physical layer in the 5 GHz band; ISO/IEC 880211:1999/Amd 1:2000(E); IEEE Std 802.11a-1999 ,2000. [104] IEEE standard for information technology- telecommunications and information exchange between systems- local and metropolitan area networks- specific requirements Part II: wireless LAN medium access control (MAC) and physical layer (PHY) specifications; IEEE Std 802.11g-2003, 2003. [105] IEEE P802.11, The Working Group for Wireless LANs, official webpage: http://grouper.ieee.org/groups/802/11/ [106] Pau, G.; Maniezzo, D.; Das, S.; Yujin Lim; Pyon, J.; Heeyeol Yu; Gerla, M.; A cross-layer framework for wireless LAN QoS support, Proc. Information Technology: Research and Education ITRE, Aug. 2003. [107] Hannikainen, M.; Design of Quality of Service Support for Wireless Local Area Networks, PhD Thesis, Tampere University of Technology, publication 386, 2002, 200 pages. [108] Lindgren, A.; Almquist, A.; Schelen, O.; Evaluation of Quality of Service schemes for 802.11 wireless LANs, Proc. 26th Annual IEEE Conference on Local Computer Networks, LCN’01, Nov. 2001. [109] Draft Supplement to International standard [for] Information Technology, Telecommunications and Informational Exchange between LAN/MAN Specific Requirements; IEEE 802.11e/D2.0, Nov. 2001. [110] IEEE P802, Status of Project 802.11e, official webpage: http:// grouper.ieee.org/groups/802/11/Reports/tge update.htm
REFERENCES
[111] IPERF: The TCP/UDP Bandwidth Measurement Tool, http://dast.nlanr.net/Projects/Iperf/
83
official webpage:
[112] Helix Universal Server, official webpage: http://www.realnetworks.com/ resources/contentdelivery/server/simulatedlive.html [113] Ethereal: A Network Protocol http://www.ethereal.com/
Analyzer,
official
webpage:
[114] NLANR/DAST: Network Performance Measurements Tools, official webpage: http://dast.nlanr.net/NPMT/ [115] Marsh, I.; Li, F.; Karlsson, G.; Wide Area Measurements of VoIP Quality, QofIS’03: Stockholm, LNCS 2811, Oct. 2003. [116] 3GPPTR 43.051; Technical Specification Group GSM/EDGE Radio Access Network; Overall description, Stage 2. [117] Halonen, T.; Melero, J.; Romero, J.; Gsm, Gprs and EDGE Performance: Evolution Towards 3G/UMTS, Wiley, 2002. [118] Kuzmanovic, A.; Knightly, E.; Low-Rate TCP-Targeted Denial of Service Attacks (The Shrew vs. The Mice and Elephants), Proc. ACM SIGCOMM’03, Karlsruhe, Germany, Aug. 2003. [119] Alberto’s thesis, 2004. [120] Moby Dick – Mobility and Differentiated Services in a Future IP Network; IST-200025394, FP5 Project; official webpage: http://www.ist-mobydick.org/ [121] INTERMON – Advanced architecture for INTER-domain quality of service MONitoring, modelling and visualisation; IST-2001-34123, FP5 Project; official webpage: http://www.ist-intermon.org/ [122] TEQUILA – Traffic Engineering for Quality of Service in the Internet, at Large Scale; IST-2000-16419, FP5 Project; official webpage: http://www.ist-tequila.org/ [123] Blondia, C.; A discrete-time batch Markovian arrival process as B-ISDN traffic model, Belgian J. of Operations Research, Statistics and Computer Science, vol. 32, 1993. [124] Lucantoni, D.M.; New results on the single server queue with a batch Markovian arrival process, Stochastic Models, vol. 7, no. 1, pp. 1 – 46, 1991. [125] Neuts, M.F.; A versatile Markovian point process, Journal on Applied Probability, vol. 16, pp. 764 – 779, 1979. [126] Moltchanov, D.; Koucheryavy, Y.; Harju, J.; Superposed and per-process analysis in sigmaD-BMAP/D/1/K queing system, HET-NETs, Ilkley, UK, July, 2003.
84
REFERENCES
[127] Hofkens, T.; Spaey, K.; Blondia, C.; Transient Analysis of the D-BMAP/G/1 Queue with an Application to the Dimensioning of a Playout Buffer for VBR Video, Networking 2004, May 9 – 14, Athens, Greece. [128] Blondia, C.; A discrete time Markovian arrival process, RACE document, PRLB 123 0015 CC CD, August 1989. [129] Fuhrman, S.; Le Boudec, J.-Y.; Burst and cell models for ATM buffers, ITC13, pp.975 – 980, Copenhagen, Denmark, June 1991. [130] Heffes, H.; Lucantoni, D. M.; A Markov Modulated Characterization of Packetized Voice and Data Traffic and Related Statistical Multiplexer Performance, IEEE Journal on Selected Areas in Communications, vol. 4, no. 6, 1986, pp. 856 – 868. [131] Saito, H.; Teletraffic Technologies in ATM Networks, Artech House, 1994. [132] MPEG traffic achieve, available at: http://www.info3.informatik. uniwuerzburg.de/MPEG/traces/ [133] Lombardo, A.; Morabito, G.; Schembra, G.; An accurate and treatable Markov model of MPEG video traffic, Proc. of IEEE INFOCOM 1998, pp. 217 – 224. [134] Moltchanov, D.; Koucheryavy, Y.; Harju, J.; A model of single smoothed MPEG traffic source based on the D-BMAP arrival process with limited state space, Proc. of ICACT, pp. 57 – 63, Korea, February 2003. [135] Floyd, S.; Jacobson V.; Random Early Detection gateways for Congestion Avoidance, IEEE/ACM Transactions on Networking, vol.1, no.4, August 1993, pp. 397 – 341. [136] Jacobson, V.; Congestion Avoidance and Control, Computer Communication Review, vol. 18, no. 4, pp. 314 – 329, Aug. 1988. [137] Jacobson, V.; Modified TCP Congestion Avoidance Algorithm, Technical Report, end2end interest mailing list, April 30, 1990. [138] Makknar, R.; Lambardis, I.; Salim, J.; Seddigh, N.; Nandy, B.; Babiarz, J.; Empirical study of buffer management scheme for DiffServ assured forwarding PHB, Proc. of ICC2000, pp. 632 – 637, August 2000. [139] Meine, B.; Fast algorithms for the numerical solutions of structured Markov chains, Ph.D. Thesis, Departments of Mathematics, University of Pisa, 56125, Italy. [140] Moltchanov, D.; Per-source performance evaluation in discrete time DBMAP/ D/1/K/N system, International seminar Telecommunication networks and teletraffic theory, St.-Petersburg, pp. 132 – 143, January 2002. [141] Takine, T.; Sengupta, B.; Hasegawa, T.; An analysis of a discrete-time queue for broadband ISDN with priorities among traffic classes, Communications, IEEE Transactions on ,Volume: 42 ,Issue: 234 ,FEBRUARY/MARCH/APRIL 1994, pp. 1837 – 1845.
REFERENCES
85
[142] Akar, N.; Oguz, N.C.; Sohraby, K.; TELPACK: an advanced teletraffic analysis package, Communications Magazine, IEEE ,Volume: 36 ,Issue: 8 ,Aug. 1998, pp. 84 – 87.
Publications
Publication 1 Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Analytical Estimation of EF PHB Service Parameters for Aggregated MPEG Traffic,” in Proc. of the 16th Nordic Teletraffic Seminar, NTS’02, Espoo, Finland, August 2002, pp. 279 – 290. c Copyright 2002 by NTS’02.
An analytical estimation of EF PHB service parameters for aggregated MPEG traffic Koucheravy Y., Moltchanov D., Harju J. Institute of Communication Engineering, Tampere University of Technology, P.O. Box 553, FIN-33101, Tampere, FINLAND {yk, moltchan, harju}@cs.tut.fi http://www.cs.tut.fi/tlt
Abstract The differentiated services (DiffServ, [1]) expedited forwarding (EF, [2]) per-hop behavior (PHB) is targeted on applications which need strict guarantees on end-to-end delay and very low loss probability. It makes EF PHB an appropriate choice for loss-free timely delivery of delay and loss intolerant traffic. It is expected that the substantial part of such traffic will be generated by video-on-demand (VoD) services. In order to provide transmission service which is based on EF PHB to VoD traffic, several traffic conditioning functions have to be implemented within the DiffServ ingress network nodes. These conditioning functions are based on traffic profiles. In this paper we show how to compute EF PHB traffic profiles for VoD traffic, which are based on simple token bucket mechanism. Further, a tandem queuing network that models the process of EF traffic treatment within the DiffServ ingress network node is considered. Based on this, the simple analytical approach for worst-case delay evaluation of the EF traffic treatment within the DiffServ ingress network node is presented. Finally, under several assumptions the DiffServ domain’s end-to-end worst-case delay of VoD traffic is evaluated.
1 Introduction VoD service is expected to be a one of the very popular applications in broadband multi-service networks. From the traffic transmission point of view, VoD service can be classified as one of the most delay and loss intolerant client-server application. Traffic generated by such sort of applications need strict quality of service (QoS) guarantees in terms of bandwidth, losses, delays and jitter. In order to grant loss-free timely delivery of this type of traffic an appropriate network transmission facility should be provided. Currently, Internet does not provide any QoS guarantees to its applications. It provides the best-effort service only which does not satisfy the demands of real-time applications. As a result, the work on IP QoS has been initiated by Internet Engineering Task Force (IETF). IETF has proposed two QoS frameworks: Integrated Services (IntServ, [3]) and Differentiated Services (DiffServ, [1]). IntServ
approach is based on connection admission control (CAC) procedures and uses explicit resource reservation technique [3]. IntServ can provide deterministic QoS guarantees and for this purpose it requires a signaling protocol in order to inform the appropriate network elements about the necessary resource reservation [4]. On the other hand, DiffServ employs quite different approach. DiffServ aggregates flows with the same QoS requirements and assigns to them the same treatment at the routers along the path of aggregated traffic flow. The service differentiation is achieved by implementing several simple per-hop behaviors (PHBs) in the routers. As a result, DiffServ provides probabilistic guarantees to aggregated traffic flows inside the network and uses a sort of static CAC algorithms which is based on service level agreements (SLAs) between subscribers and service providers, or between two service providers. DiffServ approach is more preferable compared to IntServ because of its simplicity and high scalability [1]. In this paper we assume that the network, in addition to the best-effort service, implements DiffServ QoS framework. IETF DiffServ working group has standardized two PHB identification code groups. These are Assured Forwarding PHB (AF PHB, [5]) and Expedited Forwarding PHB (EF PHB, [2]). The first one is designed for a range of applications which need very different QoS requirements. There are four classes of PHB identification codes within AF PHB group. Moreover, within the each class there are three distinct DiffServ codepoints (DSCP) with different packet drop precedence associated with each DSCP. The EF PHB group is targeted on applications which require strict guarantees on end-to-end delay and very low loss probability. Transmission services which are based on EF PHB look like as end-to-end “pipe” of constant bandwidth between two applications [2]. Thus, within the DiffServ compliant domain applications can use three different types of service: best-effort, AF PHB and EF PHB. It is widely stated that in order to grant loss-free timely delivery, MPEG traffic should have the necessary amount of reserved bandwidth and/or some type of priority in service. To handle such type of traffic within the DiffServ domains appropriately we propose to use EF PHB group. The modern compression algorithms mostly operate in variable bit rate (VBR) mode rather than in constant bit rate (CBR) mode. VBR mode allows very high compression ratios while the quality of decoded picture remains almost the same during the playback. At the other hand, EF PHB is used to construct transmission services with constant end-to-end bandwidth. From the networking resources point of view, it seems to be inappropriate to transmit VBR compressed video over CBR channel, but currently this is almost only way to transmit compressed video over Internet with some type of guarantees. In order to provide transmission service which is based on EF PHB to behavior aggregate, several traffic conditioning functions must be implemented within the DiffServ ingress network nodes. The traffic conditioning functions are based on traffic profiles. One of the most popular traffic profile used in the current Internet is based on a token bucket policing mechanism. In accordance with this profile each DSCP is assigned to a particular token bucket, and during transmission all packets marked with this DSCP should be measured against this token bucket meter. The estimation of token bucket parameters is a crucial task in context of configuration of network nodes along the path of EF behavior aggregate. These parameters are used in conditioning functions of DiffServ ingress and/or egress network nodes. One of the major advantages of the services which are based on EF PHB consists in their ability to bound the end-to-end delay. The transmission path through the DiffServ
domain consists of DiffServ ingress network node, a number of interior nodes and DiffServ egress network node. Therefore, using a decomposition method we can represent the transmission path of the behavior aggregate through the domain as a sequence of queuing systems and estimate the end-to-end delay bound. The full paper is organized as follows. Section 2 deals with VoD service peculiarities. VoD service configuration and aggregated MPEG traffic modeling issues are considered there. In Section 3 the attention is paid to analysis of EF traffic treatment within the DiffServ ingress node. In Section 4 the DiffServ domain’s end-to-end delay of EF behavior aggregate is evaluated.
2 VoD service 2.1 VoD service configuration Let us consider the network configuration of VoD service. Assume that there are N neighboring Internet domains making up a chain. Further assume that these domains belong to different large internet service provides (ISPs). Every ISP has its own DiffServ implementation in accordance with IETF specifications [1]. In order to provide QoS guarantees to their users, let us assume that these ISPs have installed SLAs between each other. Such configuration is depicted in Fig. 1. We believe that this configuration may be the case in Internet and we called it the “remote” VoD configuration. In this configuration the ISPs may not have its own VoD server or dedicated connection to VoD content provider. Moreover, the nearest VoD server can be located far from the home domain of the VoD user. Anyway, the desired QoS requirements must be provided between the VoD service entities.
...
1
ISP 1
ISP N VoD user
DiffServ domain
SLA
...
SLA
DiffServ domain
VoD Server
M VoD AP
VoD user
Fig. 1 “Remote” VoD service configuration. We also assume that there are several users in ISP N who wish to use VoD service, however, only ISP 1 has a dedicated connection to VoD content provider. In this case VoD service should be provided across all N domains belonging to different ISPs.
2.2 VoD server functional description Let us consider the VoD system functional diagram that is depicted in Fig. 2. MPEG storage stores information of MPEG sequences. These sequences are represented by files. Note that the structure of the MPEG storage system has a lot of peculiarities and is outside the scope of this paper. Rate information block keeps information about the predefined rates of each MPEG sequence. This information is available from the GoP smoothing algorithm. It has been shown that such algorithm smoothes both burstiness and high peak rate found in MPEG sequences and does not have significant influence on MPEG time structure [6]. During the transmission of video sequence the rate information block enforces the server software to use predefined rates instead of real frame rates. Then, the smoothed MPEG
sequence is packetized into IP packets. In order to achieve synchronized transmission we propose to use RTP over UDP protocols configuration. VoD system Rate information block Smoothing
RTP UDP IP MPEG storage Network interface
Fig. 2 VoD server functional description. To proceed further, let us introduce the several assumptions about the VoD server configuration. As long as we adapted the “remote” VoD configuration, we can assume that length of all IP packets belonging to different microflows within the behavior aggregate are constant and their value lies between acceptable bounds. Thus, the packet service time within the DiffServ network node is a constant value and equal to the time to transmit one packet on the outgoing link. This is true at least for DiffServ ingress network node. In addition, we also assume that domains do not implement IP packets segmentation procedures along the path between the VoD server and the end-user equipment. In this case the service time of VoD IP packets is constant in each DiffServ network node along the path to destination.
2.3 Aggregated traffic modeling Modeling the VBR compressed video traffic has become an important issue since it provides the starting point in both theoretical analysis and engineering design. The source models of different types of videotraffic are needed to design and study the performance of the network and also to predict the QoS that a particular video application may experience at different levels of network congestion [7]. An emerged theory of measurement-based traffic modeling [8,9,10] allows us to recognize and define the major statistical characteristics which have a strong influence on queuing systems performance parameters. In accordance with [8,9] the major impact on performance parameters of queuing systems is produced by the probability distribution function (PDF) of number of arriving packets and the behavior of its autocorrelation function. Moreover, it was stated [10,11,12] that the behavior of autocorrelation function is of paramount importance in context of queuing systems performance. It was also shown [10,11] that the fairly good approximation of empirical data can be produced when the both the autocorrelation function and the PDF of the model match well their empirical counterparts. It should be noted that the modeling of MPEG traffic is a complicated issue in network engineering. Every time we start to construct the traffic model we should keep in mind the application area of our model. The main application area of traffic models are the network node performance evaluation and the evaluation of the end-to-end performance of the modeled data transmission. Performance evaluation includes the derivation of several meaningful parameters which include mean delay, mean loss and, probably, their PDFs. We can carry out such types of performance evaluation by two different approaches. On the one hand, we can introduce a network node queuing model or a whole network queuing model
and derive the necessary parameters analytically. On the other hand, we can carry out the simulation studies using the one of the existing network simulators. Keeping in mind the previous considerations we have several classifications of traffic models. First of all, based on particular properties of the models, we should distinguish between analytically tractable and analytically intractable models. Analytically tractable models allow us to produce results mathematically. Moreover, in most cases it is possible to produce software implementations of these models. Thus, with analytically tractable models we also can carry out simulation studies. In contrast, analytically intractable models are limited to simulation studies only. Based on the type of queuing system and on the properties of the model, analytically tractable models can be classified into continuous-time and discrete-time. For example, in the case of ATM network nodes, when the time slot sizes are fixed, the queuing system is discrete-time. It is appropriate for such type of queuing system to construct a model of arriving traffic which also yields discrete-time nature. On the other hand, in continuous-time traffic models we should explicitly define the interarrival time distribution. It should be noted that the experience accumulated during the past years in analysis of different types of discrete time queuing systems may significantly help us to evaluate the Internet services performance. If we consider the whole Internet, it is hard to define the service time distribution. This is because the sizes of IP packets vary significantly for different applications. The introduction of DiffServ and IntServ QoS frameworks and the structure of the Internet as a composition of a large number of separately administrated networks can allow us to investigate the performance parameters of several services analytically. The packet stream at the network interface between the VoD system and the DiffServ ingress network node consists of multiplexed traffic from a specified number of smoothed MPEG traffic sources. We called it the “MPEG traffic behavior aggregate”. In this paper we propose to model the multiplexed traffic from a specified number of smoothed MPEG traffic sources by the Markov modulated arrival processes. Particularly, these are continuous-time Markov modulated Poisson arrival process (MMPP) or discrete-time batch Markovian arrival process (D-BMAP). It was shown by S.-Q. Li [10] that the MMPP process matches the mean arrival rate of the single MPEG traffic source and its autocorrelation function well. Later, the same conclusions were made for D-BMAP process [6,13]. It should be noted that the D-BMAP process allows us to use two approaches in order to define packets arrivals. In each state of the modulating Markov chain we may use analytical distribution. Moreover, in special cases we may prefer to use an empirical histogram of relative frequencies as the PDF of number of arriving packets. In [13], it was shown that the D-BMAP process can capture the histogram of relative frequencies of real MPEG sources. The latter gives us an assurance that all statistical characteristics of the D-BMAP process match their empirical counterparts well. Indeed, this is the true because the stationary processes can be described by its PDF and the autocorrelation function. The authors in previous work [6] have found that the D-BMAP process can also model the multiplexed traffic from a specified number of MPEG sources. It captures the histogram of relative frequencies and the autocorrelation function of multiplexed MPEG traffic well. To have the assurance that the used model is close enough to the empirical data, further, in this paper we use D-BMAP process.
The construction of Markov modulated processes from the empirical data involves the inverse eigenvalue problem. It is known that the general solution of inverse eigenvalue problem is not exist. However, up to date several papers addresses the solution of this problem when some limitation on the form of eigenvalues are set [6,10,13].
3 DiffServ ingress network node modeling and evaluation 3.1 Functional description To achieve loss-free timely transmission, compressed video traffic should have the necessary amount of reserved bandwidth and/or some type of priority in service. In accordance with IETF specification [2] EF PHB should have a predefined minimum amount of forwarding resources such as well-defined minimum departure rate (bandwidth) and the amount of buffer space at each node along the path of EF behavior aggregate. Therefore, in the case of EF PHB a necessary amount of resources in terms of bandwidth (minimum departure rate) and buffer space can be assigned to VoD traffic. Moreover, the process of EF traffic treatment should be independent from other traffic treatment within the network nodes. DiffServ ingress node VoD server
Traffic Conditioning Functions
EF Queue
1 2 Network interface Fig. 3 Functional blocks of DiffServ ingress network node serving EF behavior aggregate.
In accordance with DiffServ specifications [1] interior network nodes do not implement any traffic conditioning functions. All these functions must be performed by ingress network nodes. Let us consider the DiffServ ingress network node which implements EF PHB in addition to the best-effort service. In accordance with EF PHB specification [2] such node must implement two major functions: conditioning the EF behavior aggregate via policing and/or shaping and configuring node so that the EF behavior aggregate has a well-defined minimum departure rate. The functional blocks of DiffServ ingress network node serving the EF behavior aggregate is shown at Fig. 3 where the “1” is a traffic conditioning block, and the “2” is an EF queue at the output port of ingress network node.
3.2 Token bucket parameters In order to provide transmission services which are based on EF PHB, several traffic conditioning functions must be implemented within the DiffServ ingress network nodes. These traffic conditioning functions are based on traffic profiles. Traffic profiles specify the temporal properties of traffic stream selected by a classifier [1]. One of the most popular traffic profiles used in current Internet is based on a token bucket policing mechanism. In accordance with this profile each DSCP is assigned to a particular token bucket and during the packets transmission all packets marked with this
DSCP should be measured against this token bucket meter. Token bucket traffic profile defines rules for determining whether a particular packet is in-profile or out-of-profile. Outof-profile packets are those packets which arrive when insufficient tokens are available in the bucket. Different conditioning actions may be applied to in-profile and out-of-profile packets. Token bucket can be used for two purposes. Firstly, it shapes incoming traffic up to the some horizon. Shaping bounds of the simple token bucket are given by two parameters: bucket depth (b) and token rate (r). Token bucket also can serve as traffic marker - DSCP of packets which do not conform to the token bucket specification (r,b) can be changed. Note that the bucket depth and token rate are often given in bytes rather than in packets. This is because the packet lengths can be variable. In our case when all packets have the same length we measure the bucket depth and token rate in packets without the loss of generality. Recently, it was shown [14] how the token bucket parameters for IntServ's Tspec can be derived. We show how token buckets parameters can be evaluated in the case of services which are based on EF PHB and used for VoD traffic transmission. Server
EF behavoir aggregate
(b-1) waiting positions
Served EF behavoir aggregate
Fig. 4 The token bucket equivalent queuing system. Assume that the EF behavior aggregate after entering the DiffServ ingress network node is policed and shaped via simple token bucket mechanism and then it enters the corresponding EF queue. It is known that the token bucket can be represented by the equivalent queuing system with single server (Fig. 4). The length of the buffer in such queuing system equals to the token bucket depth b and the service rate equals to the token rate r. The probability of loss in this queuing system corresponds to the probability of remarking. Note that remarked packets will not enter the EF queue. To provide loss-free transmission of VoD traffic through the DiffServ domain we should provide such (r,b) that the probability of packets remaking is equal to zero. This probability can be derived from the analysis of simple queuing system D-BMAP/D/1/b, where D-BMAP process is the aggregate arrival process from a number of smoothed MPEG sources composing the EF behavior aggregate and b is the capacity of the system, measured in packets. The service time in such system is slotted and the slot length is given by ∆ = S / B where S is the length of the IP packet in bits and B is the token rate measured in bits per second. Token buckets allow packets remarking in accordance with negotiated traffic profiles. Remarking is highly undesirable for MPEG traffic transmission over EF PHB service. Let us consider the token bucket with parameters (r,b) which allows packets remarking at the DiffServ ingress network node. In this case a certain part of arriving packets is remarked to other DSCP. A number of packets which can be remarked depends on (r,b) parameters of the token bucket and on the arrival process as well. The DSCP to which these packets are remarkered may represent AF PHB classes or even DSCP corresponding to the best-effort service. Therefore, microflows can experience unpredictable delays, reordering of packets and ever losses which are caused by traffic conditioning functions. Therefore, the token
bucket mechanism must not allow the packet loss (remarking) and the delay of the packets within the token bucket should be bounded by some value. The probability of v packets loss in D-BMAP/D/1/b system is given by next expression [15]:
f L (v ) = Pr{L = v} =
∑∑∑ (xki ⋅ (D(= b + v − k )ij ) , b M M
(1)
k =0 i =1 j =1
where b is the capacity of the system in packets, M is the number of states of modulating Markov chain of the D-BMAP process, xki , k ∈ {0,1,.., b} , i ∈ {1,2,.., M } are the elements of
stationary distribution of modulating Markov chain x , D(= k ) , k ∈ {1,2,..} are the matrices each of which contains elements which denote the transition of modulating Markov chain from state to state with k arrivals. Since our queuing system must not allow the loss consider the probability at least one packet loss in the D-BMAP/D/1/b system. This probability can be derived from (1) and is given by next expression [15]: f L (v ≥ 1) = Pr{L ≥ 1} =
∑∑ ∑ (xki ⋅ (D(≥ b + 1 − k ))ij ), b M M
k = 0 i =1 j =1
(2)
where D(≥ k ) , k ∈ {1,2,..} are the matrices each of which contains elements which denote the transition of modulating Markov chain from state to state with more than k arrivals. Analyzing the expression (2) we should note that the event of at least one packet loss depends on the arrival process and the capacity of the system (bucket depth). Note that the D-BMAP arrival process depends on the service time duration ∆ which, in turn, depends on the outgoing link rate (token rate). Based on these considerations, we can write next functional dependencies: Pr{L = v} = f (b, ∆, D − BMAP ), Pr{L ≥ v} = f (b, ∆, D − BMAP ). (3) Thus, using the given arrival process and by varying b and ∆ parameters we can choose the appropriate regions of token bucket parameters (r,b). Recall that these regions do not allow the loss of the packets of behavior aggregate. It should be noted that the services which are based on EF PHB must guarantee a bounded delay in addition to low losses. However, there are some areas within the chosen token bucket (r,b) parameters which allow a very high delay of certain packets. Such behavior corresponds to very high values of b. These values allow low token rates r and depend on their values. Therefore, in order to obtain the appropriate values of (r,b) we should take into account the delay. The probability that the packet will wait in D-BMAP/D/1/b system during the w time slots is given by [15]: w−1 ∞ M M 1 fV ( w) = Pr{V = w} = ∑∑ ∑∑ x ki ⋅ (D(r ) )ij ⋅ . (4) r k =0 r =1 i =1 j =1 where r is the size of arriving batch. Using the (r,b) parameters which do not allow packets losses computed at the previous step it is easy to choose the appropriate (r,b) pairs. These pairs should not allow the delay of any packet on more than w time slots within the token bucket mechanism. All of these computations can be performed by VoD server software before the beginning of SLA establishing procedure. Moreover, in the case when the ISP uses some sort
(
)
of dynamic bandwidth allocation through Bandwidth Brokers (BB) or some other devices, traffic profiles can be computed by VoD server dynamically.
3.3 EF service configuration within the ingress network node After the policing and shaping procedures the behavior aggregate enters the EF queue at the output port of the ingress network node. In accordance with specification EF behavior aggregate must have a well-defined minimum departure rate that is independent of the dynamic state of the node [2]. Therefore, we can consider EF queue within the DiffServ ingress network node isolated from the other queues within that node. The system diagram of the DiffServ ingress network node is shown at Fig. 5 where the blocks correspond to those ones depicted at Fig. 3.
b
...
DiffServ ingress node
X/D/1/K r
VoD server tb(r,b)
... K-1
Network 1 2 interface Fig. 5 System diagram of DiffServ ingress network node serving EF behavior aggregate.
If the network node is configured in such way that the minimum departure rate is equal to the token rate r, the buffer is empty at the output port of the router. In this case the packets are transmitted immediately at the outgoing link without any delays caused by buffering, and the delay within the DiffServ ingress network node is defined by the delay within the token bucket mechanism. As far as the lengths of all packets are equal and keeping in mind that each EF PHB class must have minimum amount of guaranteed departure rate, we propose to use a discrete time queuing system X/D/1/K, where K-1 is the length of EF queue and X denotes the arrival process from traffic conditioner. To define the arrival process to the EF queue at the DiffServ ingress network node we should consider the general characteristics of the output process from the D-BMAP/D/1/b queuing system which models the token bucket mechanism. It is known that the output process from the D-BMAP/D/1/b queuing system is a D-MAP process which belongs to the class of D-BMAP processes. Therefore, the buffer at the output port of the DiffServ ingress network node can be modeled by D-MAP/D/1/K queuing system. It should be noted that the derivation of the D-MAP process at the output of the D-BMAP/D/1/b queuing system is a very complicated procedure which involves a lot of mathematical calculations. Thus, in order to define the minimum departure rate for EF behavior aggregate we use approximation. In accordance with IETF specification [2], the services which are based on EF PHB should not allow packets losses while the end-to-end delay should be bounded. The latter requirement can be fulfilled only if the delay of the packets is bounded within the each network node along the path of EF behavior aggregate. In Section 3 it has been shown that the appropriate choice of token bucket parameters bounds the delay within the token bucket
mechanism in the DiffServ ingress network node. If the minimum departure rate which is assigned to the EF behavior aggregate is less than the token rate, the packets will be queued time after time. Note that the size of queue grows when short-term traffic arrival rate exceeds the minimum departure rate. The behavior aggregate’s short-term maximum arrival rate after leaving the token bucket system is equal to the token rate r and is measured in packets per second. In this case the interarrival time between any two packets is equal and is given by 1/r, where r is the token rate. Moreover, since the EF service must have well-defined minimum departure rate and the length of packets is equal we can consider the EF queue as a D/D/1/K “pseudo” queuing system where K is the capacity of the system. Based on these assumptions we can consider the DiffServ ingress network node as the tandem queuing network consisting of D-BMAP/D/1/b queuing system and D/D/1/K “pseudo” queuing systems as depicted at Fig. 6. DiffServ ingress node VoD server
D-BMAP/D/1/b
D/D/1/K
1 2 Network interface Fig. 6 Tandem queuing network which models the process of EF behavior aggregate treatment within the DiffServ ingress network node.
Note that both the packet service time and the packets interarrival times are deterministic and are given by 1/B and 1/r respectively. When r>B (1/B0. Later, this model has been extended by Elliot [7] who allowed the error free state of the Gilbert model to have probability of bit error more than zero (p0>0). The model became more flexible in terms of its application area capable of modeling more sophisticated error sequences and correlations. However, such model still has only two states of Markov chain and, therefore, the range of bit error correlations are limited. The next expansion came from Fritchman [8], who allowed an arbitrary number of error free states of Markov chain in Gilbert model and, therefore, the range of bit-error correlations was again significantly expanded. The last
extension was made in middle 90th when Fang verified [9] a versatile Markov based wireless link model by allowing an arbitrary number of error states. All abovementioned wireless link models were used in performance evaluation of link layer protocols. However, all of these studies were based on steady-state behavior of Markov chain modeling the bit errors. Such analysis, of course, gives us some basic ideas regarding the performance of link layer protocols. However, it is rarely occurs in practice in mobile networks serving the movable users since the mobility patterns and session times differ significantly from user to user and application to application. Therefore, in order to predict the network performance and to adequately predict the QoS degradation experienced by nomadic users we have to integrate the wireless link model into the user integrated traffic model of 4G All-IP users. Therefore, in general, the time varying nature of wireless link also depends on user mobility, cannot be neglected and considered independently of user traffic model in future 4G All-IP networks. The one possible way to do that is to analyze the mobility patterns of current mobile users. For example, for highly movable users we can expect some periodicity in cell boundaries crossing and, therefore, we can predict to the some extent the sequence of state visited by Markov error model. The other possibility is to incorporate the error model into the integrated traffic model. Mathematically such model can depend on or be incorporated into integrated traffic model of 4G All-IP mobile user. Finally, the performance evaluation of 4G All-IP networks can be graphically represented as shown in Fig. 4.
Unreliable medium
high-level QoS-based description of isolated IP networks or domains. For example, it can be useful to characterize the IP network in terms of loss and end-to-end delay. If the used QoS framework is DiffServ it is possible to characterize the end-toend QoS parameters based on service level agreements. The means how this information can be delivered to mobile terminal depend on QoS framework implementation in 4G AllIP networks. D. Performance Evaluation of the CN Core network performance parameters such as probability of loss and probability of certain delay within the network element or end-to-end performance parameters can be computed using the existing techniques which were successfully applied in fixed QoS-aware networks like ATM or DiffServ-enabled Internet domains. VI.
REFERENCES [1] [2] [3]
Teletraffic
Mobility
Fig. 4. User traffic model in 4G All-IP network. C. Fixed Part of the RAN While changing the path at the wireless part of the RAN, the traffic path within the fixed part of the RAN can also change due to user mobility. Therefore, the network has to be able to characterize the performance parameters of new path and to inform mobile terminals about it. This can be done by using the classic teletraffic methods presented in literature and taking into account all possible paths of user traffic. Form such point of view it is crucial to provide
CONCLUSIONS
In this paper we have identified requirements induced on user traffic model in developing 4G All-IP networks. We have also considered the performance evaluation issues of these networks. We have concluded that performance evaluation of both fixed part of the RAN and core network can be based on existing techniques but requires certain new definitions and performance measures like IP domain’s end-to-end delay etc.
[4]
[5]
[6] [7] [8]
[9]
R. Braden, D. Clark, S. Shenker, “Integrated Services in the Internet Architecture: an Overview,” RFC 1633, June 1994. S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss, “An Architecture for Differentiated Services,” RFC 2475, December 1998. J. Heinanen, F. Baker, W. Weiss, J. Wroclawski, “Assured Forwarding PHB Group,” RFC 2597, June 1999. B. Davie, A. Charny, J. Bennett, K. Benson, J. Le Boudec, W. Courtney, S. Davari, V. Firoiu, D. Stiliadis, “An Expedited Forwarding PHB (PerHop Behavior),” RFC 3246, March 2002. N. Antunes, A. Pacheco, R. Rocha, “An integrated traffic model for multimedia wireless networks,” Computer Networks, Vol. 38, pp. 25-41, 2002. E. Gilbert, “Capacity of a burst-noise model,” Bell.Syst. Tech. J., Vol. 39, pp. 1253-1265, 1960. E. Elliott, “Estimates of error rates for codes on burst-noise channel,” Bell.Syst. Tech. J., pp. 1977-1997, 1963. B. Fritchman, “A binary channel charachterization using partitioned Markov chain,” IEEE Trans. Inform. Theory, Vol. 13, pp. 221-227, 1967. H. Wang, “On verifying the first-order Markovian assumption for a Rayleigh fading channel model,” in Proc. ICUPC’94, pp. 160-164, 1994.
Publication 5 Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Performance Evaluation of Live Video Streaming Service in 802.11b WLAN Environment Under Different Load Conditions,” in Proc. of the ACM International Workshop on Multimedia Interactive Protocols and Systems, MIPS’03, LNCS 2899, Napoli, Italy, November 2003, pp. 30 – 41. c Copyright 2003
Performance evaluation of live video streaming service in 802.11b WLAN environment under different load conditions Yevgeni Koucheryavy, Dmitri Moltchanov, Jarmo Harju Institute of Communication Engineering, Tampere University of Technology, P.O.Box 553, Tampere, Finland {yk, moltchan, harju}@cs.tut.fi
Abstract. Live video streaming service, which is common nowadays in the Internet, is supposed to be very challenging and demanding service in next-generation wireless networks. However, both limited quality of service (QoS) support and unstable quality of the air interface can restrain its wide deployment. In this paper we consider live video streaming over IEEE 802.11b wireless local area network (WLAN), which is claimed to be used as a part of layered infrastructure of next-generation mobile systems to provide coverage in highly populated areas. We performed our experiments under different signal-to-noise ratios (SNRs) and different competing TCP and UDP traffic volumes. The main conclusion of our study is that despite a common belief live streaming multimedia services are not ready for wide implementation in hot-spot areas where both high traffic volume and relatively weak signal strength (less than 30 dB) may deny the service easily.
1
Introduction
Nowadays, a considerable part of research activities in telecommunications are switching towards development of next-generation IP-based wireless networks. The major motivation behind that is to introduce a common service platform and transport facilities for future composite mobile Internet. In addition to broadband wireless access to the Internet, next-generation mobile systems should be able to provide quality of service (QoS) to various applications. Nowadays, new multimedia services attracted to these networks are striving hard towards commercial market. Both limited QoS support and unstable quality of the air interface can restrain their wide deployment. The most crucial layers for QoS support in next-generation wireless networks are physical, data link and transport. In this paper we restrict our attention to transport layer only. The reason behind that is to judge whether the particular multimedia services may already be smoothly implemented on commercial basis over currently available wireless local area networks (WLANs), which are claimed to be a part of layered infrastructure of next-generation mobile systems. In this paper we consider most popular IEEE 802.11b WLAN.
In next-generation wireless networks there will be a clear separation between wireless part (Radio Access Network, RAN) and fixed one (Core Network, CN). RAN is used to hide all access specific peculiarities from the CN. Therefore, CN has a little impact on introduction of new RANs and can evolve independently. A strict separation between RAN and CN will lead mobile systems to multiaccess environment. This environment introduces an additional conceptual notion of next-generation wireless networks – an Always Best Connected (ABC) concept [1]. The ABC should allow users to choose the most suitable RAN at any instant of time during duration of a call. Particularly, this feature is claimed to be very attractive for users with complex mobility patterns. Technical implementation of ABC is to be based on intersystem (vertical) handover that should be implemented in a seamless way between any types of access networks [2, 3]. In addition to multi-access environment, it is becoming clear that nextgeneration networks will have a layered infrastructure with at least two hierarchical levels. In accordance with layered network infrastructure, there should be cells of different size (picocells, microcells, macrocells) each of which serves users in areas with different population densities. Layers with picocells or microcells are able to provide a high capacity with high bandwidth in hot-spot areas. They can serve slow mobility users with high traffic demands. It is assumed that in next-generation mobile systems this role will be assigned to WLANs. Therefore, we can state that WLANs and 3G RANs are not competitors, but complement each other to allow coverage in areas with different population densities. Live video streaming service is widely spread nowadays in the Internet and QoS perceived by end-users varies in average from good up to excellent. For example, some of live streaming videos are already available in 300Kbps and can be streamed continuously without brake in service via well-provisioned parts of the Internet. It is supposed that live video streaming service will be very challenging and demanding in next-generation mobile systems. In our study we have implemented series of performance tests of live video streaming service in 802.11b WLAN environment under different signal-to-noise ratios (SNR) and different load conditions. We concentrate our attention on tracing real traffic. Indeed, it is anticipated [4] that testing of real implementations can bring better understanding and new knowledge in the area. To date only few studies in this area are available. Authors in [5] considered the multimedia streaming service over IEEE 802.11b WLAN. They defined a number of SNR ranges and evaluated a perceived QoS provided to the user. In [6] the perceived QoS of monomedia applications under heterogenous wireless environment was evaluated. Among other conclusions, it can be found that bandwidth-greedy applications can already be implemented over current IEEE 802.11b networks. However, it may be the case that unstable nature of the wireless link along with high bandwidth competing traffic can deny the service easily. In this paper, we extend the previous results to the case of multimedia applications, different SNR ranges and different competing traffic loads. The rest of the paper is organized as follows. Testing prerequisites are considered in Section 2. Testbed configuration is given in Section 3. Carried mea-
surements and corresponding results are outlined in Section 4. Conclusions are drawn in last section.
2 2.1
Testing prerequisites IEEE 802.11b WLAN
The IEEE 802.11x specifications are wireless standards that specify an ’overthe-air’ interface between a wireless client and a base station (access point), as well as among wireless clients. The IEEE 802.11 specifications address both the physical and media access control (MAC) layers and are targeted to resolve compatibility issues between manufacturers of WLAN equipment. Approved in 1997 by the IEEE 802 committee, IEEE 802.11 uses the 2.4GHz microwave and defines two different (and mutually incompatible) methods of encoding: FHSS (Frequency Hopping Spread Spectrum) and DSSS (Direct Sequence Spread Spectrum). FHSS spreads the transmission across 75-MHz subchannels, continuously skipping between them, while DSSS breaks the band into 14 overlapping 22-MHz channels and uses one at a time. Two basic operating modes are defined: infrastructure and ad-hoc. Most dedicated hardware provides a basic service set that builds the wireless ’infrastructure’. It allows clients to roam between access points while roaming across routers is prohibited. The ad-hoc mode allows individual nodes to participate in a peerto-peer communication without an access point. The major problem with 802.11 was its relatively low throughput compared to wired networking and the mutual incompatibility of FHSS and DSSS equipment. In 1999, the IEEE 802 committee extended the specification, deciding to concentrate on DSSS. This extension, known as 802.11b, allowed more complicated encoding techniques which increased the throughput up to 5.5 Mbps. 2.2
Multimedia traffic
Multimedia applications are continuously growing in popularity. The availability of high-speed fixed access networks is the primary reason behind that. Today, it is necessary to support these services over wireless access networks. Basically, real-time multimedia traffic consists of one or more media streams and can be characterized by strict delay requirements while can tolerate some losses. It is supposed that applications emerging from the Internet will become capable of defining the required QoS level soon. However, currently in almost all networks multimedia traffic is treated similar to ordinary best effort traffic, which does not often require strict delay guarantees. Therefore, it is crucial to predict the QoS degradation that multimedia applications may experience over wireless access networks. In our paper we consider live video streaming service, which consists of both video and audio medias. Note that from the user point of view the service can be described by two phases: prefetching phase and playing phase. While in prefetching phase the application stores data and then turns into playing phase. When
application is in prefetching phase it uses all available bandwidth to prefetch data. When playing back, it restricts itself to the certain average bandwidth of combined stream (target rate).
3 3.1
Testbed Client-server streaming implementation
In our testbed we used commercial implementation of client-server streaming service. Combination of RealNetworks’ Helix server and RealNetworks’ RealOne player was chosen. We used free distribution of Helix server available at [7]. It should be noted that compared to commercial distribution there are several limitations of free one. Particularly, the number of simultaneous connections should not exceed 10 while the bandwidth should be less than 1Mbps. However, all these limitations do not add bottlenecks in our testbed since it was not necessary to stream more than one video at a time. Helix server can stream a lot of well known medias including both proprietary and standard-based ones. We have chosen real media streaming format because of the following reasons. At first, free distribution of Helix server allows all server-side capabilities only when real media format is used. Secondly, the real media format is currently very popular in the Internet because of relatively good quality of low bit rate videos. Additionally, when real media format is used free distribution of Helix server allows to serve clients with different bandwidth capabilities. Moreover, the bandwidth at which the client is served can also be changed dynamically during connection. In order to achieve that the video should be coded at different target rates each of which is specific for a certain bandwidth capability of the client. In our testbed we used live streaming service. In accordance with it the server continuously listens specific ports for a connection requests. When the request arrives server sets up RTSP connection, adds client into connection pool and then begins streaming at the rate which is the most appropriate for requesting client. However, if the bandwidth capability of the client changes the server can adapt connection by increasing or decreasing the target rate of video. The bandwidth capability of the client is indicated in the ’BANDWIDTH’ field of RTSP protocol. The ’live streaming’ means that the server transmits video from that actual point in time when RTSP connection has already been established. In our study live video streaming service was emulated. The usage of real live streaming adds unnecessary complexity to our testbed, i.e. requires to consider codec-specific peculiarities (type of source, compression and coding delays). It may lead our focus away from network-specific issues which are the main topics of our paper. Note that the Helix server allows to emulate all features of live streaming service using the ’slta’ utility, which is the part of server distribution. Protocol configuration used by Helix server and RealOne player is shown in Fig. 1. In order to facilitate live video streaming both TCP and UDP protocols are employed. RTSP over TCP (the solid line) is used at the connection establishment phase when the client poses the request on specific video. The actual
streaming is performed over UDP (the dashed line). However, sometimes when it is not possible to use RTSP/UDP combination, RTSP/TCP is employed instead.
Server's core RTSP TCP
Player's core RTSP
Network UDP
RTSP/UDP (RTSP/TCP) RTSP/TCP
UDP
TCP
Fig. 1. Client-server protocol configuration.
3.2
Testbed configuration
Our testbed has been built in such way that we were able to test and compare performance of live video streaming service under different competing traffic loads and different signal-to-noise ratios (SNR). Testbed configuration is presented in Fig. 2. In our environment we used several computers equipped with different operating systems (OS) and different access network devices. WLAN tests were carried out on a base of running implementation of 802.11b WLAN in campus area of Tampere University of Technology. To enable LAN access we used 100 Mbps Ethernet. Both WLAN and LAN are connected via ’broker-gw’ edge router. The mobile node called ’real-client’ was IBM ThinkPad PIII laptop under Win2000 OS. It was equipped with exterior Cisco’s Aironet 350 802.11b WLAN card. To ensure mobile node performance against OS-specific issues we validated all our tests with different mobile nodes. The fixed node called ’helix-server’ was desktop PIV under Win2000 OS connected to 100 Mbps Ethernet LAN. The access point was Avaya’s ORiNOCO range extender. To hide implementationspecific issues, all our test have been carried out with only one access point. To generate competing traffic we used well-known ’iperf’ client-server utility [8]. To maintain iperf server desktop PC PIII under Linux OS connected to 100 Mbps Ethernet LAN has been chosen. This node is called ’iperf-server’ in Fig. 2. The mobile node called ’iperf-client’ was Mac PowerBook G4 under Jaguar v.10.2 OS. It was equipped with interior Airport WLAN 802.11b adapter. To evaluate performance of live streaming service the fragment of high motion pre-recoded video was chosen. Due to the fact that using the ’slta’ utility we were able to simulate live streaming service continuously, it was possible to carry out tests as long as required. To encode video to real media format Helix Producer v9.0 has been used. The resolution of video was set to 240×352 pixels, while the target rates were chosen to be 56Kbps, 150Kbps and 350Kbps. Note that paths between both Helix server and ’iperf-server’ and corresponding clients are stable, and pass only one router (’broker-gw’). We do not consider
'helix-server' LAN Ethernet
WLAN 802.11b 'broker-gw' 'real-client'
'iperf-client'
'iperf-server'
Fig. 2. Testbed configuration.
’broker-gw’ as a bottleneck of our configuration since all tests were carried out when both LAN and WLAN were in totally unloaded conditions. Additionally, we have to recall that both Helix server and ’iperf-server’ were on the same Ethernet segment. Such condition cannot also be considered as the bottleneck, since the bandwidth of fixed LAN is substantially higher than that of 802.11b WLAN. Therefore, the only bottleneck in out testbed is the WLAN. The bandwidth capability of the RealOne player was set to maximum allowable (10Mbps). The connection was assumed to be ’failed’ after 30 seconds of unsuccessful attempts and the player was not allowed to prefetch data. In this paper we propose to distinguish between five SNR levels. The ranges of SNR and corresponding user-friendly channel conditions were chosen as follows (note that the other partitioning of SNR is also possible): ≤ 10dB (very bad); 10 – 20dB (poor); 20 – 30dB (fair); 30 – 40dB (good); ≥ 40 (excellent). Additionally to different SNR ranges, we performed our tests under different competing traffic volumes. We used both UDP and TCP competing traffic, which were generated by ’iperf-sever’ and received by ’iperf-client’. Traffic characteristics are presented in Table 1. One can note that in certain cases the traffic volume increases the maximum theoretical throughput of the WLAN. It was done to be able to test the performance of live video streaming service in highly overloaded network conditions. The initial window size of TCP connections was chosen to be 60Kbps. We found out that with such choice TCP connections can potentially achieve the maximum throughput.
Table 1. Parameters of competing traffic patterns.
Type
Number of streams
Target bandwidth, Mbps
Window, Kbps
UDP
4
2
UDP
4
1
– –
TCP
4
–
60
To capture traffic and obtain statistics we used Ethereal software [9] package in conjunction with post processing Perl scripts. 3.3
End-to-end performance testing
Firstly, we had to explore performance characteristics of WLAN’s access point. Several advanced UNIX-based utilities [8] have been used. We obtained end-to-end performance parameters of the path between ’iperfserver’ and ’iperf-client’ using the ’iperf’ utility. Performance parameters of ’iperf-server’ – ’iperf-client’ and ’helix-server’ – ’real-client’ paths are similar since ’iperf-client’ and ’real-client’ were on the same WLAN access point, while ’iperf-server’ and ’helix-server’ were on the same Ethernet LAN segment. To obtain stable characteristics of WLAN we had to measure them on a wide time scale. To get values for each column of Table 2 we performed 120 minutes of testing. The following parameters were of particular interest: maximum throughput of WLAN, end-to-end round trip time (RTT) and jitter. Table 2. Performance parameters of unloaded WLAN under different SNR ranges.
Parameters
SNR ≤ 10
Max throughput, Mbps 0.22
4
10 – 20
20 – 30
30 – 40
≥ 40
0.35
3.84
3.86
3.93
Min RTT, ms
5.79
4.02
3.09
3.32
3.23
Avg. RTT, ms
12.41
8.20
6.31
4.84
3.76
Max RTT, ms
371.59
107.94
43.53
34.26
12.29
Jitter, ms
135.92
31.02
4.06
4.02
4.47
Results
In our testbed environment we evaluated the performance of live video streaming service in WLAN environment under different SNR levels and different compet-
ing traffic loads. Traffic was generated by user of ’real-client’ node by posing the request on live streaming video from ’helix-server’. The summary of statistics under different SNR ranges and unloaded network condition is presented in Table 3. Despite TCP’s 3-way handshake procedure was successful and mean throughput was measured to be 0.22Mbps (Table 3), one can note that the RTSP connection has not been established when SNR was ≤ 10 dB. It stems from the fact that under ≤ 10dB condition the quality of the air interface was very unstable and, therefore, there were often bandwidth renegotiation performed by RTSP. After the server had failed to establish connection in 30 seconds the connection request was rejected by the client. Table 3. Summary of statistics under unloaded network condition and different SNR ranges. Parameters
SNR ≤ 10
10 – 20
20 – 30
30 – 40
≥ 40
3-way handshake, ms
0.0063
0.0055
0.0034
0.0026
0.0019
Conn. est. phase, ms
–
5.109
4.790
4.605
1.356
Avg. packet size, bytes –
1036.92
1063.34
1065.66
1069.33
Avg. packets per second –
42.80
40.28
41.56
41.16
Avg. throughput, Mbps –
0.355
0.343
0.354
0.352
We have to note that when SNR was in 10 – 20dB range the RTSP connection was successfully established and it did not take substantially more time compared to 20 – 30dB range. It should also be noticed from Table 3 that when SNR increases, the average size of IP packet gets larger. At the same time, to maintain the constant mean rate as required by our live streaming video the number of packets per second also increases. Note that the time taken by 3-way handshake procedure has often been claimed as a drawback for real-time services [10]. One can realize that this time is relatively small compared to the time of RTSP connection establishment. The latter one is at least one thousand times higher. We found that the most critical point in live streaming service is RTSP connection establishment phase. Indeed, there were cases when the SNR was fluctuating a lot while the RTSP connection was setting up. It caused a lot of errors at the wireless link which result in frequent bandwidth negotiations. Therefore, sometimes client’s connection establishment timer had expired before the connection was set up. Then, we provide the same tests given the different load conditions of the WLAN (Table 1). The results are given in Table 4 where in each cell the first value is for the case of UDP 4×2Mbps competing traffic, the next one is for UDP 4×1Mbps and the last one is for TCP 4×60Kbps. The competing traffic adds additional (to that given by SNR) fluctuations to the bandwidth available at the path between client and server and makes the bottleneck at the WLAN.
Table 4. Summary of statistics under load conditions and different SNR ranges (UDP 4×2Mbps/UDP 4×1Mbps/TCP4×60Kbps). Parameters
SNR 30 – 40
≥ 40
3-way handshake, ms
– /9.893/0.250
– /0.016/0.032
Conn. est. phase, ms
– /28.52/24.341
– /7.61/1.303
Avg. packet size, bytes
–/–/–
– /1037.78/1060.37
Avg. packets per second
–/–/–
– /39.95/40.51
Avg. throughput, Mbps
–/–/–
– /0.332/0.344
We have to note that the RTSP connections have not been established under considered types of competing traffic when the SNR were in 00 – 10dB, 10 – 20dB and 20 – 30 ranges. These ranges are omitted in Table 4. The cause was that TCP was not able to complete 3-way handshake procedure, which is necessary to establish RTSP connection. The live video streaming service did not work when SNR was ≥ 40dB and the network was loaded by four UDP traffic source each of which targeted on 2Mbps. Particularly, the client permanently failed to perform TCP’s 3-way handshake. The duration of 3-way handshake procedure when SNR was in 30 – 40 range and the network was loaded by four 1Mbps UDP source was roughly ten times higher than that of the network in unloaded conditions (Table 3). At the same time duration of RTSP connection establishment phase was much greater. However, the perceived quality of live video streaming service was good and can be roughly compared to the perceived quality in the network in unloaded conditions. Almost similar observations were made when the network was under load of four competing TCP connections each of which had 60Kbps initial window size. However, one can note that the duration of RTSP connection establishment phase was smaller compared to the previous case. It can be explained by behavior of competing TCP connections whose bandwidth changes dynamically depending on losses. Note that from the other side UDP connection tries to get as much as it needs. The quality of the picture was sometimes slightly deteriorated, i.e. some frames were blurry and truncated. However, the picture quality was acceptable. As was expected live streaming service did not operate when the networks was loaded by four UDP sources each of which tried to get 2Mbps. When the target bandwidth of each competing source has been decreased to 1Mbps Helix server and RealOne player have completed 3-way handshake and established RTSP connection. However, it took much longer time compared to unloaded network conditions. Additionally, due to high bandwidth fluctuations the server was not able to use live video streaming over UDP and actual streaming was performed over TCP. The perceived quality was unacceptably bad: there always were long pauses without picture and even without voice, most frames were corrupted. Despite of that the service operated without brake.
One can note that durations of both 3-way handshake and connection establishment phase were smaller when the network was loaded by four TCP connections. However, even worse picture and voice quality were perceived. In this case the server was also not able to use RTSP over UDP and streaming was again performed over TCP. The throughput graphs for each case are given in Fig. 3, Fig. 4 and Fig. 5.
5
Conclusions
Summarizing, we conclude that it is not possible to support multimedia services like live video streaming on commercial basis over current evolution of IEEE 802.11b WLAN. Despite a common belief live streaming services are not ready for wide implementation in hot-spot areas where both high traffic volume and relatively weak signal strength (less than 30 dB) may deny the service easily. Our measurements have shown that live video streaming service cannot be successfully implemented over wireless medium. The usage of TCP at the connection establishment phase may easily deny the the service even the network in unloaded conditions. The live streaming service performs well in presence of any type of considered competing traffic, only when excellent channel conditions (greater than 40dB) are met. However, the perceived quality becomes unacceptable in presence of any type of competing traffic with 30 – 40dB. High volume of competing traffic easily denies service when the signal strength is under 30dB. The major problem is that the client cannot set up RTSP connection with the server. Frequent bandwidth fluctuations, caused by both SNR and competing traffic loads, stimulated numerous bandwidth negotiations, and therefore, connection establishment’s timer often expires before RTSP connection sets up. However, these bandwidth fluctuations do not actually indicate that the streaming service fails due to scarce of the bandwidth, since the actual streaming is performed over bandwidth greedy UDP protocol. Additionally, in those cases when video streaming is performed over TCP the QoS becomes unacceptable.
References 1. M. Droma, I. Ganchev, G. Morabito, R. Narcisi, N. Passas, S. Paskalis, V. Friderikos, A. Jahan, E. Tsontsis, C. Bader, J. Rotrou, and H. Chaouchi. Always best connected enabled 4G wireless world. In IST Mobile and Wireless Communications Summit, 2003. 2. F. Fitzek, A. Kopsel, M. Krishnam, and M. Reisslein. Providing application-level QoS in 3G/4G wireless systems: A comprehensive framework based on multi-rate CDMA. IEEE Wireless Communications, 9(2):42–47, April 2002. 3. G. Leijonhufvud. Multi access networks and always best connected (ABC). In Proc. of MMC workshop, Berlin, Germany, November 2001. 4. National science foundation workshop on network research testbeds. Final report. Available at: http://gaia.cs.umass.edu, NSF, 2002.
5. T. Kuang and C. Williamson. A measurement study of realmedia audio/video streaming traffic. In Proceedings of SPIE ITCOM, pages 68–79, Boston, USA, July 2002. 6. Y. Koucheryavy, D. Moltchanov, and H. Jarmo. Impact of mobility on entertainment services’ performance in heterogeneous wireless environment. In Submitted to ATNAC’2003, Melbourne, Australia, December 2003. 7. Helix server. Available at: http://www.realnetworks.com/. 8. NLANR. Network performance and measurements tools. Available at: http://dast.nlanr.net/npmt/. 9. Etherial software. Available at: http://www.ethereal.com/. 10. J. Kurose and K. Ross. Computer networking: A top-down approach featuring the Internet. Addison Wesley, 2nd edition, 2003.
Throughput, Kbps
Throughput, Kbps
5
4 .10
5
5
3 .10
5
2 .10
5
1 .10
4 .10
5
3 .10
5
2 .10
5
1 .10
0
0
75
150
225
0
300
0
75
150
Time, s.
225
300
Time, s.
(a) ≥ 40dB
(b) 30 – 40dB
Throughput, Kbps
Throughput, Kbps
5
4 .10
5
5
3 .10
5
2 .10
5
1 .10
4 .10
5
3 .10
5
2 .10
5
1 .10
0
0
75
150
225
300
0
0
75
150
Time, s.
(c) 20 – 30dB
225
300
Time, s.
(d) 10 – 20dB
Fig. 3. Throughput under different SNR rages and unloaded network condition.
Throughput, Kbps
Throughput, Kbps
5
4 .10
5
5
3 .10
5
2 .10
5
1 .10
4 .10
5
3 .10
5
2 .10
5
1 .10
0
0
75
150
225
0
300
0
75
150
Time, s.
225
300
Time, s.
(a) UDP 4×1Mbps
(b) TCP 4×60Mbps
Fig. 4. Throughput under different load conditions and ≥ 40dB.
Throughput, Kbps
Throughput, Kbps
5
4 .10
5
5
3 .10
5
2 .10
5
1 .10
4 .10
5
3 .10
5
2 .10
5
1 .10
0
0
75
150
225
Time, s.
(a) UDP 4×1Mbps
300
0
0
75
150
225
300
Time, s.
(b) TCP 4×60Mbps
Fig. 5. Throughput under different load conditions and 30 – 40dB.
Publication 6 Y. Koucheryavy, D. Moltchanov, and J. Harju, ”Impact of Mobility on Entertainment Services’ Performance in Heterogeneous Wireless Environment,” in Proc. of the Australian Telecommunications, Networks and Applications Conference, ATNAC’03, Melbourne, Australia, December 2003, CD ISBN 0-646-42229-4. c Copyright 2003
Impact of Mobility on Entertainment Services’ Performance in Heterogeneous Wireless Environment Y. Koucheryavy, D. Moltchanov, J. Harju Institute of Communication Engineering Tampere University of Technology Tampere, Finland {yk,moltchan,harju}@cs.tut.fi
Abstract — Emerging multimedia services are becoming very attractive for next-generation wireless networks. However, both limited quality of service (QoS) support and lack of bandwidth at the air interface restrain their wide deployment in current evolution of third generation (3G) wireless systems. Mp3-based entertainment applications, such as file downloading and streaming which are common nowadays in the Internet, are supposed to be very challenging and demanding services for nextgeneration wireless networks. In this paper we consider mp3 traffic delivery using two wireless access technologies, which are claimed to be used in next-generation wireless networks, wireless local area network (WLAN) and general packet radio service (GPRS). We provide our experiments under different user’s mobility patterns and compare the QoS of these services provided over wireless access with local area network (LAN) access technology. Additionally, we practically investigate specific issues arising in QoS assurance for both mp3 streaming and downloading services. 4G All-IP network, WLAN, GPRS, multimedia services, testbed, mobility patterns
I.
INTRODUCTION
Nowadays, a considerable part of research activities in telecommunications are switching towards development of third and fourth generation (3G/4G) IP-based mobile networks. The major motivation behind that is to introduce a common service platform and transport facilities for future composite Internet mobile network. New multimedia services attracted for next-generation wireless networks, such as downloading and streaming, are striving hard towards commercial market. Both limited QoS support and scarce bandwidth of wireless environment can restrain their wide deployment in 3G/4G. The most critical layers for QoS support in next-generation wireless networks are physical, data link and transport. In this paper we restrict our attention to transport layer only. We intend to check whether the particular multimedia services can already be smoothly implemented on commercial basis over currently available wireless access networks or not.
In this paper we consider two access networks which are claimed to be used in next-generation wireless mobile systems. These are wireless local area network (WLAN, IEEE 802.11b) and current state (2.5G) of 3G evolution – general packet radio service (GPRS). Note, the latter is an overlay network, and it is already implemented in most GSM networks. Recently only a few studies have addressed performance of WLAN and GPRS under different signal-to-noise ratios (SNR), mobility patterns and load conditions. In [1] performance of real-time streaming service over WLAN has been investigated under different SNR. Authors in [2] considered performance of synthetic TCP load in GPRS network, while performance of real services was not taken into account. In [3] we considered performance of real time video streaming services in WLAN environment under different ranges of SNR ratios and concurrent load conditions. However, taking into account mobility of the user one cannot expect that the mobile terminal experiences almost the same SNR range during the whole duration of a session. Due to this fact it is needed to explicitly consider the mobility behavior of the user. Additionally, nextgeneration mobile networks introduce an additional notion – an Always Best Connected (ABC) concept [4]. The ABC will allow users to choose the most suitable RAN at any instant of time during the whole duration of a session. Particularly, this feature is claimed to be very attractive for users with complex and variable mobility behavior. It is frequently claimed that user services should be smoothly implemented in heterogeneous environment. In our study we extend previous work by implementing series of tests for both wireless access networks and compare obtained parameters with fixed local area network (LAN, IEEE Ethernet) access, which is currently de-facto standard in multimedia networking. Indeed, it is highly anticipated [5] that testing of real implementations can bring better understanding and new knowledge in the area. It is necessary to mention here that WLAN and GPRS are not competitors, but associates because in next-generation wireless networks the different roles are assigned to different access networks. We concentrate our attention on tracing of real traffic, but not simulated one. We consider two mp3-based entertainment applications: mp3 file downloading and streaming as they are
supposed to be very challenging and demanding services for next-generation wireless networks. From traffic transmission point of view, mp3 downloading service has essential peculiarity like relatively small time of connection/flow (transmission phase) duration. Indeed, traffic load simulators allow us to obtain some type of stationary traffic processes. However, taking into account average sizes of mp3 files, MPEG-based compression techniques and unpredictable nature of music sources we can assert that mp3based application is intend to produce traffic which cannot be stochastically described by stationary stochastic processes. Even in those cases when, based on certain available information about codecs and music sources, one can expect stationary behavior of mp3 traffic, we still not able to prove it statistically because of short duration of mp3 sessions. Second type of mp3 applications we considered here is a streaming-based one. Internet Live Radio stations are widely spread nowadays in the global network and QoS perceived by end-user varies in average from good up to excellent. For example, some of stations are already available in 128 Kbps stereo format and can be streamed continuously for several days without brake in connection via public Internet. The rest of the paper is organized as follows. Some general prerequisites are considered in Section II. Testbed configuration is given in Section III. Carried measurements are outlined in Section IV. In Section V we discuss obtained results. Conclusions are drawn in last section. Appendix to the paper contains graphical representation of obtained results. II.
TESTING PREREQUISITIES
A. GPRS and WLAN GPRS phones' data exchange is unsymmetrical. Data uploading (e.g. sending an e-mail message) is slower than downloading (e.g. reading an e-mail message). The higher number indicates the phone's ability to employ channels for downloading while the smaller one indicates the ability to employ channels for uploading. Thus, GPRS (3+1) phone can send data by one and receive it by three channels. However, downloading data by three channels does not mean that the data transmission rate grows three times. 40 kbps and other data transmission rates are distributed between all employed channels. Initially, GPRS supports up to 8 channels on each bearer frequency. In order to better utilize the channel’s capacity, four different coding methods are used (correspondingly CS-1 9.05, CS-2 13.4, CS-3 15.6 and CS-4 21.4 kbps per one channel). Currently, most commercial networks support only two coding schemes (CS-1 and CS-2) and, therefore, theoretical GPRS’ rate cannot be achieved. It may be the case that all channels will never be used for GPRS data transfers since GSM voice calls employ the same channels. The other reason is that it would require greater computing efficiency (which results in greater power consumption) that phones lack. It is known that WLANs were initially developed to match requirements of non real-time services like file transfers, while cellular networks are targeted on constant-rate voice communications. In order to enable IP-based multimedia
services for nomadic users, it is needed to study a system architecture that combines short-range broadband wireless access (IEEE 802.11b) with cellular access technologies (GPRS). The main reason to consider such heterogeneous environment is that practically it is almost impossible to define a RAN that combine all advantages of different access technologies [6]. B. Multimedia Traffic Multimedia applications are continuously growing in popularity. The availability of high-speed access networks is the primary reason behind that. Today, it is strategically necessary to support these services over wireless networks. Basically, multimedia traffic consists of one or more media streams and can be characterized by strict delay requirements while can tolerate some losses. It is supposed that applications emerging from Internet will become capable of defining the required QoS level soon. However, currently within almost all networks multimedia traffic is treated similar to ordinary best effort data traffic, which often do not require strict delay bounds. Therefore, it is crucial to predict the QoS degradation that may be experienced by multimedia applications over wireless access networks. In our paper we consider two entertainment applications: mp3 file transfer and mp3-based Internet Live Radio. Depending on bandwidth requirements, these applications fall into two major categories: limited bulk transfer (mp3 downloading) and controlled transfer (Internet Live Radio). Note that from user point of view both services behave quite similar and can be described by two phases: prefetching phase and playing phase. While in prefetching phase application stores data and then turns into playing phase. When both applications are in prefetching phase they use all available bandwidth to prefetch data. However, playing back, mp3 player continues downloading at maximum available bandwidth while Internet Live Radio restricts itself to the certain necessary rate. We also would like to note that it is almost impossible to characterize mp3 traffic analytically. It stems from both relatively short duration of mp3 session and unpredictable nature of mp3 content. Because of these reasons we cannot model mp3 traffic as a stationary process which is often used for traffic modeling, since it is not possible to judge whether the traffic is stationary or not. It is also too complicated to model Internet Live Radio traffic mathematically since we are dealing with source-controlled service. Therefore, real measurements are the only way to characterize these services. C. TCP-related Issues Logically, streaming services have to be implemented over UDP/RTP/RTCP set of protocols. However, we found that all Internet Live Radio stations are currently based on TCP protocol. One possible reason behind this fact is that the overall Internet quality has increased significantly each year. Indeed, using ordinary LAN access we traced short prefetching and very rare re-prefetching periods. It allows to confirm the thesis that the overall quality in public Internet is growing in terms of average delay and loss [7].
As far as TCP is a transport layer protocol used in implementations of downloading and streaming services of mp3-based applications, we cannot neglect wireless TCP issues in our study. There are a lot of studies devoted to TCP performance evaluation in wireless environment. For example, it is highly anticipated that isolation of wireless link from the rest of the network is promised approach [8]. The other approaches propose some improvement on end-to-end basis. However, most of proposed approaches are verified on simple simulation studies and do not have real implementations. Therefore, currently wireless TCP implementation cannot be easily included into testbed of real environment. Additionally, we note that practically we already use ordinary ’fixed’ TCP implementations over wireless links, for example, for Internet Live Radio service over WLAN. Note that in this scenario the perceived QoS and performance are good in non-nomadic and limited nomadic cases. Based on these arguments in our study we admitted the usage of ordinary ’fixed’ TCP implementation provided by operating systems we used in testbed. III.
TESTBED
A. Overall Configuration Our testbed has been built in such way that we were able to test and compare different access networks performance for different workloads and mobility patterns. We used three access networks WLAN, GPRS and fixed Ethernet LAN. Note that the Ethernet LANs are currently seen as a de-facto standard in multimedia networking. We chose two types of traffic workloads, which are basically assumed to be demanding in next-generation wireless networks. These are limited bulk and controlled transfers. Testbed configuration is presented in Fig.1. In our testbed environment we used several computers equipped with different operating systems (OS) and different access network devices. The mobile node called ‘mobila’ was Mac PowerBook G4 under Jaguar v.10.2 OS. It was equipped with WLAN 802.11b (interior) and Bluetooth facilities (used to connect computer with GPRS adapter namely GPRS phone). To ensure mobile node performance against OS-specific issues we validated all our tests with different mobile node which was based on IBM ThinkPad PIII laptop under Win2000 OS and equipped with exterior Cisco’s WLAN 802.11b card and same Bluetooth adapter. To emulate the mp3 file server we used desktop PC PIII under Win2000 OS connected to 100 Mbps Ethernet LAN. This node is called ‘mp3-source’ in Fig. 1. Again, to ensure fixed node performance against OS-specific issues we validated our tests with different fixed nodes, which were based on different PC PIII with Linux OS RH7.2. This is because the end-user perceived QoS may vary significantly from one OS to another. Additionally, it was motivated by the fact that different operating systems give slightly different performance results. The detailed analysis of this fact is out of scope of this paper, but in our tests we have decided to use a mobile-fixed couple Mac PowerBook – PC Linux.
To evaluate the performance of live mp3 streaming service we chose one of the stable Internet Live Radio. This machine is called ‘mp3-live’ in our testbed (Fig. 1). It is clear that paths between the end user and both mp3 file server and Internet Live Radio node are different. To overcome this obstacle we performed traffic tests of each destination over relatively large time periods. We found that the both paths are stable, and therefore, can be used in our testbed. mp3-live SGSN
GGSN GPRS CN
GPRS RAN
WLAN 802.11b
mobila
Pubilc Internet
broker-gw LAN Ethernet mp3-source
Fig. 1 Testbed configuration. In our testbed environment we evaluated both mp3-based services under different user’ mobility patterns. We use term mobility pattern here to denote global behavior of user. We define three global mobility patterns. These are: fixed user, walking user and driving user. Since WLANs are assumed to be only accessible in relatively small hot-spot areas, they can serve both fixed and walking users. GPRS’s access network serves all three types of users. LAN is only accessible for fixed user. Different types of mobility covered by different access networks, which are presented in Table I. TABLE I DIFFERENT TYPES OF MOBILITY COVERED BY DIFFERENT ACCESS NETWORKS
Type of mobility Fixed Walking Driving
Type of access network LAN WLAN GPRS YES YES YES NO YES YES NO NO YES
B. Tools and Environment WLAN tests were carried out on a base of running implementation of WLAN 802.11b in campus area of Tampere University of Technology. In order to enable GPRS access we used tandem configuration of Nokia 7650 terminal and Mac PowerBook G4 connected via Bluetooth. We consider that Bluetooth access is not a bottleneck in our testbed, because both devices were located near by each other, i.e. signal strength and channel error rate were in acceptable bounds. In our testbed we used GPRS connection via one of the commercial Finnish GPRS networks, which is currently offer GPRS (3+1) time slots allocation with CS-2 coding scheme. Therefore, theoretically our GPRS is capable to achieve 36 kbps downlink and 12 kbps uplink transfer rates. To enable LAN access we used 100 Mbps Ethernet. Both WLAN and LAN are connected to campus network via
‘broker-gw’ edge router (Fig.1). In our tests this point is used as a reference one. To capture generated traffic and to obtain statistics we used Ethereal software [9] package in conjunction with post processing Perl scripts. C. End-to-end Performance Testing Before performance evaluation of mp3-based services we had to explore performance characteristics of each access network. Several advanced UNIX-based utilities [10] were employed. We obtained end-to-end performance parameters via three different access networks, i.e. for ‘mobila’-‘mp3source’ route via WLAN, GPRS and LAN. The values for these parameters are presented in Table II, where “T. – E.R.” denotes the path between terminal and ‘broker-gw’ edge router. The following parameters are of particular interest: maximum throughput of access network, end-to-end round trip time (RTT), loss probability, jitter, and RTT between nomadic node ‘mobila’ and ‘broker-gw’ edge router. Using the latter parameter it is possible to characterize the particular access network because behind the edge router the traffic generated by end node traverses through the fixed part of campus Internet, which gives quite low delays and error probability. In order to characterize GPRS access network we used GGSN as an edge router. As we can see the differences between end-to-end RTT and RTT between end node and edge router are relatively small. Note the fact that in the case of GPRS such approach gives us an estimation of joint access network and core network parameters. However, since the GPRS core network is built on a base of high-speed wired technology for which the values of delays and probability of loss are much smaller than in GPRS access network, we can neglect it. To obtain characteristics of each access network we had to measure them on a wide time scale. To get values for each column of the Table II as well as for other measurements’ results throughout the rest of the paper we performed our tests five times. Duration of each test is 120 minutes. TABLE II ACCESS NETWORKS PARAMETERS
Parameters Max throughput, Mbps Min RTT, ms Avg. RTT, ms Max RTT, ms Loss probability Jitter, ms Min RTT (T. – E.R.) Avg. RTT (T. – E.R.) Max RTT (T. – E.R.)
Access network LAN WLAN GPRS 95.6 3.96 21.2E-3 0.215 2.39 582.39 0.43 6.76 705.97 4.56 63.45 1174.29