Quality of Service Oriented Traffic Engineering

Home , Teletraffic engineering, Traffic mix

Quality of Service oriented Trafﬁc Engineering Methods for Multi-Service Cellular Networks

Dienstgüte-orientierte verkehrstheoretische Methoden für die Bereitstellung unterschiedlicher Dienste in zellularen Netzen

Der Technischen Fakultät der Universität Erlangen-Nürnberg zur Erlangung des Grades

DOKTOR INGENIEUR vorgelegt von

LARISSA N. POPOVA

Erlangen – 2009 Als Dissertation genehmigt von der Technischen Fakultät der Friedrich-Alexander-Universität Erlangen-Nürnberg

Tag der Einreichung: 07. Dezember 2009 Tag der Promotion: 12. February 2010 Dekan: Prof. Dr.-Ing. Reinhard German Berichterstatter: Prof. Dr.-Ing. Villy Baek Iversen Prof. Dr.-Ing. Wolfgang Koch To my husband Denis and my sons Konstantin and Maxim

Acknowledgment

First of all, I would like to thank my supervisor Prof. Wolfgang Koch, for giving me the opportunity to pursue my Ph.D. at his Institute of Mobile Communication and for providing an environment where it was a pleasure to work. I am grateful to him for the fruitful discussions on my work, for his continued support and faith in the success of my research. I am indebted to Prof. Villy Baek Iversen for his kind-hearted mentoring, for reviewing my thesis, and for making my memorable stay at Technical University of Denmark, Lyngby, possible. I am very grateful to Dr. Wolfgang Gerstacker for his interest in my work, for his great encouragement, and for finding time to participate in the defense of this work. I would also like to thank all my colleagues at the Telecommunications Laboratory in Erlangen for creating pleasant and humorous atmosphere. Special thanks go to my roommate Armin Schmidt, for interesting scientific and non- scientific discussions, as well as for his technical LaTeX support. Finally, I would like to give thanks to my parents, Liudmila and Nikolay for their continued love and support. And my deepest thanks I owe to my husband, Denis, who tireless supported me to the realization of my thesis, for his patience and love.

Contents

Introduction 1

Contribution of the Thesis 3

Einführung 5

Zusammenfassung 7

1. Universal Mobile Telecommunications System (UMTS) 9 1.1. SystemArchitecture ...... 9 1.2. WCDMAConcept ...... 10 1.3. RadioResourceManagement...... 12 1.4. Techniques for Radio Resource Utilization ...... 13 1.4.1. PowerControl ...... 13 1.4.2. Handover Control ...... 14 1.4.3. Call Admission Control ...... 14 1.4.4. Congestion Control ...... 14 1.5. Summary ...... 15

Part I. Traﬃc Modeling for Wireless Networks 17

2. Traffic Concepts and Models 19 2.1. MathematicalBackground: BasicModel ...... 19 2.1.1. Birth-DeathProcess ...... 19 2.2. Markovian Queuing Systems ...... 21 2.2.1. M/M/1System...... 22 2.2.2. Erlang’s Loss System ...... 24 2.2.2.1. Blocked-Call-Cleared – Erlang B ...... 25 2.3. Traffic Models for Wireless Networks ...... 26 2.3.1. General Remarks on Existing Traffic Models: Drawbacks and Chal- lenges ...... 26 2.4. Summary ...... 29

3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio Interface 31 3.1. Motivation ...... 31 3.2. SingleCellCapacityModel ...... 32 3.3. Connection dependent Model with Transmission Rate Reduction Policy . . 33 3.4. Multi-CellCapacityModel...... 37 viii Contents

3.4.1. Restricted Network Accessibility - Modeling of Other-Cell Inter- ference...... 38 3.4.2. Proposed algorithm ...... 39 3.5. AnalyticalSystemPerformance...... 43 3.5.1. Performance evaluation analysis ...... 43 3.6. Summary ...... 45

4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA Radio Interface 47 4.1. MotivationandRelatedWork...... 47 4.2. Multi-CellTwo-LevelCapacityModel ...... 48 4.2.1. Connection Level ...... 49 4.2.2. A Novel Model for Packet Level Performance Evaluation ...... 49 4.2.2.1. Blocked-Call-Held (BCH) Model (Fry-Molina Model) at packetlevel ...... 50 4.2.2.2. Blocked-Call-Interfered (BCI) ...... 52 4.2.3. Performance Measures ...... 54 4.3. PerformanceResults ...... 56 4.3.1. Impact of the Trafﬁc Mix ...... 56 4.3.2. Analyzing trends by load change ...... 58 4.3.3. Adjustment of loss rates at packet level ...... 59 4.3.4. Impact of the Activity Factor ...... 60 4.4. Summary ...... 61

5. Extended Analysis of Two-Level Performance Model for Multi-Rate Delay System with WCDMA Radio Interface 63 5.1. MotivationandRelatedWork...... 63 5.1.1. On the Generality of the Algorithm ...... 63 5.2. Uniﬁed Analytical Trafﬁc Model ...... 64 5.3. SystemPerformanceAnalysis...... 66 5.3.1. CallAdmissionandHandlingPolicy ...... 66 5.3.2. Comparison of Blocked-Call-Buffered (BCB) model with Blocked- Call-Held(BCH) ...... 67 5.3.3. Behavior of the System with Soft Blocking ...... 67 5.3.4. Probability of Getting Service without Priority Classes ...... 68 5.4. Summary ...... 70

Part II. Traﬃc Management for Wireless Networks 73

6. Mobile-to-Mobile: A Novel Concept for Spectrum Eﬃcient Data Transfer in WCDMA 75 6.1. Motivation ...... 75 6.2. Cooperative Communications in Wireless Networks (General Concepts, Strategies,Principles) ...... 77 6.3. Mobile-to-Mobile(M2M)Concept ...... 80 6.4. ConceptAnalysisandEvaluation...... 83 6.4.1. Model Characteristics and Assumptions ...... 83 Contents ix

6.4.2. RadioInterfaceRestrictions ...... 83 6.4.3. M2M Propagation Model ...... 84 6.4.3.1. Pathloss and Shadowing ...... 84 6.4.3.2. Fading ...... 86 6.5. M2M-GroupOrganizationPolicy...... 86 6.6. CooperativeDataTransferPolicy...... 90 6.7. Performance Evaluation of M2M Data Transfer: Verification of Basic Func- tionality...... 94 6.7.1. Simulation Scenarios ...... 94 6.7.2. Traffic Model: Pure M2M File Transfer ...... 94 6.7.3. Performance Measures ...... 95 6.7.4. Comparison of M2M File with conventional UMTS Data Transmis- sion...... 95 6.7.5. ImpactofMulticastTechnique...... 99 6.7.6. Impact of Group Update Interval ...... 100 6.7.7. Effect of the Restrictions in the Group Organization Policy . . . . . 101 6.8. Direct Mobile-to-Mobile Data Transfer for mixed traffic scenario with ser- vicedifferentiation ...... 103 6.8.1. Traffic Model: Incorporating Speech User Population into the Model103 6.8.2. Simulation Scenarios ...... 103 6.8.3. Impact of Cross-Traffic on the M2M Performance: Uplink Interfer- ence...... 104 6.8.4. Comparison of M2M File Dissemination and Conventional UMTS Data Transmission for Mixed Traffic Scenarios ...... 106 6.9. Dependability of Mobile-to-Mobile Data Transfer ...... 109 6.9.1. Practicallyrelevantschedulingpolicy ...... 109 6.9.2. Impact of Group Size on Inter-Group Interference ...... 110 6.9.3. Effect of Restricted base station (BS) Support ...... 112 6.10.Summary ...... 117

7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer119 7.1. MotivationandRelatedWork...... 119 7.2. NetworkCodingBasics...... 121 7.3. network coding (NC) applied to M2M Data Transfer ...... 123 7.4. NumericalResults...... 127 7.4.1. Simulation Environment ...... 127 7.4.2. Trafﬁc Scenario ...... 128 7.4.3. Performance Measures ...... 129 7.4.4. Performance Comparison of Simple M2M Algorithm with network coding – mobile-to-mobile (NC-M2M) File Dissemination ...... 129 7.4.5. Impact of Extended M2M User Availability and Users’ Mobility . . 130 7.4.6. Steady State System Performance ...... 132 7.4.7. Further Beneﬁts of NC-M2M File Dissemination: Released Uplink . 132 7.5. Enhanced Network Coding for Operation on Data of Arbitrary Size . . . . 134 7.5.1. Generations: Optimized Packet Combination ...... 134 7.6. NumericalResults...... 135 x Contents

7.6.1. Memory Requirements and Computational Complexity ...... 136 7.6.2. Analysis of Relationship between Different Figures of Merit . . . . 137 7.6.2.1. Generation Distribution Strategy ...... 137 7.6.2.2. Generation Size and File Size ...... 138 7.6.2.3. Quality of Service (QoS) Requirements ...... 140 7.7. Efﬁcient Large File Distribution in UMTS supported by Network Coded M2M Data Transfer with Multiple Generations ...... 142 7.7.1. Comparison of NC-MG-m2m File Sharing with Replicate-and-Forward M2M Data Dissemination ...... 142 7.7.1.1. Practical Issues ...... 144 7.8. Summary ...... 144

8. Conclusions 147

A. Kendall Notation 149

B. Binomial-Poisson-Pascal traﬃc (BPP) Paradigm 151

C. Convolution Algorithm 155

D. Advanced Propagation Models 157

List of Acronyms 161

List of Symbols 163 PartI...... 163 PartII...... 164

List of Figures 165

List of Tables 169

Bibliography 171 Introduction

The rapid growth of wireless multimedia services and their demand for high data rates put considerable load on the valuable and limited resources of cellular wireless networks. The third generation of wireless networks makes use of the packet-switching technology to achieve higher efficiency and better utilization of the scarce radio resources. However, new opportunities entail new challenges. As the capabilities and flexibility of networks increase, the demand for radio resources becomes less homogeneous. In order to provide services in a heterogeneous environment of 3G systems more efficiently, it is necessary to predict the limits of the traffic load such that upper bounds for probabilities of call blocking and packet dropping for all classes of demands can be maintained. Typically, the downlink is the potential bottleneck, since all data transmissions have to be organized by providing individual links from a base station to each user that required a data service. As a result, the radio interface quickly becomes saturated, which leads to degradation or even loss of the service. The problem of congestion on the typically overloaded downlink channels in UMTS can be solved by exploiting the usually underused uplink frequency band for direct mobile-to-mobile communication on the currently free uplink channels for the distribution of popular content in a non-real time multicast manner.

Contribution of the Thesis

The utilization of the air interface is one of the key optimization problems of cellular radio networks. In this thesis the radio access part of Universal Mobile Telecommuni- cations System (UMTS) has been investigated. We model the UMTS air interface to introduce efficient analytical algorithms for system performance estimation and propose methods for solving some of the radio resource optimization problems within the network. The work consists of two independent, yet closely related parts. The first part focuses on the development of efficient algorithms for analytical performance prediction of multi-service UMTS networks, and is dedicated to the analysis of the performance of the proposed algorithms. Algorithms that allow for leveraging the UMTS air interface resources constitute the second part of the thesis. The monograph is organized as follows. In Chapter 1 a brief introduction of UMTS is given. The basic principles of its air interface (Wideband Code Division Multiple Access (WCDMA)) are explained and the most important algorithms for radio resource management (RRM) are introduced. Chapter 2 introduces some fundamentals of teletraffic theory and gives an overview of existing traffic models for wireless networks, followed by a definition of the challenges of traffic engineering in third generation wireless networks. In Chapters 3-5 the novel analytical algorithms for system performance estimation are introduced and the traffic management techniques are applied to the UMTS air interface and incorporated into the proposed analytical approaches to specify how resources have to be adequately engineered to meet quantitative performance objectives. In Chapter 6 algorithms for boosting the spectral efficiency of UMTS are proposed. Our concept is based on the uplink/downlink traffic imbalance in UMTS. It enables direct mobile-to-mobile (M2M) data transfer allocating users to temporarily unused uplink channels. To further improve the system performance network coding technique is embedded into the M2M algorithm in Chapter 7 as a solution to the scheduling problem in the distributed dynamic environment of large-scale UMTS networks. Finally, the content of this monograph is summarized.

Einführung

Drahtlose Kommunikationsnetze der dritten Generation nutzen die Paketvermittlungs- technik um eine höhere Effizienz der Übertragung und eine bessere Ausnutzung der oh- nehin knappen Funkressourcen zu erzielen. Durch die gesteigerte Flexibilität und die erweiterten Fähigkeiten der Netze wurden die Anforderungen an die Funkressourcen inhomogener. Um in einer heterogenen Umgebung unterschiedlicher 3G-Systeme effizienter Dienste anbieten zu können, ist es erforderlich, die Grenzen der zu erwartenden Verkehrslast abzuschätzen, damit z. B. obere Schranken für die Blockierwahrscheinlich- keit und die Paketverlustwahrscheinlichkeit aller Dienstarten eingehalten werden kön- nen. Typischerweise stellt der Downlink den Engpass eines solchen Systems dar, da alle Übertragungen durch individuelle Verbindungen von der Basisstation zu jedem einzel- nen, einen Dienst anfordernden Benutzer koordiniert werden müssen. Aufgrund dieser Tatsache kann es schnell zu einer Überlastung der Luftschnittstelle kommen was da- zu führt, dass die vorgegebene Dienstgüte nicht mehr gewährleistet werden kann, bis hin zum Verlust der Verbindung. Effiziente und präzise analytische Algorithmen müs- sen bereitgestellt werden, um eine gewisse Dienstgüte selbst unter unterschiedlichsten Übertragungsbedingungen sicherstellen zu können. Solche Algorithmen zielen auf die Analyse und Optimierung der Kapazität und der Qualität eines Funknetzes ab. Das häufig auftretende Problem überlasteter Downlink-Ressourcen kann unter Verwendung einer neuen Strategie für die Informationsverteilung in einem Mobilfunknetz gelöst werden, die auf der Einbettung einer Peer-to-Peer Datenübertragung in die hierarchi- sche Architektur von UMTS beruht. Damit wird der kooperative Austausch von Daten zwischen mobilen Endgeräten ermöglicht. Das Konzept macht sich die ungleichmäßi- ge Auslastung der Up- und Downlink-Frequenzen in 3G-Netzen zunutze und erhöht die spektrale Effizienz von UMTS durch direkte Mobile-to-Mobile Datenübertragungen und dynamische Zuweisung von Nutzern zu zeitweise ungenutzten Uplink-Kanälen.

Zusammenfassung

Die bessere Ausnutzung der Funkressourcen ist eines der zentralen Optimierungspro- bleme bei zellularen Funknetzen. In dieser Arbeit wird die Luftschnittstelle von UMTS betrachtet, um einerseits leistungsfähige analytische Algorithmen vorzustellen, die eine schnelle Berechnung der Blockierwahrscheinlichkeiten und der resultierenden Netzwerk- Kapazität für eine gegebene zellulare Netzstruktur ermöglichen, und andererseits Me- thoden zu entwickeln um einige Optimierungsprobleme hinsichtlich der Funkressour- cenzuteilung innerhalb des Netzes zu lösen. Diese Arbeit besteht aus zwei voneinander unabhängigen, jedoch eng miteinander verwandten Teilen. Das Hauptaugenmerk des ersten Teils der Arbeit liegt auf der Entwicklung leistungsfä- higer verkehrstheoretischer Algorithmen für die analytische Abschätzung der Kapazität eines UMTS Netzes, das unterschiedliche Dienstarten unterstützt. Weiterhin werden die vorgestellten Methoden bezüglich ihrer Leistungsfähigkeit untersucht. Der zweite Teil dieser Arbeit beschäftigt sich mit Algorithmen die es ermöglichen, die Luftschnittstelle von UMTS effizienter zu nutzen. Die Arbeit ist wie folgt gegliedert. Ein kurzer Überblick über UMTS findet sich in Kapitel 1. Die Grundprinzipien der WCDMA-Luftschnittstelle werden erklärt und die wichtigsten Algorithmen des Radio Ressource Management werden eingeführt. Einige Grundlagen der Verkehrstheorie werden in Kapitel 2, zusammen mit bereits bestehenden Modellen für drahtlose Netze, eingeführt. Darauf folgt eine Definition der Herausforderungen der Verkehrstheorie in Mobilfunknetzen der dritten Generation. Die Kapitel 3 bis 5 führen neuartige Algorithmen zur Abschätzung der Systemleis- tungsfähigkeit ein. Weiterhin werden Verkehrsmanagementmethoden auf die UMTS- Luftschnittstelle angewandt und die bisher vorgestellten analytischen Ansätze um diese Methoden erweitert, was es ermöglicht festzulegen wie Ressourcen geeignet zugeteilt werden können um quantitative Leistungskriterien zu erfüllen. In Kapitel 6 werden Algorithmen vorgestellt, die es erlauben die spektrale Effizi- enz in UMTS deutlich zu steigern. Das entsprechende Konzept basiert auf dem Un- gleichgewicht des Datenverkehrs im UMTS Uplink/Downlink. Es ermöglicht die direkte Mobile-zu-Mobile (M2M) Datenübertragung, indem Benutzern vorübergehend unge- nutzte Uplink-Ressourcen zugewiesen werden. Zur weiteren Verbesserung des Durchsatzes im System wurde im Kapitel 7 zusätzlich Network Coding zur Lösung des Scheduling-Problems in großen dynamischen Funknet- zen miteinbezogen. Letztendlich folgt eine Zusammenfassung der Arbeit.

Chapter 1. Universal Mobile Telecommunications System (UMTS)

In the following, the UMTS architecture is brieﬂy introduced and a more detailed overview of UTRAN is given, including essential network elements and the interfaces between them. Furthermore, various aspects of the network are discussed, such as network capacity, radio resource management, etc.

1.1. System Architecture

UMTS evolved from GSM (Global System for Mobile Communications) and utilizes the well-known architecture of GSM, but also introduces some new elements. As shown in Figure 1.1 taken from [9], the network elements are grouped by functionality into the UMTS Terrestrial Radio Access Network (UTRAN), that handles all radio related functions, and the Core Network (CN). The CN is the long-range network that transports user data to its destination. It contains an amount of switching systems (like Mobile Services Switching Center (MSC) for interworking with location databases), as well as gateways to other networks, like the Public Switched Telephone Network (PSTN) or the Internet. The CN supports two types of switching: circuit and packet switching. Both parts of the CN use the same Radio Access Network (RAN). The Gateway Mobile Switching Center (GMSC) provides edge functionality within a PSTN for circuit-switched calls. It terminates the PSTN signaling and trafﬁc formats and converts them to protocols used in mobile networks. For mobile calls, it interacts with the Home Location Register (HLR) to obtain routing information. The Gateway GPRS Support Node (GGSN) supports the edge routing function of the GPRS (General Packet Radio Service) network. The GGSN performs the task of an Internet Protocol (IP) router to external packet data networks, e.g., Internet. The interface between CN and RAN (Iu-interface) is divided into an interface for the circuit-switched part of the CN (IuCS-interface) and an interface for the packet- switched part (IuPS-interface). The RAN contains two types of nodes: Radio Network Controller (RNC) and Node B (often called base station), and is responsible for all radio related tasks like radio resource and connection management. The RNC controls resource management for one or more Node Bs, that are connected to it. The RNC is 10 Chapter 1. Universal Mobile Telecommunications System (UMTS)

the service access point for the services UTRAN provides to the CN, e.g. management of connections to user equipment (UE). The interface between RNC and Node B is called

the Iub-interface. The UE, which will be called mobile terminal (MT) in the following, is connected to a Node B via the Uu-interface. The Serving GPRS Support Node (SGSN) (part of CN) keeps track of the location of an individual MT and performs security functions and access control. A Node B supplies one or more radio cells and converts the

data ﬂow between Iub and Uu interfaces. Besides the main transfer functionality, a Node B performs some basic radio resource management operations, like inner loop power control, and handles the radio measurements for handover decision, load estimation in the cell, or admission control as requested by the RNC. System information messages are distributed by Node B according to the schedule given by the RNC. The coordination between Node B and RNC is performed on a master-slave basis.

Radio Access Network (RAN) Core Network (CN)

UTRAN Circuit-switched domain

¨ ©

PSTN

¥¦

¡ ¢ £ ¤

¥ ¦

¡ ¢ £ ¤

¡¢£ ¤

¥ ¦

Figure 1.1.: UMTS architecture [9].

1.2. WCDMA Concept

For UTRAN Direct Sequence Wideband Code Division Multiple Access (DS-WCDMA) has been chosen as the basic radio access technology. Using this radio transmission standard all users in the network are allowed to transmit on the same frequency at the same time but separated by codes. This section discusses the fundamental WCDMA concepts, like spreading and scrambling, power control, and soft handover.

Spreading As mentioned above WCDMA is a common technique used in the UTRAN air interface to distinct the user’s physical channels in both uplink and downlink. Data symbols of users are made separable by multiplying them with spreading (also called channelization) codes. This process is called spreading, because the original narrowband signal is spread over a wider frequency band depending on the system chip rate, i.e., the number of chips per time interval. In UMTS, the chip rate is 3.84 Mcps, which corresponds approximately to a system bandwidth of 5 MHz. After spreading has taken place, each data symbol is represented by 1.2. WCDMA Concept 11

a number of chips. The ratio between the chip rate and the bit rate is called the spreading factor. It implies the number of chips that are used to spread one data

symbol.

# $ % & # $ % ' # $ % # $ %) (

Figure 1.2.: OVSF code tree, SF denotes the Spreading Factor.

The spreading sequence is known at both the transmitter and the receiver. The data can be recovered at the receiver by correlating the received signal with the spreading sequence of the data stream to be recovered. Different spreading sequences should be (at least approximately) mutually orthogonal (with zero cross- correlation). To achieve orthogonality between codes with different spreading factors, the tree structured orthogonal variable spreading factor (OVSF) technique can be used. This technique was originally proposed in [3] and in Figure 1.2 the tree-construction principle is depicted.

Scrambling In addition to spreading, the transmitter also performs the scrambling operation. Scrambling is done after spreading, in order to randomize the chip sequence resulting from spreading, and to identify the source of each signal (MT, Node B). Randomization prevents the reception of correlated signals, which would lead to signiﬁcant performance losses due to interference. Scrambling codes are quasi-random chip sequences. Figure 1.3 shows the schematic illustration of scrambling and spreading in WCDMA. 12 Chapter 1. Universal Mobile Telecommunications System (UMTS)

spreading code C1

DPDCH 1

spreading code C2 scrambling code S DPDCH 2 Σ

spreading code Ck

DPDCH k

Figure 1.3.: Scrambling and Spreading in WCDMA, DPDCH denotes Dedicated Physical Data Channel.

1.3. Radio Resource Management

The capacity of a cell in a WCDMA network essentially depends on the orthogonality and number of spreading codes used. Perfectly orthogonal codes guarantee that the different physical channels do not cause mutual interference. In this case, the capacity of a radio cell is determined by the number of mutually orthogonal codes. However, in a real world system, the combination of spreading and scrambling code does not yield a fully orthogonal sequence, which means every user causes some interference to other users. Thus, the wireless interference in a WCDMA system is the main limiting capacity factor. Therefore, efficient management of radio resources is the most important task in such a system. The group of control mechanisms, which are responsible for utilization of the radio interface is called Radio Resource Management (RRM). Their tasks are to supply optimal coverage, to maintain the planned capacity, to adequately engineer system resources to meet quantitative performance objectives in terms of Grade of Service (GoS) requirements, and at the same time to provide connection quality which meets some given Quality of Service (QoS) characteristics. GoS and QoS concepts have different viewpoints. While the QoS concept views the situation from the customer’s perspective, the GoS concept considers the network aspects. In case a user requests a service from a system where all channels are currently occupied, congestion occurs. GoS is expressed in relation to how frequently the system congestion is permitted to exceed some predefined value. The traditional measure used in conjunction with congestion is the call blocking probability – the probability that a user cannot establish a call. Call blocking might occur when all user channels are busy or due to insufficient network capacity. The call is not accepted and is either rejected or allowed to wait in a queue (depending on the type of the system). Obviously, the larger 1.4. Techniques for Radio Resource Utilization 13

the value of GoS, the better is the service. Since the traffic volume is continuously growing, to maintain the GoS value within reasonable limits, the network is initially sized to have a much better GoS than the recommended one. Quality of Service (QoS) is defined in ITU-T Recommendation E.800 as the collective effect of service performance, which determines the degree of user satisfaction regarding undesired effects such as call dropping, packet delays, etc., concerning a particular service. By means of sophisticated RRM algorithms, like load control, power control and handover the systems can be kept in a stable operational state for all traffic and radio environment conditions.

1.4. Techniques for Radio Resource Utilization

All functions for handling the radio interface resources of UMTS can be expressed by the term Radio Resource Utilization (RRU). Those, which are relevant for this thesis are described below.

1.4.1. Power Control

A mobile communication radio environment is constantly and rapidly changing, e.g. due to fast and slow fading, as well as due to the interference from users in the cell.

The received ratio of bit energy to noise power spectral density, Eb/N0 is an important

parameter in communication systems. If the received Eb/N0 drops below an acceptable level and does not improve during a predetermined period of time, the call will be dropped. Thus, for an efﬁcient operation of the interference limited WCDMA network it is crucial that the transmit power of all users is controlled. A group of measures is therefore introduced within the RRU to minimize the interference level in the system. They are summarized as power control (PC) and consist of open-loop PC, inner-loop PC (also called fast-closed-loop PC), and outer-loop PC in both uplink and downlink. Open-loop PC makes initial estimates of the transmit powers needed in the uplink and the downlink for each MT, before it accesses the network, and is only important for the initialization of the call. The estimation is based on path loss calculations in the downlink direction. The other two loops are active for the complete duration of the

call. The target of the inner-loop PC in WCDMA is to keep the Eb/N0 constant. Due to the fast feedback loop (1500 Hz) this is fairly successful. It means that for a chosen service, given channel conditions, and a required BLER (block error rate), the received power on the radio channel divided by the interfering power is approximately constant [46]. The outer-loop power control is needed to keep the quality of communication at the required level by setting the target carrier-to-interference ratio C/I (average received power of the useful signal to that of all relevant interfering signals) for the fast-closed-loop power control both in the uplink and the downlink. The frequency of the outer loop power control is typically 10-100 Hz. 14 Chapter 1. Universal Mobile Telecommunications System (UMTS)

1.4.2. Handover Control

Another algorithm for radio resource utilization of UTRAN is the soft handover (SHO). SHO has been developed to provide a seamless user migration from one cell to another. SHO is different from traditional hard handover. For hard handover a decision is made on whether to handover or not, and the MT communicates with only one Node B. However, in WCDMA systems, the user can get two or more simultaneous connections with different Node Bs. Hard handover happens at a distinct time, while soft handover lasts for a period of time. Moreover, in SHO the user has a gain from diversity. The diversity gain provided by SHO is called macro diversity. This means that one or more (up to 3, usually) additional signals can be received at a time, which offers protection against fast fading by selecting the best signal or combining the signals in a suitable way. The user transmits the same signal to all Node Bs it is connected to. Furthermore, the Node Bs transmit their received signals to the RLC (Radio Link Control), which combines them or selects the best one. If the signal strength on one radio link rapidly changes, the RNC instantly switches to a signal from one of the other Node Bs with signiﬁcantly better quality. This prevents an interruption in data ﬂow. Macro diversity can be exploited for decreasing the uplink interference, which leads to a higher system capacity.

1.4.3. Call Admission Control

Since WCDMA allows many users to transmit simultaneously within the same frequency band, admitting a new call will always increase the interference level in the system, reducing the cell capacity for new calls and even leading to dropping of established connections. The objective of the Call Admission Control (CAC) is to admit or reject new or handover call requests, based on the actually free capacity, the current traffic load situation in the system and the desired QoS of the new calls as well as calls in progress. The CAC is proactive, its goal is to prevent unacceptable degradation of QoS and to achieve a trade-off between system capacity and user satisfaction. Compared to the straightforward CAC algorithms in TDMA systems, where the system capacity is fixed (predefined number of radio channels), in cellular networks with WCDMA radio interface it is much more difficult to predict the effect of an admitted new calls on the active calls due to uncontrollable fading variations and wireless interference.

1.4.4. Congestion Control

Even with an efﬁcient admission control algorithm an overloaded situation in the system may occur. When an overload is encountered, the output powers are rapidly increased by the fast closed-loop power control until one or several transmitters have reached their maximum output power. The connections unable to achieve their required quality are considered useless and are only adding interference to the system. 1.5. Summary 15

This is an unacceptable system behavior. Hence, a procedure to remove the congestion is needed. The method is called congestion or load control, and its objective is as quickly as possible to monitor the traffic load in the network, detect traffic overload, and react in order to get the system back to a feasible traffic load, which is defined by the radio planning. This can be achieved e.g. by slightly degrading the QoS in the overloaded cell during the time it takes to resolve the congestion.

1.5. Summary

In this chapter an overview of the Universal Mobile Telecommunications System has been given. The basic principles of its air interface, namely the spreading and scrambling techniques of WCDMA used to generate the transmission signals, have been brieﬂy explained. Furthermore, several schemes for mitigating wireless interference within the framework of radio resource management have been introduced: Power control for minimizing the interference that individual mobile terminals (MTs) cause, Handover control to allow MTs a smooth transition between two cells to provide macro diversity, Call Admission Control to constrain the maximum number of simultaneous connections, and Congestion Control for appropriately reacting to suddenly occurring overload situ- ations. In this work, all proposed algorithms and schemes focus speciﬁcally on the UMTS air interface.

Part I.

Traﬃc Modeling for Wireless Networks

Chapter 2. Traﬃc Concepts and Models

Teletraffic theory quantitatively studies the relationship between communications systems, traffic and performance. Queuing theory, a subset of teletraffic theory, deals with the mathematical modeling of traffic and the analysis of the statistical behavior of telecommunication systems. The basic goal is to provide reliable methods for determining the cost-effectiveness of various network sizes and configurations. In the following some fundamental concepts of queuing theory will be introduced and examples will be given.

2.1. Mathematical Background: Basic Model

Queuing systems model processes, where users generate traffic flows, which can either be circuit-switched (speech services) or packet-switched (data services). An example of a queuing system is illustrated in Figure 2.1 and can be described as follows: Ser- vice requests enter the system in order to be processed. If the required resources are available, the requests are admitted and leave the system after successful handling. In case a new request arrives and finds all resources busy, it enters a queue and waits until sufficient capacity becomes available to serve its request. The standard notation to identify the main elements that define the structure of a queuing system is the Kendall notation A/B/N/n/S/X , where A denotes the distribution of the interarrival times and B the distribution of the service times; N specifies the number of servers, n the maximum number of waiting requests in the finite case (number of servers plus the capacity of the queue), S is the number of users (passive and active) and X defines a queuing discipline. For a detailed description of Kendall notation see Appendix A.

2.1.1. Birth-Death Process

Certain types of queuing systems can be modeled by birth-death processes. A birth- death process is a process, where the population with size i (the number of currently active users) may increase or decrease according to certain rules. Speciﬁcally, when the population size is i, the only possible state transitions are from i into i − 1 (death) or from i into i + 1 (birth). 20 Chapter 2. Trafﬁc Concepts and Models

Arriving Served requests requests Queue with N waiting N servers requests

Figure 2.1.: A queuing system with arrival and departure of service requests.

λ1

0 1

µ1

Figure 2.2.: Birth-Death process.

The following state-transition diagram (Figure 2.2) is called a birth-death Markov chain. Directed branches represent transitions between the states. The birth-transition N intensity is described by λi ≥ 0 for i ≥ 0, i ∈ (user arrival rate) and the death- N transition intensity is designated by µi ≥ 0 for i > 0, i ∈ (user service rate). Teletraffic theory relies on the concept of stochastic processes. It is usually assumed that interarrival times are independent and identically distributed [44]. In most cases, the interarrival time distribution is assumed to be exponential, in which case the arrival process is a homogeneous Poisson process In Figure 2.3 a realization of a Poisson process is shown. Poisson arrivals occur completely at random in time. Mathematically, the process is defined as a collection X (t) : t ≥ 0 of random variables, where X (t) is the number of events that have occurred up to time t. A Poisson process is a pure birth process and is characterized by λ (i. e. number of events per time unit) such that the number of events X (t) in a finite time interval of length t follows a Poisson distribution with parameter λt. This relation is given as

e−λt (λt)k Pr {X (t)= k} = , k = 0, 1, . . . , n (2.1) k! A Poisson process has several important properties:

Stationarity The distribution function of the number of observations in a time interval depends only on the length of the interval and not on its position in time. 2.2. Markovian Queuing Systems 21

0 t

X (t)

0 t

t1 t2 ∼Poisson (λt1) ∼Poisson (λt2)

Figure 2.3.: Poisson arrivals

Memorylessness A Poisson process X (t) satisﬁes the Markov property memorylessness, which states that

Pr{X (ti+1)= xi+1|X (ti)= xi,..., X (t1)= x1}

= Pr{X (ti+1)= xi+1|X (ti)= xi} (2.2)

for any choice of time instants ti, i = 1, . . . , n, where ti < ti+1.

This should be understood as follows: the future state of the process at time tn+1

is independent of its past history (earlier time instants t1,..., tn−1) and depends

solely on its present state at time tn (this follows from properties of the exponential distribution).

Ordinarity means that the normalized probability of the occurrence of more than one event in an very small interval h goes to 0:

lim Pr k events appear during h, k > 1 = 0 (2.3) h→0 PASTA (Poisson Arrivals See Time Averages) An arriving event sees the system as if it observes the system at a random instant of time.

2.2. Markovian Queuing Systems

The traffic intensity in the network, which is defined as ρ = λ/µ may change over time. However, for practical calculations of the system performance measures, we are interested in traffic in statistical equilibrium, which means that the traffic intensity is 22 Chapter 2. Traffic Concepts and Models

rather stable during the period of observation. Furthermore, the traffic process must be observed for a time interval sufficiently large to neglect the effect of a transient period, assuming we start with an empty system. This is in agreement with conditions during busy hour (the busy hour is the 60-minute period during the day when the traffic volume reaches its maximum) and therefore allows for approximating the actual traffic by means of a stationary traffic intensity model.

2.2.1. M/M/1 System

The simplest queuing system model is M/M/1, a single-server queue with randomly distributed service requests. The deﬁning characteristics of the system are:

• The time intervals between successive events are exponentially distributed, with

λi = λ.

• The service time is an exponentially distributed (Markovian) random variable,

with µi = µ.

• The number of arrivals during a ﬁnite time interval follows a Poisson distribution.

• The arrival process and the service process are independent of each other.

The capacity of the queue is inﬁnite and the service discipline is First-In-First-Out (FIFO) (see Appendix A). The M/M/1 queue is a birth-death process with the birth rate and death rate constant, which is easy to deal with mathematically. A very useful relation between E(S), the long-term average number of users in a system, E(W), the long-term average time a user spends in the system (waiting time plus service time) and λ, the average number of users arriving at the system per unit time is given by Little’s law [8]:

E(S)= λE(W). (2.4)

The above mentioned properties include the Markov (memoryless) property, which makes it possible to describe the state of the system at arbitrary points in time with a single variable, which holds the number of users in the system. Markovian systems can be directly mapped to a continuous time Markov chain (CTMC) [6]. The CTMC for M/M/1 is depicted in Figure 2.4.

λ λ λ λ

0 1 2 3 µ µ µ µ

Figure 2.4.: M/M/1 system.

A system transition from state to state is described by a Markov process [6]. The process starts in state 0 (the system is empty) and moves successively from one state up 2.2. Markovian Queuing Systems 23

to another with rate λ, and down with rate µ. A Markov chain is completely deﬁned, when all states and state transition rates are given. This concludes the description of the M/M/1 queuing system. However, up to now we have not yet studied the behaviour of the queue, since the

steady-state probabilities pi must be determined. The steady-state probability is the probability of observing the system in state i at a random point in time. This probability does not depend on the states the chain was in before the current state. If cut between two arbitrary states i and i + 1, i ≥ 0, the transitions from each side of the cut to the other are the state transitions i → i + 1 and i + 1 → i.

“The system is in statistical equilibrium, if the trafﬁc process changes from state i → i + 1 the same number of times as it changes from state i + 1 → i” [37].

We thus have the following equilibrium equation or balance equation, which balances the average intensities of leaving and entering a state in equilibrium:

λ · pi = µ · pi+1, i ≥ 0 (2.5)

The condition for such a system to be stable (underlying CTMC to be ergodic) is µ ≥ λ [6]. By substituting λ and µ by ρ = λ/µ, which is called offered trafﬁc, Equation (2.5) can be rewritten as

pi+1 = ρ · pi, i ≥ 0. (2.6)

The offered trafﬁc is the total number of call attempts (successful plus unsuccessful). It is dimensionless, however, it is often speciﬁed in the pseudo unit erlang. Expressing all

state probabilities by p0 we get

i pi = ρ · p0 (2.7)

and by requiring, that the sum of the probabilities must be equal to one, the ﬁnal result is

(1 − ρ)ρi p = , i ≥ 0. (2.8) i λ

Hence, by solving the balance equations, the steady-state probabilities pi can be obtained and the performance measures, expressed through these state probabilities, can be calculated. Relevant performance measures in the analysis of queuing models are:

• The probability distribution of the waiting time and the sojourn time of a customer.

• The distribution of the number of users in the system.

• The number of congested call attempts occuring during the observation period. 24 Chapter 2. Trafﬁc Concepts and Models

2.2.2. Erlang’s Loss System

The M/M/N/n queuing system with arrival rate λ and service rate µ can be also modeled as a birth-death process. The system is truncated and has no waiting room if n = N. Such a system corresponds to a loss system with

λi = λ, i ≥ 0, (2.9)

µi = iµ, 0 ≤ i ≤ N. (2.10)

The state transition diagram for a system with a limited number of servers and N = n is depicted in Figure 2.5.

λ λ λ

0 1 2 N − 1 N

µ 2µ Nµ

Figure 2.5.: State diagram for Erlang’s loss system.

Based on the equilibrium equations λ · pi =(i + 1) · µ · pi+1, i = 1, 2, . . . the following steady-state probabilities can be derived:

λ p = p 1 0 µ λ 1 λ 2 p = p = p 2 1 2µ 0 2 µ . .

λ 1 λ i p = p = p i i−1 iµ 0 i! µ . .

λ 1 λ N p = p = p N N−1 Nµ 0 N! µ Using the normalization condition:

N N 1 λ i 1 = p = p , (2.11) i 0 µ i=0 i=0 i! X X 2.2. Markovian Queuing Systems 25

We get for p0:

1 = p0 i . (2.12) N 1 λ i=0 i! µ P Thus, by letting ρ = λ/µ all states probabilities can be expressed by p0,

ρi p = · p . (2.13) i i! 0

2.2.2.1. Blocked-Call-Cleared – Erlang B

In a truncated system, when a particular user requests service and all of the servers are already in use, the user is denied access to the system and thus blocked. This type of truncation is called Blocked-Call-Cleared (BCC), see Figure 2.6. The following assumptions are made:

• There are memoryless arrivals of requests, implying that all users may request service at any time.

• The number of arrivals during any time interval is Poisson distributed.

• There is an inﬁnite number of users.

• The duration a user occupies a server is exponentially distributed, so that longer calls are less likely to occur than shorter calls (there is no limit for the call length, but the distribution is not heavy-tailed).

• There is a ﬁnite number of servers available in the network.

• Requests are not allowed to queue if there is no server available when the request arrives.

• Blocked requests (or calls) are lost.

• Lost calls are not retried.

The quantity of interest is pN , the probability that a user cannot establish a call due to insufﬁcient system capacity. Such a system is often denoted as M/M/N-loss system and leads to the derivation of the Erlang B formula, which can be used for calculating the capacity of a telephone network required to carry the user service requests. The capacity of such networks is traditionally measured in channels. The following equation is called Erlang-B formula; it determines the probability that a call is blocked and yields a measure for the Grade of Service (GoS):

ρN B(N, ρ)= N! , (2.14) N ρi i=0 i! P 26 Chapter 2. Trafﬁc Concepts and Models

User 1 Carried traffic User 2 User 3 Complete loss User 4 User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Capacity N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time Figure 2.6.: Example of Blocked-Call-Cleared (BCC) model.

where B is the blocking probability and N is the number of (servers) channels in the system. During the proportion of time where all channels are busy all further service requests are blocked and lost.

2.3. Traﬃc Models for Wireless Networks

This section gives a brief overview of the existing traffic models for wireless networks. We identify challenges of aligning traditional traffic models to the specific characteristics of 3rd generation mobile radio systems.

2.3.1. General Remarks on Existing Traﬃc Models: Drawbacks and Challenges

As the capacity and flexibility of wireless communication networks increases and the amount of services they provide expands, the demand for radio resources becomes less homogeneous. For successful planning of wireless networks it is necessary to predict the boundaries of traffic load, in order to maintain upper bounds for probabilities of call blocking, packet dropping, and packet delay. Traffic theory currently plays a minor role in the design of third generation cellular radio networks. Network dimensioning of UMTS is generally based on rough estimations rather than on realistic analytical models for the expected traffic behavior. Consider- able effort is spent on the design of a variety of QoS mechanisms. The role of the latter is to ensure an appropriate level of service for different user classes in UMTS 2.3. Traffic Models for Wireless Networks 27

networks. Nevertheless, there still remains the necessity of applying appropriate traffic- performance models, if the objective is to ensure that QoS meets certain targets for a given population of users. Traffic models like Erlang’s loss system are established models, which are generally used for dimensioning of fixed networks and can quite accurately model single-service circuit-switched traffic, where the capacity is hard limited (e.g. by the number of traffic channels). The performance evaluation of multi-service wireless networks is, however, a more complex problem. Non-Poisson traffic characteristics and dynamic behavior of user transmission rate have to be considered. Furthermore, in UMTS all users utilize the same frequency channel. If the resulting co-channel interference is high, consequently the carrier-to-interference ratios are low. If the capacity is limited by the amount of interference in the radio interface, it is by definition a soft capacity, since there is no single fixed value for the maximum network capacity. The above mentioned specific characteristics of multi-rate cellular traffic violate a number of Erlang-B assumptions. Therefore, the straightforward application of traditional traffic models for performance estimation to 3rd generation wireless mobile systems like UMTS is not recommended, because it would yield too pessimistic system performance results. In the following, a brief overview of the state of the art of traffic engineering for UMTS networks is presented. The load of existing UMTS networks is still too low to anticipate the actual busy hour traffic values. Therefore, research and engineering is focused on finding an appropriate technique based on analytical modeling or simulation of the functionality of the UTRAN, in order to propose a simple and accurate algorithm for better design and performance estimation of 3G mobile networks. Methodologies to analyze the UTRAN performance by suitable simulation tools have been introduced in [28]. Also Lassila and Virtamo [47] provide efficient methods for simulation of Poisson and non-Poisson multiple traffic streams for multi-service loss models. The algorithm proposed in [25] is based on the event scheduling approach and simulates traffic models of a Full Availability Group (FAG) with and without reservation of resources. FAG is the traffic model that uses complete sharing resource policy. Each newly arriving user has unrestricted access to all free resources, regardless of the service-class [35]. For an example of FAG with multi-rate traffic, see Figure 2.7. Although simulation tools produce good results, they need a huge amount of computation time and many input parameters, most of which cannot be easily inferred from real measurements. Compared to simulations, analytical methods for system performance analysis make it possible to get a quick and reliable result and also provide a cost effective tool for network dimensioning and understanding the relationship between the system, traffic, and performance. In 1981 Kaufman and Roberts published a method which can be used for computing the blocking probabilities of the multi-dimensional Erlang loss models [43]. The algorithm is a simple one-dimensional recursion, applicable to a multi-rate system, re- 28 Chapter 2. Traffic Concepts and Models

FAG

1 service-class 1: λ1, µ1 2 service-class 2: λ2, µ2 3

service-class n: λ , µ n n N

Figure 2.7.: An example of Full Availability Group (FAG): every inlet has access to every outlet with unlimited access to resources, n is the number of services and N denotes the number of servers.

gardless of the dimensionality (number of distinct service classes) of the system. The basic model of such a system is the FAG. However, later in 1989, claims have been made about the insufficient accuracy of the Erlang B formula for modelling call blocking for multi-service wireless networks in [48]. In [40], the applicability of the Erlang B distribution to model cellular network traffic has been questioned. In [70, 71, 66, 16] the Kaufman-Roberts recursion scheme has been extended in order to calculate the blocking probability in limited-availability groups with multiple traffic streams. This extension comprises the introduction of state dependent transition probabilities. In [69] and later in [39] state dependent blocking probabilities caused by wireless interference have been introduced for calculating the performance of cellular systems using a WCDMA radio interface. In the above mentioned work, the other-cell interference is modeled as a log-normal random variable and the blocking probabilities are computed using an appropriately modified Kaufman- Roberts recursion. In [72] the uplink performance of the UTRAN has been estimated. To determine the blocking probability of a new call appearing in the cell under consideration, the load generated by calls in neighboring cells has been taken into account. All above mentioned models are based on classical traffic models, adopted for cellular radio networks. However, the specifics of such important features of UMTS like traffic- adaptive transmission rate have not been included into the models so far. To our best knowledge, there are no results available for explicit theoretical modeling and analysis of the complex behavior of multimedia traffic in wireless heterogeneous environments. One of the objectives of this thesis is to develop an accurate mathematical algorithm, which can operate with a limited amount of input data, in order to evaluate the impact of new services on the behavior of 3G networks. Those services exhibit traffic patterns that have not yet been thoroughly investigated. However, explicit knowledge of their characteristics is crucial for an efficient performance prediction of multi-service systems with WCDMA radio interface. 2.4. Summary 29

2.4. Summary

In this chapter, fundamental mathematical concepts of the teletraffic theory have been introduced. The notion of a queuing systems, which have been described by Markov- and Poisson processes, along with their specific properties have been presented. Next, the adaptation of some traffic models to wireless networks, especially with regard to UMTS, has been given. Though there are some models existing in the literature, their accuracy and efficiency still leaves room for improvement, since some specifically interesting features of WCDMA are still disregarded.

Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio Interface

3.1. Motivation

We concentrate our research on the UMTS Terrestrial Radio Access Network (UTRAN), where all users served by a given cell make use of the same frequency channel and differentiation of the transmitted signals is possible only by application of orthogonal codes [30]. In order to provide service in the heterogeneous environment of a multi- service system using WCDMA radio interface more efficiently, it is desirable to be able to estimate the limits of the traffic load such that upper bounds of probabilities for call blocking, packet dropping and packet delay can be maintained for all classes of demands. In general, applications and services can be divided in two groups: real-time demanding and non-real time demanding. The main distinguishing feature between these service classes is their delay sensitivity; the real-time class is highly delay sensitive, while the non-real time applications such as file transfer protocol (FTP) and email are delay insensitive and can tolerate delay variations. Most multimedia applications from the latter group do not continuously require dedicated channels for the entire duration of the connection and have variable bit rate (VBR) with packet transmission at peak rates. The connections have average rates less than the peak rates and we may over- book the system and admit new calls even if their peak rates cannot be simultaneously accommodated. We may still maintain an acceptable performance level of the system due to statistical multiplexing among services. However, such flexibility makes the Call Admission Control (CAC) procedure more complex. To address this problem, traffic engineering operates at both connection or flow level and packet level. Traffic flows at connection level are characterized by a requests arrival process. The relevant performance measures are blocking probabilities of new and handover calls. CAC, handover rate control and congestion control guarantee that these probabilities stay within a predefined range. At packet level, a connection is characterized by the call handling process, e.g. the activity factor (fraction of time the call Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 32 Interface

is in active (on) state) and the data rate requirement during on periods. Quality of Service (QoS) measures at packet level are packet loss probability and packet delay. In the following we introduce a new efﬁcient analytical method for uplink connection level performance estimation in multi-service WCDMA networks by taking the key features of multi-service radio networks, namely wireless interference and the dynamic behavior of the user transmission rate, into account.

3.2. Single Cell Capacity Model

In Wideband Code Division Multiple Access (WCDMA) wireless systems, accurate signal reception is possible only when the ratio of energy per bit to noise power spectral-

density Eb/N0 is appropriate (Eb: average received bit energy, N0: power spectral den-

sity of white Gaussian noise). The Eb/N0 for the jth user is calculated as (considering the uplink)

E W Pj b = · (3.1) N R · ν I − P 0 j j j total j

where Pj is the received signal power from the jth user, W – the chip rate, νj – activity

factor of the jth user, R j – service bit rate of the jth user, Itotal – total received wideband power including thermal noise power. The average power of the jth user is determined as

1 Pj = Itotal = L j · Itotal . (3.2) 1 + W (Eb/N0)j ·R j ·νj

We deﬁne L j as a load factor of one connection,

1 L j = . (3.3) 1 + W (Eb/N0)j ·R j ·νj

The total load for the uplink is then

ηU L = L j, (3.4) j=1 X

where N is the number of active connections in the cell for all services. When ηU L approaches 1, the cell has reached its pole capacity. The above relation is, however, only true if we consider a single isolated cell. In the real-world UMTS scenario (comprising multiple cells), knowledge about idle spreading codes in the cell is not sufﬁcient to determine the maximum network capacity, i. e., the number of simultaneous users a system can support while maintaining the QoS requirements. 3.3. Connection dependent Model with Transmission Rate Reduction Policy 33

However, to help the reader get acquainted with the topic, in this section we restrict ourselves to a single isolated cell, thus eliminating all possible impairments from neighboring cells, e.g. wireless interference, etc. First we will deal with hard blocking, where capacity is only limited by the amount of spreading codes. We consider a Full Availabil- ity Group (FAG) (see Chapter 2) which is offered n multi-service traffic streams, i. e. none of the service classes has priority over another class. In multi-service systems, different service classes have different data rate requirements. Thus, it would be beneficial for the teletraffic calculations to specify some unit channel in order to express data rates as a multiple of this unit. The following definition of a channel used in our model is also given in [39]. The authors of [39] denote the service-related entities product in Eq. (3.3) by κ:

E κ = b · R · ν . (3.5) j N j j 0 j Then the maximum number of channels of service j in one isolated cell is:

W N = η 1 + . (3.6) max,j U L κ j

Then Rs,j, let us call it equivalent bandwidth per channel for service j, can be expressed as

2 W 1 κj Rs,j = = κj − . (3.7) Nmax,j ηU L κj + W !

This equation allows the conversion from power-interference-based capacity to channel-based capacity. The channel unit, to be referred as basic bandwidth unit (BBU), is the greatest common divisor of equivalent bandwidths of all call streams offered to a system. The higher the required granularity, the smaller the BBU has to be chosen. The reduction of the UMTS network model to a circuit-switched form is an important contribution. The analogy allows several important operating strategies and analysis techniques from the circuit-switched literature to be directly applied to a WCDMA system.

3.3. Connection dependent Model with Transmission Rate Reduction Policy

The following assumptions are made for the system under consideration:

• The network serves n independent classes of Poisson trafﬁc streams with arrival

rates λ1, λ2,..., λn. Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 34 Interface

• Each call is assigned a single or multiple channels (BBUs) depending on the parameters of the particular service class. The conversion from power-interference- based capacity into channel-based capacity is done as in [39], as also explained in the previous section. A channel unit is chosen to be a data rate of 12.2 kbit/s, corresponding to the highest speech rate in UMTS.

• Service class k is characterized by its data rate of dk = mk × 12.2 kbit/s, k =

1, 2, . . . , n, and mk is a positive integer. We choose the BBU so that each services class’ demand is an integer multiple of the BBU.

• The service time for calls of a particular service class is exponentially distributed.

• A call holds the requested resources for a period of time, called mean holding time

(or service time) τk. The resources are released with a service rate µk = 1/τk.

• The mean trafﬁc offered to the system for service class k is ρk = λk/µk, which is the mean number of connections.

• Calls of different service classes get different resource allocation (user transmission rate and service time), which depend on the total number of already occupied channels. This means, that a new call of service k can be admitted into the system with a transmission rate that differs from the initially requested one. We call this call admission strategy Transmission Rate Reduction Policy (TRRP).

• The pole capacity of the system (here: cell) is N BBUs.

We, thus, have a Multiservice Loss Model (MLM), and the objective is to determine the call blocking probability and average user transmission rate for each service class in the system. The above mentioned model can be described by a multi-dimensional

Markovian process of type u(t)= {i1(t), i2(t),..., in(t)}, where each dimension repre-

sents one trafﬁc class and ik(t) is the number of currently active users of service class k. The state-transition-rate diagram of a two-dimensional Markov chain is shown in Figure 3.1. The theory, which deals with Multiservice Loss Models states that there exists a

unique stationary distribution for the stochastic process u(t), thus P(i1, i2,..., in) is

the steady-state probability, where {i1, i2,..., in} belongs to the state space S deﬁned as follows:

S = i1, i2,..., in mkik ≤ N (3.8) ( = ) k 1 X

where 0 ≤ ik ≤ N, and mk denotes the number of channels for service-class k, requires to establish one connection, mj ≥ 1; both are integer [74]. The values of P(i1, i2,..., in) can be shown to be composed of a product of the state probabilities of the trafﬁc streams of each service class in isolation, i. e., before truncation (it is known that the trafﬁc 3.3. Connection dependent Model with Transmission Rate Reduction Policy 35

PSfrag

λ2

i, j + m1 i + m2, j + m1

(i+m2)µ2 m2

(j+m1)µ1 (j+m1)µ1 λ1 λ1 m1 m1

λ2

i, j i + m2, j

(i+m2)µ2 m2

Figure 3.1.: A fragment of a Markov chain for a two service system.

streams are independent before truncation) and thus, given by the product form solution:

P(i1, i2,..., in)= P(ik). (3.9) k=1 Y

With n incoming Poisson ﬂows and mean offered trafﬁc ρk for service class k, we get the multi-dimensional Erlang B formula:

i1 i2 in 1 ρ1 ρ2 ρn P(i1, i2,..., in)= · · ··· , (i1, i2,..., in) ∈ S, (3.10) G i1! i2! in!

where G is a normalization constant:

i i ρ 1 ρ 2 ρin G = 1 · 2 ··· n . i ! i ! i ! (i ,i ,...,i )∈S 1 2 n 1 2Xn

Each trafﬁc stream is characterized by its individual blocking probability Bk. As mentioned earlier, for the time being we are interested in hard blocking only. The hard blocking occurs only in border states, thus

i i 1 ρ 1 ρ 2 ρin B = 1 · 2 ··· n (3.11) k G i ! i ! i ! (i ,i ,...,i )∈D 1 2 n 1 2 Xn k

where Dk ⊂ S are the states with blocking for class k deﬁned by

Dk = i1, i2,..., in N − mk < mj ij ≤ N . (3.12) ( = ) j 1 X

Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 36 Interface

However, calculating the blocking probabilities by using the above equations is numerically intractable and not suitable for multi-dimensional scenarios in large-scale networks, since the size of the transition matrix will grow exponentially with the number of service classes. In fact, the blocking probability for an arriving call does not only depend on the number of users in the network but also on the resources they occupy. Consequently, it is sufficient to determine the steady state distribution of the resource occupancy. We aggregate a multi-dimensional service process with different multi-rate traffic streams by a two-dimensional Markov chain [43, 65], since the product form solution allows for this. Furthermore, we adopt the Kaufman-Roberts recursion [43, 65] based on recurrence determination of the channel occupancy distribution P(r), where P(r) is the probability n of r busy BBUs in the FAG (r is the global state, e. g. r = k=1 ik, to which various multichannel traffic streams are offered, and resulting in P 1 n P(r)= ρ m P(r − m ), (3.13) r k k k k=1 X where P(r)= 0 for r < 0. In order to extend our algorithm to support dynamic user transmission rates, we incorporate TRRP into the model by using a modified analytical approach for estimating the blocking probability in Asynchronous Transfer Mode (ATM) networks proposed in [52]. In this threshold model for blocking avoidance, new calls of certain service classes can get service with requests different to the initial resource and service time requests, which are state dependent. After some modifications to (3.13) and the addition of the expression from [52] we get:

1 n P(r)= ρ m δ (r)P(r − m )+ r k k k k k=1 X 1 n sk ρ m δ (r)P(r − m ) (3.14) r kl kl kl kl k=1 l=1 X X for r = 1, . . . , N, where δ (r) = 1 and δ = 0, when 1 ≤ r ≤ N (for service classes k kl without rate reduction policy), or r ≤ J + m (for service classes that have the possi- kl k bility to be admitted with reduced transmission rate), otherwise δ (r)= 0 and δ = 1; k kl s is the number of possible variations of the transmission rate, and J is the adjusted kl threshold. The algorithm enables a performance assessment for each individual service class. 3.4. Multi-Cell Capacity Model 37

For calculating the total blocking probability for a service class while taking TRRP into account we modify the limits of the summations in the following expression:

− N mks B = 1 − P(r). (3.15) ks r=0 X This is the probability of a call of service class k to be blocked with its lowest possible transmission rate. In the next section we discuss the extensions to the multi-cell case.

3.4. Multi-Cell Capacity Model

As mentioned before the key element that negatively affects the performance of radio networks is wireless interference. WCDMA allows many users to transmit simultaneously within the same frequency band, which means that from a single-user perspective, all other users’ signals are noise. Every time a new call is accepted, the interference level increases, and consequently the signal-to-noise ratios for all other users decrease. Before admitting a new connection in the system, CAC ensures that a new call will not adversely affect the quality of the existing connections; a new call should not be accepted by the system if this call increases the noise for any other existing call above the tolerable level. The probability for admitting a new request of a speciﬁc service class k in the cell considered depends on the current state of the cell, the noise-rise of the new connection, and the other-cell interference. Consequently, only a fraction of all links can be active simultaneously at any given instant. Thus, taking other-cell interference into consideration, the uplink load factor must be rewritten as

N N 1 ηU L =(1 + α) · L j =(1 + α) · W , (3.16) j=1 j=1 1 + X X (Eb/N0)j ·R j ·νj where mean other-cell interference power α = (3.17) mean own-cell interference power

The interference ratio α depends on cell environment, type of antenna used, trafﬁc load, etc. The interference reduces the cell capacity by a factor 1 , which means that if 1+α we have a single isolated cell capacity (pole capacity) of Ns, then when we go multi-cell, we get a reduced capacity

1 N = N . (3.18) m 1 + α s

A system whose capacity is limited by wireless interference is called a soft capacity system. The interference contribution from other cells is typically around 35-40%. The Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 38 Interface

principle of a system with a soft capacity is explained in Figure 3.2 and should be understood as follows: the less interference coming from the neighboring cells, the more capacity is available in the middle cell [30].

Cell 2 Cell 3 Cell 1

Cell 7 Cell 4

Cell 6 Cell 5

Figure 3.2.: Capacity (interference) sharing between cells in WCDMA.

It is commonly assumed that the maximum possible usage of radio interface resources without violating the QoS agreement is at 50-80 % [30].

3.4.1. Restricted Network Accessibility - Modeling of Other-Cell Interference

From now on we will deal with soft capacity networks, where the admission of new requests does not only depend on the current load situation in the cell under consideration but also on the load conditions in the neighboring cells. Due to the frequency reuse factor of 1 in WCDMA, all users in both the considered cell and the neighboring cells contribute to the total interference level, influencing the link quality and determining the capacity of the cell. However, defining an adequate model for wireless interference is not a straightforward task. Most interference-based investigations deal with some fixed interference factor, which is expressed as the ratio between other-cell interference and own-cell (from the users in the same cell) interference [78, 77]. In [68] the focus is on deriving an analytical method to compute the other-cell interference in a large Universal Mobile Telecommuni- cations System (UMTS) network. The proposed model is based on an iterative solution of a fixed-point equation which describes the interdependence of the interference levels in neighboring cells. The authors assume the power of the other-cell interference to be a log-normal distributed random variable and validate their assumption by the proposed iterative approach. 3.4. Multi-Cell Capacity Model 39

In [72] the authors model a FAG, using the idea of the switching node in Asyn- chronous Transfer Mode (ATM) networks, servicing a mixture of multi-rate trafﬁc streams. For a thorough description of the switching node see [73]. The level of wireless interference is expressed by a load factor. It was assumed that each call generates load ′ L j in the access cell and load L j = L j in the neighboring cells. In our work we model the other-cell interference power as a log-normal distributed random variable with parameters µ = α · N as proposed in [69]. α+1 To validate the model we have performed numerous simulations by using a MATLAB based WCDMA Network Simulator (RANSim), which has been developed at the Insti- tute of Mobile Communications, Friedrich-Alexander-University Erlangen-Nuremberg. The simulator is intended to resemble the properties and peculiarities of the UMTS air interface as close as possible. The simulated network consists of regular hexagonal cells. A wrap-around technique is used to avoid border effects. The sampled values for the other-cell interference powers taken from the simulation, for low, medium and high offered load, match well with whose drawn from a theoretical log-normal distribution. The most important parameters used in the simulations are summarized in Table 3.1. This good ﬁt is visualized in Figure 3.3, where density-histogram overplots and a quantile-quantile plot for the other-cell interference distribution are shown.

Table 3.1.: Main simulation parameters. Trafﬁc and environmental settings Users are equally distributed over all cells Traf. load (mean. num. of users/cell) 10 (low), 30 (med.), 50 (high) Number of radio cells 21 Antenna type omni-directional Cell radius 200 m Radio interface and algorithmic settings Distance attenuation 3 Shadowing standard deviation 6 Power control according to 3GPP Downlink Soft Handover Power Synchronization Simulation step size 1 radio frame (0.01 s) Number of steps 40 000-120 000

3.4.2. Proposed algorithm

The system under consideration is a UMTS network. We assume the overall system being in statistical equilibrium. For such a system, any cell is statistically identical to all other cells. We also conjecture homogeneous trafﬁc conditions, i. e., statistically the same amount of trafﬁc is uniformly loaded in each cell. Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 40 Interface

Density/Histogram overplot, other−cell interference vs. lognormal distribution (low load) Density/Histogram overplot, other−cell interference vs. lognormal distribution (medium load) 250 300

250 200

200

150

150 Frequency Frequency 100

100

50 50

0 0 0 0.5 1 1.5 2 2.5 0.5 1 1.5 2 2.5 3 3.5 4 4.5 −15 −15 other−cell interference (mW/Hz) x 10 other−cell interference (mW/Hz) x 10 (a) Density-Histogram overplot, low load (b) Density-Histogram overplot, medium load

−4 Density/Histogram overplot, other−cell interference vs. lognormal distribution (high load) x 10 QQ Plot, other−cell interference vs. lognormal distribution (high load) 250 4.5

4.4

200

4.3

150 4.2

Frequency 4.1 100 Lognormal Quantiles (mW/Hz)Lognormal Quantiles

50 3.9

0 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 1.5 2 2.5 3 3.5 4 4.5 5 5.5 −15 −15 other−cell interference (mW/Hz) x 10 Quantiles of interference samples (mW/Hz) x 10 (c) Density-Histogram overplot, high load (d) Empirical QQ-plot of the quantiles of the sampled values for other-cell interference (x-axis) versus theoretical quantiles from a log-normal distribution (y-axis).

Figure 3.3.: Density-histogram overplots and quantile-quantile plot for other-cell interference versus log-normal distribution for WCDMA system with low, medium and high trafﬁc load. 3.4. Multi-Cell Capacity Model 41

• The service processes in the cells are dependent, and the current trafﬁc load in one cell inﬂuences the service probability in other cells.

• Blocking occurs if the total interference power exceeds a predeﬁned level or if the cell load exceeds the maximum threshold. In this case admission control rejects the request for a new connection.

• The pole capacity of each cell is N BBUs.

• Other model characteristics used in the algorithm remain unchanged (see Sec- tion 3.3).

The algorithm is novel due to its ability to analytically deal with two distinct aspects associated with QoS: acceptable interference level and average user transmission rate on one hand, and blocking probability of new calls on the other hand. In order to determine the call blocking probability and the average user transmission rate for each service class, we revisit the analysis from [39], where the following recur- rent formula has been proposed for estimation of the WCDMA uplink performance:

1 n P(r)= ρ m P(r − m )(1 − b ), (3.19) r k k k r−mk k=1 X where P(r) is the probability of r channels being busy in the cell under consideration,

ρk = λk/µk and b is the state dependent blocking probability (soft blocking), which characterizes the current other-cell interference situation. Let us consider a part of the state-transition-rate diagram of the underlying Markov chain (Figure 3.4a), constructed for a system supporting two trafﬁc classes as proposed in [39]. All transitions in the continuous time Markov chain (CTMC) are triggered according to Poisson processes. The impact of state dependent (soft) blocking can be taken into

account by modifying arrival rates λk in each state by some passage factor. Thus, for a

new call arriving in system state r with resource requirement mk the passage factor is deﬁned as

′ ′ 1 − br,k = Pr{r < N − r ∧ r + 1 < N − (r + 1)∧

′ ′ r + 2 < N − (r + 2) ∧ ... ∧ r + mk − 1 < N − (r + mk − 1)} (3.20)

where r′ is the interference from other cell expressed in equivalent acpBBU. The pas-

sage factor 1 − br,k denotes the probability that the neighbor cell interference is less than the available capacity in the cell under consideration (N − r), which consequently means, that a new call of service class k will be not blocked in state r. This call is only accepted if all channels required to serve it are available. The check for available resources is done sequentially, as can be seen from Equation (3.20) and Figure 3.4a. Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 42 Interface

(1 − bi+j+1)(1 − bi+j+2)λ1

i, j + 1 i + 2, j + 1

i+2 µ 2 1

(1 − bi+j)λ2 (j + 1)µ2 (1 − bi+j+2)λ2 (j + 1)µ2

(1 − bi+j)(1 − bi+j+1)λ1

i, j i + 2, j

i+2 µ 2 1

(a) A fragment of a two-dimensional Markov chain in WCDMA with state-dependent blocking and two service-classes with

m1 = 2 and m2 = 1.

(1 − b )λ (i+j+m2)+m1−1 1

i, j + m2 i + m1, j + m2

i+m1 µ1 m1

(j+m2)µ2 (j+m2)µ2 (1 − bi+j)λ2 (1 − bi+m +j)λ2 m2 1 m2

(1 − b )λ (i+j)+m1−1 1

i, j i + m1, j

i+m1 µ1 m1

(b) Modiﬁed two-dimensional Markov chain in WCDMA with state-dependent blocking.

Figure 3.4.: Comparison of two state transition diagrams of a two-dimensional Markov chains constructed for WCDMA network with state-dependent blocking.

We appropriately extended this model to study a realistic UTRAN environment. For this we assume that the multiple channels required to serve a new call of service class k are occupied simultaneously and therefore the interference remains the same during this allocation process for all channels, which corresponds to the real situation in the WCDMA radio interface. The modiﬁed state-transition diagram is shown in Figure 3.4b. 3.5. Analytical System Performance 43

With the above mentioned assumption, the service probability for a new service re-

quest with resource requirement mk is deﬁned as

′ 1 − br,k = Pr{r < N − (r + mk − 1)}. (3.21)

Furthermore, we keep in mind, that UMTS supports adaptive data rates for service- classes, depending on the load of the own cell and the current interference situation in the other cell(s). Thus, by merging Equation (3.19) with Equation (3.14) (describing the threshold model), where a new call of service k can be admitted in the system with a different transmission rate than the initially requested one, we obtain the following algorithm:

1 n P(r)= ρ m δ (r)P(r − m )(1 − b )+ r k k k k r−mk,k k=1 X (3.22) 1 n s ρk mk δk (r)P(r − mk )(1 − br−m ,k ) r l l l l kl l k=1 l=1 X X for r = 1, . . . , N, where δ (r) = 1 and δ = 0, when 1 ≤ r ≤ N (for service classes k kl without rate reduction policy), or r ≤ J + m (for service classes that have the possi- kl k bility to be admitted with reduced transmission rate), otherwise δ (r)= 0 and δ = 1; k kl s is the number of possible variations of the transmission rate, and J is the adjusted kl threshold.

3.5. Analytical System Performance

Some numerical examples are presented in this section, in order:

• to demonstrate the ﬂexibility and the accuracy of the analytical method used to evaluate the teletrafﬁc performance of WCDMA systems,

• to study how various parameters affect the performance measures of interest,

• to quantify the benefits of the proposed traffic management strategy by calculating the maximum number of users that a UMTS cellular network can handle for a given Grade of Service (GoS) under certain system conditions, such as network topology, relevant cell scenario, realistic user/traffic distribution profile, etc.

3.5.1. Performance evaluation analysis

We consider a FAG in a WCDMA system, supporting two trafﬁc classes: speech calls

with a data rate d1=12.2 kbit/s and video streaming with d2=128 kbit/s; Eb/N0 targets are 4 dB and 1.5 dB, respectively. We investigate the system performance for a mixed service scenario with 70 % speech calls and 30 % data calls. The maximum number of Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 44 Interface

channels is 315, the other-cell interference factor α = 0.55 and the offered trafﬁc for the above mentioned trafﬁc-mix scenario is ρ = 35 erlang, with the resource require-

ments m1 = 1 BBU and m2 = 10 BBUs, respectively. The other-cell interference power is a log-normal distributed random variable. Under realistic assumptions, the mean 2 −15 other-cell interference power is E{| Iother | } = 5 × 10 mW/Hz with a variation coefﬁ- cient of 1, corresponding to medium trafﬁc load. The main system parameters are also summarized in Table 3.2.

Table 3.2.: Main simulation parameters. Type of the service k Speech Data

Data rate, dk (kbit/s) 12.2 128 Eb/N0 (dB) 4 1.5 Corresponding number of required BBUs, mk 1 10 Chip rate of spreading signal (Mchip/s) 3.84 Maximum number of channels/cell 315

Obviously, systems with a complete sharing policy the service classes with low transmission rates profit from a lower blocking probability in comparison to high-rate service classes. A very undesirable situation is, however, when the blocking probability of high- rate service classes exceeds the GoS constraint, while the blocking probabilities of other classes are much lower than the prescribed constraints. In the following we examine the effect of applying different admission criteria to different traffic streams in order to give satisfactory performance in terms of both blocking probability and satisfactory user transmission rate for each traffic class. Specifically, the blocking probability for the speech service class and the blocking probability for the data service class are expected to be in the range of 1-2 % and 10-11 %, respectively. For this, we assume that speech is a constant bit rate (CBR) service, while data calls allow for adaptive data rates, depending on the load of the system. Each new user of the data class has a resource requirement d = 128 kbit/s that will be satisfied if r ≤ J (in k kl our case l = 1) and a reduced rate d = 64 kbit/s, otherwise. Figure 3.5 demonstrates kl the system performance versus the threshold for a mixed-service scenario with 70 % speech calls and 30% data calls. If no threshold is applied, the blocking probability for the algorithm given in [39] without TRRP is obtained (right points of the curves). This is the worst-case blocking. From the figures one can see that setting the threshold to a low value leads to a decrease in service probability of speech calls and to an unnecessary degradation of the transmission rate for the data service class, whereas a gain in terms of service probability for data calls is hardly observable. In turn, the overes- timation of the threshold (too high threshold) has a negative impact on the blocking probability of data calls. The optimal threshold for the above mentioned scenario is J = 200, corresponding to a transmission rate of a data call of about 77 kbit/s. In Table 3.3 we compare the performance of the algorithm of [39] (without TRRP) with that of our algorithm. The results indicate a significant system performance im- 3.6. Summary 45

Blocking probability for speech (70%) and video streaming (30%) Data rate for video streaming N=315, ρ= 35 Erlang) 130 data speech 120 0.25

110 0.2

100 0.15 90

Blocking Probability 0.1 80 Data rate for video streaming (kbit/s) rate forData video streaming 0.05 70

0 60 0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350 Threshold (number of channels) Threshold (number of channels) (a) Blocking probability for speech and video (b) Average data rate for video streaming streaming

Figure 3.5.: Blocking probability and data throughput for speech and video stream versus the threshold.

Table 3.3.: System performance with/without TRRP and reservation

Different system scenarios Bk (%) Speech Video Streaming Without TRRP 2.27 27.13 With TRRP 2.27 15.16 With TRRP and channel reservation 2.83 11.45

provement which can be attained by TRRP. Furthermore, through the incorporation of a reservation scheme in our algorithm we can fulfill the specified GoS demands. Obvi- ously, service classes with higher resource requirements are less likely to be served as the low rate calls from e.g. the speech service class. To compensate for this imbalance we set an upper bound for the number of BBUs that may be assigned to speech calls. Namely, new service requests from the speech service class will be blocked if this service class has exhausted its quota (the population of BBUs, they are allowed to be used). It is clear, that a decreased blocking probability of data calls is achieved at the expense of some degradation of service probability for speech traffic; due to the reservation, data calls get more access to the available resources.

3.6. Summary

In this chapter we have introduced an efﬁcient and accurate algorithm for estimating the GoS at connection level in the UMTS uplink for a multi-service scenario, taking into account the main properties of the network, like wireless interference and dynamic user transmission rates. Chapter 3. Connection Level Performance Modelling for Multi-Rate Loss System with WCDMA Radio 46 Interface

The algorithm presented here is not only useful for performance prediction of UMTS. A traffic management technique has been applied to UMTS and incorporated into the analytical approach to specify how resources have to be adequately engineered to meet quantitative performance objectives. For this we have investigated the joint behavior of the admission control characteristics and the threshold and reservation schemes for UTRAN in terms of their agreement with the GoS requirements. We have demonstrated, that in case of a properly selected admission policy a reduction of data call loss from about 28 % for the no-threshold case to about 11% can be achieved. Although our results have been presented for a two-class system, with our algorithm we can easily extend our model to more service classes, as well as to multiple admission thresholds. Though, due to the TRRP the proposed model looses the product form solution, the accuracy of our approximation has been verified by simulations and is found to be rather satisfactory. The verifications have been made within the scope of a student research project; the results can be partly found in [17]. In the next section we will discuss the extension to a joint connection- and packet-level traffic model and analyze the interaction between these two service levels. Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA Radio Interface

4.1. Motivation and Related Work

Most studies dealing with Call Admission Control (CAC) algorithms only take connection level performance measures into consideration [63, 64, 54]. This approach is correct and sufficient while analyzing the performance of circuit-switched services. Opti- mization problems of QoS in wireless networks accommodating constant bit rate (CBR) services have been extensively studied [83, 55]. In this case, the packet-level QoS is assured as long as the required bandwidth is guaranteed, and the problem becomes the same as for connection-oriented circuit-switched services. Thus, the joint optimization two-level QoS problem can be reduced to an analysis of the system behavior at connection-level. On the other hand, work addressing packet-level QoS aspects, usually do not consider algorithms for applying appropriate traffic performance relations to attain QoS targets, given a specific kind of user population. In [10, 18] considerable effort is made to design a variety of QoS strategies and mechanisms. Currently there are only a few proposals for simultaneously addressing both, connection- and packet-level QoS metrics. New call blocking probabilities, handover dropping probabilities and packet losses have been calculated in [34] based on joint connection and packet-level QoS. However some key features of wireless networks, such as interference limited soft capacity, has not been included in the model so far. In the above mentioned studies, system performance has been analyzed using suitable simulation tools. An analytical approach for performance estimation in packet- oriented networks with on-off traffic sources has been proposed in [53]. Connection request arrivals are modeled by batched Poisson processes and at packet level the lost packet cleared model (equivalent to Blocked-Call-Cleared (BCC)) is assumed. The model has been analyzed for two levels in a separate manner assuming product form, i.e. service independence. However, the reduced load due to lost packets, i.e. reduction Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 48 Radio Interface

of the mean holding time, is not taken into account at connection level. Thus the model is an approximation. The motivation behind our work is to investigate the performance prediction and the QoS provisioning problem in wireless packet-oriented cellular networks by taking the interaction between two service levels (admission control procedure and call handling process) into account. The algorithm discussed in this chapter deals with the problem of coupling connection and packet level QoS characteristics in an analytical way. The traffic is modeled as multi-rate Binomial-Poisson-Pascal traffic (BPP) (see Ap- pendix B) at connection level and as on-off traffic at packet level. The Blocked-Call- Held (BCH) scheme is used for packet level analysis. This scheme is introduced in Section 4.2.2.1. The probability of blocking a new connection and the probability of packet losses are our primary GoS/QoS metrics for the connection level and the packet level, respectively.

4.2. Multi-Cell Two-Level Capacity Model

In our approach, the network serves n independent classes of BPP trafﬁc streams at connection level. The pole capacity of the system (cell) is N bandwidth units, to be

referred to as channels. A stream of type k requires mk basic bandwidth units (BBUs). −1 The mean service time for a stream of type k is µk and is exponentially distributed. However call length distributions maybe non-exponential. It is known that the model considered is insensitive to the distribution of the service time [37].

Trafﬁc stream k is characterized by its mean offered trafﬁc ρk and peakedness Zk,

(k = 1, . . . , n). For a Poisson arrival process ρk = λk/µk is the offered trafﬁc measured in number of connections and the peakedness equals one. For the Binomial (Engset)

case the arrival rate is (Sk − rk) γk, when rk sources are busy.

For the Pascal case, the arrival rate is (Sk + rk) γk when rk sources are busy. For a

linear state-dependent Poisson arrival process, the offered trafﬁc is ρk = S (1 − Zk),

where Sk is the number of trafﬁc sources. The peakedness is Zk = 1/(1 + βk), where

βk = γk/µk, and γk is the arrival rate of an idle source. Mathematically, we can deal

with Pascal traffic with the same formulæ as for Engset by letting Sk and βk be negative. For Engset traffic, peakedness is less than one (smooth traffic), whereas for Pascal traffic peakedness is greater than one (bursty traffic). We have the following relations between the two representations:

β 1 ρ = S · Z = , (4.1) 1 + β 1 + β 1 − Z ρ β = S = . (4.2) Z 1 − Z 4.2. Multi-Cell Two-Level Capacity Model 49

To describe the trafﬁc stream at packet level, a two-state on-off-model is used for each service class. The user of service class k can alternate between the active on-mode,

requiring transmission data rate dk and consuming radio resources or the inactive off-

mode, using no resources. An activity factor νk (0 < νk ≤ 1) defines the proportion of time stream k stays in on-mode. The service time for packet calls (packet data session) is composed of multiple packet data calls with periods of inactivity in between. Therefore, the service time distributions are assumed to be any general time distribution −1 with a mean value of µk , independent of the arrival process. We denote the traffic load, which reflects the load of all users which are simultaneously in on-mode, as the effective system load. The effective system load depends on the activity factor of each particular stream class.

4.2.1. Connection Level

At connection level in our model each connection is accepted with a certain peak rate dk

(dk = mk · 12.2 kbit/s), where mk is a positive integer and denotes the number of BBUs necessary to support the required data rate for a particular service class. At this level, new connections will experience blocking, with a probability that depends on both the required data rate and the state of the system. A new call attempt is blocked if the actual load of the system surpasses a certain level. At this level we use the BCC model (see Figure 2.6). In general this level depends on packet level averages during a

certain time interval. As data services at packet level often have average rates (dk · νk)

lower than peak rates, dk, it is reasonable to overload the system at connection level by admitting more service requests, than the system can simultaneously handle. A certain Quality of Service (QoS) can still be maintained, however. This system may be evaluated for BPP multi-rate trafﬁc streams using either the convolution algorithm allowing for minimum and maximum allocation for each stream (see Appendix C) or the generalized state-based algorithm allowing for trunk reservation [36]. We have applied the convolution algorithm to the model to deﬁne the feasible state space at packet level. By convolution we can obtain the marginal state probability

distribution of each service class. For each trafﬁc stream we have evaluated time Ek,

call Bk, and traffic Ck congestion [37]. These congestion measures are identical for Poisson arrival processes, because the traffic intensity is independent of the state of the system (PASTA-property), but for Engset and Pascal traffic the relevant measure is the

trafﬁc congestion Ck, i. e. the proportion of offered trafﬁc which is lost. Congestion at this level may be denoted as hard blocking.

4.2.2. A Novel Model for Packet Level Performance Evaluation

As mentioned earlier, the capacity of the system is not limited by the number of spreading codes but by multiple access interference (MAI). A user produces interference only if it actually transmits data. This feature of Wideband Code Division Multiple Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 50 Radio Interface

Access (WCDMA) can be interpreted as an inherent but uncontrolled statistical multiplexing of on-off-sources that leads to a higher resource utilization but also to some uncertainty for network planning which we will discuss next. We exploit the fact that the data service is composed of multiple packet data calls, which are interspersed with inactivity periods of the transmission. These services are interpreted as on-off-sources, transmitting at peak rate during on periods and with zero rate during off periods. At packet level, there is no feedback to the user and thus a packet will be transmitted independently of whether it is (partly) lost. This is modeled by Fry-Molina’s BCH model (here call means packet) [23].

4.2.2.1. Blocked-Call-Held (BCH) Model (Fry-Molina Model) at packet level

The Blocked-Call-Held (BCH) model, also called Fry-Molina Model, regards pure loss systems, however, differs principally from the Blocked-Call-Cleared (BCC) model. The fundamental difference between the BCC and the BCH models is shown in Figure 4.1. A new data call which cannot be serviced immediately due to insufficient capacity is held for a period of time equal to its original service time. If during the service time of a “held” request the system has evolved to a state where sufficient capacity is available, the call will be assigned the available number of BBUs for its remaining service time. After the service time has expired, the part of the call which has not been processed yet is lost and cleared from the system. The corresponding state transition diagram is presented in Figure 4.2. The BCH model is based on the non-truncated Poisson arrival process. The service time is negative exponentially distributed with mean value µ. In fact, the Fry-Molina model considers a pure loss system. However, as one can infer from the state diagram, it can be interpreted as a waiting system with N channels and an infinite queue with timeout, as shown in Figure 4.3. Nevertheless, BCH systems are distinctly different from conventional waiting systems, where the requests are queued and get the service as soon as resources become available. The time a request/call spends in such a system is the sum of the waiting time and the service time. In the BCH model, the timeout δ is exponentially distributed with a mean value equal to the mean service time; the total time spent in the system is independent of the waiting time and is determined by the required service time. Since the exponential distribution is memoryless, the system can be exactly analyzed using the above interpretations. Thus, the distribution of a call’s remaining service time is identical to the distribution of its total service time. The state probabilities for BCH can be calculated as follows:

P(r) for 0 ≤ r < N P(r)= ∞ (4.3) ( j=N P(j) for r = N,

where r is the number of busy BBUsP and N is the total number of BBUs. 4.2. Multi-Cell Two-Level Capacity Model 51

User 1 Carried traffic User 2 User 3 Complete loss User 4

User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Real capacity Carried traffic: 20 BBUs N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time (a) Blocked-Call-Cleared model

User 1 Carried traffic User 2 User 3 Shared loss each channel carries 3/4 User 4

User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Real capacity Carried traffic: 29 BBUs N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time (b) Blocked-Call-Held model

Figure 4.1.: Blocked call models. Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 52 Radio Interface

λ λ λ

0 1 2 N − 1 N

µ 2µ Nµ (N + 1)µ λ

N + 1

(N + 2)µ λ

N + 2

(N + 3)µ λ

Figure 4.2.: State transition diagram for the Molina model.

Inﬁnite Arriving queue calls N channels

Timeout δ

Figure 4.3.: Model of the Molina loss system regarded as a waiting system with timeout.

The call congestion in the Molina system can be deﬁned as the probability that a call does not obtain an idle channel during its sojourn time. In our models we used a modiﬁed version of BCH – Blocked-Call-Interfered (BCI).

4.2.2.2. Blocked-Call-Interfered (BCI)

In Blocked-Call-Interfered (BCI) all packets, which are blocked due to wireless interference, will be lost. At packet level, losses occur due to overload in multiplexing streams (other users in own cell) and interference from neighboring cells. The interference power from neighbor cell is modeled as a stochastic variable with Log-Normal distribution as described in Section 3.4.1. For a given number of accepted heterogeneous connections we evaluate the loss due to interference at packet level using multi-rate BCH models and incorporate other-cell-interference. The main differences between the BCH (from the previous subsection) and the BCI model are demonstrated in an illustrative example in Figure 4.4. The BCI model we used, has no immediate feedback from the packet level to the connection level about lost data. 4.2. Multi-Cell Two-Level Capacity Model 53

User 1 Carried traffic User 2 User 3 Shared loss each channel carries 3/4 User 4

User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Real capacity Carried traffic: 29 BBUs N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time (a) Blocked-Call-Held model

User 1 Carried traffic User 2 User 3 Complete loss User 4

User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Real capacity Carried traffic: 20 BBUs N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time (b) Blocked Call-Interfered model

Figure 4.4.: Comparison of BCH and BCI models. Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 54 Radio Interface

This loss should be below a certain threshold to ensure the QoS, and, we may by averaging the loss at packet level give feedback to the connection level. The blocking probability at connection level should not depend on ﬂuctuations of the interference during the transmission of a single packet, but on changes on a time scale of the same order as the mean connection holding time. Blocking at packet level is called soft blocking. The main performance measure for the BCI model is the system throughput, which can be directly derived from the rate of successful packet transmissions. To calculate the performance measures for each service at packet level we used the recursion formula for multi-rate loss systems from [38]:

m ρ r − m 1 − Z P (r) = k · k · P(r − m ) − k · k · P (r − m ) · (1 − b ) (4.4) k r Z k r Z k k r−mk,mk k k

where r is the number of busy channels in the system, Pk(r) denotes the contribution

of stream k to the global state probability P(r), mk is the number of BBUs required for a user of service class k; and b is a state dependent blocking probability (soft r,mk blocking), which characterizes the current other-cell interference level. Each trafﬁc

stream is characterized by its mean offered trafﬁc ρk and peakedness Zk.

4.2.3. Performance Measures

In the following the deﬁnitions of the main performance measures are given. We obtain individual performance measures for each service class, both at connection and at packet level. We investigate the trade-off between these two levels to meet the Grade of Service (GoS) requirements for cellular networks with a WCDMA radio interface by means of several case studies. The performance measures below are connection level capacity constraints.

Traffic Congestion, (Ck) The fraction of offered traffic which is not carried, despite possibly several attempts. When dealing with the traffic concept, a clear distinction between offered traffic and carried traffic should be made. The offered traffic is the traffic offered to a system in accordance with a defined theoretical description of the traffic case [35]. The offered traffic is usually equivalent to the average number of call attempts per mean holding time (τ = 1 ); ρ = λ 1 = λ . µ µ µ The traffic carried is the traffic effectively handled by a system. When the capacity of the system is infinite, the offered traffic is equal to the traffic carried.

Call Congestion, (Bk) The fraction of all call attempts, which observes all BBUs busy.

Time Congestion, (Ek) The fraction of time when all BBUs are busy. 4.2. Multi-Cell Two-Level Capacity Model 55

The calculation of the mentioned congestion measures is done by using the following equations.

Time congestion Ek for stream k is deﬁned as

1 E = P k (r), (4.5) k ik G r∈S XEk where

SEk = (ik, r)|(ik > Nk − mk) ∨ (r > N − mk) , r = ik (4.6) k X and ik denotes the number of channels occupied by service class k. The summa-

tion is extended to all states SEk , where service requests from class k are blocked:

(ik > Nk − mk) defines the states, where traffic stream k has exhausted its quota (each service-class k has a finite BBU population, N denotes its upper bound), and k r > N − mk corresponds to the states with less than mk idle channels in the system. G is a normalization constant. The call congestion Bk for traffic stream k is defined as the ratio between the number of blocked call attempts for stream k and the total number of call attempts for stream k.

k λk(ik) · P (r) SEk ik Bk = N r (4.7) Pk λ (i ) · P k (r) r=0 ik=0 k k ik P P Yk, is known as the carried trafﬁc for stream k:

Nk r Y = i · P k (r) (4.8) k k ik r=0 i =0 X Xk

The traffic congestion, Ck, is usually calculated using the offered ρk and carried Yk traffics (both measured in connections or BBU channels). The traffic congestion is given by the following ratio:

ρk − Yk Ck = (4.9) ρk

Performance measures at packet level are:

Packet (dropping) loss probability: The calculated capacity at packet level is the maximum number of packet data calls that could be concurrently accommodated in a cell with their peak data rate. The capacity is subject to the constraint that interference does not exceed a predeﬁned threshold. If this constraint is violated, packet dropping will occur. The ratio of dropped packets is called packet dropping probability. Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 56 Radio Interface

Table 4.1.: Service modeling parameters Stream k = 1: Poisson k = 2: Binomial k = 3: Pascal Total (speech) (web browsing) (e-mail)

λk 1 – –

Sk ∞ 5 −37 γ 0 4 − 1 k 3

µk 1 1 1 1 Z = γ 1 0.2 1.5 k 1+ µ mk 1 10 2

ρk = S(1 − Z)mk 50.5 40 37 127.5

νk 0.5 0.3 0.7 ρ = ρν 25.25 12 25.9 63.15 cp k

Average service data rate: The mean value of the individual data rates of each service class.

4.3. Performance Results

In the following some numerical examples of the proposed method for performance estimation in wireless packet-oriented networks are presented. The first case shows how performance measures for a particular traffic mix are obtained. The second one demonstrates the effects of changing the offered traffic load. The two last examples investigate the QoS level by changing the packet drop level and the activity factor, respectively.

4.3.1. Impact of the Traﬃc Mix

Performance measures for a particular traffic mix are presented below. Our calculations are based on the traffic parameters defined in Table 4.1. The service mix has been specified in the following way. Three services (n = 3) were defined being

• speech or radio stream (regular narrow-band: Z1 = 1, m1 = 1, ν1 = 0.5),

• web browsing (regular broadband: Z2 = 0.2, m2 = 10, ν2 = 0.3),

• e-mail (bursty narrow-band: Z3 = 1.5, m3 = 2, ν3 = 0.7). We deﬁne a BBU equal to the amount of resources needed to carry one speech call. The maximum cell capacity is n = 128 BBUs, which is the pole capacity of the cell. 4.3. Performance Results 57

Table 4.2.: Performance results of the trafﬁc scenario from Table 4.1. Stream k=1: Poisson k=2: Engset k=3: Pascal Total (speech) (web browsing) (e-mail) Connection level results

Ek 0.049 0.467 0.098 Ck 0.049 0.096 0.146 Bk 0.049 0.348 0.102 Yk 48.02 36.14 31.59 115.75 Packet level results p ρk = Yk·νk 24.008 10.843 22.111 56.962 p Yk 20.699 9.012 18.964 48.676 p− p p ρk Yk Ck = p 0.138 0.169 0.142 ρk Y pth 23.643 10.496 21.73 55.869 k p p− th pth ρk Yk Ck = p 0.015 0.032 0.017 ρk

In the example we show the system behavior when the offered load is close to pole cell capacity (127.5 erlang) and each service class contributes approximately the same amount of traffic. The other-to-own-cell interference ratio is α = 0.55. The mean value of the other-cell-interference is chosen equal to its variance to indicate that the other- cell load is of medium level with average fluctuations. Calculations have been done for a scenario with small fluctuations (variance equals 1/10 of the mean value) and for large fluctuations of the interference power (the variance is two times larger than the mean value) as well. However, all further calculations are presented under the assumption of a mean value equal to the variance. Performance results of the setup given above are presented in Table 4.2 and should be read as follows:

Ek, Ck and Bk denote time, traffic and call congestion at connection level. Yk is the p carried traffic at connection level, ρk = Yk·νk is the traffic accepted at packet level p (carried traffic at connection level scaled down by the activity factor, νk). Yk is the p p p p carried traffic at packet level. Ck = (ρk − Yk )/ρk is the traffic congestion at packet level. pth Yk is the carried traffic for the BCH model, where the loss level is identified by some predefined threshold and without neighbor-cell-interference. The BCH model should pth p pth p be used in non-wireless environment. Ck = (ρk − Yk )/ρk is the proportion of lost traffic at packet level. The traffic congestion expresses the proportion of information lost at packet level. Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 58 Radio Interface

120 92 1st, radio stream, BCH 91 2nd, web-browsing, BCH 100 90 3rd, email, BCH 1st, radio stream, BCH 80 2nd, web-browsing, BCH 89 3rd, email, BCH Total 88 60 87

40 86

85 20

Carried traffic on connection level (Erlang) 84 Percentage of carried traffic on packet level 0 83 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50 Offered traffic on connection level for service 1 (Erlang) Offered traffic on connection level for service 1 (Erlang) (a) (b)

10 92 9 91 8 90 7 89 6 88 5 1st, radio stream, BCH 2nd, web-browsing, BCH 87 4 3rd, email, BCH 86 3

85 1st, radio stream, BCH Average service data rate/12.2 kbit/s 2 2nd, web-browsing, BCH Percentage of carried traffic on packet level 84 3rd, email, BCH 1

83 0 0 5 10 15 20 25 30 5 10 15 20 25 30 35 40 45 50 Offered traffic on packet level for service 1 (Erlang) Offered traffic on connection level for service 1 (Erlang) (c) (d)

Figure 4.5.: Carried trafﬁc at connection/packet level as a function of the offered traf- ﬁc at connection/packet level, respectively. The system pole capacity is 128 channels.

4.3.2. Analyzing trends by load change

A goal of this analysis is to study how system performance changes when the load is altered. The system load is changed by keeping the offered traffic load from two service classes fixed while varying the third traffic class. In the graphs presented below, web-

browsing and email traffic are fixed and the speech traffic load (ρ1) is increased from 8 to 48 erlang, while the other two services have the same load as shown in Table 4.1. Similar calculations for other services have been performed as well. Figure 4.5a presents the relation of offered traffic and carried traffic at connection level for each service. Obviously, the results are self-explanatory. These plots have been included for scenario visualization purposes. Figure 4.5b relates the behavior of the carried traffic at packet level to the offered traffic at connection level, for each service class. Increasing the load at connection level reduces the carried traffic at packet level due to increased interference. At packet level, usually applications can tolerate a particular amount of lost traffic (protection by 4.3. Performance Results 59

forward error correction (FEC) or automatic repeat request (ARQ) mechanisms either at physical layer or at application layer – like transmission control protocol (TCP) segment retransmission, or Moving Picture Experts Group (MPEG) frame interpolation). If this threshold is exceeded, the session will most probably be terminated by the network for not satisfying the QoS requirements, or packets will become extensively delayed. Figure 4.5c demonstrates the dependences of the carried traffic for each service at packet level to the accepted traffic at connection level. Figure 4.5c represents the same data as in Figure 4.5b, however note that the x-axis has been transformed. Namely, in Figure 4.5b the x-axis shows the offered load for speech service, while in Figure 4.5c each curve follows its own accepted load at packet level. Figure 4.5c should be read as follows. When the load of the speech service is increased, the accepted traffic at packet level also increases (from left to right), and at the same time the probability of packet loss increases. In the meantime, web-traffic and email services have the same offered load at connection level, but due to higher blocking probabilities, the accepted load at packet level is reduced (curves move from right to left) and at the same time less traffic is carried at packet level (curves move downwards). Finally, Figure 4.5d shows the average throughput for each individual service class, expressed in number of BBUs. From the figure it can be seen that the mean service data rate is lower than requested (10, 2 and 1 BBU for web-browsing, e-mail and speech service classes, respectively), which is obvious, due to the lost traffic at packet level.

4.3.3. Adjustment of loss rates at packet level

As mentioned above, it is desirable to improve the average service data rate at packet level, as it directly results in higher user satisfaction. This means that the performance shown by the curves in Figure 4.5d should be closer to the target values (required data rates). The way to achieve this is to accept less traffic at connection level. In the previous scenario users have been allowed to load the network up to the pole capacity (128 channels) (due to the activity factor, the network at packet level is never loaded to the pole capacity). Figure 4.6 presents results with tightened admission control at connection level, where the cell capacity at 70 % of the system pole capacity, while keeping the same offered traffic as in the previously shown analyses. After such an adjustment, the total carried traffic of all 3 service classes at connection level has been reduced from 115 to 85 erlang (compare Figure 4.6a to Figure 4.5a), assuming the offered traffic load for the speech service is fixed at 48 erlang (heavy load). Packet level performance has been improved from 86.5 % to 91.5 % for speech service, and from 83 % to 90 % for bursty e-mail service, see Figure 4.6b. If services are able to tolerate up to 10 % of packet loss, such an adjustment would ensure that sessions are not dropped (or delayed) at the expense of reducing the total accepted traffic by 20 erlang. Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 60 Radio Interface

90 93

80 1st, radio stream, BCH 92.5 2nd, web-browsing, BCH 1st, radio stream, BCH 70 3rd, email, BCH 2nd, web-browsing, BCH 3rd, email, BCH 92 60 Total

50 91.5

40 91

30 90.5 20 90 Carried traffic at connection level (Erlang) level atconnection Carried traffic

10 level at packet traffic carried of Percentage

0 89.5 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50 Offered traffic at connection level for service 1 (Erlang) Offered traffic at connection level for service 1 (Erlang) (a) (b)

10 93

9 92.5 8 1st, radio stream, BCH 92 7 2nd, web-browsing, BCH 3rd, email, BCH 6 91.5 5

91 4

3 90.5 Average service data rate/12.2kbit/s Averageservice 1st, radio stream, BCH 2 Percentage of carried traffic at packet level at packet traffic carried Percentage of 90 2nd, web-browsing, BCH 3rd, email, BCH 1

89.5 0 0 5 10 15 20 25 5 10 15 20 25 30 35 40 45 50 Offered traffic at packet level, Erlang Offered traffic at connection level for service 1 (Erlang) (c) (d)

Figure 4.6.: Analysis of the system performance with tightened admission control at connection level (70 % of the pole capacity, 90 channels).

4.3.4. Impact of the Activity Factor

As previously mentioned, modelling and analysis of the interaction between two service levels (connection and packet level) allows for the assessment of the multiplexing gain and evaluation of the impact of neighbor cell interference. The multiplexing gain is achieved by interleaving on-off-trafﬁc sources; the data services considered in our work can dwell in two states: idle, consuming no resources and active. A connection is established and the source is transmitting at full power, for the given RF conditions.

These states are represented in our models by the activity factor νi. Figure 4.7 shows the behavior of the trafﬁc performance when the activity factor of the second service (broadband web-browsing) is 0.6 instead of 0.3, and the bursty third service has its activity factor reduced from 0.7 to 0.3. At the same time, the system capacity is limited to 78% of its pole capacity (100channels). 4.4. Summary 61

100 93 1st, radio stream, BCH 90 92.5 2nd, web-browsing, BCH 80 1st, radio stream, BCH 3rd, email, BCH 2nd, web-browsing, BCH 70 92 3rd, email, BCH 60 Total 91.5 50 91 40

30 90.5 20 90

10 Percentage of carried traffic on packet level Carried traffic on connection level (Erlang)

0 89.5 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50 Offered traffic at connection level for service 1 (Erlang) Offered traffic on connection level for service 1 (Erlang) (a) (b)

93 10

1st, radio stream, BCH 9 92.5 2nd, web-browsing, BCH 3rd, email, BCH 8

92 7 1st, radio stream, BCH 6 91.5 2nd, web-browsing, BCH 5 3rd, email, BCH 91 4

90.5 3

90 Average service data rate/12.2 kbit/s

Percentage of carried traffic on packet level 1

89.5 0 0 5 10 15 20 25 5 10 15 20 25 30 35 40 45 50 Offered traffic on packet level for service 1 (Erlang) Offered traffic on connection level for service 1 (Erlang) (c) (d)

Figure 4.7.: Impact of activity factors on the system throughput.

4.4. Summary

In this chapter a novel analytical method for performance evaluation of wireless packet- oriented networks has been introduced. The method is based on the combined analysis of connection- and packet levels, which have been evaluated by the Blocked-Call- Cleared (BCC) and Blocked-Call-Interfered (BCI) strategies. The advantage of using the BCI model at packet level is that the load of the system only depends on the blocking at connection level, not on the blocking at packet level. Thus, we may calculate the performance measures in one step. In other works as [53], the BCC model is used at both, the connection level and the packet level. The model has been analyzed for two levels in a separate manner assuming product form, i.e. service independence. Then the model becomes an approximation since the reduced load due to the lost packets, i.e. reduction of the mean holding time, is not taken into account at connection level. Therefore, the BCI model is more realistic to be used at packet level, rather than at the connection level. The presented work is novel and differs from other analytical solutions for QoS estimation of packet-oriented networks in the way performance results have been evalu- Chapter 4. Joint Connection and Packet Level Analysis for Multi-Rate Loss System with WCDMA 62 Radio Interface

ated. Both, the specifics of UTRAN, e. g. the joint behaviour of two QoS provisioning levels, as well as the wireless interference, a realistic model for packet-call duration, etc., have been taken into account. An example has been presented to show how the proposed methodology can be applied to traffic engineering. Numerical results indicate that the user admission control procedure and the call handling process are not independent and that their interaction has a significant effect on the QoS. In contrast to the widespread view, that analytical approaches for modelling a radio interface are too simplified and yield inaccurate results, the analytical method proposed here demonstrates the flexibility and the accuracy of this approach. Of course, some degree of simplification is necessary. For this, the dominant aspects of the real quantities have been identified and represented in the selected model and less important aspects have been neglected. Chapter 5. Extended Analysis of Two-Level Performance Model for Multi-Rate Delay System with WCDMA Radio Interface

5.1. Motivation and Related Work

The work presented is this Chapter is an extension of the previous work (see Chapter 4), in which an analytical approach for modeling and analysis of interaction between two service levels (connection level and packet level) in packet-oriented multi-service wireless systems was introduced. This work proposes a more general and more detailed performance analysis of wireless large-scale queueing networks. A new uniﬁed analytical model, which combines both loss and queuing systems with multi-rate trafﬁc streams, has been applied for theoretical performance study of multi-service wireless systems with WCDMA radio interface. Examples for evaluation of realistic cases are given.

5.1.1. On the Generality of the Algorithm

The number of generalizations of the traditional teletraffic paradigms, which we can make by applying a new unified analytical model to large-scale systems, renders system performance analysis quite diversified. The key aspects of the algorithm are the following:

• It supports the coupling between connection level and packet level QoS characteristics for system performance analysis.

• Suitable for representation and analysis of all classical Markovian trafﬁc models.

• It supports both real-time delay sensitive services as well as non-real time background services. Chapter 5. Extended Analysis of Two-Level Performance Model for Multi-Rate Delay System with 64 WCDMA Radio Interface

• Generalization of scheduling policy and service priority discipline to performance evaluation of trafﬁc streams with different QoS provisioning problems (parameters).

5.2. Uniﬁed Analytical Traﬃc Model

The analytical model, used in our work, is a generalization of a two-level top-down analytical model proposed in [36] for performance analysis of multi-rate loss systems. At connection level the network serves n independent classes of Binomial-Poisson-

Pascal (BPP) traffic streams. Traffic stream k is characterized by the offered traffic ρk

and peakedness Zk. For each service class, a two-state on-off model is used to describe the accepted traffic streams at packet level. At connection level the offered traffic is handled according to a Blocked-Call-Cleared model (BCC) [37]. At packet level the more complex traffic behavior is evaluated by Blocked-Call-Buffered (BCB) strategy. The main differences between the Blocked-Calls-Held (BCH) model, which was used in Chapter 4 and in [5] for traffic modeling at packet level, and the extended new model are demonstrated with an illustrative example in Figure 5.1. At packet level the system has N + c channels; N represents the number of effective channels (real capacity of the system), and c reflects the buffering capacity. Thus the service rate of each accepted connection is reduced and the transmission time increased when some of the buffers are occupied. Each user is assigned one or multiple channels (BBUs), depending on the service class of the source traffic. All users in this simple example are subject to a capacity sharing algorithm. When all streams are single-slot this is the Processor Sharing algorithm. To incorporate a buffer into the model and to adapt the algorithm for representation of all classical Markovian traffic models the recursion formula for multi-rate loss systems [39] has been appropriately modified:

r mk ρk r − mk 1 − Zk Pk(r)= max , 1 · · · P(r − mk) − · · Pk(r − mk) · N r Zk r Zk § ª (5.1) (1 − b ) r−mk,mk

where r is the number of busy BBUs in the system, Pk(r) denotes the contribution of

stream k to the global state probability P(r), c is the buffer length, mk is the number of BBUs required by stream k, and b is a state dependent blocking probability. The r,mk main difference between the two models is the system throughput improvement, which we achieve by incorporating the buffer scheme into the algorithm. 5.2. Uniﬁed Analytical Trafﬁc Model 65

User 1 Carried traffic User 2 User 3 Shared loss each channel carries 3/4 User 4

User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Real capacity Carried traffic: 29 BBUs N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time (a) Blocked-Call-Held model

User 1 Carried traffic User 2

User 3 Buffered Traffic User 4

User 5 User 6

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time

Buffer capacity c=2

Real capacity Carried traffic: 32 BBUs N=3

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Time (b) Blocked-Call-Buffered model

Figure 5.1.: Illustration of the main difference between Blocked-Call-Held and Blocked- Call-Buffered Chapter 5. Extended Analysis of Two-Level Performance Model for Multi-Rate Delay System with 66 WCDMA Radio Interface

The basic performance measure of interest for the Blocked Call Buffered model is the mean carried trafﬁc. The total carried trafﬁc (non-delayed and delayed) for stream k is given by:

N N+c

Yk = r · Pk(r)+ N · Pk(r) (5.2) r=0 r=N+1 X X Data throughput is deﬁned here as the number of carried trafﬁc units, for example measured in bytes.

5.3. System Performance Analysis

We concentrate our analysis on a multi-service wireless queueing network with WCDMA radio interface. We investigate the behavior of multiple traffic flows under more realistic assumptions, e.g. taking the influence of packet delay on the QoS characteristics and the effect of soft blocking into account. The influence of some system parameters on network performance characteristics and on individual QoS indicators for each service-class has been already discussed in the previous Chapter. We present further performance analysis of WCDMA networks. In particular, by generalizing the traditional teletraffic paradigm we extend a previous relationship between loss and delay systems and traffic specifications. Although the presented study is for- mulated and numerically analyzed for a specific network type, qualitative conclusions are easily extended to a broader scope of networks.

5.3.1. Call Admission and Handling Policy

In this subsection, the equations and performance analysis for the non-priority case, i.e., no differentiation between services classes, are given. We introduce a buffers scheme in order to maximize the average system performance as well as individual users’ throughput. The data calls admitted into the system at the connection level according to BCC (Blocked Call Cleared) strategy can, at the packet-level be: off, i.e. in idle mode, or on, i.e. in busy mode. If a sufficient number of channels are available, then packets are transmitted without delay. If buffers are occupied, then the available capacity is shared in a reversible way and all packets are delayed. Our performance measures at packet level are mean carried traffic, mean packet queue length, overall mean waiting time, and packet dropping probability. The performance metrics at connection level are primarily given by call blocking probability and mean admitted traffic volume. 5.3. System Performance Analysis 67

Table 5.1.: Cell parameters and service trafﬁc description Stream k = 1: Poisson k = 2: Engset k = 3: Pascal (radio stream) (web browsing) (email)

Peakedness, Zk 1 0.2 1.5 Number of required BBUs, mk 1 10 2 Offered trafﬁc, ρk (Erl) 4(8)128 40 37 Activity factor, νk 0.5 0.3 0.7

5.3.2. Comparison of Blocked-Call-Buﬀered (BCB) model with Blocked-Call-Held (BCH)

An extensive variety of scenarios has been analyzed in order to obtain insight into the effect of the different parameter settings on the system performance measures. In the following we restrict ourselves to the most meaningful ones. We examine the performance of a single cell within a multi-cell environment. The cell capacity is fully shared between different service classes. The maximum number of effective channels in a cell at connection level is N=128. At the packet level the upper bound for packet admission is reduced to 42 channels. The buffer size is fixed to 20 channels. We relax the classical Poisson assumption for the traffic streams and apply the theoretical model sketched above to the following application classes in the WCDMA network: low-rate real-time services like radio streaming and delay insensitive applications like e-mail or web-browsing. Table 5.1 shows the most important parameters of the traffic streams under investigations. We start our analysis with Figure 5.2, where the system performance versus offered traffic load for the particular service-mix scenario (according to Table 5.1) is demonstrated. The y-axis shows the carried traffic in a WCDMA network, which was modeled by using the BCB algorithm. The performance is then compared with that of the BCH approach, where the level for service loss was defined by some hard threshold value (see analyzes in Chapter 4 for more details). The system load was varied only for the 1st service class by keeping a fixed value of offered traffic load for the two other services. According to the obtained results the system based on the Blocked-Call-Buffered model demonstrates essentially improved performance characteristics in comparison with the Blocked-Call-Held model, with regard to the mean carried traffic over the various parameters of offered traffic. In the graph positive effects of the buffer can be recognized in terms of outage (packet dropping) reduction within the WCDMA system.

5.3.3. Behavior of the System with Soft Blocking

In wireless queuing networks with WCDMA radio interface a quite important point to analyze is how the trafﬁc loads of other-cells inﬂuence the performance results of the Chapter 5. Extended Analysis of Two-Level Performance Model for Multi-Rate Delay System with 68 WCDMA Radio Interface

System Performance with different Packet Admission Policies 40 1st, radio stream, BCB 2nd, web−browsing, BCB 35 3rd, email, BCB 1st, radio stream, BCH 30 2nd, web−browsing, BCH 3rd, email, BCH

10 Carried traffc on packet level (Erlang)

0 0 20 40 60 80 100 Offered traffic to connection level for service 1 (Erlang)

Figure 5.2.: Impact of buffer scheme on the packet-level system performance.

System Performance with/without wireless interference System Performance with/without wireless interference (heavy loaded other cells) (low loaded other cells) 40 40 1st, radio stream, BCB 1st, radio stream, BCB 2nd, web−browsing, BCB 2nd, web−browsing, BCB 35 3rd, email, BCB 35 3rd, email, BCB 1st, radio stream, BCB−interf 1st, radio stream, BCB−interf 30 2nd, web−browsing, BCB−interf 30 2nd, web−browsing, BCB−interf 3rd, email, BCB−interf 3rd, email, BCB−interf

25 25

20 20

15 15

10 10 Carried traffc on packet level (Erlang) Carried traffc on packet level (Erlang)

5 5

0 0 0 20 40 60 80 100 0 20 40 60 80 100 Offered traffic to connection level for service 1 (Erlang) Offered traffic to connection level for service 1 (Erlang)

Figure 5.3.: Performance comparison of system with and without soft blocking for heavy loaded other cells (left) and low loaded other cells (right).

cell under consideration. For this purpose, the other-cell traffic scenario was varied by keeping other network characteristics constant (Figure 5.3). In the graphs, the variations of the mean carried traffic for 3 service classes are shown in dependency on other-cell interference scenarios. An interesting phenomenon can be observed: the performance curves do not de- grade monotonically by introducing state-dependent blocking. We conjecture that this is due to some fluctuations of the other-cell user density, and, consequently, variable interference power from neighboring cells (low load leads to higher variation of the interference power, etc.).

5.3.4. Probability of Getting Service without Priority Classes

Another important performance characteristic is QoS of data calls, expressed in terms of the expected dwell time. 5.3. System Performance Analysis 69

The proportion of time new calls (packets) of the stream k get full service at the time of arrival is deﬁned as:

N−mk

Esk = P(r), k = 1, 2, . . . , n. (5.3) r=0 X

In a similar way Edk is calculated as proportion of time new calls of the service class k are delayed: N+c−mk

Edk = P(r), k = 1, 2, . . . , n. (5.4) r=N−m +1 Xk The time congestion Ebk of the stream k is deﬁned as:

N+c

Ebk = P(r), k = 1, 2, . . . , n. (5.5) r=N+c−m +1 X k

Obviously, Esk + Edk + Ebk = 1. Figure 5.4 demonstrates the sojourn time distribution of trafﬁc streams as a function of the offered trafﬁc load for service 1, which is varied between 8 and 104 Erlang. The left-hand graph shows the probability of immediate service, i.e., the proportion of time new calls get full service immediately upon admission. The right-hand plot depicts system performance in terms of probability of establishing a connection with some delay.

Proportion of time new calls are served immediately Proportion of time new calls are delayed 1 1 1st, radio stream, BCB−interf 0.9 0.9 2nd, web−browsing, BCB−interf 3rd, email, BCB−interf 0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3 Probability of delayed service Probability of immediate service 0.2 0.2 1st, radio stream, BCB−interf 0.1 2nd, web−browsing, BCB−interf 0.1 3rd, email, BCB−interf 0 0 0 20 40 60 80 100 0 20 40 60 80 100 Offered traffic to connection level for service 1 (Erlang) Offered traffic to connection level for service 1 (Erlang)

Figure 5.4.: Expected dwell time, probability of immediate service (left) and probability of delayed service (right).

The mean queue length of stream k measured in trafﬁc units becomes:

N+c

Lk = (r − N) · Pk(r) k = 1, 2, . . . , n (5.6) r=N+1 X and is shown in Figure 5.5. The above-mentioned sojourn time averages are also summarized in Table 5.2. Chapter 5. Extended Analysis of Two-Level Performance Model for Multi-Rate Delay System with 70 WCDMA Radio Interface

Mean Transfer Queue Length 4 1st, radion stream, BCB−interf 2nd, web−browsing, BCB−interf 3.5 3rd, email, BCB−interf

2.5

1.5 Mean Queue Length

0.5

0 0 20 40 60 80 100 Offered traffic to connection level for service 1 (Erlang)

Figure 5.5.: Mean Transfer Queue Length.

Table 5.2.: Time Average Performance Measures

Offered Trafﬁc for Service 1, (Erlang) 8 24 40 56 72 88 104 Es1 1.0 1.0 1.0 0.99 0.9 0.62 0.31 −1 Ed1 (x10 ) 0.0 0.0 0.0 0.05 0.93 3.73 6.9 E b1 0.0 0.0 0.0 0.0 0.0 0.0002 0.0035 Es2 0.98 0.98 0.989 0.993 0.997 0.999 0.999 −1 Ed2 (x10 ) 0.003 0.006 0.0 0.0 0.0 0.0 0.0 E b2 0.013 0.013 0.01 0.006 0.002 0.0007 0.0001 Es3 0.93 0.94 0.97 0.99 0.99 0.99 0.99 −1 Ed3 (x10 ) 0.66 0.5 0.23 0.07 0.02 0.005 0.001 −3 E b3 (x10 ) 0.28 0.07 0.009 0.0008 0.0001 0.0 0.0

5.4. Summary

In this chapter we have presented analyses of resource allocation schemes for large- scale queueing wireless networks using a WCDMA radio interface, in terms of performance prediction, optimization of system parameter and network dimensioning. To model the complex multi-rate multimedia traffic and eventually evaluate the performance of the system reliably, an efficient algorithm from generalized teletraffic theory is used. The proposed generalized algorithm can be used to derive different system performance indicators in a highly efficient way. The obtained results help to work out and further understand the typical behavior of multi-service systems under different air interface conditions and various traffic and scheduling scenarios. It allows for analyzing the system performance under consideration of important UMTS Terrestrial Radio Ac- cess Network (UTRAN) features, such as channel reservation, generalization of traffic arrival characteristics, handover, as well as packet-service delay. The findings of the application of the above mentioned model to WCDMA networks show the advantage of using the Blocked-Call-Buffered model as a more realistic approach for modelling system processes on the rather complex network packet-level. Although we concentrated our research on UTRAN and considered on-off variable variable bit rate (VBR), the principle can be used for performance prediction and opti- 5.4. Summary 71

mum design of arbitrary systems, such as ATM, Multiprotocol Label Switching (MPLS) and more.

Part II.

Traﬃc Management for Wireless Networks

Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Eﬃcient Data Transfer in WCDMA

6.1. Motivation

Currently the telecommunication infrastructure is characterized by a rapid growth of new high data fixed-line services. Such observations motivate operators of wireless networks to extend their business models by offering new data applications, such as real-time streaming, distributed video conferences, as well as downloading popular movies or music files. However, the basic constraint of hierarchically constructed cellular systems, namely, the limited capacity of the radio interface, remains a challenge for wireless networks. Furthermore, it should be kept in mind, that the most data applications in 3G networks differ fundamentally from the requirements of the speech service with respect to the downlink/uplink traffic asymmetry. Unlike the two-way conversational service-class, which requires strict adherence to downlink/uplink balance, the highly asymmetric traffic characteristics of data services (the demand in the downlink is likely to be several times higher than that in the uplink) lead to a quick saturation of the downlink resources. Typically, the downlink is the potential bottleneck, since all data transmissions have to be organized by providing individual links from a base station (BS) to each user required data service. As a result, the radio interface quickly becomes saturated, which leads to degradation or even loss of the service. In general, services can be divided in two groups: real-time and non-real time demands. The main distinguishing factor between these service classes is their delay sensitivity. The so-called streaming class as a subgroup of the former is very delay sensitive, while the background class belongs to the latter and is delay insensitive. Although the data content of the latter service type, like digital camera images, demo versions of programs for popular handset games and mp3 files, is interesting for many users, conventional solutions for capacity utilization such as multicast or broadcast are not applicable since users may request services at different times. 76 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

So far an appropriate solution which is capable of exploiting the potentials of wireless cellular networks to enable new services and also to motivate users to adapt to the new technology is not yet in place. The above considerations stimulate research for new paradigms with the need to reexamine prevailing principles in the architecture of cellular systems, as well as in policies in handling of user service requests. Currently, there are two main concepts by which the performance of wireless systems can be enhanced: complementary networks with multi-mode mobile terminals (MT) or cooperative networks, where the same radio interface can be used for different networks. Both focus on the efficient realization of a flexible radio interface and network architecture. The main idea behind the first approach is to exploit available technologies. By implementing several radio interfaces in one device, the network operator can provide users with a seamless connectivity to data, i.e., users can proceed to be active as they roam between networks, which is very important for real-time data session continuity. Consider for example a DVB-H (Digital Video Broadcasting) enabled UMTS (Univer- sal Mobile Telecommunication System) phone. Combining the technologies broadcast TV and cellular wireless is beneficial, since UMTS resources can be used to provide a feedback channel to support interactive services. Such an interworking between networks allows the portable use of new services which neither cellular wireless networks like UMTS nor digital radio networks like DVB-H can offer on their own. An alternative way to support an expanding variety of data applications in the 3G of cellular radio networks is to extend their existing radio interface, thereby enabling it to support different network architectures in a dynamic way. One of the interesting research directions is to consider the relationship between the two network concepts, client-server and peer-to-peer. The client-server system model has a centralized structure, where clients communicate only with the server and never with each other. A typical example for client-server networks is a cellular network. On the other hand, the peer-to-peer system model allows a direct communication between users in an ad hoc manner with minimum infrastructure. Each user offers both client and server functionality (e.g. fixed-line peer-to-peer Internet protocols like Bittorrent [13], eDonkey [50], etc.). Typically, these two network approaches were considered as competing, though, given the significant support of both concepts by the industry, most recent visions tend to regard them as complementing each other. Recently, even more progressive studies are investigating the effect of the synergetic cooperation between the above mentioned networks by using a unified radio interface. These visions are illustrated in Figure 6.1, where the differences of the contending network solutions are highlighted [19]. This chapter is organized as follows. In the next Section we discuss what value peer- to-peer multihop communications can add to mobile cellular networks and vice versa and survey the most interesting case studies related to complementary networks. Than 6.2. Cooperative Communications in Wireless Networks (General Concepts, Strategies, Principles) 77

Figure 6.1.: Visions of the contending network solutions.

we propose a new technique for the integration of peer-to-peer applications into cellular wireless networks.

6.2. Cooperative Communications in Wireless Networks (General Concepts, Strategies, Principles)

What is the actual benefit of cooperation in wireless networks? One of the primary bases for network cooperation is to fully exploit the available technologies, increase efficient usage of frequency spectrum, as well as reduce infrastructure costs. The motivation for cooperation between cellular wireless and peer-to-peer networks is the predicted ability of peer-to-peer systems to complement conventional cellular networks in areas with poor coverage, as well as in high user density areas. Owing to direct communication between MTs, there are substantially more sender-receiver pairs than in conventional cellular networks, where the data transmission is organized by providing individual links from the base station to each user. Thus, such a hybrid network structure is capable of increasing the number of MTs that can be simultaneously handled in peer-to-peer mode. In addition, the peer-to-peer network approach provides a further advantage. Due to the short range between MTs, the interference is expected to be lower, which leads to an increase in capacity and an improved QoS. On the other hand, also peer-to-peer networks can benefit from the existing infrastructure of cellular networks. Peer-to-peer networks have to tolerate an increased communication traffic so that they will be able to use a decentralized structure: distributed peers generate a considerable amount of signalling traffic for coordination between them. This drawback can be mitigated by taking advantage of already existing infras- 78 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

tructure of cellular communication systems. For example, network operators know the location, online status and service agreement of the mobile users. They can utilize this knowledge by providing each new user with information about all active users in the system in the range of tens of meters, to determine the potential cooperative community. Such an explicit use of the hierarchical cellular infrastructure leads to more efficient and reliable routing and significantly lowers the protocol complexity. Complementary and cooperative network approaches have been recently proposed for various wireless communication systems with the purpose to improve network coverage as well as individual and aggregate throughput of users. The feasibility of the extended eDonkey peer-to-peer file sharing Internet protocol in a GPRS (General Packet Radio Service) environment was investigated in [33] and appropriately modified for UMTS radio networks in [32]. The main focus lies on resource mediation (functions to locate resources or entities) and resource control mechanisms (functions to permit, prioritize, and schedule the access to resources) by using different strategies for data caching in the wired part of the network. The specialized fixed- network cache peer stores popular files in the network in order to avoid downloading content from the BS which has been already downloaded. An interesting idea for traffic balancing was also proposed in [82]. The main focus lies on improving call blocking probability for circuit-switched traffic by diverting traffic from a congested cell to a neighboring lightly loaded cell, introducing strategically located stationary ad-hoc relays. The concept is referred to as iCAR (integrated Cellular and Ad Hoc Relay System). However, the placement of fixed relays in specific location is necessary for successful use of the algorithm, i.e., at every border of two adjacent cells, it is not always possible due to legal issues, e.g., government laws and regulations, or security of private companies. Moreover, the deployment of additional equipment always incurs an extra cost. In our proposed concept the mobile terminals (MTs) operate in peer-to-peer communication mode for cooperative distribution of popular content, i.e., they act as both client and server. Similar techniques have been used in [51] and [76], where the authors allow MTs to act as ad-hoc relay stations in order to improve their data throughput. In [76] it was shown how UMTS capacity can be improved by embedding WLAN (Wireless Local Area Network) systems, whereas the authors in [51] propose a unified cellular and ad- hoc network architecture (UCAN) with combination of EV-DO (Evolution-Data Only)1 and IEEE 802.11b (Wi-Fi) networks for an ad-hoc data transmission mode. The results indicate noticeable throughput gains. However, two radio interfaces have been used to switch between networks, which implies that additional resources from alternative network(s) are taken. Further references can be found also in [11]. The key difference between the cited above studies and our work lies in the basic principle of capacity improvement techniques in wide-area cellular networks. In all mentioned above network scenarios an additional frequency band of some local-area

1Integral part of the CDMA2000 (Code Division Multiple Access) family of 3G standards. 6.2. Cooperative Communications in Wireless Networks (General Concepts, Strategies, Principles) 79

networks was used in order to achieve an improvement in individual and aggregate throughput of cellular users. The approach in [27] aims at exploiting the advantages of Intelligent Relaying (IR) techniques with the purpose to achieve an increased range of high data rate services in the UMTS network. The communications relaying protocol ODMA (Opportunity Driven Multiple Access) proposed for the UTRA-TDD (Time Division Duplex) mode is used [1]. The IR technique is illustrated in Figure 6.2.

Figure 6.2.: The Intelligent Relaying technique.

According to the maximum transmission bit rate that can be achieved between BS and MT, the cell coverage area is divided into two regions: high bit rate area and low bit rate area. The ODMA mechanism increases the transmission bit rate for the MTs in the low bit rate area by relaying the information transmitted between a BS and some particular MT in the low bit rate area over a number of hops, which retransmit data on behalf of this MT. The performance results, obtained via analytical models and simulation experiments proposed in this study, indicate a significant improvement of the transmission rates for the packet requests in the UMTS network supported by the ODMA mechanism in case of low load traffic scenarios. However, the maximum transfer rate is bounded by the rate of the link between a MT and its corresponding BS. Furthermore, since on average more radio resources are consumed to enable transmissions via the IR technique, it has adverse effects on QoS for other packet transmissions (i.e. higher packet blocking probability) in heavy loaded networks. In our work, instead of involving external resources, e.g., borrowing frequency, we utilize the available frequency spectrum within cellular networks which is typically underused in the uplink with the purpose to improve the system performance, e.g., the number of simultaneously served users, individual and aggregate throughput and QoS. Unlike to try to balance traffic by rerouting connections from overloaded to underloaded cells, as it was proposed in iCAR, we leverage the specific property of UMTS FDD (Frequency Division Duplex) mode, namely, its paired spectrum allocation principle. UMTS FDD uplink and downlink is designed to operate in two specified symmetri- 80 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

cal frequency bands. Using our knowledge about highly asymmetric traffic distribution of data services between uplink and downlink we exploit the normally underused uplink frequency band for a group-based cooperative mobile-to-mobile data dissemination on the currently free uplink channels. To the best of our knowledge, the closest work to ours is [26], where the traffic load imbalance in the UMTS uplink/downlink has been exploited as well. The idea of this study is to put an UTRA-TDD (Time Division Duplex) link into the underused UTRA- FDD (Frequency Division Duplex) band. On the border of an UTRA-FDD macro cell an ad-hoc pico cell with MTs, operating in TDD mode, is organized. One of these MTs is designated to act as a gateway to the BS and, thus, to the backbone network. It was shown that such an architecture has a potential to improve the system throughput, as well as to increase service probability on the cell border. However, a shortcoming of the proposed solution [26] is the strong dependency between the two system modes and corresponding requirements of efficient coordination of their functionality. For example, the amount of FDD uplink resources to be used for the TDD mode has to be determined. Moreover, special consideration needs to be given to the functionality of the gateway MT. To avoid collisions between several gateway MTs and to arbitrate their interference to the MTs in the pico cells, time slot synchronization is required and the maximum number of gateway MTs has to be controlled. The above mentioned drawbacks can be circumvented by our new solution, where a unified radio interface is used for both conventional cellular and peer-to-peer system modes. Furthermore, the above mentioned studies aim at fitting the QoS to service categories where each user is individual with personal and somewhat unique needs. This makes the users’ incentives to cooperate not necessarily apparent. Our hybrid network architecture is developed for efficient distribution of popular non-real time data content, which is interesting for many users. Thus, users are strongly motivated to act in a cooperative manner in order to reconstruct content of common interest effectively. Our algorithm yields significant relative cell throughput gain in number of MTs, that can be simultaneously served, as well as individual user throughput gains by exploiting the underloaded uplink frequency band, without requiring additional equipment or resources from other (external) networks.

6.3. Mobile-to-Mobile (M2M) Concept

To overcome the capacity limitations of UTRA-FDD2 and to release overall downlink capacity of the system we developed a hybrid technique of efﬁcient distribution of popular non-real-time data contents in order to optimize the data availability to users in high user density areas (hotspots). It is an appealing scenario for e.g., airport lounges, rail-

2FDD mode of WCDMA 6.3. Mobile-to-Mobile (M2M) Concept 81

way stations, shopping malls, where users increasingly demand ubiquitous data availability. We leverage a specific property of the UMTS FDD mode, namely its paired spectrum allocation principle. UMTS in FDD mode operates in two distinct symmetrical frequency bands per cell and operator. The proposed concept is based on cooperative exchange of data among users in (direct) mobile-to-mobile (m2m) mode on currently unused UMTS uplink channels and is realized by integrating a peer-to-peer technique into the existing cellular structure of UMTS networks. Although, due to the distributed time of users’ service requests, conventional solutions for capacity utilization such as multicast or broadcast are not applicable, one can handle a background user as a member of a distinct group with some common characteristics. Our concept properly exploits the above mentioned setup for background service types. Namely, the users that are interested in downloading a popular file, for example a movie trailer or the latest computer games, form a mobile cooperative community (groups). By using the fact that traffic load of multimedia services is asymmetrically distributed between uplink and downlink they contribute their own normally only partly used uplink capacity for providing the packets of the content they have to other users in the group (in their coverage range) in multicast mode on the typically underused uplink carrier frequencies. Instead of serving all background users which are interested in downloading a popular content simultaneously by transmission of the complete data file to each MT in- dividually, the original popular file, which is available somewhere in the network, is divided into m logical packets, each with an individual ID. Each logical packet is then distributed by the BS only once to one of the above mentioned users, in general each packet to a different user, by organizing dedicated channels for a short period of time. After that, none of the users has the complete desired file, but only a small random set of single packets. Nevertheless, in summary all packets, i.e., one complete copy of the file, are now present within a radio cell. A user who received a packet from the BS can act as a server for that particular packet on his own, currently not used uplink channel. Hence, the users start to cooperate with each other for exchanging packets, being organized into dynamically reshaped multiple groups with the purpose to reconstruct the original file of interest. The mobile terminals are assumed to be able to receive in both uplink and downlink bands and each new user brings further uplink resources into the system. Now the reasonable question can arise, what are the incentive s for users to cooperate? Why not use conventional solutions for capacity utilization such as multicast or broadcast? Although background services such as digital camera images, mp3 file or movie trailers downloads are interesting for many users, the mentioned above conventional methods are not applicable since users request service at different times. The mobile users 82 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

prefer to access content on-demand rather than following a fixed schedule. As a result, after distributing the file by organizing, e.g., a broadcast channel, there will be quite many users in the network without any information about the popular file (due to the distributed time of users’ service requests), and users with a complete copy of data. The latter will leave the system immediately after finishing their download due to the lack of incentives to relay their data on behalf of the less "lucky" mobile terminals, which came into the network later during the last phase of the file broadcasting. In our concept, the fact that none of the users receives the complete file directly from the BS via e.g., a broadcast channel, mitigates the lack of user’s willingness to cooperate, thus providing strong motivation for file-sharing-participants to become a server for distribution of the packets of the particular popular content. It is a kind of so-called enlightened self interest. A user decides to send the packet to another user in the group for his own potential advantage. Investing some battery life, he may not have an immediate gain, but looks ahead to a point in time he will have packets to receive from other users in the group. By adopting this cooperative data transfer among users, the BS does not need to provide conventional unicast links of long duration to each user, saving valuable downlink resources, consequently. Thus, a major part of the traffic is shifted away from the downlink, making the released downlink capacity available for other (e.g. real-time demand) services. To enable the cooperative peer-to-peer technique in wireless networks we revisit the idea of a mesh architecture for the fixed-line Internet [13] and appropriately extend it for the UTRA-FDD. To illustrate the main concept of the m2m technique, consider the example in Figure 6.3. MTs voluntarily participate in file sharing via direct mobile- to-mobile data transfer with the purpose to reconstruct the original popular content, which is distributed in the network.

Figure 6.3.: M2M Concept.

Upon arrival each new user, which is interested in download of the popular content, establishes contact to the Node B/RNC (Radio Network Controller) in order to get an authorization to participate in m2m ﬁle transfer and to get information about nearby located m2m users. All authorized m2m users must allow using of their uplink capacity 6.4. Concept Analysis and Evaluation 83

for providing the packets of the content, he has requested, to other m2m users, which are interested in it. The contents liable to cost can be also distributed with m2m strategy, might be, however, protected, for example, via DRM (Digital Rights Management). In order to reduce the transmission of identical packets on the network links the m2m transfer is performed in multicast mode on the normally only partly used uplink carrier frequency, whereas receivers in the group switch to listen on the uplink. The proposed file sharing technique is hybrid: the search of the participants in m2m file transfer, who are looking for the same content is centralized (BS) and the transfer is distributed (m2m). Thus, by combining the intelligence of the cellular network (fixed infrastructure and centralized control of BS) and the flexibility and distributed nature of peer-to-peer networks, we utilize normally underused uplink capacity in UTRA-FDD and in turn, improve the system capacity in terms of cell throughput (in both number of MTs that can be simultaneously handled and in amount of data), individual users’ data throughput, QoS, etc.

6.4. Concept Analysis and Evaluation

In this work we evaluate the performance of a novel m2m algorithm for a UTRA-FDD scenario, taking into account the speciﬁcs of UTRAN, e.g., wireless interference, a realistic propagation model, etc. However, the principle can be applied to other systems as well.

6.4.1. Model Characteristics and Assumptions

In contrast to Internet peer-to-peer applications in a cellular system the number of users who can cooperate with each other is limited by the transmit power of the MT and its coverage range, which will be typically less than a cell. Therefore, in case of wireless cooperative community formation, mobile communities are location and radio propagation dependent. Thus, to cooperate with each other, m2m users must be organized into groups with nearby located users. The decision to which group a user should be assigned is based strictly on current propagation conditions, described in the Section Group Organization Policy.

6.4.2. Radio Interface Restrictions

Mobile networks differ from the wireline systems mainly by the high costs for air transmission and by the mobility of the users. In the following some restrictions and potential solutions of implementing a peer-to- peer technique in UMTS Radio Access Network (UTRAN) are listed.

• One of the main restrictions (problem) of UTRAN is wireless interference. In order to avoid interference from an MT transmitting in m2m mode on other signals 84 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efﬁcient Data Transfer in WCDMA

at the Node B receiver, the transmit power is set to the minimum, which is -44 dBm according to 3GPP speciﬁcations [2].

• Compared to fixed-line networks the air interface of wireless systems has a relatively limited transmission capacity. In conventional wired peer-to-peer applications distributed peers generate a considerable amount of signalling traffic for coordination between them. Search requests are flooded to (potentially) all peer- users, which become critical point in case of wireless, because search means flood- ing scarce wireless resources.

• Furthermore, the effect of user blocking, as a consequence of congestion control, is expected to occur more often than in fixed-line environments in both directions. Thus, in order to increase the efficiency of uplink bandwidth usage it is necessary to maximally reduce the signalling information. This can be done by taking advantage of already existing infrastructure of cellular communication systems. For example, network operators know the location, online status and service agreement of the mobile users. They can utilize this knowledge by providing each new user with information about all active users in the system in the range of tens of meters, to determine the potential cooperative community. Such an explicit use of the hierarchical cellular infrastructure leads to more efficient and reliable routing and significantly lowers the protocol complexity.

• The limitation of battery capacity of the handsets or laptops results in a lower online time compared to the ﬁxed-line desktop PCs. Thus, a well designed ﬁle sharing organization between mobiles is essential.

• There is a need for appropriate propagation models for wireless peer-to-peer links.

6.4.3. M2M Propagation Model

Successful data transmission in wireless systems depends on radio propagation conditions. Thus, an appropriate radio propagation model is very important to predict the link quality within the network. The characteristic of the m2m channels is signiﬁcantly different from conventional Node B - MT links because of the low antenna heights of both receiver/transmitter and a relatively short propagation distance. In this work, a modiﬁed radio propagation model tailored for low antenna heights, for both, transmitter and receiver was used.

6.4.3.1. Pathloss and Shadowing

Pathloss and Shadowing: In our work the pathloss between each MT-MT pair is calculated using advanced propagation model COST 231-Walﬁsh-Ikegami-model (COST- WI)[14]. This model is applicable for systems with low base station antenna heights. 6.4. Concept Analysis and Evaluation 85

COST-WI can be used with reasonable accuracy when the following restrictions are met:

• Carrier frequency lies between 800 MHz and 2000 MHz

• Antenna height of the base station (BS) is in the range 3 . . . 50 m

• Antenna height of the mobile terminal (MT) is in the range 1 . . . 3 m

• Distance between the BS and MT is restricted to 0.02 . . . 5 km

For a detailed description of the COST-WI see Appendix D. The shadowing process is characterized by a lognormal distribution. In the simulations we used advanced distance dependent shadowing from [80]. In this paper the distance dependent standard deviation is modelled using the following function:

−(d − d ) σ (d)= S · 1 − exp 0 for d ≥ d (6.1) Shadow D 0 S where d is a distance between a MT-MT pair, S is the maximum standard deviation,

DS is the growth distance factor (in meters), d0 = 10 m. Appropriate values for the estimated parameters have been chosen according to [80]. Figure 6.4 shows a scatter plot for the received power depending on transmitter/ receiver separation, also taking shadow fading into account (a minimum transmit power of -44 dBm was assumed). As one can see from the plot, the values of the received powers lie in a range acceptable for UMTS systems. Furthermore, the predicted values for distance attenuation have been compared to the results obtained from the measurement-based model for low antenna heights at both sides [80]. The direct comparison shows a good suitability of the COST-WI model for the present application.

Received power samples for pairs of m2m−sender/receiver −40

−60

−80

−100

−120 Received signal power (dBm) −140

−160 0 20 40 60 80 100 120 140 160 Distance between mobile pairs (m)

Figure 6.4.: Scatter plot of received signal power for m2m links as a function of pathloss and shadowing, constant transmit power of −44 dB. 86 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efﬁcient Data Transfer in WCDMA

6.4.3.2. Fading

All mobile terminals in the simulations presented herein are modeled as pedestrians according to the International Telecommunication Union (ITU) Pedestrian A channel type (power delay proﬁle). Hence, the position of each mobile can be assumed as nearly constant during one radio frame (10 ms). Additionally, in the proposed work it is assumed that each continuous m2m connection is established for the same short duration (one frame, see Subsection 7.6 for more details). Due to the slow mobile speed, also the Doppler frequency is low, which in turn means, the channel coherence time is larger than the duration of one radio frame. Therefore, constant multipath fading within a radio frame (block-fading) can be assumed.

Change in received signal power for one transmitter/receiver pair depending on pathloss with shadowing and multipath fading during 60 slots (1frame=15 slots) −80 pathloss/shadowing pathloss/shadowing incl. fading

−85

−90

Received signal power (dBm) −95

−100 0 15 30 45 60 Time (slots)

Figure 6.5.: Changing radio propagation conditions for an arbitrarily chosen transmitter/receiver pair caused by multipath fading, pathloss and shadowing during one radio frame (resolution is 1 slot).

See also Figure 6.5 from the simulation which shows the changing radio propagation conditions for an arbitrarily chosen transmitter/receiver pair in the course of several frames. From the plot one can see the progression of the received power caused by multipath fading, pathloss and shadowing during one radio frame (resolution is 1 slot), which shows that the wireless channel remains nearly constant.

6.5. M2M-Group Organization Policy

In the following general assumptions for the organization of dynamic groups in m2m ﬁle sharing are presented.

• Each new user looks for a group to join when trying to receive popular content.

• MTs form groups, which satisfy the following conditions:

10lg PTX −Lij − Λij ≥−112dBm |

PTX = PTX ,min, ∀i, j ∈ group, 6.5. M2M-Group Organization Policy 87

where Lij is the pathloss between MTs i and j and Λij is the random variable describing the shadowing process. The receiver sensitivity is set to the maximum, which is -112 dBm according to 3GPP speciﬁcations [2].

• The simplest way to inform a new MT about all other MTs requesting the same content in its coverage range is to transmit "Hello" packets by all MTs periodically. But this procedure puts considerable load on signalling channels. Thus, taking into account that a wireless system is limited by the available frequency spectrum, it would be more efﬁcient if upon arrival each new m2m user contacts a BS that provides information about all other m2m users already in the system in the range of tens of meters, to determine the potential members of the group. This can be done by using e.g. GPS (Global Positioning System) or widely used triangulation techniques (see e.g. [75].

• Information about the link quality can be derived from e.g., "Hello" packets, periodically transmitted by MTs.

• Only MTs assigned to a multiple member group send a "Hello" packet to get an appropriate information about the pathloss to any other MTs from its group. In such a way the signalling information between stand-alone MTs can be reduced.

• In case no appropriate group for a new user is found it forms a group with a stand-alone user only.

• Users dynamically join and leave the group at any time (battery life, handover), representing a relatively loosely coupled formation.

• Users move according to a Gaussian random walk, with given average speed and average acceleration.

• Groups are periodically updated and reshaped in order to check the positions of MTs and their radio propagation characteristics on the one hand, and to track and authorize new m2m users in the group in case they fulﬁll the above mentioned "join-group" criteria on the other hand.

• The size of the groups is limited. Why does it make sense to limit a group’s size? All members within a group have to be in one another’s radio range, due to the limited transmission range in wireless environments, it is desirable to constrain the maximum group size.

• Each MT can be a member of only one group at a time.

• The group members are not restricted to belong to the same cell. However, the BSs, to which these MTs are assigned must belong to the same Radio Network 88 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efﬁcient Data Transfer in WCDMA

Controller (RNC). Otherwise, the grouping is restricted to members of the same radio cell.3

Figure 6.6 visualizes the basic order of message sequences during the very ﬁrst stages of the cooperative m2m data exchange.

Figure 6.6.: Signalling message sequence for setup of communication between m2m users.

3A strictly RNC-based group organization policy is chosen in order to avoid possible collisions of signalling information in practical implementation of the ﬁle sharing protocols. 6.5. M2M-Group Organization Policy 89

The following pseudo-code shows the algorithm for dynamic group organization. Algorithm 1: CREATE_GROUPS

1 clear_old_groups, set_max_member; 2 set_minTXpower, set_receiver_sensitivity; 3 calc_distance(MT(m2m)), calc_pathgain(MT(m2m)); 4 calc_signal_strength(MT(m2m)); 5 forall MT(m2m) ∈ system do /* check only active m2m users */ 6 for i = 1 to sum(user_m2m) do 7 if active(MT(i)) == false then 8 continue; 9 end /* current candidate not yet grouped */ 10 if not_grouped(MT(i)) == true then /* create NEW group */ 11 new_group(MT(i)); /* check remaining m2m users */ 12 for k = i + 1 to sum(user_m2m) do 13 if active(MT(k)) == false then 14 continue; 15 end /* candidate not member of other groups */ 16 if not_grouped(MT(k)) then /* max. group size not yet reached */ 17 if sum(new_group) < max_member then /* link quality to all group members "ok" */ 18 if link_to_all(MT(k),new_group) == true then /* → candidate joins group */ 19 add_to_group(MS(k),new_group); 20 end 21 end 22 end 23 end 24 end 25 end 26 end

As already mentioned, in the simulation mobile terminals move like pedestrians. Due to their low velocity the propagation conditions of each mobile can be assumed as nearly constant during one radio frame (10ms). So it is sufﬁcient to trigger further the group update once per second. See also Figure 6.7 from the simulation, which depicts the received power for an arbitrary transmitter/receiver pair, during the interval be- 90 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efﬁcient Data Transfer in WCDMA

tween two consecutive updates of the link quality prediction to reshape groups, which is performed every frame. As all one can see from the plot, changes within one second (100 frames) are just minor.

Received signal power for an arbitrary MT−MT pair (distance ~6 m), group update interval 1 frame −70

−75

−80 Received signal power (dBm)

−85 50 100 150 200 250 300 Time (frames)

Figure 6.7.: Received power for an arbitrary transmitter/receiver pair, during the interval between two consecutive updates of the link quality prediction to reshape groups (update interval 1 frame).

6.6. Cooperative Data Transfer Policy

Using the m2m service, each authorized m2m user must allow the use of his currently not occupied uplink capacity by providing the packets of the content he requested to other m2m users which are interested in it (service level agreement). Data Exchange Policy We assume that each user knows about the packets he has downloaded and the IDs of the packets that are available at its neighbors (the information can be sent in "Hello" packets). 4

• Initially, the original ﬁle is available in the core network only.

• In order to avoid packet collisions, to reduce the transmission of identical packets on the network links and in turn to increase the efﬁciency of usage of scarce and asymmetrically loaded uplink/downlink, the m2m data transfer is performed in the multicast mode among users within a group on the uplink carrier frequency, whereas receivers in the group switch to listen on the uplink. Such a parallel packet downloading policy improves the performance of the system in terms of number of simultaneously served m2m users.

• Identiﬁcation of the sender is done using a unique scrambling code. The MTs must be able to receive in both uplink and downlink bands.

4However, this assumption is not strictly required and is relaxed in further investigation presented in the Subsection 7.9. 6.6. Cooperative Data Transfer Policy 91

• BS supports intra-group data transfer with signalling information, such as "change your mode", "listen for this scrambling code", "listen on the uplink frequency". In case members of some group are assigned to different BSs, they will be coordi- nated (supported with necessary information) via the RNC that is responsible for the set of BSs.

• Due to the RNC-restricted group organization policy the probability of some collisions caused by using the same uplink channel in two or more neighboring groups is very low. Moreover, in the unlikely event that an uplink channel collision occurs, the probability that these groups are in close vicinity is negligible.

• To make m2m transmission as effective as possible, the data exchange algorithm, performed at the BS, ﬁnds an appropriate "sender" candidate, based on a local "most-utile-packet" scheme, in order to maximize the number of users for which the packet can be useful.

• No physical data channel on the BS is needed to control the data exchange process in the group.

• After the best candidate is found, the admission control procedure veriﬁes whether the system has enough uplink capacity to accept the connection. In case more than one relevant "sender" candidate within a group was declared by the data exchange algorithm the sender will be chosen at random. If a sender has more than one packet to send, the packet with the lowest ID will be distributed ﬁrst.

• In order to avoid collisions on the uplink channels of each group, the BS allows only one MT within a group to transmit in a given time interval. As a result of BS signalling, each MT knows whether it has to transmit or to listen. This procedure is performed framewise, which means, that in each group only one sender can be active within a frame.5

• Additionally, due to the low antenna heights of both, transmitter and receiver, and their limited transmit power the inter-group interference is minimized as well.

• Users who have not found any useful packets within a speciﬁed time interval try to connect to the BS for packet delivery. The following ﬂow chart 7.1 visualizes the data exchange mechanism.

The pseudo-code shown in Algorithm 2 mimics the organization of m2m data exchange in terms of frequency usage. Lines 3 to 6 reﬂect, that whenever Node B is distributing data into the radio cell, the active mobile stations receive it on Universal Mobile Telecommunications System

5In this work we have assumed that the size of a logical packet is equal to one UMTS radio frame. However, for other packet sizes the algorithm can be executed as well. 92 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efﬁcient Data Transfer in WCDMA

Algorithm 2: SEND_RECEIVE_MODES (Switching between sending and receiving on uplink and downlink frequencies for m2m users)

1 mode = set_mode(distribute_packets,setup_data_transfer,sent_request,step); 2 forall MS(m2m) ∈ system do 3 if mode == distribute_packets then 4 for i = 1 to sum(user_m2m) do 5 if linkadmission(MS(i),BS) == true then 6 DOWNLINK_RECEIVE(MS(i),data); 7 end 8 end 9 end 10 if mode == setup_data_transfer then 11 for i = 1 to sum(groups) do 12 for j = 1 to members(group(i)) do 13 if linkadmission(group(i)) == true then 14 if type(MS(i)) == ’sender’ then 15 UPLINK_SEND(data); 16 end 17 if type(MS(i)) == ’receiver’ then 18 UPLINK_RECEIVE(data); 19 end 20 end 21 end 22 end 23 end 24 if mode == sent_request then 25 for i = 1 to sum(user_m2m) do 26 if waiting_counter(MS(i)) ≥ max_wait then 27 if linkadmission(MS(i), BS) == true then 28 DOWNLINK_RECEIVE(MS(i), data); 29 end 30 end 31 end 32 end 33 end 6.6. Cooperative Data Transfer Policy 93

Figure 6.8.: Flow chart of m2m ﬁle sharing.

(UMTS) downlink frequency, comparable to reception of data transmitted from BS to users in conventional unicast UMTS operation mode. Lines 10 to 18 resemble the part of data exchange within the m2m groups, with line 13 assuring that data transmission is performed only in case of successful channel assignment from link admission control algorithm. As already pointed out, the UMTS uplink frequency serves for m2m data transfer, thus in line 14 and 17 the distinction between sender and (multicast) receiver(s) in uplink is made. In lines 24 to 28 the request of a mobile station with expired timeout for packet delivery from the BS is shown. In case of successful link admission, the MS will receive the desired packet from the BS on the downlink frequency. On the one hand the last point of the download policy gives the MTs better chances of finding missing packets and finishing the download faster, on the other hand, however, it puts additional load on the downlink resources. Another cause of performance degradation is already mentioned above wireless interference problem. To overcome these problems: 1) to avoid packet collisions caused by wireless interference from other groups, 2) to the enable a more efficient data exchange by minimizing probability of serving m2m requests via Node B, it is necessary to estimate the optimal group size, which will be discussed in the next section. 94 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

6.7. Performance Evaluation of M2M Data Transfer: Veriﬁcation of Basic Functionality

In this section we present some numerical results to highlight the effectiveness of the proposed m2m technique. We used a time-driven simulator (RANSim) and carried out comprehensive simulations. The simulator is intended to resemble the properties and peculiarities of the UMTS air interface as close as possible. The simulated network consists of regular hexagonal cells. A wrap-around technique is used to avoid border effects.

6.7.1. Simulation Scenarios

Particularly in hotspot environment users increasingly demand ubiquitous data availability. Thus, the main focus of our analysis lies in the optimization of data availability to users in hotspots, e.g. airports, railway stations. We assume that there are mobile specific content types, like mp3, to be distributed with m2m strategy. First, we focus on the download data volume characteristics when only one popular file is dispersed with the m2m technique. We assume that there is mobile specific content type of 500 Kbyte size to be distributed with the m2m strategy6.

6.7.2. Traﬃc Model: Pure M2M File Transfer

We study the performance of the proposed m2m concept in a UMTS network with dynamic user arrival pattern. User arrivals are modelled by a Poisson process and the MTs are randomly distributed over the cell area. The number of users varies between approx. 10, 30 and 50 m2m users per radio cell. This means, that the overall number of users in the whole simulation area is always ﬁxed for each particular scenario, while the sum of mobile stations in each cell might differ slightly from one cell to another due to the random distribution of the users’ initial locations, their mobility during the simulation, and handovers. The most important parameters used in our simulations are summarized by Table 6.1. Obviously, the table includes no arrival rates for new m2m users. This is valid, since all simulations in this section are based on the principle to work off a certain initial population and to track for this population statistical values for performance measures of the m2m algorithm. In the following sections we will allow new m2m users to arrive to the system during the simulation run. For the basic functionality studies and performance comparisons, however, it was sufﬁcient to restrict m2m users to the initial population.

6The sizes of popular mp3 ﬁles used in our simulation have been taken from [31] 6.7. Performance Evaluation of M2M Data Transfer: Veriﬁcation of Basic Functionality 95

Table 6.1.: Main simulation parameters. Trafﬁc and environmental settings Traf. load (max. num. of m2m users/cell) 10 (low), 30 (med.), 50 (high) Maximum m2m group size 7 Antenna type omni-directional Cell radius 50 m User proﬁle Pedestrian Radio interface and algorithmic settings File size 500 kbyte Required gross data rate 60 kbit/s Maximum user data rate with 1/2-rate coding 30 kbit/s Size of logical packets for m2m data 225 bit (coded) Receiver sensitivity −112 dBm Transmission power in m2m mode −44 dBm

Eb/N0 target 3 dB Inner loop power control for m2m sender OFF Simulation step size 1 radio frame (0.01 s) Number of steps 40 000-120 000 Group update period 100 radio frames (1 s)

6.7.3. Performance Measures

We consider four performance measures.

• Released Downlink Capacity: data volume reduction in the downlink.

• Download time gain: download time reduction. We define the download time as the time window, in which the user receives the complete file. The criterion for the download time gain is the 90% quantile of finished downloads in the system.

• Overall System Throughput: Service probability gain (increase in the number of served users (%)).

• Effective data rate of m2m users.

• Relative losses of link quality: amount of corrupted data in %.

6.7.4. Comparison of M2M File with conventional UMTS Data Transmission

In order to validate the advantages of using m2m mode in the UMTS network for distribution of popular non-real time content in terms of its basic functionality, we compared its performance with that of the conventional UMTS (unicast) mode, where a continuous transmission of data is organized by individual links from Node B to each user. We list now the assumptions and parameter settings we employed in our simulations: 96 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efﬁcient Data Transfer in WCDMA

• We simulate a population of users with Poisson distributed service requests, that are interested in downloading a popular ﬁle. This means, that a transient period is included to simulate a "warm up" phase of the m2m mechanism, where not all m2m users are active.

• The MTs are distributed randomly over the cell area. Each MT has a given constant speed during a simulation session. All users in this part of investigation are assumed to be pedestrians.

• The main focus of our analysis lies in the optimization of data availability to users in hotspots; the chosen radius of the cell is 50 m.

• We assume that there is mobile speciﬁc content type of 500 Kbyte size to be distributed with the m2m strategy.

• The size of the groups is restricted to 7 members for the time being.

• The MTs depart from the system immediately after finishing their download. How- ever, this assumption is not strictly required and can be relaxed. Then, users which have already finished their downloads may stay in the system for a while to speed up the distribution of missing packets to still not finished users. This is, however, a kind of so-called enlightened self interest. Therefore, external reward schemes might be needed to motivate the users to stay and help, e.g. upload credits. We will investigate this case later.

• The size of a logical packet is equal to one UMTS radio frame, i.e., 225 encoded and spread bits.7

• An appropriately modiﬁed radio propagation model for low antenna heights for both, transmitter and receiver was used. In order to avoid interference from MTs transmitting in m2m mode on other signals at the BS receiver, the transmit power is set to the minimum, which is -44 dBm according to 3GPP speciﬁcations [2].

• The quality of the wireless channel in each group remains constant within a radio frame, but can very from frame to frame (block-fading channel).

• If the packet is incorrect after detection, we declare a packet loss.

• The BS/RNC responsibilities are 1) to distribute at least one complete copy of the original ﬁle in every radio cell (time interval between BS "packet upload sessions" is 10 radio frames), 2) to support the data exchange process in the m2m groups with signalling information, 3) to serve timeout requests from MTs.

7Depending on the coding scheme and spreading factor this leads to corresponding packet sizes of information symbols. 6.7. Performance Evaluation of M2M Data Transfer: Veriﬁcation of Basic Functionality 97

• The simulation time is 400-1200 s and we collect data framewise.

The most important parameters for conventional UMTS data mode simulation are given in Table 6.2.

Speech Data1 Data2 Data3 Data4 Spreading Factor 128 128 64 16 8 Overall data rate (kbps) 30 30 60 240 480 Effective data rate (kbps) 12.2 24 45 210 453 RLC payload size per transport block (bit) 122 80 150 350 320 L1 code word size (bit) 130 100 200 400 360 Outer-loop BLER target 0.01 0.01 0.1 0.1 0.1 Max. transport blocks/frame uplink 1 3 3 6 13 Offset Pdata/Pcontrol at max.rate (dB) 5.46 5.46 9.54 11.48 13

Eb/N0 target uplink (dB) 3.0 3.0 3.0 2.8 2.8 Max. transport blocks/frame downlink 1 3 3 6 13 Max. DPDCH bits/slot 32 16 30 140 302 Number of DPCCH bits/slot 8 4 10 20 18

Eb/N0 target downlink (dB) 3.0 3.0 2.0 2.0 2.0 Max. DL power per channel type 1 1 3 5 10

Table 6.2.: Channel types and parameters as used in simulations

Figure 6.9 shows the system performance versus the offered traffic load. The left graphs in Figure 6.9 demonstrate the efficiency of the m2m file sharing mechanism for low, medium and high traffic load in terms of the download time reduction. The right graphs depict the m2m performance gain in terms of released downlink resources. Regarding merely the file download time, the data transmission in m2m mode is clearly inferior to conventional UMTS mode for a low traffic scenario. Obviously, with only a few users in the system, the file sharing process yields sub-optimal efficiency, since it may take several unsuccessful attempts until a user finds a group to join and to receive data from. In contrast, a small number of users in conventional unicast UMTS mode can be served with a comparatively high data rate, leading to shortened download durations. However, another important performance measure, namely, the number of reserved downlink channels in m2m mode keeps far below those in unicast UMTS mode, demonstrating substantial gain in terms of releasing overall downlink capacity. As one would expect, the higher the traffic load, the more efficient is the performance of m2m file sharing. Obviously, with increased traffic load m2m users have better opportunities to find the content of interest, which positively influences the file download time, on the one hand and increases downlink throughput capacity on the other hand, making the released BS capacity available to other services. 98 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Comparison of download times (low load) Comparison of reserved downlink channels (low load) 80 UMTS Conventional 1 m2m 70

60 0.8

50 0.6 40

0.4 30

Reserved downlink channels 20

Probability of completed downloads 0.2 10 m2m UMTS Conventional 0 0 0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400 Time (seconds) Time (seconds) (a) (b)

Comparison of download times (medium load) Comparison of reserved downlink channels (medium load) 180 m2m 1 160 UMTS conventional 0.9 gain: 7% 140 0.8

0.7 120

0.6 100

0.5 80 0.4 60 0.3 40 Probability of completed downloads

0.2 Number of reserved downlinkchannels

0.1 UMTS Conventional 20 m2m 0 0 0 100 200 300 400 500 600 0 100 200 300 400 500 600 Time (seconds) Time (seconds) (c) (d)

Comparison of download times (high load) Comparison of reserved downlink channels (high load) 1 180 UMTS Conventional UMTS Conventional 0.9 m2m 160 m2m gain: 24% 0.8 140 0.7 120 0.6 100 0.5 80 0.4 60 0.3 40 Probability of completed downloads 0.2 Number of reserved downlinkchannels

0.1 20

0 0 0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800 Time (seconds) Time (seconds) (e) (f)

Figure 6.9.: Performance comparison of conventional and m2m data transfer mode in terms of file download time and reserved downlink capacity for WCDMA system with low, medium and high traffic load. 6.7. Performance Evaluation of M2M Data Transfer: Verification of Basic Functionality 99

We observe that at least 90 % of all users experience a reduction of the download time for the complete ﬁle of up to 24 %. Furthermore, up to 91 % of the downlink capacity is released in a UMTS network, supported by m2m data transmission mode compared to the conventional UMTS mode. In UMTS mode, in contrast, more active users cause trafﬁc congestion of the air interface which leads to more channel requests being denied by the link admission control mechanism.

Table 6.3.: Overall downlink throughput gain: Data volume in downlink in m2m mode for different trafﬁc scenarios and released downlink capacity gain (for users within one cell). Load low medium high Data volume in DL in conventional mode (MB/cell) 5.73 16.51 34.44 Data volume in DL in m2m mode (MB/cell) 1.77 2.26 2.91 Released Downlink capacity (%) 69.07 86.31 91.55 Download time reduction (%) - 7 24

Table 6.3 shows some numerical values for the overall downlink throughput gain and relative download time reduction for complete file download in m2m network mode compared with conventional UMTS data transmission. The rows “Data volume in DL in conventional mode (MB/cell)” and “Data volume in DL in m2m (MB/cell)” show how many Mbyte of data had to be sent via the downlink channels in order to distribute the data file of 500 kbyte size to the users within one cell. We examine now some parameters and their influence on the system performance.

6.7.5. Impact of Multicast Technique

As the required data rate for each m2m request is always the same it is correct to evaluate the effectiveness of the proposed algorithm with respect to the relative service probability gain by using our knowledge about the number of successfully delivered packets in each multicast group every frame. Figures 6.10-6.12 demonstrate the benefit of using multicast in terms of the relative gain in the number of MTs, that can be supported by m2m algorithm. For different m2m traffic scenarios (low, medium, high) the instantaneous number of transmitting MTs and corresponding number of receiving multicast MTs is shown along the time of the simulation. The gap between uplink channels and multicast receivers is getting bigger with increasing traffic load, resulting in higher numbers of multicast receivers served by one uplink channel. This is obvious, since communication opportunities depend much more on user density. Thus, the probability for user encounters and as a result, for cooperative communication is higher in a network with higher user density. Table 7.11 presents the average number of multicast receivers in a group for the traffic scenarios mentioned 100 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Multicast Efficiency (low load) Multicast Efficiency (medium load) 150 150 uplink channels uplink channels multicast receivers multicast receivers

100 100

50 50 Uplink channels and multicast receivers Uplink channels and multicast receivers

0 0 0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450 Time (seconds) Time (seconds) (a) low m2m trafﬁc load (b) medium m2m trafﬁc load

Figure 6.10.: Instantaneous number of transmitting MTs and corresponding number of receiving multicast MTs, respectively (low and medium trafﬁc load).

above. Comparing our performance results with those of the conventional UMTS mode, one can observe a dramatic increase in service probability by using the proposed m2m technique.

Multicast Efficiency (high load) 250 uplink channels multicast receivers 200 Figure 6.11.: The average number of multicast receivers/group for dif- 150 ferent m2m trafﬁc scenarios

100 Load Users per one Service prob. UL-channel gain (%) m2m (low) 1.91 46.92 50

Uplink channels and multicast receivers m2m (medium) 2.90 65.51 m2m (high) 4.41 77.20

0 0 100 200 300 400 500 600 700 Time (seconds)

Figure 6.12.: Instantaneous number of transmitting MTs and corresponding number of receiving multicast MTs, respectively (high trafﬁc load).

6.7.6. Impact of Group Update Interval

Former performance analyzes were based on the assumption that the m2m-groups are updated every second (100 radio frames). We now examine the impact of an increased group update interval on the performance of the m2m file dissemination algorithm. The left graph of Figure 6.13 shows performance of m2m algorithm versus file download time. We observe a substantial increase in file download durations for an extended group update interval (from 1 s to 20 s). 6.7. Performance Evaluation of M2M Data Transfer: Verification of Basic Functionality 101

Comparison of download times (high load) Multicast Efficiency for different m2m group update intervals (high load) 300 group update interval 1 s uplink channels (1 s) 1 group update interval 10 s multicast receivers (1 s) 0.9 group update interval 20 s 250 uplink channels (10 s) multicast receivers (10 s) 0.8 uplink channels (20 s) multicast receivers (20 s) 0.7 200 0.6 0.5 150 0.4 0.3 100

Probability of completed downloads 0.2 Uplink channels and multicast receivers 0.1 50

0 0 100 200 300 400 500 600 700 800 0 20 40 60 80 100 120 140 160 180 200 Time (seconds) Time (seconds) (a) File download time for different group update (b) Multicast efﬁciency intervals

Figure 6.13.: Impact of extended group update interval on the system performance (high trafﬁc load).

The right plot demonstrates the benefit of using multicast in terms of the relative gain in number of MTs, that can be supported by m2m algorithm for different m2m group update intervals in high loaded UMTS system. From the Figure one can see that increasing the group update interval significantly decreases the span between uplink channels and corresponding number of multicast receivers. This results in an lower number of users served by one occupied uplink channel and can be explained as follows. If the radio propagation conditions are estimated more frequently in order to organize the m2m groups, the information about the link quality between potential group members can be kept up-to-date and consequently more accurate. This in tern means, that the probability, that the multicast signal will be received error-free at the MTs is higher, which yields to increased effectiveness of the proposed multicast within a group on the one side, and to the reduced file download time on the other side.

6.7.7. Eﬀect of the Restrictions in the Group Organization Policy

For the results in previous sections it was assumed that the m2m groups can consist of users from two or more neighboring cells. Nevertheless further investigations on the proposed m2m concept such as the actual implementation of the file sharing protocol or determination of the Node B/RNC responsibilities in supporting the m2m data transfer might yield the necessity for a strictly cell-based group organization policy. Figure 6.14 shows the effect of the restriction in the group organization policy on the performance of the proposed m2m technique in a medium and high loaded UMTS system. From the Figure 6.14 one can see that the restriction of the grouping to members of the same radio cell in high loaded system leads to an increase of file download time in order of 10% compared to the more relaxed group organization policy and puts some additional load to the downlink resources. In turn, in case of medium traffic load the 102 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Comparison of download times for different group organization policies (high load) Comparison of download times for different group organization policies (medium load) not restricted not restricted 1 restricted to same cell 1 restricted to same cell 0.9 0,9 0.8 0,8 0.7 0,7 0.6 0,6 0.5 0,5 0.4 0,4 0,3 0.3

Probability of completed downloads 0.2

Probability of completed downloads 0,2 0,1 0.1 0 0 0 50 100 150 200 250 300 350 400 450 500 0 100 200 300 400 500 600 700 Times (seconds) Time (seconds) (a) medium trafﬁc load (b) high trafﬁc load

Figure 6.14.: Impact of policy change "Group members from same radio cell only" on the ﬁle download time.

Table 6.4.: Impact of the restriction in the group organization policy on downlink throughput (per cell), service probability gain and number of multicast receivers per group.

Groups restr. to one cell Groups not restricted Data volume in DL (MB) 3.47 2.26 Users per one UL-chan. 2.041 2.90 Service prob. gain (%) 50.57 65.51

(a) medium trafﬁc load

Groups restr. to one cell Groups not restricted Data volume in DL (MB/cell) 7.1 2.91 Users per one UL-chan. 2.66 4.41 Service prob. gain (%) 62.35 77.20

(b) high trafﬁc load

performance of the proposed m2m technique is affected only moderately and the loss is not so pronounced (see Table 6.4). Intuitively, with an increased number of m2m users in the network more larger groups could be formed across the cell border and the probability of potential group members being rejected increases, yielding comparatively smaller and less efﬁcient m2m communities. 6.8. Direct Mobile-to-Mobile Data Transfer for mixed trafﬁc scenario with service differentiation 103

6.8. Direct Mobile-to-Mobile Data Transfer for mixed traﬃc scenario with service diﬀerentiation

The results presented in the previous section demonstrate the performance of the m2m algorithm for a simple scenario with m2m file sharing participants only, where inter- group interference could be the sole source of disturbance in communication between m2m-users. In this section we analyze the applicability of the m2m technique in a real world scenario in the presence of speech-traffic (cross-traffic) by addressing efficiency and reliability issues. The present section extends the basic m2m model by introducing a hybrid case of traffic. We consider an UMTS system scenario with coexisting conventional UMTS speech users and m2m file sharing participants and analyze the cross-traffic impact. Since the speech users operate on the same uplink frequency as the m2m users, the signals they generate in the uplink are a potential source of disturbance for the m2m signals, that can lead to some performance degradation of the m2m data transfer. In order to cover as many traffic scenarios as possible and to find out the degree of performance degradation of the m2m algorithm due to increased uplink interference, we experimented with different traffic load scenarios for both m2m and speech users.

6.8.1. Traﬃc Model: Incorporating Speech User Population into the Model

We simulate a network area populated by MTs of two service classes. We consider the mixed traffic scenarios, where the m2m file sharing participants coexist with the UMTS speech users. In this scenario we assume that service requests for speech users are served in conventional mode, where discontinuous transmission is organized by providing individual links from the Node B to each user. The network architecture is assumed to support two alternative modes of serving user requests (m2m network mode and conventional UMTS mode). The transmission in conventional mode complies with 3GPP specifications for UTRA-FDD [2].

6.8.2. Simulation Scenarios

In this thesis we outline only a few of the simulated mixed traffic scenarios to demonstrate the reliability of the proposed m2m concept and its robustness to an additional wireless interference, caused by speech users. The following simulation setups have been analyzed: Scenario 1: The number of speech users per cell is kept constant (approx. 3 speech user per cell), while the load of m2m users is varied. Scenario 2: The amount of speech traffic varies proportionally to the m2m traffic load, with the mean number of speech users being 20% of the mean number of m2m 104 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

users in a cell. For low load, this results in 2, for medium load in 6 and for high load in 10 active speech service sessions per radio cell. Scenario 3: The same characteristics as Scenario 2, however, with increased intensity of 40% for speech users, i.e. 4 (low load), 12 (medium load) and 20 (high load) active speech users per radio cell. The service requests from the speech users are always of a high priority. In the case of an incoming voice call at MT, which is in m2m data transmission/ reception mode, the m2m mode will be stopped and the mobile terminal may switch to the conventional UMTS voice mode. Only a buffering of all already received packets is necessary before leaving the m2m mode temporarily. This will not be a problem since 1) the packets can be distributed in an arbitrary order 2) the procedure of the packet distribution is performed within speciﬁed time interval (frame-wise in this work, which means one packet per frame). After the voice call has been terminated, the m2m algorithm may be resumed in the considered terminal. The parameters of interest are the released downlink transmission capacity, time for download of compete ﬁle, effective m2m-user data rate, as well as the service probability gain. Possible collisions between users of different service classes creating wireless interference have been taken into account.

6.8.3. Impact of Cross-Traﬃc on the M2M Performance: Uplink Interference

In previous studies where we analyzed the efficiency of the m2m file dissemination technique in a pure m2m traffic environment, we assumed that the quality of the wireless channel in each group remains constant within a radio frame, but can vary from frame to frame.

• Now due to the interference from speech users, operating in conventional unicast UMTS mode, the characteristics of the wireless channel in each group can vary from slot to slot.

• Information about the quality of the multicast signal within the group is obtained based on the ratio of the average received power of the useful signal to that of all relevant interfering signals (C/I) on a slot-by-slot basis.

• If the packet is incorrect after detection, we declare a packet loss. It should be understood as follows, up to a third of the frame (1 frame=15 slots) is corrupted, caused by an unacceptable C/I level, transmitted data is still assumed to be re- coverable (due to channel coding).

The following flow chart 6.15 visualizes the mechanism of m2m file sharing with cross-traffic in presence. 6.8. Direct Mobile-to-Mobile Data Transfer for mixed traffic scenario with service differentiation 105

Figure 6.15.: Flow chart of m2m ﬁle dissemination for mixed trafﬁc scenario.

The performance of m2m file sharing is evaluated by comparing the distribution of download times for a complete file download for two uplink interference scenarios. One m2m scenario involves only inter-group interference while the second scenario consists of both m2m inter-group interference and interference from speech traffic which operates on the same uplink frequency as the m2m users. Figures 6.16 and 6.17 show results from simulations with the above mentioned parameter settings for speech traffic load. The medium and high load for m2m traffic scenarios are investigated for each speech traffic setting. Figures 6.16a and 6.20a illustrate the influence of the speech traffic interference on the performance of m2m concept in terms of download time increment. Since for voice applications, the uplink is needed, the uplink capacity available for the m2m data dissemination is time-varying depending on intensity of the cross-traffic. Intuitively, the higher the speech traffic load the higher the interference in uplink. From the graphics one can see that the download time increases up to about 25%. This results in decreased service probability for m2m users on one hand and in high retransmission probability (because of amount of corrupted blocks) on the other hand. Figures 6.16b-(d) demonstrate the impact of uplink interference on efficiency of m2m data transfer for different cross-traffic scenarios. Relative losses of link quality (number of corrupted data in %) for various speech-traffic intensities are also shown in Table 6.5. 106 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Table 6.5.: Impact of uplink interference for different cross/m2m trafﬁc scenarios (erroneous data %). Scenario Erroneous data (%) Pure m2m trafﬁc Scenario 1 Scenario 2 Scenario 3 Low (m2m) load 2.07 7.14 4.21 9.29 Medium (m2m) load 5.20 8.76 13.71 18.29 High (m2m) load 8.89 13.23 16.22 22.46

Interestingly, from Figure 6.20b one can see that with growing cross-traffic intensity the number of reserved downlink channels for m2m mode is reducing. This phenomenon can be explained as follows. The increased number of service requests by speech users leads to rejection of some uplink connections for data transfer in m2m mode. This results in an increased packet reception interval at the MTs. Although the number of packet requests from m2m users directly to the BS occur more frequently, additional downlink channels cannot be allocated either due to exhausted downlink capacity caused by speech users, resulting in significant download time increment. In Figure 6.18 the multicast efficiency of m2m data transfer for different cross traffic scenarios has been compared. The effect remains. Due to the limitation of the link admission for m2m-senders (bounded by the uplink capacity) we observe decreased service probability of the proposed m2m algorithm. Nevertheless, the direct performance comparison of m2m file distribution with those of conventional data transfer for a hybrid traffic scenario has shown that the performance gap between two alternative network modes of serving user requests becomes bigger to the disadvantage of unicast mode (see 7.8.4).

6.8.4. Comparison of M2M File Dissemination and Conventional UMTS Data Transmission for Mixed Traﬃc Scenarios

The direct performance comparison of m2m file distribution with those of conventional data transmission demonstrate a substantial overall downlink throughput gain, obtained by using the m2m technique. This can be seen from Figure 6.19, where the system performance versus download time is demonstrated for both system modes for different cross traffic scenarios. With increased amount of cross-traffic, which is according to our simulation scenarios always the high prioritized speech service, a significant performance degradation of data transfer in terms of download time can be observed in conventional UMTS mode. Fur- thermore, the number of MTs that can be supported by using m2m mode in UMTS system is three time higher than in case of conventional UMTS network. In fact, the more speech users desire a service, the more downlink channels they occupy. Since the number of users that can be admitted in the system is finite, the network will first serve high prioritized speech calls at the expense of service denial 6.8. Direct Mobile-to-Mobile Data Transfer for mixed traffic scenario with service differentiation 107

Comparison of download times for different cross traffic scenarios Impact of uplink interference on m2m data transmission (high m2m load) (high m2m load, scenario 1) 1 250 cross traffic, scenario 1 erroneous 0.9 cross traffic, scenario 2 successful cross traffic, scenario 3 0.8 total without cross traffic 200 0.7

0.6 150 0.5

0.4 100

0.3

0.2 m2m data (packets per second) 50 Probability of completed downloads 0.1

0 0 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 Time (seconds) Time (seconds) (a) File download time for different cross trafﬁc (b) Impact of an additional uplink interference scenarios and high m2m load. caused by speech trafﬁc of low intensity.

Impact of uplink interference on m2m data transmission Impact of uplink interference on m2m data transmission (high m2m load, scenario 2) (high m2m load, scenario 3) 250 250 erroneous erroneous successful successful total total 200 200

150 150

100 100

m2m data (packets per second) 50 m2m data (packets per second) 50

0 0 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 Time (seconds) Time (seconds) (c) Impact of an additional uplink interference (d) Impact of an additional uplink interference caused by speech trafﬁc of moderate intensity. caused by speech trafﬁc of high intensity.

Figure 6.16.: Impact of an additional uplink interference caused by speech traffic of various intensities on the file download time (high m2m traffic load).

Comparison of download times for different traffic scenarios Comparison of reserved downlink channels for different cross traffic scenarios (medium m2m load) (medium m2m load) 1 25 without cross traffic without cross traffic 0.9 cross traffic, scenario 1 cross traffic, scenario 2 cross traffic, scenario 2 cross traffic, scenario 3 0.8 20 cross traffic, scenario 3 0.7

0.6 15

0.5

0.4 10

0.3

0.2 5 Probability of completed downloads Number of reserved downlink channels 0.1

0 0 0 100 200 300 400 500 600 0 100 200 300 400 500 600 Time (seconds) Time (seconds) (a) File download time for different cross trafﬁc (b) Impact of cross trafﬁc on the number of re- scenarios and medium m2m load. served downlink channels.

Figure 6.17.: Impact of cross traffic on the M2M performance (medium m2m traffic load). 108 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Comparison of multicast efficiency for different cross traffic scenarios (high m2m load) 300 m2m uplink senders/pure m2m scenario m2m multicast receivers/pure m2m scenario 250 m2m uplink senders/cross traffic, scenario 3 m2m multicast receivers/cross traffic, scenarios 3

200 ∆=78 multicast receivers 150

100

50 Uplink channels and multicast receivers

∆=15 uplink senders 0 0 50 100 150 200 250 Time (seconds)

Figure 6.18.: Impact of cross trafﬁc on multicast efﬁciency (fragment).

Comparison of download times for different cross traffic scenarios (high m2m load, file size 500 kb) 1 gain: 32% 0.9 gain: 51% 0.8

0.7

0.6

0.5

0.4

0.3

m2m, scenario 1

Probability of completed downloads 0.2 m2m, scenario 2 0.1 UMTS conventional, scenario 1 UMTS conventional, scenario 2 0 0 200 400 600 800 1000 1200 1400 Time (seconds)

Figure 6.19.: Performance comparison of conventional and m2m data transfer mode in terms of download time, high m2m load.

for some data users. All this results in degradation of the service probability for non- real time data traffic and increases the file download time for admitted data users substantially. Some numerical values for the overall downlink throughput gain with respect to complete file download in m2m network mode while introducing a hybrid traffic case with service differentiation are compared in Table 6.6 with values for conventional UMTS data transmission. The direct comparison of the m2m performance results with those for the conventional UMTS data transmission demonstrates an overall gain in released downlink capacity of up to 83%, obtained by using the m2m technique. The table also shows how many Mbyte of data had to be sent via the downlink channels in order to distribute the data file of 500 Kbyte size to the users within one cell. The last four rows in the table show the relative gain in the number of MTs, that want to download the popular content and can be supported by the m2m concept in the presence of speech-traffic for various different traffic scenarios. 6.9. Dependability of Mobile-to-Mobile Data Transfer 109

Table 6.6.: Overall downlink throughput gain: Data volume in downlink in m2m mode for different traffic scenarios and service probability gain (for users within one cell) Load Low Medium High Data volume in DL in conventional mode (MB/cell) 5.7 16.5 34.4 Data volume in m2m mode w/o cross-traffic (MB/cell) 1.7 2.2 2.9 Released Downlink capacity (%), (w/o cross-traffic) 69.1 86.3 91.5 Scenario 1: Cross-traffic 3 users/cell (MB/cell) 2.0 3.3 5.9 Released Downlink capacity (%), Scenario 1 64.5 80.0 82.8 Scenario 2: Cross-traffic 20 % of the mean number of m2m user/cell (MB/cell) 2.0 3.4 6.3 Released Downlink capacity (%), Scenario 2 65.4 79.2 81.6 Scenario 3: Cross-traffic 40 % of the mean number of m2m user/cell (MB/cell) 2.1 3.7 7.2 Released Downlink capacity (%), Scenario 3 63.5 77.6 79.0 Service probability gain (%), (w/o cross-traffic) 46.9 65.5 77.2 Service probability gain (%), Scenario 1 30.1 52.2 66.5 Service probability gain (%), Scenario 2 33.0 47.2 64.2 Service probability gain (%), Scenario 3 29.2 44.6 61.2

6.9. Dependability of Mobile-to-Mobile Data Transfer

6.9.1. Practically relevant scheduling policy

In our basic m2m concept an ideal packet scheduling was employed, based on local "most-utile packet" distribution scheme, assuming perfect knowledge of the current state of each file download performed in the neighborhood (info is periodically sent by m2m users in "Hello" packets). However, such an assumption is quite unrealistic, since finding an optimal realizable packet scheduling algorithm is a quite complex problem in a large scale distributed network. In such an environment the mobile terminals (MTs) have only local information about the transfers performed within their local groups and very incomplete knowledge about the global network state. Moreover, the amount of signalling information necessary to coordinate the data exchange process within a group cannot be neglected when this packet distribution scheme is used. We investigate now the performance of m2m data transfer under more realistic assumptions. For this we assume that neither BS, nor MTs have an information about the packets available in their neighborhood. The data exchange algorithm determines the sender at random. The packet to be distributed is decided at random too, among the packets available at the currently active MT-sender. We called our new and more realistic version of m2m algorithm random-m2m file dissemination technique. 110 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Comparison of reserved downlink channels Number of transmitting MTs and corresponding number of multicast receivers over time (high load, file size 500 kb) (high load, file size 500 kb) 180 350 UMTS Conventional m2m uplink channels/random file dissemination 160 m2m/most−utile packet transfer policy m2m multicast receivers/random file dissemination 300 m2m/random packet transfer policy m2m multicast receivers/most−utile file dissemination 140 m2m uplink channels/most−utile file dissemination 250 120

100 200

80 150 60 100 40 Number of reserved downlink channels

Uplink channels and multicast receivers 50 20

0 0 0 100 200 300 400 500 600 700 800 900 1000 0 50 100 150 200 250 Time (seconds) Time (seconds)

Figure 6.20.: Performance comparison of the different ﬁle transfer strategies in terms of reserved downlink channels and multicast efﬁciency (UMTS conventional, most-utile m2m and random m2m algorithms, group size 7).

While comparing the multicast efficiency of the "most utile packet" distribution scheme with modified random packet distribution policy (see Figure 6.20, right plot), one can see some decrease in multicast efficiency for the "random" case. This is evident, since by using the random packet distribution, the probability in each time interval for m2m receivers to get an innovative packet is lower, than by applying the "most-utile packet" distribution scheme. However, as one can see from Figure 6.20 (left plot) the degradation of the released downlink capacity is negligible. Moreover, a direct comparison to the conventional system mode in terms of resource consumption still shows a substantial gain in released downlink capacity by applying the random m2m algorithm.

6.9.2. Impact of Group Size on Inter-Group Interference

We examine now some parameters and their influence on the system performance. We consider again the same m2m scenario with maximum group size 3, 7, 10 and 13 and investigate the effect of the group sizes on the performance of the proposed technique for low, medium and high traffic load. In Figure 6.21a the influence of the group sizes on the file download time for a high loaded UMTS system is illustrated. With increased number of multiple coexisting groups (group size 3) in the network some performance degradation of the m2m technique is observed. This effect is influ- enced by admission control, wireless interference and user mobility. Obviously, with increasing the number of groups the number of sender candidates that can be admitted is bounded by the uplink capacity. This results in a rejection of some link admission requests of sender candidates. Besides, if the group size is too small and the mobility of users is low (as assumed in this work), the probability to find a missing packet in each frame is quite low; the number of packet requests from m2m users to the BS increases and puts additional load on the downlink resources. On the other hand, a larger group size brings about a higher multicast efficiency. The number of senders necessary to distribute a file among all MTs is lower, hence, 6.9. Dependability of Mobile-to-Mobile Data Transfer 111

Comparison of download times for different group sizes (high load) Multicast efficiency for different group sizes (high m2m load) 1 200 group size 3 m2m uplink channels/group size 3 0.9 group size 7 180 m2m multicast receivers/group size 3 group size 10 m2m uplink channels/group size 13 0.8 group size 13 160 m2m multicast receivers/group size 13 0.7 140

0.6 120

0.5 100

0.4 80

0.3 60

0.2 40 Probability of completed downloads Uplink channels and multicast receivers 0.1 20

0 0 0 200 400 600 800 1000 1200 1400 0 20 40 60 80 100 120 140 160 180 200 Time (seconds) Time (seconds) (a) Impact of the m2m group size on the ﬁle down- (b) Impact of the m2m group size on multicast ef- load time. ﬁciency (fragment).

Figure 6.21.: System Performance for different group sizes (high m2m trafﬁc load).

Table 6.7.: Impact of uplink interference for different group sizes and trafﬁc scenarios (erroneous data %). Scenario Erroneous data (%) Group size 3 Group size 7 Group size 10 Group size 13 Low (m2m) load 5.16 3.28 2.07 2.07 Medium (m2m) load 13.70 5.56 5.00 4.83 High (m2m) load 25.32 13.77 10.01 9.67

for the a fixed number of members in the system large groups consume less uplink capacity. Equivalently, with the same uplink resource consumption more members will get service when the group size is larger (see Figure 6.21b). At the same time inter-group uplink interference is reduced. Indeed, when groups are large, the total amount of uplink interference is lower. Since all MTs in the close vicinity of the transmitting MT most probably belong to the same group, no inter-group interference occurs. Another important performance criterion is the average rate of the successfully delivered packets. For systems with small group size we observed enormous m2m uplink interference, which leads to significant performance degradation and in turn to an increase of download time. Figure 6.22 demonstrate the impact of the uplink interference on m2m data transmission for group sizes 3, 7 and 13 in a high loaded UMTS system (effective user data rate is 30 kbit/s). The probability of losses of link quality (number of corrupted data in %) for different group sizes and traffic scenarios are shown in Table 6.7. 112 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Impact of uplink interference on m2m data transmission, group size 3 (high load) Impact of uplink interference on m2m data transmission, group size 7 (high load) 90 150 erroneous erroneous 80 successful successful total total 70

60 100

30 50 m2m data (packets per second) 20 m2m data (packets per second)

0 0 0 200 400 600 800 1000 1200 1400 0 100 200 300 400 500 600 700 800 900 1000 Time (seconds) Time (seconds) (a) Performance of the m2m data transmission for (b) Performance of the m2m data transmission for group size 3 group size 7

Impact of uplink interference on m2m data transmission, group size 13 (high load) 180 erroneous 160 successful total 140

120

100

m2m data (packets per second) 40

0 0 100 200 300 400 500 600 700 800 Time (seconds) (c) Performance of the m2m data transmission for group size 13

Figure 6.22.: Impact of group size on inter-group interference for high loaded UMTS system.

6.9.3. Eﬀect of Restricted BS Support

Since the main goal of our m2m concept is to release maximum possible overall downlink capacity, it is preferable to further relieve the BS "duties". To achieve this goal, we introduced a tighter file dissemination policy by restricting the active assistance from the BS. Up to now the responsibility of the BS includes the continuous packet distribution support. Namely, after a complete copy of the file is available in each radio cell, the BSs continue to periodically ("packet upload session" is every n-th time step) distribute packets to active m2m users; one randomly chosen packet per MT. This procedure provides additional support for the m2m algorithm, which assures faster data dissemination in the network and consequently reduced download times for the complete file. The disadvantage is, however, the unnecessary allocation of radio resources in the downlink. Now the algorithm operates with restricted BS support. 6.9. Dependability of Mobile-to-Mobile Data Transfer 113

• With immediate BS deactivation after one copy of the ﬁle is distributed within its radio cell.

• BS supports and serve only timeout requests from MTs.

• Instead of the periodical dissemination of additional packets from the BS, the m2m-data transfer will be performed every n-th time step.

• BS further supports the data exchange process in the m2m groups with signalling information.

We want to investigate the degree of influence of such a tighter file dissemination policy on the performance of m2m algorithm. We quantify dependability of the system under consideration to restricted BS support in terms of possible prolongation of file download time and in terms of the downlink resource consumption. As one can see from Figure 6.23, where the performance results have been presented for the system with pure m2m traffic scenario (file of size 500 kbyte), no performance degradation in terms of download time can be observed. By restricting the BS support a substantial reduction in the number of reserved downlink channels has been achieved, as expected. Clearly, now even more traffic is shifted away from the BS and handled by m2m-enabled MTs, saving transmissions on the valuable downlink frequencies, at the same time increasing the efficiency of the cooperative m2m data dissemination (see Figure 6.23d, where the m2m packet rate progression over time is plotted). The above mentioned findings are suitable for explaining the effect that no additional time is necessary to guarantee the reception of the complete file for all users. Figures 6.24 and 6.25 demonstrate the impact of the restricted BS support on the performance of the medium loaded network for the mixed traffic scenario (the popular content’s size is 40 kbyte). Here, the situation is somewhat different. Although the main goal to further release downlink resources is met, it was achieved at the cost of longer file download times. The observations indicate that the deactivation of the periodical distribution of additional packets by the BS leads to lengthened (up to 8 %) file reception duration. This can be explained as follows. First, the probability for some currently active senders to find all uplink channels busy in a particular time step (due to the occupation of the channels by speech users) is much higher, than in the pure m2m traffic scenario. As a result the time intervals between packet receptions at the MT-receivers are getting larger. Furthermore, without additional support from the BS, the data transfer performed predominantly by MTs is more sensitive to uplink interference, caused by cross traffic. Since the transmission power of the m2m-senders is set to its minimum (to avoid disturbance of other signals at the BS receiver) and the speech users, in turn, can send with higher power to reach their BSs, errors in and losses of m2m data occurs more frequently. 114 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Comparison of download times for different BS support policies (high m2m load) Comparison of reserved downlink channels for different BS support policies (high load) 1 35 m2m/restricted BS support m2m/BS support 0.9 m2m/BS support m2m/restricted BS support 30 0.8

0.7 25

0.6 20 0.5 15 0.4

0.3 10 0.2 Probability of completed downloads

Number of reserved downlink channels 5 0.1

0 0 0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 Time (seconds) Time (seconds) (a) Impact of restricted BS support on the ﬁle (b) Impact of restricted BS support on downlink download time. resource consumption.

Multicast efficiency for different BS support policies (high m2m load) Error−free m2m data transmission (high m2m traffic load) 160 140 m2m uplink channels/BS support m2m/restricted BS support 140 m2m multicast receivers/BS support m2m/BS support m2m uplink channels/restricted BS support 120 m2m multicast receivers/restricted BS support 120 100 100 80 80 60 60 40 40 m2m data (packets per second)

Uplink channels and multicast receivers 20 20

0 0 0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 Time (seconds) Time (seconds) (c) Impact of restricted BS support on multicast ef- (d) Impact of restricted BS support on the m2m ﬁciency. data rate progression.

Figure 6.23.: Impact of restricted BS support on the performance of system with high m2m trafﬁc load, ﬁle size 500 kbyte.

Consequently, as can be seen from Figures 6.24c and 6.25c a decreased rate of successfully delivered packets is noticeable. 6.9. Dependability of Mobile-to-Mobile Data Transfer 115

Comparison of download times (medium m2m load, scenario 1) Comparison of reserved downlink channels (medium m2m load, scenario 1) 1 12 m2m/BS support 0.9 m2m/restricted BS support 10 0.8

0.7 8 0.6

0.5 6

0.4 4 0.3

0.2 Probability of completed downloads 2 Number of reserved downlink channels 0.1 m2m/BS support m2m/restricted BS support 0 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Time (seconds) Time (seconds) (a) Impact of restricted BS support on the ﬁle (b) Impact of restricted BS support on downlink download time. resource consumption.

m2m data rate progression over time (medium m2m load, scenario 1) 10 m2m/BS support 9 m2m/restricted BS support 8

4 Data rate (kbit/s) 3

0 0 10 20 30 40 50 60 70 80 90 100 Time (seconds) (c) Impact of restricted BS support on the m2m data rate progression.

Figure 6.24.: Impact of restricted BS support on the performance of system for mixed traffic scenario with service differentiation (file size 40 kbyte, medium m2m load, cross-traffic scenario 1). 116 Chapter 6. Mobile-to-Mobile: A Novel Concept for Spectrum Efficient Data Transfer in WCDMA

Comparison of download times (medium m2m load, scenario 2) Comparison of reserved downlink channels (medium m2m load, scenario 2) 1 12 m2m/BS support 0.9 m2m/restricted BS support 10 0.8

0.7 8 0.6

0.5 6

0.4 4 0.3

0.2 Probability of completed downloads 2 Number of reserved downlink channels 0.1 m2m/BS support m2m/restricted BS support 0 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Time (seconds) Time(seconds) (a) Impact of restricted BS support on the ﬁle (b) Impact of restricted BS support on downlink download time. resource consumption.

m2m data rate progression over time (medium m2m load, scenario 2) 10 m2m/BS support 9 m2m/restricted BS support 8

4 Data rate (kbit/s) 3

0 0 10 20 30 40 50 60 70 80 Time (seconds) (c) Impact of restricted BS support on the m2m data rate progression.

Figure 6.25.: Impact of restricted BS support on the performance of system for mixed traffic scenario with service differentiation (file size 40 kbyte, medium m2m load, cross-traffic scenario 2). 6.10. Summary 117

6.10. Summary

In this section we have shown how cooperative behavior of wireless users can substantially improve the efficient usage of frequency spectrum in cellular networks. We introduced a new hybrid technique for efficient distribution of popular non-real- time data content in order to optimize the data availability to users in hotspot scenarios. The proposed concept is based on integrating a peer-to-peer technique into the existing cellular structure of the UMTS network in order to realize a direct m2m cooperative data exchange by dynamically allocating users to temporarily unused UMTS uplink channels. Our concept is based on the uplink/downlink traffic imbalance in 3G wireless networks and clearly confirms the social principle "real egoistic behavior is to cooperate". Simulation results demonstrated that the proposed cooperative approach is capable to improve the performance of a cellular wireless system considerably. It has been shown that a cellular network, supported by our cooperative solution significantly outperforms a network in conventional UMTS mode and might be a promising alternative for distribution of content in cellular radio networks like UMTS.

Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

In the previous sections a cooperative concept was proposed, where the users, who are interested in fast downloading of a popular content (new movie trailers or music files) form loosely coupled groups where the members directly cooperate with each other. The primary goal of the above mentioned work was to release the downlink of UTRA-FDD and consequently reduce the likelihood of bottlenecks by increasing the efficiency of usage of the available frequency spectrum. This has been done by shifting the non-real time traffic away from the downlink, utilizing the released capacity for providing better Quality of Service (QoS) for real-time services, and is accomplished by dynamically allocating m2m users to temporally underused uplink channels. For this purpose the popular file is divided into small logical packets, which are distributed packet by packet in arbitrary order among the m2m users in multiple groups on their own uplink channels in a multicast mode. It has been shown that the proposed cooperative m2m strategy significantly outperforms the conventional UTRA-FDD mode for download of popular content. An ideal packet scheduling was employed, assuming perfect knowledge of the current state of each file download performed in the neighborhood.

7.1. Motivation and Related Work

However, ﬁle distribution by exploiting direct mobile-to-mobile data transfer in wireless networks needs some additional technique to improve dependability of information dissemination. Finding an optimal realizable packet scheduling algorithm is a quite complex problem in a large scale distributed network. In such an environment the mobile terminals (MTs) have only local information about the transfers performed within their local groups and very incomplete knowledge about the global network state. To circumvent the above mentioned difﬁculties and to speed up the distribution of the packets, in this section an extended algorithm is presented, where a new scheduling criterion for packet dissemination is applied. 120 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

In particular, we extend the m2m algorithm by incorporating network coding in the packet scheduling process. Network coding in general is an efficient packet distribution strategy [4, 20]. Each user in the network encodes all the packets he has already received by performing algebraic operations on them, treating Bytes of data packets as elements in a certain base field. A user forms a linear combination of these elements and relays the resulting packet, thus, increasing the information content of each transmission, compared to simple replicate-and-forward concept of the data transmission. Due to the random nature of the coefficients of the encoded packets, each encoded packet will most likely contain information that is useful for each user, improving dependability of packet distribution, consequently. This means that the performance of the system depends less on the scheduling mechanism and there is no need for a global scheduler-coordinator. Most of the previous work on network coding has focused on the theoretical aspects of this technique [49, 45, 21]. There is only a small number of publications available which evaluate the benefits of network coding for special applications. One of the most representative efforts to combine the theory of network coding and practical design of large scale networks was introduced in [42], where the case of a wireless mesh network with unicast traffic was analyzed. Practical schemes were proposed in [24], where the gain achieved by using network coding in a fixed, large, unstructured overlay network with bidirectional unicast traffic was assessed. In [84], the mutual exchange of information, corresponding to different content, between two nodes in an ad-hoc network is investigated. The main focus lies on duplex traffic streams, i.e., the nodes send packets to each other in opposite directions by using exactly opposite routing paths. The packets in the intermediate hops are XOR-ed together and broadcast to their next hop. Network coding was performed only for a small subset of the forwarded packets. In [41] it was shown how network performance in terms of throughput can benefit from opportunistic coding (i.e., intelligent mixture of the packets) and listening. However, error-free wireless channels were assumed.

Our contribution differs from previous work in the following aspects: • We consider a hybrid cellular based peer-to-peer network employing network coded data distribution.

• Network Coding is applied in order to ease the problem of packet scheduling in the distributed dynamic environment of a wireless large scale network. We call our concept NC-m2m.

• Instead of looking for the best scheduling with a simple packet replication we use network coding to obviate the need for centralized knowledge of the global state of the network.

• Unlike other content distribution schemes, our approach takes into account the main features of UTRAN, i.e., collisions between users of different service classes creating wireless interference. 7.2. Network Coding Basics 121

• We demonstrate that the NC-aided m2m ﬁle sharing technique signiﬁcantly outperforms our previously proposed m2m algorithm with a simple replicate-and- forward routing scheme in terms of download time.

• We show that the network, supported by the NC-based m2m algorithm for ﬁle distribution, is quite dependable and robust to sudden departure (due to battery life, handover, loss of interest in the content) of ﬁle sharing participants.

• Furthermore, in our extended m2m technique, signiﬁcantly less radio resources as in the earlier simple m2m algorithm need to be spent in both downlink and uplink in order to achieve the same aggregate system throughput.

7.2. Network Coding Basics

In this Section the basic concept of network coding will be explained. Conventionally, independent data streams share network resources, but never the information they transmit/receive. Consider, e.g., some routing process in the network. Conventionally, a router forwards and routes information; each information unit on an output link must be a copy of information, which came earlier on an input link. network coding (NC) is a relatively novel field emerging from information theory that changes these assumptions and promises optimal utilization of the network resources. The principle behind network coding is to allow network entities (nodes) to perform some computation. Generally, network coding is the transmission, mixing and remix- ing (encoding) of information (packets) at nodes, in such a way that the transmitted information can be successfully unmixed (decoded) at its final destination. The benefit of coding over routing is illustrated by the following example, see Fig- ure 7.1. Consider the multicast of two data packets a and b, from the source A in the

communication network to receivers F1 and F2. A communication network is mapped to a ﬁnite directed graph, where multiple edges from one node to another are allowed. A directed edge is called a channel in the communication network. All links in this network have capacity one. If the network nodes can only replicate- and-forward the information (packets), then obviously the link D − E is the bottleneck and must be timeshared between the two sessions. Suppose packet a would be sent ﬁrst

through the D − E; then the receiver F1 receives packet a twice and does not know b at

all. Sending packet b poses a similar problem for the receiver F2. In this case, routing is insufﬁcient because no routing scheme can transmit packets a and b simultaneously to both destinations. Let us now explain the idea of network coding. Again two packets a and b must

be multicast from the source A to both receivers F1 and F2. Now intermediate node D transmits a ⊕ b instead of a and b in sequence. As a result, both receivers can now recover the packets of interest, while the number of transmissions is reduced. The beneﬁt of coding over routing is obvious. 122 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

Figure 7.1.: Flow chart of m2m ﬁle sharing.

Similar to erasure decoding, successful reception of information does not depend on receiving specific packet content but rather on getting a sufficient number of independent combinations of the packets. Network Coding was first considered in [4], where it has been proved that maximum throughput in a network can be always achieved using Network Coding.

Consider again the graph with a sending node A and two receiving nodes F1 and F2.

Let R(A, F), where F = {F1, F2}, be an achievable information rate from A to F1 and F2 (see Figure 7.2). The value of any A− F cut, which partitions the network into two subsets, where the first one contains source A and the second contains any single receiver, taken from F, is defined as the sum of the capacities (max-flow) of the links from nodes in first subset to nodes in a second one. The minimum of the values of all such cuts is an upper bound

for the information rate R from A to receivers F1 and F2

R(A, F) ≤ min MinCut(A, Fk) k = 1, 2, (7.1) Fk∈F and is denoted as MaxFlow(A, F ). Thus, h = min MinCut(A, F ) is the maximum k Fk∈F k capacity for a defined source node and a set of receivers in an arbitrary directed graph, and can be always achieved in any network using network coding, as proved also in [4]. Linear NC is, in general, similar to an XOR operation with the difference that the XOR is replaced by a linear combination of the data, interpreted as numbers over some F F finite field 2s , where s is the word length (in bits) of the numbers contained in , so that there is always a limited number of coefficients for encoding. 7.3. NC applied to M2M Data Transfer 123

Figure 7.2.: Flow chart of m2m ﬁle sharing.

As it has been proved later in [45] linear coding is sufﬁcient to achieve the maximum capacity. This is the simplest way of network coding, so that coding can be implemented at low computational cost.

7.3. NC applied to M2M Data Transfer

Large-scale wireless networks with their complicate routing schemes represent a quite suitable environment for applying the network coding technique. A cellular network can be regarded as a collection of directed wireless links. Thus, a directed graph, consisting of a vertex set (network entities) and an edge set (links between entities) can be mapped onto cellular networks as well. The idea of linear Network Coding will then be applied to encoding for m2m data transfer in the UMTS system. In our basic m2m concept the data exchange policy was based on a local "most-utile packet" distribution scheme [58, 57]. For this, it was assumed that the users make a packet scheduling decision based on the local information about packets available in the neighborhood (info is periodically sent by m2m users in "Hello" packets). However, such an assumption is quite unrealistic in a large distributed wireless network. As the number of users increases, it becomes difficult to find (to ensure) optimal scheduling of distributing packets between users. Moreover, this packet distribution scheme has much redundancy, when considering the global state of the whole network. Since the "most- utile" packet exchange policy is employed in each group independently it could happen that the most beneficial packet in a particular group is well mapped in other groups, 124 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

i.e., in the rest of the network, and the local scheduling decision will be suboptimal therefore. By using a more practically relevant scheduling policy, namely, random-m2m file dissemination algorithm, the packet to be distributed is decided at random. The obvious disadvantage of such a packet distribution policy is that multiple copies of the same packets traverse through the network, wasting scare wireless resources. We discuss in the next sessions how network coding can help to better exploit scares wireless resources and to conserve e.g. battery life of MTs. Furthermore, we show the benefits of network coding in terms of network reliability and robustness of the proposed m2m file dissemination technique.

• In order to minimize redundancy of the packet distribution and to enhance the ef- ﬁciency of intra-group multicast sessions the simple replicate-and-forward packet routing policy, proposed in our previous work, is replaced by Network Coding (NC).

• In our new version of the m2m ﬁle sharing algorithm, m2m users independently combine original packets they have already got via algebraic operations and send the corresponding results to their group neighbors.

• Due to the random nature of the encoded packets coefﬁcients, any randomly picked packet processed by NC will most likely contain information that is useful for each of the m2m participants in the group.

• Forming the linear combination of the packets increases the reliability of the system. This means that the probability for any packet to become rare or even unavailable because of sudden user departures is minimized.

• The data exchange algorithm determines the sender candidate at random. There is no need for a scheduler any more.

• The coding scheme is assumed to be predetermined. Coefﬁcient vectors specify- ing the generated linear combinations are sent with each data packet, allowing decoding at the MTs.

To illustrate the main concept of the extended NC-m2m algorithm, consider the example in Figure 7.3. Mobile terminals voluntarily participate in ﬁle sharing via direct mobile-to-mobile data transfer with the purpose to reconstruct the original popular content. In each step, the respective user sends the linear combination of all packets he has received so far. The following Initialization procedure is performed:

Each active m2m user creates two zero matrices X = 0m,w, G = 0m,m, where m is the number of logical packets, the source ﬁle is divided into, and w is the length of a logical packet in number of Bytes. X will be called the information matrix or decoding matrix. 7.3. NC applied to M2M Data Transfer 125

F Figure 7.3.: NC-m2m Concept (for simplicity, 2 is used in the ﬁgure.)

It stores the received encoded data (vectors) and is used for decoding later on. We call G encoding matrix. In this matrix the corresponding coefﬁcient vectors will be stored. The rank ν of G is initialized with zero. Initially, no user has any information (packets) of the popular content; the logical packets are available in the core network only. The original packets are periodically distributed by the BS packet by packet to active m2m users in order to generate one complete copy of the content within a radio cell. A user which received a packet from the BS acts as a server for the obtained packets. The packet transfer is performed on the user’s own, currently not used uplink carrier frequency and receivers switch to listen on the uplink. To avoid collisions on the uplink channels and to minimize the interference among m2m users, the MTs are organized into multiple groups, based on their locations and radio propagation conditions. The group organization policy, as well as the ﬁle transfer policy are performed as described in Sections 6.5-6.7, but also in [56, 61]. The following algorithm is executed:

Step 1 The packet dj with elements dj,l (j ∈{1, . . . , m}, l ∈{1, . . . , w}). received

by a MT from its BS is stored in the ﬁrst all-zero row ν + 1 of X (Xν+1,l = dj,l ). The

corresponding row ν + 1 in G (Gν+1,j = g j, j ∈{1, . . . , m}) will be a unit vector with the 1 at the ith position, which means that the currently received information vector is not encoded.

Step 2 An MT, which is entitled to send in the current frame1, uses the information stored in X to create a new encoded packet x′. For results shown in this work all oper- F ations are done in the ﬁnite ﬁeld 28 . In general, for the network under consideration

11 UMTS radio frame=0.01 sec. 126 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

the ﬁeld size must be sufﬁciently large [20].

The MT-sender has to execute the following steps:

• Create a vector g of uniformly distributed random coefﬁcients gi, where

F i ∈{1, . . . , ν} and gi ∈ 28

We call it local encoding vector.

• Compute the new encoded vector, which we call encoded packet,

ν ′ x = gk · xk k=1 X

where xk is the kth row of matrix X. The encoding is done separately for each Byte position in a packet, using the same encoding vector.

• x′ is broadcast to the group members together with the corresponding encoding ′ vector gν+1, where

′ ′ ′ ′ gν+1 = g1 · g1 + g2 · g2 + ... + gν · gν .

We call g′ the global encoding vector, it represents the encoded packets in terms of the source packets. The MT-receiver(s) has to execute the following steps:

Step 3 If a packet x′ has been received by a MT, it is stored in the row ν + 1 of X, and the corresponding encoding vector g′ is stored in G. This vector is also included in each coded packet’s header, as a tag. The overhead this incurs is negligible if packet

size (m log2 q bits) is sufficiently large, where q is the field size. The benefit of the tag is profound, this gives the ability to be completely decentralized: receivers can compute G and decode without knowledge about network topology. G is transformed to lower triangular form, using Gaussian elimination. If the received packet was innovative, the number of non-zero rows in G and therefore ν will increase by one. If ν = m (G has full rank) the user has collected enough linearly independent combinations to completely reconstruct the original file. Thus, file reconstruction can be accomplished by simply solving a system of linear equations. The matrix G is invertible with high probability, since all of the coefficients of all the local encoding vectors in the network are chosen randomly, independently, and uniformly from the sufficiently large field F. The field size, used for investigation in this work (28) is sufficient, as it has been proven in [29] and [67]. 7.4. Numerical Results 127

Step 4 Users, which have not found any useful packets within a speciﬁed time interval, try to connect to the BS for packet delivery.

Additionally, our elimination routine outputs the IDs of all packets dj, which are guaranteed to increase ν and therefore can be ordered in Step 4. All encoding/decoding operations are performed in the same way for each m2m user. The network coding scheme implemented in our algorithm is linear, which makes encoding/decoding procedure quite simple.

7.4. Numerical Results

In this section we present numerical results to highlight the effectiveness of the proposed NC-m2m technique.

7.4.1. Simulation Environment

To obtain performance results and to demonstrate the effectiveness of the proposed network coding based m2m technique, we have developed a simulation model based on realistic assumptions. Our reference scenario consists of a UMTS network supported by m2m data transmission mode. We consider mixed trafﬁc scenarios, where conventional UMTS speech users coexist with m2m ﬁle sharing participants. The simulated network is divided into regular hexagonal cells. Since the main focus of our analysis lies in the optimization of data availability to users in hotspots, e.g. airport lounges, railway stations, shopping malls, the radius of the cell is assumed to be 50 m.

• We simulate a population of m2m users with Poisson distributed service requests, that are interested in downloading a popular ﬁle. Each user moves independently with the same speed. In our simulation, the minimal and maximal speed were 3 m/s and 30 m/s, respectively.

• The size of the mentioned above content is 40 Kbyte; content is distributed with the cooperative m2m strategy.

• Users dynamically join and leave the group at any time due to battery life, handover. They can also leave the system before they have ﬁnished the download process (e.g. loss of interest in the content), representing a relatively loosely coupled formation.

• The group size decision for the simulated scenarios is made in favor of group size 13. The previous analyses have shown, that this choice is near optimal with respect to the acceptable level of uplink interference between groups. Therefore, for the following studies the group size 13 will be used. 128 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

7.4.2. Traﬃc Scenario

We study the following trafﬁc scenarios:

1. Scenario 1: The number of speech users per cell is kept constant (approx. 3 speech user per cell), while the load of m2m users is varied (low (10 users/cell), medium (30 users/cell), high (50 users/cell)). All users are assumed to be pedestrian.

2. Scenario 2: The amount of speech trafﬁc varies proportionally to that of the m2m trafﬁc, with the average number of speech users being 20% of the average number of m2m users in a cell. All users are assumed to be pedestrian.

3. Scenario 3: The same intention as the previous setting with 20% speech users load, increasing the velocity of m2m users from 3 m/s to 30 m/s.

4. Scenario 4: The same as Scenario 2, but with dynamic arrivals of new m2m users.

The speech users operate on the same uplink frequencies as m2m participants, com- posing a potential source of m2m signal disturbance. The transmission in conventional mode complies with 3GPP speciﬁcations for UTRA-FDD [2].

• During the simulation all m2m users in the network use the same packet distribution policy, either transmission of original packets or dissemination of encoded packets using linear network coding.

• Network coding is done only at the MTs, who participate in the m2m ﬁle sharing.

• In order to avoid interference from MTs transmitting in m2m mode on other signals at the BS receiver, the transmit power is set to the minimum, which is -44 dBm according to 3GPP speciﬁcations [2].

• Without loss of generality, we assume that the transmission rate is 1 packet/frame. Thus, depending on the coding scheme and spreading factor the packet length becomes appropriately large.

• The characteristics of the wireless channel in each group can vary from slot to slot (fast fading) due to the cross-trafﬁc users, which, as we have already mentioned before, transmit on the same uplink frequency.

7.4.3. Performance Measures

Network performance for mentioned above scenarios is evaluated in terms of usage of network resources. The beneﬁts are an improved system throughput, released overall downlink/uplink capacity and reduced ﬁle download time. We evaluate the system performance according to the following measures:

• Overall downlink/uplink throughput gain: data volume reduction in the downlink/uplink.

• Service probability gain: increase in the number of served users (%).

7.4.4. Performance Comparison of Simple M2M Algorithm with network coding – mobile-to-mobile (NC-M2M) File Dissemination

We first study the situation where MTs depart from the system immediately after finishing their downloads and we don’t consider arrivals of new users. All users are assumed to be pedestrians. We will discuss possible benefits of the NC-supported m2m file dissemination, in terms of operational complexity in a dynamically changing environment of wireless networks. We start our analysis with Figure 7.4, where we compare the performance of a UMTS network, supported by the NC-m2m algorithm to that of a system with a simple replicate-and-forward random-m2m algorithm. The upper graphs of Figure 7.4 demonstrate the benefit of the NC-m2m file distribution mechanism for medium traffic load (Scenario 1) and high traffic load (Scenario 2) in terms of the download time reduction. The bottom graphs depict the performance gain of the NC-m2m algorithm in terms of released downlink resources. The results show that the NC-m2m algorithm performs significantly better than m2m cooperative schemes without employing NC. In particular, we observe that the users experience a reduction of the download time for the complete file of up to 32% in case of NC. The reason is that each encoded packet can be useful to any user with a high probability. In fact, the network is dynamic, with users and links being added and removed in an ad hoc way. Mobiles failure, link failure, and packet loss can occur. However, decoding can be robust, provided that the network which is the traverse by the received packets

maintains MinCut(A, Fk) = h, which is always the case if the NC supported m2m ﬁle dissemination technique is used. 130 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

Comparison of download times (file size 40 kb, medium m2m load, scenario 1) Comparison of download times (file size 40 kb, high m2m load, scenario 2)

NC−m2m NC−m2m 1 m2m 1 m2m

gain: 32% gain: 31% Probability of completed downloads Probability of completed downloads

0 0 0 10 20 30 40 50 60 70 0 20 40 60 80 100 120 Time (seconds) Time (seconds) (a) Comparison of download times of m2m with (b) Comparison of download times of m2m with NC-m2m ﬁle transfer (medium m2m load, Sce- NC-m2m ﬁle transfer (high m2m load, Sce- nario 1). nario 2).

Comparison of reserved downlink channels (file size 40kb, high m2m load, scenario 2) Comparison of reserved downlink channels (file size 40 kb, medium m2m load, scenario 1) 12 12 NC−m2m NC−m2m m2m m2m 10 10

8 8

6 6

4 4 Number of reserved downlink channels Number of reserved downlink channels 2 2

0 0 0 10 20 30 40 50 60 70 80 0 20 40 60 80 100 120 Time (seconds) Time (seconds) (c) Comparison of reserved downlink channels (d) Comparison of reserved downlink channels of of m2m with NC-m2m ﬁle transfer (medium m2m with NC-m2m ﬁle transfer (high m2m m2m load, Scenario 1). load, Scenario 2).

Figure 7.4.: Performance comparison of simple replicate-and-forward m2m ﬁle transfer with NC-m2m for medium m2m trafﬁc load, Scen.1 (left) and high m2m load, Scen.2 (right)).

7.4.5. Impact of Extended M2M User Availability and Users’ Mobility

We investigate now a more strict file sharing policy. Namely, the finished users have to stay in the system for a certain period of time allowing the m2m algorithm to continue sharing the file with not yet finished users. We are especially interested in the distribution of finished MTs in different states of the download job progress. The observations, visualized in Figure 7.5, indicate that if we encourage the users, who have completed their downloads, to behave "altruistically" by staying online for a while, this can help the other file sharing participants in the last states of the download progress to speed up the reception of their still missing packets. It is quite evident, that 7.4. Numerical Results 131

such a modified policy gives still not finished users better opportunity to get missing packets directly from staying completed users, releasing, in turn, more BS capacity. However, external reward schemes might be needed to motivate finished users to participate in this policy and help. Strategies such as offering bonus systems (e.g. upload credits) to network subscribers who choose to allow using their uplink capacity to provide the contents to other not jet finished users or implementing restrictions on the amount of traffic that a particular terminal is permitted to distribute, could mitigate lack of user’s willingness to cooperate.

Comparison of download times (medium m2m load, scenario 2) Comparison of download times (medium m2m load, scenario 3)

NC−m2m / users staying 1 NC−m2m / users leaving 1 m2m / users staying m2m / users leaving gain: 33%

Probability of completed downloads Probability of completed downloads NC−m2m / users leaving / vehicular NC−m2m / users leaving / pedestrian m2m / users leaving / vehicular m2m / users leaving / pedestrian 0 0 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 time (seconds) time (seconds)

Figure 7.5.: Impact of cooperative behav- Figure 7.6.: Impact of users’ velocity on ior of finished users on the the download times, medium file download times, medium m2m load, file size 40 kbyte, m2m load, file size 40 kbyte, Scenario 3)). Scenario 2.

User’s mobility has a significant effect on the performance of the proposed m2m algorithm. We increased the users’ speed from 3 m/s to 30 m/s (Scenario 3) in order to investigate the influence of user mobility on the performance of the proposed NC-m2m algorithm. The results indicate a significant effect of the user mobility on the system performance. We observe a substantial gain of up to 33 % in download time reduction in case of fast moving MTs compared to MTs with low velocity in a UMTS system with medium traffic load (see Figure 7.6). This is reasonable, because the higher the mobility of the users, the more frequently the groups are updated and reshaped. Thus, the probability for some user to find packets of interest during a short time interval is increased. As can be seen from the Figure 7.6 NC-m2m also benefits much more from user mobility; already for medium load shown in Figure 7.6 the performance gain due to users’ mobility is essential. 132 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

7.4.6. Steady State System Performance

The results presented in the previous subsection analyze the performance of the NC- m2m algorithm for a snapshot simulations. In this section we investigate the efﬁciency of the NC-m2m scheme in a network scenario with dynamic new users’ arrivals (Sce- nario 4). Users join and leave the network, but the population of the system remains approximately constant. The number of reserved downlink channels in the steady state of the system versus time is depicted in Figure 7.7.

Comparison of reserved downlink channels (medium m2m load, scenario 4) 11 NC−m2m / users staying / new user arrivals 10 m2m / users staying / new user arrivals

4 Number of reserved downlink channels 3

2 0 50 100 150 200 250 Time (seconds)

Figure 7.7.: Number of reserved downlink channels in medium loaded UMTS system in the steady state versus time.

Some numerical values for the overall downlink throughput gain with respect to amount of data which had to be sent via the downlink channels in order to distribute the data ﬁle of 40 Kbyte size to the users within one cell are presented in Table 7.1.

Table 7.1.: Overall downlink throughput gain, Scenario 1. Load medium high Data vol. in DL in m2m mode (kbyte/cell) 122.8 130.3 Data vol. in DL in NC-m2m mode (kbyte/cell) 69.6 71.9 Data vol. in DL in NC-m2m mode (stay) (kbyte/cell) 64.3 64.7

Three packet distribution strategies have been compared: the simple replicate-and- forward m2m algorithm and the NC-based m2m scheme w/o and with finished "helpers" (the finished users stay in the system for a certain period of time and help to speed up the distribution of missing packets to still not finished users). Again, the network, where the finished users stay in the system for a certain period of time performed better.

7.4.7. Further Beneﬁts of NC-M2M File Dissemination: Released Uplink

Another performance indicator of interest is the released uplink capacity. We compare the random-m2m algorithm with NC-M2M ﬁle dissemination. As it was already 7.4. Numerical Results 133

mentioned in previous sections non of these algorithms has explicit information about network topology and only limited knowledge about their neighborhood. Figure 7.8 demonstrates the positive impact of the NC-based m2m file dissemination in terms of number of m2m transmission needed to disseminate a popular file among the users on the uplink channels. The problem is actually equivalent to a simple variation of the coupon collector problem. Results in the literature establish that O (m) transmissions per MT, where m is the number of MTs, will be sufficient in case of using network coding [15]. To accomplish the same task, using the random-m2m algorithm O (n log n) transmissions will be needed. This is due to the fact, that by simply applying replicate-and- forward scheme, the packets are not distributed evenly. As a result, some packets are sparsely available in the network, and it is difficult for m2m-receivers to get them. In turn, when NC is applied each m2m-sender in a group saves a number of transmissions, by combining original packets into network-coded packets, which most probably can be useful to any user in the group. Indeed, if we would track the data reception at each MT and collect the observations in a matrix, the matrix will with high probability have full rank. From Figure 7.8 one can observe the 45% of released uplink capacity.

Comparison of reserved uplink channels (medium m2m load, scenario 4)

30 NC−m2m / users leaving / new user arrivals m2m / users leaving / new user arrivals

10 Number of reserved uplink channels 5

0 0 50 100 150 200 250 Time (seconds)

Figure 7.8.: Impact of the NC-m2m technique on the uplink resource consumption, medium trafﬁc load.

NC enables higher goodput while reducing the number of transmissions which directly results in less radio resource consumption and prolonged battery life for the MTs. Moreover, by reducing the number of transmissions network coding assisted m2m ﬁle dissemination reduces also the trafﬁc congestion in network. 134 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

7.5. Enhanced Network Coding for Operation on Data of Arbitrary Size

As it could be seen from the previous section network coding ensures optimal network resource utilization. In many applications, however, the files to be distributed are too large to be processed with the straightforward realization of network coding. Combining the entire file to construct a single encoded block is impractical due to high encoding and decoding costs not compatible with limited processing capabilities of mobile devices. Furthermore, the amount of data on which a MT can perform coding operations is constrained by the only small-to-moderate storage available at a mobile device. In this section we discuss how an efficient NC-m2m file dissemination can be achieved even with strong memory constraints.

7.5.1. Generations: Optimized Packet Combination

In order to operate on files of virtually arbitrary size the network coding technique must be modified. One way to achieve this, is to organize the packets of the information source (popular file) into groups of packets, called generations with f packets per generation, where f will be the generation size. It determines the size of matrices the receiver needs to buffer, and later invert to decode the information. Hence, there exists one matrix per generation in each receiver and only packets of the same generation will be combined (encoded). To keep track of packets in the same generation, each packet is tagged with its generation number z (one byte is sufficient for this purpose), where z ∈{1, . . . , u}, and u is the total number of generations the file is divided into (see also [12, 81]). Figure 7.9 demonstrates the basic concept of handling generations in NC-based m2m file dissemination mode.

• The original packets are periodically distributed by the BS packet by packet to active m2m users in order to generate one complete copy of the content within a radio cell.

• Packets arrive in arbitrary order at a MT. After arrival the packets are sorted by generation number and put into the MT buffer. A generation number is appended to the packet headers.

• If the MT is entitled to send in the current frame it generates a new encoded packet as a linear combination from packets tagged by generation number one, or number two, or number x, depending on the generation distribution strategy (GDS), which will be discussed in detail in section 7.6. Thus, the coding is performed only inside the generations. Packets are encoded symbol-wise. 7.6. Numerical Results 135

• To trace the ﬁle download progress at each MT, each arriving packet is added to a generation buffer, according to its generation number. A generation number is appended to the packet header. Gaussian elimination is then performed on those encoding matrices, that correspond to generation buffers that have new data available. This allows to store the encoding matrices in row reduced eche- lon form and to check whether a new packet is innovative immediately after its reception.

• A packet received by a MT is innovative if it increases the rank of the corresponding encoding matrix of packets at this MT.

• Non-innovative packets will be ignored.

• The BS supports MTs with the signalling information and allows only one MT within the group to transmit in a given time interval, avoiding collisions within a particular group.

The concept presented above renders the NC-based m2m algorithm applicable for real networks, where the optimization of the data availability of non-real time services was the main focus. However, note that this concept is not sufficient for delay sensitive services like video- streaming. Since the packets on wireless links are subject to random delays, the asynchronous distribution of the packets will be a problem for later decoding. Thus, in order to decode the packets within a prescribed time window, it is necessary to assure near- constant end-to-end delays. A possible generation management method for real-time applications is presented in [12]. To make network coding practical for services that are not resilient to delay, the authors suggest synchronizing packets within a generation by buffering, additionally taking into account random delays. New packets arriving at a node will be put into a common buffer sorted by generation number with a current generation at the top of the queue. The current generation is periodically advanced and the old is flushed from the buffer according to a flushing policy. All packets from the previous generation, which arrive after a predefined time-span will be discarded. In [81] a modified generation management technique has been introduced by using a hashing function to determine which generation each packet is associated with. The choice of the hash function affects the number of generations, transmitted data is divided into. We will show later, that the generation size plays a very important role in achieving adequate performance of NC-based algorithms.

7.6. Numerical Results

The following sections deal with performance evaluation and analysis of m2m data dissemination in UMTS supported by network coding with multiple generations (NC- MG). 136 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

Figure 7.9.: The basic concept of handling generations for the network coding based m2m ﬁle dissemination scheme.

Our work distinguishes itself from other works in this area by considering a wide variety of views in the analyzes. The major goal is to analyze the trade-off between ﬁle size, size of generations, memory requirements, computational complexity and Quality of Service (QoS) like transmission delay, etc. System performance for generation-based NC-m2m data dissemination is then compared with that of simple straightforward network coding in terms of the above mentioned characteristics. We start out with an outline of some fundamental properties of multi-generation NC, namely, computational complexity and memory requirements at the MT.

7.6.1. Memory Requirements and Computational Complexity

As mentioned earlier one of the major issues that arise in a system which applies network coding is extensive usage of quite limited memory resources in mobile devices. 7.6. Numerical Results 137

Evidently, large ﬁles must be split into a larger number of packets, which, by applying straightforward NC (with linear combining of all packets available at the MT), results in a m × m decoding matrix, (where m is the number of packets in the ﬁle). Thus, the effective memory required in each MT increases with O (m2), which leads to a quick memory exhaustion of mobile devices.

m 2 m2 · u = ≤ m2, with 1 ≤ u ≤ m (7.2) u u As can be seen from the simple calculation in Equation (7.2), the arrangement of the packets into groups, can reduce the memory complexity, if the number of generations u is chosen depending on the total file size m and the fixed generation size f . With u ∼ m f memory complexity becomes O (m), which means, that by applying generation-based NC, memory consumption increases linearly with the file size. The next figure of merit we would like to analyze is computational complexity. It is well known that the Gaussian elimination procedure has a complexity of O (m3). In case of NC with multiple generations, assuming constant file size, a larger number of generations results in a smaller generation size and reduced decoding matrix size (within each generation), respectively. Thus, the computational complexity (complexity of encoding operations) is O (u· f 3), where u· f = m, which is lower than that Gaussian elimination in standard NC implementation.

7.6.2. Analysis of Relationship between Diﬀerent Figures of Merit

To obtain an extensive statistical characterization of the performance of NC-MG-m2m technique, we have conducted a substantial number of simulations with the different parameter settings. In the thesis, however, we outline only some of them.

7.6.2.1. Generation Distribution Strategy

We investigate and compare performance of ﬁle dissemination for three generation distribution strategies (GDSs): local rarest-ﬁrst, sequential and random. Each of the above mentioned strategies requires a different degree of knowledge about the network topology.

Local Rarest-First generation distribution strategy (GDS) Realization of generation based-NC m2m data dissemination by using local rarest-ﬁrst GDS requires explicit knowledge about users in each m2m group. The algorithm, which decides that packets of a certain generation will be encoded, must be supported with information regarding the packet dissemination progress throughout a group. For the local rarest-ﬁrst case the packets from the most unpopular generation (local rarest) will be combined in order to generate a new encoded packet, which the MT-sender multicasts to its group members. 138 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

However, local neighborhood information (within a m2m group only) is sufﬁcient to perform local rarest-ﬁrst GDS-based cooperative data transfer.

Sequential GDS In order to perform NC-m2m data transfer by using sequential GDS, even less information about neighboring MTs is needed. Instead, the data exchange algorithm performed at the BS decides on the number of the next generation (in sequential order) to be sent in a particular m2m-group in the current time interval. Along with the generation number, the sender identiﬁcation information is transmitted.

Random GDS The sender chooses at random the generation number of the packets to be encoded and sent from the set of non-empty generations. No additional knowledge about packets existing in nearby MTs is necessary.

Note: For all distribution strategy the following role is valid: If the generation to be sent in the speciﬁed time interval is empty the packets of some other (according to the generation distribution strategy currently in use) generation will be picked, coded and sent. We examine now some parameters of generation construction and their impact on the system performance. The performed simulations return following observations.

7.6.2.2. Generation Size and File Size

As mentioned before, the generation size has a significant impact on the performance of network coding. To achieve sufficient performance of the algorithm and at the same time assure reduced memory requirements and computational complexity, it is necessary to determine the optimal generation size for a given file size. In this section we compare the performance of the NC-m2m file dissemination algorithm for different generation sizes. For the file size fixed at 40 kbyte we distinguish small (20 packets), medium (100 packets) and large (720 packets) generation sizes. Simulations have been performed for the system with a high traffic load scenario by using sequential GDS solely. All users were assumed to be Pedestrian. No arrival of new m2m users are simulated. The most important parameters used in our simulations are summarized by Table 7.2. Figure 7.10a shows the impact of the generation size on the file download time. As can be seen from the diagram, with an increasing number of generations (in turn, decreasing the generation size) we observe NC-m2m performance degradation. This is obvious, since every time an m2m-sender picks some generation, according to the GDS, he creates an encoded packet (using the whole set of data stored therein) and sends it to its neighbors, together with the corresponding encoding vector. Some choices will be more beneficial for the group members, others will only provide them with little information. If only one generation is present the choice is always trivially optimal. As the number of generations increases, the probability of choosing a generation 7.6. Numerical Results 139

Table 7.2.: Main simulation parameters. Trafﬁc and environmental settings Traf. load (max. num. of m2m users/cell) 50 (high) Maximum m2m group size 13 Antenna type omni-directional Cell radius 50 m User proﬁle Pedestrian Radio interface and algorithm settings File size 500 kbyte Required gross data rate 60 kbit/s Maximum user data rate with 1/2-rate coding 30 kbit/s Size of logical packets for m2m data 225 bit (coded) Receiver sensitivity −112 dBm Transmission power in m2m mode −44 dBm

Eb/N0 target 3 dB Inner loop power control for m2m sender OFF Simulation step size 1 radio frame (0.01 s) Group update period 100 radio frames (1 s)

Comparison of file download times for different Comparison of number of reserved downlink channels for different generation sizes by using sequential GDS generation sizes using sequential GDS (high load, file size 40kb) 1 18 generation size 720 packets generation size 20 packets 0.9 generation size 20 packets 16 generation size 100 packets generation size 100 packets generation size 720 packets 0.8 14 0.7 12 0.6 10 0.5 8 0.4 6 0.3 4

Probability of completed downloads 0.2 Number of reserved downlink channels 0.1 2

0 0 0 10 20 30 40 50 60 70 80 10 20 30 40 50 60 70 80 Time (seconds) Time (seconds) (a) Comparison of file download times for differ- (b) Impact of different generation sizes on down- ent generation sizes and fixed 40 kbyte file link resource consumption. size.

Figure 7.10.: Impact of generation size on the performance of NC-based m2m ﬁle dissemination in high loaded UMTS system. 140 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

Comparison of download times for different GDSs Comparison of number of reserved downlink channels using different GDSs (high load, file size 40 kb) (high load, file size 40 kb, generation size 720 packets) 1 10 gen. size 720 packets/sequential GDS 0.9 9 gen. size 720 packets/random GDS gen. size 720 packets/rarest−first GDS 0.8 8

0.7 7

0.6 6

0.5 5

0.4 4

0.3 3

gen. size 720 packets/random GDS

Probability of completed downloads 0.2 2 gen. size 720 packets/rarest−first GDS Number of reserved downlink channels 0.1 gen. size 720 packets/sequential GDS 1 NC−m2m/one generaion 0 0 0 10 20 30 40 50 60 70 80 10 20 30 40 50 60 70 80 Time (seconds) Time (seconds) (a) Impact of generation distribution strategy on (b) Downlink resources consumption in NC-m2m the ﬁle download time by using large genera- mode by using different GDSs and large gener- tion size (ﬁle size 40 kbyte). ation size.

Figure 7.11.: Performance comparison of NC-MG m2m algorithm in high loaded system for different GDSs and ﬁxed generation size.

for coding and transmitting packets that is of inferior use for the receivers (compared to the optimal choice) also increases, consequently, decreasing the benefits of NC. Such an unfavorable choice of generation size coerces m2m users to request packets of the desired file from the BS more frequently. This results in the allocation of additional downlink channels, thus demanding more of the cell’s valuable downlink capacity and impairing the efficiency of the m2m concept, cf. Figure 7.10b. Furthermore, when the number of generations approaches the total number of packets, the performance of the NC-MG-m2m will converge to that of random replicate-and- forward m2m file dissemination. Thus, a single, large generation will always yield the best achievable performance; and it is desirable in multiple-generation case to apply a fairly large generation size.

7.6.2.3. Quality of Service (QoS) Requirements

To analyze the Quality of Service (QoS) resulting for the proposed generation-based NC technique, we have implemented all of the above mentioned generation distribution strategies: local rarest-first, sequential and random. Among the figures of merit we are interested in the data rate degradation caused e. g. by unfavorably chosen GDS. As one can see from Figure 7.11a the performance of generation-based NC-m2m is almost identical for all three generation distribution policies in terms of download time, with a given file size and a large generations size (two generations only). The diagram shows that all policies approach the performance of the optimal single generation case. The amount of consumed resources on the downlink is also almost the same for all three GDSs, as shown in Figure 7.11b. However, with increased number of generation and unchanged file size, the rarest-first GDS performs significantly inferior to the other 7.6. Numerical Results 141

Comparison of number of reserved downlink channels for different GDSs (high load, file size 40 kb, generation size 100 packets) Comparison of file download times for different GDSs 25 (high load, file size 40 kB, generation size 100 packets) 1 rarest−first GDS sequential GDS 0.9 random GDS relative loss: 56% 20 0.8

0.7 15 0.6

0.5 10 0.4

0.3 5

Probability of completed downloads 0.2 gen. size 100 packets/random GDS Number of reserved downlink channels 0.1 gen. size 100 packets/rarest−first GDS gen. size 100 packets/sequential GDS 0 0 0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 Time (seconds) Time (seconds) (a) Impact of generation distribution strategy on (b) Downlink resources consumption in NC-m2m the ﬁle download time by using generation mode by using different GDSs and generation size 100 packets (ﬁle size 40kbyte). size 100 packets.

Figure 7.12.: Performance comparison of NC-MG m2m algorithm in high loaded system for different GDSs and ﬁxed generation size.

two policies, while the difference in performance between these two policies is still very minor, see Figure 7.12a. These observations can be explained as follows. The gap of 56% in file download time is due to the fact that we assume a large number of generations and at the same time a relatively small file size. The performance of the rarest-first GDS, in terms of download time, then degrades, since the "rarest" generations always contain less packets at each MT than the remaining ones. Thus, the linear combination of the packets currently available at a sender candidate formed from rarest generation are less likely to be innovative for other group members. As a result the time interval between the reception of innovative packets at the MTs increases, reducing data rate and prolonging file download time, consequently. Related observations can be made from Figure 7.12b. A lack of novel packets, received by MTs in m2m mode, has to be compensated for. Users with expired waiting counters for new packets will request more data directly from their BSs. This results in unavoidable occupation of additional downlink resources. In contrast, the distribution of generations in random order prevents users from hav- ing a strict subset of the generations, which yields a better performance of m2m algorithm. The slight performance degradation of the random GDS compared to the sequential GDS is simple to explain. This is due to the fact that by using the random policy a generation can accidentally be chosen multiple times in a short time interval, thus triggering redundant data transmissions. We concentrate our further investigations therefore on sequential GDS, which ap- pears to be the most natural generation distribution strategy. 142 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

Comparison of download times of simple replicate−and−forward and NC−based m2m file sharing schemes (file size 500 kb, high m2m load, group size 13) Comparison of number of reserved downlink channels of simple replicate−and−forward and 1 NC−based file sharing schemes (file size 500 kb, high m2m load, group size 13) 14 0.9 random m2m gain: 30% NC−MG/720 packets/seq. GDS 12 0.8

0.7 10

0.6 8 gain: 31% 0.5

0.4 6

0.3 4

Probability of completed downloads 0.2 Number of reserved downlink channels 2 0.1 NC−MG/720 packets/seq. GDS random m2m 0 0 0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 Time (seconds) Time (seconds) (a) Comparison of ﬁle download times for NC-MG- (b) Downlink resources consumption in NC-MG- m2m, conventional and simple replicate-and m2m and simple random m2m mode in high forward m2m mode. loaded system.

Figure 7.13.: Performance comparison of large ﬁle distribution using NC-MG-m2m and simple random m2m algorithm in high loaded system (ﬁle size 500 kbyte).

7.7. Eﬃcient Large File Distribution in UMTS supported by Network Coded M2M Data Transfer with Multiple Generations

In this section the system performance for network coded m2m data dissemination using multiple generations (NC-MG-m2m) is compared with those of the simple replicate- and-forward m2m scheme, which has been analyzed in section 6, in terms of usage of network resources, duration of ﬁle download, as well as processing costs.

7.7.1. Comparison of NC-MG-m2m File Sharing with Replicate-and-Forward M2M Data Dissemination

Now, by using generations, files of virtually arbitrary size can be processed. In the following we analyze the dissemination of a file of 500 kbyte size, using NC-MG-m2m- data distribution strategy (Network Coding with Multiple Generations). For this, we have divided the above mentioned file into 26 generations, with 700 packets in each generation and distribute the popular content with the NC-MG-m2m algorithm using the sequential GDS. Other system parameters used in simulations remain unchanged (see Table 7.2). In Figure 7.13 and Figure 7.14 the corresponding simulation outcome is compared to the performance results of the most-utile-m2m, random-m2m and conventional packets distribution scheme for the scenario with high m2m-traffic load. The following observations can be made:

• Network coding with multiple generations (NC-MG) m2m data distribution outperforms the random m2m and conventional packet distribution signiﬁcantly, in 7.7. Efﬁcient Large File Distribution in UMTS supported by Network Coded M2M Data Transfer with 143 Multiple Generations

Comparison of number of reserved downlink channels of UMTS conventional and Comparison of download times of NC−based file sharing with conventional and simple NC−based file sharing schemes (file size 500 kb, high m2m load, group size 13) replicate−and−forward schemes (file size 500 kb, high m2m load, group size 13) 100 1 NC−MG/720 packets/seq. GDS UMTS conventional 0.9 80 0.8

0.7 60 0.6

0.5

0.4 40

0.3 NC−MG/720 packets/seq. GDS 0.2 20 Number of reserved downlink channels Probability of completed downloads most−utile m2m 0.1 random m2m UMTS conventional 0 0 0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 Time (seconds) Time (seconds) (a) Comparison of ﬁle download times for NC-MG- (b) Downlink resources consumption in NC-MG- m2m, conventional and simple replicate-and m2m and conventional UMTS modes in high forward m2m modes. loaded system.

Comparison of m2m data rate progression over time for different file distribution strategies (high m2m load) 30 NC−MG/720 packets/seq. GDS most−utile m2m 25 random m2m

Data rate (kbit/s) 10

0 0 100 200 300 400 500 600 700 Time (seconds) (c) Impact of NC-MG-m2m algorithm on the m2m data rate progression.

Figure 7.14.: Performance comparison of large ﬁle distribution in high loaded system using NC-MG-m2m, simple random-m2m and conventional UMTS network mode.

terms of both file download time and released downlink capacity. We can observe a gain of up to 30% at the 90% quantile of completed downloads in the high loaded pure m2m system, compared to the random m2m file distribution, and up to 85 % compared to the conventional file transfer mode.

• NC-MG m2m data distribution algorithm reduces the number of (uplink) m2m transmissions by a factor of log m [22] (compared to random replicate-and-forward), which results e.g. in a reduced battery consumption at the MTs.

• In comparison to the most-utile m2m policy, NC-MG m2m achieves a slightly better performance.

• Although the gain is not signiﬁcant, we save the amount of signalling information we would otherwise need to transmit in case of the most-utile m2m transfer to 144 Chapter 7. Enhanced Scheduling by Network Coding Supported M2M Data Transfer

keep each MT updated with the knowledge about all packets that exist in its neighborhood.

7.7.1.1. Practical Issues

First, to support the proposed m2m system mode within the UMTS network from a practical point of view, namely to make MTs able to operate in both uplink/downlink frequency bands, some hardware changes will be necessary. Providing for MTs signal processing capabilities for receiving on the uplink frequencies is one possibility to realize the m2m algorithm. Even though RF hardware is not cheap and this solution would increase the cost for the mobile terminals at the moment, it is known that prices for electronic devices have been decreasing at an enormous pace ever since. Second, the application of network coding entails further functionality extensions. However, nowadays the processing becomes cheaper, too. Furthermore, as we have seen from the simulation results, NC can be implemented at a lower computational cost, by using multiple generations.

7.8. Summary

The aim of analysis performed in this section was to further improve the interaction between cellular and peer-to-peer networks by generalizing the traditional scheduling paradigm. A network coding technique was embedded as a solution to the scheduling problem in the distributed dynamic environment of wireless large-scale networks. We investigate the performance of the system in terms of improvement in information distribution and dependability of information distribution among m2m users and show how, and in which terms network coding can thereby help. We have been especially interested in the distribution of the ﬁle download times for m2m users in different states of the download progress. Numerical results reveal the following ﬁndings:

• The NC-m2m (network coding – mobile-to-mobile) technique outperforms the performance of the replicate-and-forward m2m solutions proposed in Section 6, obviating the need for proper scheduling for non-real time ﬁle sharing applications.

• The proposed NC technique requires only local neighborhood information to perform the cooperative data transfer and therefore avoids ﬂooding the network with signalling information.

• NC-m2m ﬁle dissemination outperforms even most-utile m2m algorithm, which is supported with perfect knowledge about the network topology.

• Simulations demonstrate the enhanced performance of the ﬁle distribution in terms of ﬁle download time. Furthermore, the obtained results highlight that 7.8. Summary 145

network coding based m2m data transfer allows distribution of popular ﬁles to a large number of users while placing minimal bandwidth requirements on the central server, reducing at the same time the number of redundant uplink transmissions.

We have conducted a performance evaluation and analysis of m2m data dissemination for ﬁles of virtually arbitrary size in the UMTS system supported by network coding with multiple generations (NC-MG). In order to achieve a trade-off between memory requirements and computational complexity on the one side and sufﬁcient performance characteristics of the NC-MG algorithm on the other side, we have investigate different generation distribution strategies (GDSs). The following conclusions can be drawn from the numerical results:

• Size and composition of generations (distribution strategy) have a significant influence on the performance of the proposed file dissemination technique.

• The smaller the number of packets in a generation, the lower the beneﬁts of NC.

• With appropriately chosen generation size the MT is able to hold several generations in memory and to operate on ﬁles of large size, consequently.

• By disseminating the popular content with the NC-MG-m2m algorithm using the sequential or random GDS, only simple operations on the generations without any extra coordination or special knowledge about the network topology are necessary.

• This means that the NC-MG-m2m algorithm provides performance beneﬁts even in scenarios where the knowledge of the MTs regarding the state of other MTs is constrained; receivers can decode even if the network topology is unknown.

We have demonstrated that the NC-MG-m2m algorithm is a viable solution for an efﬁcient and reliable dissemination of any media ﬁle or streams at low computational cost.

Chapter 8. Conclusions

In this thesis the radio access part of UMTS has been investigated. The UMTS air interface has been modeled in order to introduce efficient analytical algorithms for system performance estimation, as well as to propose methods for solving some of the radio resource optimization problems within the network. In the first part of this work analyses of resource allocation schemes for packet- oriented large-scale queuing wireless networks with WCDMA radio interface have been presented. Performance prediction, optimization of system parameters and network dimensioning have been addressed. To model the complex multi-rate multimedia traffic and, consequently, to evaluate the system reliably, a new unified analytical model, which combines both loss and queuing systems with multi-rate traffic streams has been applied for theoretical performance study of mentioned above radio systems. The interaction between two service levels, connection level and packet level, has been taken into account as well. The obtained results can help to work out and further understand the typical behavior of multi-service systems under different air interface conditions and various traffic and scheduling scenarios. The proposed generalized algorithms can be used to derive different system performance indicators in a highly efficient way. The material presented in this part of the thesis can be also found partly in [63, 5, 60]. In the second part of the thesis concepts of cooperation in heterogeneous networks have been discussed as a solution to the downlink capacity exhaustion problem. We have highlighted the possible benefits of the interaction between cellular and peer-to- peer networks in a complementary manner. A novel hybrid technique for efficient distribution of popular non-real-time data content in order to optimize the data availability to users in hotspot scenarios has been introduced. The proposed concept based on direct mobile-to-mobile (m2m) data exchange has been realized by embedding peer-to- peer file sharing applications into the hierarchical architecture of UMTS networks. The algorithm exploits the uplink/downlink traffic imbalance, typical for 3G radio networks, and boosts the spectral efficiency of UMTS by dynamically allocating m2m users to temporarily unused uplink frequency bands. Simulation results have demonstrated that the proposed cooperative solution significantly outperforms the conventional UMTS mode. The results have indicated a substantial increase of service probability and overall sys- 148 Chapter 8. Conclusions

tem throughput, as well as a substantial reduction of the expected file download time. The publications [58, 57, 59, 56] provided a basis for this chapter. To further improve the interaction between cellular and peer-to-peer networks, properly designed scheduling of the packet transfer is of great importance. To generalize the traditional scheduling paradigm a network coding technique was embedded. We have investigated the performance of the system in terms of dependability of information distribution among m2m users. Simulations have demonstrated the enhanced performance of the file distribution in terms of file download time. Furthermore, the obtained results have highlighted that network coding based m2m data transfer allows the distribution of popular files to a large number of users, while placing minimal bandwidth requirements on the central server. The proposed concept might be a promising alternative for distribution of content in cellular radio networks like UMTS. The corresponding sections of the monograph elaborated on the findings presented in[61, 62]. Hybrid networks are expected to play a key role in supporting a large range of wireless services as a scalable mobility solution. Even though cooperation in wireless communications has not reached its full maturity yet, its realm is already broad. Appendix A. Kendall Notation

There is a standard notation for classifying queuing systems into different types. This was proposed by D. G. Kendall. Systems are described by the notation: A/B/N/n/S/X , where A denotes the distribution of the interarrival times of users and B the distribution of the service times; N specifies the number of servers, n the maximum number of waiting requests in the finite case (number of servers plus the capacity of the queue), S is the number of users and X defines a queuing discipline. A and B can take any of following distribution types:

M – Markovian Exponential time intervals (Poisson arrival process, exponentially distributed service times).

D – Deterministic Constant time intervals.

Ek – Erlang-k Erlang distributed time intervals.

G – General Arbitrary distribution of time intervals.

A queuing discipline X determines the manner in which the server handles users’ calls. It deﬁnes the way they will be served, the order in which they are served, and the way in which resources are divided between the users[7]. Here are details of four queuing disciplines:

FIFO: First-In-First-Out This principle states that users are served one at a time and that the user that has been waiting the longest is served ﬁrst. It is called a fair queue or an ordered queue, and this discipline is often preferred in real-life when users are human beings. It is also denoted as FCFS (First Come-First Served).

SIRO: Service In Random Order All users waiting in the queue have the same probability of being chosen for service. This is also called RANDOM or RS (Random Selection). For the above mentioned disciplines the total waiting time and thus the mean waiting time for all users is constant. The queuing discipline only decides how waiting time is allocated to the individual users.

RR: Round Robin A user served is given at most a ﬁxed service time (time slice or slot). If the service is not completed during this interval, the customer returns to the queue which is FCFS. 150 Appendix A. Kendall Notation

PS: Processor Sharing Capacity is shared equally between users and they all effectively experience the same delay depending on the actual service time.

Priority: Users with high priority are served ﬁrst.

In queuing theory it is generally assumed that the total offered trafﬁc is independent of the queuing discipline. For more work concerning this topic see [37]. Appendix B. Binomial-Poisson-Pascal traﬃc (BPP) Paradigm

Erlang’s classical loss system can be generalized to state dependent Poisson arrival processes, which include three models, called Binomial-Poisson-Pascal trafﬁc (BPP) trafﬁc models: • Poisson case (Erlang model)

• Binomial case (Engset model)

• Pascal case (Negative Binomial) Erlang model The arrival process is a Poisson process with intensity λ. A traffic stream is characterized by its mean offered traffic ρ = λ , (µ = 1/τ, τ is the mean hold- µ ing time) and peakedness Z = 1. In Erlang’s loss model with Poisson arrival processes, the mean offered traffic is equivalent to the average number of call attempts per mean holding time. The peakedness is defined as the ratio between variance and mean value of the state variables and equals one for the Erlang model. The most important performance measures are time congestion E, call congestion B and traffic congestion C. For systems with Poisson arrival processes all mentioned above performance measures are equal due to the PASTA property (Poisson Arrivals See Time Averages). PASTA property means that the call intensity is independent of the state, and an arriving event sees the system as if it observes the system at a random instant of time.

Engset model For the Binomial case the arrival rate per idle source and the system’s limited number of sources S are given. Each individual source has constant intensity γ of call arrivals and switches between the states idle and busy. A source is idle during a time interval, whose duration is exponentially distributed with arrival intensity γ. The source is busy during an exponentially distributed time interval with intensity (service intensity) µ. The corresponding illustration is depicted in Figure B.1. The offered trafﬁc per idle source consequently is β = γ/µ. When the source is busy the arrival intensity is zero. Thus, the arrival process becomes state- dependent; if at a given point in time i sources are busy, then the arrival intensity 152 Appendix B. Binomial-Poisson-Pascal trafﬁc (BPP) Paradigm

State ...... Busy ...... Time ...... Idle ......

......

...... 1...... 1...... µ . γ . . . . arrivaldeparture arrival Figure B.1.: Each individual source is either idle or busy, and behaves independently of all other sources.

equals (S − i)γ. The total offered traffic is ρ = S · β . The offered traffic per idle 1+β source is a difficult concept to deal with because the proportion of time when a source is idle depends on the congestion. The number of calls offered by a source depends on the number of channels. The channel-limit is assumed to be N. The state transition diagram for the Engset model with S > n is shown in Figure B.2.

Sγ (S − 1)γ (S − i)γ (S − N + 1)γ

0 1 i N − 1 N

µ 2µ iµ (N − 1)µ Nµ

Figure B.2.: State transition diagram for the Engset case with S > N, where S is the number of sources and N is the number of channels.

As mentioned before, finite source traffic is characterized by the number of sources S and the offered traffic per idle source β. However, in practice using offered traffic ρ and peakedness Z are preferable for characterization of finite source traffic. Below the relations between these two parameters are introduced in order to express Z. The peakedness for Engset traffic is less than one (smooth traffic) and is:

1 Z = < 1, 1 + β

σ2 S · α · (1 − α) β Z = = , α = µ S · α 1 + β

Z = 1 − α, 153

A Z = 1 − , S

therefore

A = S · (1 − Z),

A S = . (1 − Z)

Negative Binomial Mathematically, we can deal with Pascal traffic with the same for- mulæ as for Engset by letting S and β be negative. The model has a parameter S for defining of sources, the number of which must be limited for this model, if i sources are busy, the arrival intensity equals (S + i)γ, the peakedness is Z > 1 (bursty traffic). The corresponding state transition diagram Figure B.3 is shown below.

Sγ (S + 1)γ (S + i)γ (S + N − 1)γ

0 1 i N − 1 N

µ 2µ iµ (N − 1)µ Nµ

Figure B.3.: State transition diagram for the Pascal case (truncated negative binomial)

For more detailed information see [37].

Appendix C. Convolution Algorithm

The convolution algorithm can be used to evaluate multi-dimensional Markov models. The main idea is to calculate each dimension separately and later to convolve one by one into the full system. Here follows the detailed structure of the algorithm:

• Step 1: Calculate the state probabilities of each trafﬁc stream as if it is alone in the system. We consider classical loss systems; for trafﬁc stream k:

Pk = pk(0), pk(1),..., pk(nk) (C.1) From importance are the relative values of pk(r), therefore set qk(0)= 1 in order

to ﬁnd the values of qk(r) relative to qk(0). To avoid numerical problems the relative state probabilities are normalized:

Nk qk(r) pk(r)= , Qk = qk(r), (C.2) Qk r=0 X

where Nk is the maximum allowed number of free channels for service-class k, k = 1, 2, . . . , n and r is the total number of busy channels.

• Step 2: The next step is to calculate the aggregated state probabilities for the total system excepting trafﬁc stream k by successive convolutions (convolution operator *):

Qn/k = P1 ∗ P2 ... ∗ Pk−1 ∗ Pk+1 ∗ ... ∗ Pn (C.3)

First convolve P1 and P2 and obtain P12, which is convolved with P3 and etc. The used formula is deﬁned as:

1 u

Pk ∗ Pj = pk(0) · pj(0), pk(ik) · pj(1 − ik),..., pk(ik) · pj(u − ik) , (C.4) i =0 i =0  Xk Xk    156 Appendix C. Convolution Algorithm

where

• Step 3:

Calculate performance measures like: time congestion Ek, call congestion Bk and

trafﬁc congestion Ck of stream k. The result from the convolution:

Qn = Qn/k ∗ Pk

is r r Q (r)= Q (r − i ) · p (i )= pk (r), (C.6) n n/k k k k ik i =0 i =0 Xk Xk where pk (r) is the global state probability of the mixed trafﬁc streams, where k ik denotes the trafﬁc stream, r is the total number of busy channels, and ik is the number of occupied channels by stream k.

In order to evaluate Ek, Bk and Ck the steps number 2 and 3 are repeated for every trafﬁc stream.

For more detailed information see [37]. Appendix D. Advanced Propagation Models

The COST 231-Walﬁsh-Ikegami-model (COST-WI) is a computationally fast empirical Prediction Model for Urban Scenarios. This empirical model is a combination of the models from J. Walﬁsch and F. Ikegami. It was further developed by the COST 231 project and allows the improved path-loss estimation by considering more data to describe the character of urban environments [14]. The model is extended for base station antenna heights well below the roof-top using an empirical function based on measurements. The model differentiates line-of-sight (LOS) and non-line-of-sight (NLOS). In the LOS case a simple propagation loss formula different from free space loss is applied. The loss is based on measurements, which have been performed in Stockholm:

d f L (dB)= 42.6 + 26 log + 20 log for d ≥ 20m (D.1) b km MHz where the ﬁrst constant is determined in such a way that Lb is equal to free-space loss for d = 20 m. In the NLOS-case the loss is composed of the free-space loss L0, multiple screen diffraction loss Lmsd, roof-top-to-street diffraction and scatter loss Lrts.

L0 + Lrts + Lmsd for Lrts + Lmsd > 0 Lb = (D.2) (L0 for Lrts + Lmsd ≤ 0

The free-space loss is given by

d f L (dB)= 32.4 + 20 log + 20 log (D.3) 0 km MHz

The term Lrts represents the coupling of the wave propagation along the multiple-screen path into the street where the mobile terminal is located.

w f ∆h L = −16.9 − 10 log + 10 log + 20 log mobile + L (D.4) rts m MHz m Ori 158 Appendix D. Advanced Propagation Models

where ∆hmobile = hRoof − hmobile and LOri is a street-orientation function:

−10 + 0.354 ϕ for 0◦ ≤ ϕ< 35◦ deg L = 2.5 + 0.075 ϕ − 35 for 35◦ ≤ ϕ< 55◦ (D.5) Ori deg 4.0 − 0.114 ϕ − 55 for 55◦ ≤ ϕ ≤ 90◦ deg  where ϕ is the angle between incidences coming from base station and road in de- grees.

The multi-screen diffraction loss Lmsd is an integral for which the Walﬁsch-Bertoni model [79] approximates a solution for the cases when the base station antenna height is greater than the average roof-top. COST 231 extended this solution to the cases base station antenna height is lower than the average rooftop by including empirical functions:

d f b L = L + k + k log + k log − 9 log (D.6) msd bsh a d km f MHz m with

−18 log(1 + ∆hBase ) for h > h m Base Roof Lbsh = (D.7) (0 for hBase ≤ hRoof

The term ka represents the increase of the path loss for the base station antennas

below the roof tops of the adjacent buildings, kd and kf control the dependence of the multi-screen diffraction loss versus distance and radio frequency, respectively.

54 for hBase > hRoof k = 54 − 0.8 ∆hBase for d ≥ 0.5 km and h ≤ h (D.8) a  m Base Roof 54 − 0.8 ∆hBase d for d < 0.5 km and h ≤ h  m 0.5km Base Roof 

18 for hBase > hRoof k = (D.9) d  ∆hBase 18 − 15 for hBase ≤ hRoof  hRoof  159

where ∆hBase = hBase − hRoof

0.7 f − 1 for medium sized city and suburban = − + 925 MHz centres with medium tree density kf 4  f (D.10) 1.5 − 1 for metropolitan centres  925 MHz  The COST-WI model is restricted to:

• Carrier frequency lies between 800 MHz and 2000 MHz

• Antenna height of the base station (BS) is in the range 3 . . . 50 m

• Antenna height of the MT is in the range 1 . . . 3 m

• Distance between the BS and MT is restricted to 0.02 . . . 5 km

List of Acronyms

ARQ automatic repeat request ATM Asynchronous Transfer Mode BBU basic bandwidth unit BCB Blocked-Call-Buffered BCC Blocked-Call-Cleared BCH Blocked-Call-Held BCI Blocked-Call-Interfered BPP Binomial-Poisson-Pascal trafﬁc BS base station CAC Call Admission Control CBR constant bit rate CTMC continuous time Markov chain CN Core Network FAG Full Availability Group FEC forward error correction FIFO First-In-First-Out FTP ﬁle transfer protocol GDS generation distribution strategy GoS Grade of Service HLR Home Location Register ITU International Telecommunication Union IP Internet Protocol M2M Mobile-to-Mobile MAI multiple access interference MLM Multiservice Loss Model MPEG Moving Picture Experts Group MPLS Multiprotocol Label Switching MT mobile terminal NC-MG network coding with multiple generations PSTN Public Switched Telephone Network RNC Radio Network Controller RRM Radio Resource Management QoS Quality of Service MSC Mobile Services Switching Center SGSN Serving GPRS Support Node TCP transmission control protocol TRRP Transmission Rate Reduction Policy UMTS Universal Mobile Telecommunications System UTRAN UMTS Terrestrial Radio Access Network 162 List of Acronyms

VBR variable bit rate WCDMA Wideband Code Division Multiple Access List of Symbols

Part I

λ number of users/calls arriving at the system per unit time (arrival rate) µ service rate τ mean holding time (service time) ρ offered trafﬁc N number of servers (channels) in system

Ns capacity of isolated cell (pole capacity) Nm cell capacity in multi-cell environment c capacity of the buffer (buffer length) S number of users in the system (active and passive) W sum of waiting time and service time E(·) expectation n number of service-classes ik(t) number of currently active users of service class k pi steady-state probability mk number of BBUs each service-class requires to establish one connection dk data rate of service-classes k Bk blocking probability of service-class k r number of busy BBUs in the system G normalization constant J threshold for data rate reduction b state dependent blocking probability (soft blocking)

Zk peakedness of service-class k Eb average received bit energy N0 power spectral density of white Gaussian noise α interference ratio

Pj received signal power from the jth user W chip rate

νj activity factor of the jth user Eb average received bit energy R j service bit rate of the jth user Itotal total received wideband power including thermal noise power L j load factor of one connection ηU L average received bit energy γk the arrival rate of an idle source Yk carried trafﬁc for stream k Ck trafﬁc congestion Esk probability of immediate service 164 List of Symbols

Edk probability of delayed service Ebk probability of blocked service

Part II

m the number of logical packets the source ﬁle is divided into w length of a logical packets in number of Bytes d distance between MT-MT pair

PTX transmit power Lij pathloss between MTs Λ random variable describing the shadowing process ⊕ modulo 2 addition A sending node

F1, F2 receiving nodes h maximum capacity of a directed graph F 2s finite field X decoding matrix or information matrix G encoding matrix x ′ encoded packet g vector of uniformly distributed random coefficients g′ global encoding vector u number of generations List of Figures

1.1. UMTS architecture [9]...... 10 1.2. OVSF code tree, SF denotes the Spreading Factor...... 11 1.3. Scrambling and Spreading in WCDMA, DPDCH denotes Dedicated Phys- icalDataChannel...... 12

2.1. A queuing system with arrival and departure of service requests...... 20 2.2. Birth-Deathprocess...... 20 2.3. Poissonarrivals ...... 21 2.4. M/M/1system...... 22 2.5. State diagram for Erlang’s loss system...... 24 2.6. Example of Blocked-Call-Cleared (BCC) model...... 26 2.7. An example of Full Availability Group (FAG): every inlet has access to every outlet with unlimited access to resources, n is the number of services and N denotesthenumberofservers...... 28

3.1. A fragment of a Markov chain for a two service system...... 35 3.2. Capacity (interference) sharing between cells in WCDMA...... 38 3.3. Density-histogram overplots and quantile-quantile plot for other-cell interference versus log-normal distribution for WCDMA system with low, mediumandhightrafﬁcload...... 40 3.4. Comparison of two state transition diagrams of a two-dimensional Markov chains constructed for WCDMA network with state-dependent blocking. . 42 3.5. Blocking probability and data throughput for speech and video stream versusthethreshold...... 45

4.1. Blockedcallmodels...... 51 4.2. State transition diagram for the Molina model...... 52 4.3. Model of the Molina loss system regarded as a waiting system with timeout. 52 4.4. Comparison of BCH and BCI models...... 53 4.5. Carried trafﬁc at connection/packet level as a function of the offered trafﬁc at connection/packet level, respectively. The system pole capacity is128channels...... 58 4.6. Analysis of the system performance with tightened admission control at connection level (70 % of the pole capacity, 90 channels)...... 60 4.7. Impact of activity factors on the system throughput...... 61

5.1. Illustration of the main difference between Blocked-Call-Held and Blocked- Call-Buffered...... 65 5.2. Impact of buffer scheme on the packet-level system performance...... 68 166 List of Figures

5.3. Performance comparison of system with and without soft blocking for heavy loaded other cells (left) and low loaded other cells (right)...... 68 5.4. Expected dwell time, probability of immediate service (left) and proba- bilityofdelayedservice(right)...... 69 5.5. MeanTransferQueueLength...... 70

6.1. Visions of the contending network solutions...... 77 6.2. The Intelligent Relaying technique...... 79 6.3. M2MConcept...... 82 6.4. Scatter plot of received signal power for m2m links as a function of pathloss and shadowing, constant transmit power of −44 dB...... 85 6.5. Changing radio propagation conditions for an arbitrarily chosen transmitter/receiver pair caused by multipath fading, pathloss and shadowing during one radio frame (resolution is 1 slot)...... 86 6.6. Signalling message sequence for setup of communication between m2m users...... 88 6.7. Received power for an arbitrary transmitter/receiver pair, during the interval between two consecutive updates of the link quality prediction to reshapegroups(updateinterval1frame)...... 90 6.8. Flow chart of m2m file sharing...... 93 6.9. Performance comparison of conventional and m2m data transfer mode in terms of file download time and reserved downlink capacity for WCDMA systemwithlow,mediumandhightrafficload...... 98 6.10.Instantaneous number of transmitting MTs and corresponding number of receiving multicast MTs, respectively (low and medium traffic load). . . 100 6.11.The average number of multicast receivers/group for different m2m traf- ficscenarios ...... 100 6.12.Instantaneous number of transmitting MTs and corresponding number of receiving multicast MTs, respectively (high traffic load)...... 100 6.13.Impact of extended group update interval on the system performance (hightrafficload)...... 101 6.14.Impact of policy change "Group members from same radio cell only" on thefiledownloadtime...... 102 6.15.Flow chart of m2m file dissemination for mixed traffic scenario...... 105 6.16.Impact of an additional uplink interference caused by speech traffic of various intensities on the file download time (high m2m traffic load). . . . 107 6.17.Impact of cross traffic on the M2M performance (medium m2m traffic load)...... 107 6.18.Impact of cross traffic on multicast efficiency (fragment)...... 108 6.19.Performance comparison of conventional and m2m data transfer mode in terms of download time, high m2m load...... 108 6.20.Performance comparison of the different file transfer strategies in terms of reserved downlink channels and multicast efficiency (UMTS conventional, most-utile m2m and random m2m algorithms, group size 7). . . . . 110 6.21.System Performance for different group sizes (high m2m traffic load). . . 111 6.22.Impact of group size on inter-group interference for high loaded UMTS system...... 112 List of Figures 167

6.23.Impact of restricted BS support on the performance of system with high m2m traffic load, file size 500 kbyte...... 114 6.24.Impact of restricted BS support on the performance of system for mixed traffic scenario with service differentiation (file size 40 kbyte, medium m2mload,cross-trafficscenario1)...... 115 6.25.Impact of restricted BS support on the performance of system for mixed traffic scenario with service differentiation (file size 40 kbyte, medium m2mload,cross-trafficscenario2)...... 116

7.1. Flow chart of m2m file sharing...... 122 7.2. Flow chart of m2m file sharing...... 123 F 7.3. NC-m2m Concept (for simplicity, 2 is used in the figure.) ...... 125 7.4. Performance comparison of simple replicate-and-forward m2m file transfer with NC-m2m for medium m2m traffic load, Scen.1 (left) and high m2m load, Scen.2 (right))...... 130 7.5. Impact of cooperative behavior of finished users on the file download times, medium m2m load, file size 40 kbyte, Scenario 2...... 131 7.6. Impact of users’ velocity on the download times, medium m2m load, file size 40 kbyte, Scenario 3))...... 131 7.7. Number of reserved downlink channels in medium loaded UMTS system inthesteadystateversustime...... 132 7.8. Impact of the NC-m2m technique on the uplink resource consumption, mediumtrafficload...... 133 7.9. The basic concept of handling generations for the network coding based m2m file dissemination scheme...... 136 7.10.Impact of generation size on the performance of NC-based m2m file dissemination in high loaded UMTS system...... 139 7.11.Performance comparison of NC-MG m2m algorithm in high loaded system for different GDSs and fixed generation size...... 140 7.12.Performance comparison of NC-MG m2m algorithm in high loaded system for different GDSs and fixed generation size...... 141 7.13.Performance comparison of large file distribution using NC-MG-m2m and simple random m2m algorithm in high loaded system (file size 500 kbyte).142 7.14.Performance comparison of large file distribution in high loaded system using NC-MG-m2m, simple random-m2m and conventional UMTS net- workmode...... 143

B.1. Each individual source is either idle or busy, and behaves independently ofallothersources...... 152 B.2. State transition diagram for the Engset case with S > N, where S is the number of sources and N is the number of channels...... 152 B.3. State transition diagram for the Pascal case (truncated negative binomial) 153

List of Tables

3.1. Mainsimulationparameters...... 39 3.2. Mainsimulationparameters...... 44 3.3. System performance with/without TRRP and reservation ...... 45

4.1. Service modeling parameters ...... 56 4.2. Performance results of the trafﬁc scenario from Table 4.1...... 57

5.1. Cell parameters and service trafﬁc description ...... 67 5.2. Time Average Performance Measures ...... 70

6.1. Mainsimulationparameters...... 95 6.2. Channel types and parameters as used in simulations ...... 97 6.3. Overall downlink throughput gain: Data volume in downlink in m2m mode for different traffic scenarios and released downlink capacity gain (foruserswithinonecell)...... 99 6.4. Impact of the restriction in the group organization policy on downlink throughput (per cell), service probability gain and number of multicast receiverspergroup...... 102 6.5. Impact of uplink interference for different cross/m2m traffic scenarios (erroneousdata%)...... 106 6.6. Overall downlink throughput gain: Data volume in downlink in m2m mode for different traffic scenarios and service probability gain (for users withinonecell) ...... 109 6.7. Impact of uplink interference for different group sizes and traffic scenarios(erroneousdata%)...... 111

7.1. Overall downlink throughput gain, Scenario 1...... 132 7.2. Mainsimulationparameters...... 139

Bibliography

[1] 3GPP Technical Speciﬁcation Group Radio Access Networks. Opportunity Driven Multiple Access. 3G TR 25.924.

[2] 3GPP Technical Speciﬁcation Group Radio Access Networks. UE Radio Transmis- sion and Reception, March 2008. TS 25.101 version 7.3.0.

[3] F. Adachi, M. Sawahashi, and K. Okawa. Tree-structured Generation of Orthogo- nal Spreading Codes with Different Lengths for Forward Link of DS-CDMA Mobile. Electronics Letters, 33:27–28, 1997.

[4] R. Ahlswede, N. Cai, S. Y. R. Li, and R. W. Yeung. Network information ﬂow. IEEE Trans. Information Theory, 46, July 2000.

[5] V.Benetis, L. Popova, and V.B. Iversen. Joint Connection and Packet Level Analysis in W-CDMA Radio Interface. In Proc. of the 3rd Workshop on Wireless and Mobility, Euro-NGI, Springer, LNCS 4396, Dec. 2006.

[6] G. Bolch, S. Greiner, H. de Meer, and K. S. Trivedi. Queuing Networks and Markov Chains. Wiley Interscience Publication, 2006.

[7] S. J. Bose. An Introduction to Queuing Systems. Kluwer/Plenum Publishers, 2002.

[8] M. Bossert and M. Breitbach. Digitale Netze. B. G. Teubner Stuttgart/Leipzig, 1999.

[9] G. Bostelmann and R. Zarits. UMTS Design Details & System Engineering. Tech- nical report, INACON GmbH, 2002.

[10] J. M. Capone and I. Stavrakakis. Delivering Diverse delay/dropping QoS Require- ments in a TDMA environment. In Proc. of IEEE ACM MobiCom, Budapest, Hun- gary, Sept. 1997.

[11] D. Cavalcanti, D. Agrawal, C. Cordeiro, B. Xie, and A. Kumar. Issues in integrating cellular networks, WLANs, and MANETs: a futuristic heterogeneous wireless network. IEEE Wireless Communications Magazine, 12, 2005.

[12] P. A. Chou, Y. Wu, and K. Jain. Practical network coding. In Proc. of Allerton Conference on Communications, Control and Computing, Monticello, IL, USA, Oct. 2003. 172 Bibliography

[13] B. Cohen. Incentives Build Robustness in BitTorrent. In Workshop on Economics of Peer-to-Peer Systems, Berkeley, CA, USA, May 2003.

[14] E. Damosso. Digital Mobile Radio towards future generation systems, COST Re- port 231. In Proc. of IEEE PIMRC, Taipei, Taiwan, Oct. 1996.

[15] S. Deb and M. Medard. Algebraic gossip: A network coding approach to optimal multiple rumor mongerling. In Proc. of 42nd Annual Allerton Conference on Communications, Control and Computing, Monticello, IL, USA, Oct. 2004.

[16] P.Fazekas, S. Imre, and M. Telek. Performance evaluation of multimedia services in cellular networks. Transactions of the Society for Computer Simulation Interna- tional, 78:268–277, Apr. 2002.

[17] A. Ferizi. Simulation Model for Performance Analysis of UMTS-Networks with variable Data Rate. Technical report, Student Research Project-2006-03, Institute of Mobile Communication, Friedrich-Alexander-University, Erlangen-Nuremberg, 2006.

[18] N. R. Figueira and J. Pasquale. Providing QoS for Wireless Links: Wireless/Wired Networks. IEEE Pers. Communications, 6(5):42–51, Oct. 1998.

[19] F.H. P.Fitzek and M. D. Katz, editors. Cooperation in Wireless Networks: Principles and Applications. Springer Verlag, 2006.

[20] C. Fragouli, J. Y. L. Boudec, and J. Widmer. Network Coding: An Instant Primer. Technical report, LCA-Report-2005-010, 2005.

[21] C. Fragouli and E. Soljanin. Information Flow Decomposition for Network Coding. IEEE Trans. of Information Theory, 52, Mar. 2006.

[22] C. Fragouli and E. Soljanin. Network Coding Applications. Now Publishers Inc, January 2008.

[23] T. Fry. Probability and Its Engineering Use, page 462. Princeton, New York, 1928.

[24] C. Gkantsidis and P.R. Rodriguez. Network Coding for Large Scale Content Distri- bution. In Proc. of IEEE INFOCOM, Miami, FL, USA, 2005.

[25] M. Glabowski and A. Kaliszan. Simulator of Full-Availability Group with Band- width Reservation and Multi-Rate Bernoulli-Poisson-Pascal Trafﬁc Streams . In Proc. International Conference on Computer as a Tool, EUROCON 2007, pages 2271–2277, Warsaw, Poland, Sept. 2007.

[26] H. Haas and G. J. R. Povey. A Capacity Investigation on UTRA-TDD Utilising Underused UTRA-FDD Uplink Resources. In IEE Colloquium on UMTS Terminals and Software Radio, Glasgow, UK, Apr. 1999. Bibliography 173

[27] T. J. Harrold and A. R. Nix. Performance Analysis of Intelligent Relaying in UTRA TDD. In Proc. of IEEE VTC Fall, Vancouver, BC, Canada, Sept. 2002.

[28] N. Hassanein, A. Oliver, N. Nasser, and E. Elmallah. Uplink QoS-Aware Admission Control in WCDMA Networks with Class-Based Sharing. In Proc. of QSHINE, Oct. 2004.

[29] T. Ho, M. Medard, R. Koetter, D. R. Karger, M. Effros, J. Shi, and B. Leong. A random linear network coding approach to multicast. IEEE Trans. of Information Theory, 52, Oct. 2006.

[30] H. Holma and A. Toskala. WCDMA for UMTS. John Wiley & Sons, Ltd, 2000.

[31] T. Hossfeld, K. Tutschku, and F. Andersen. Mapping of File-Sharing onto Mobile Environments: Feasibility and Performance of eDonkey with GPRS. In Proc. of IEEE WCNC, New Orleans, USA, Mar. 2005.

[32] T. Hossfeld, K. Tutschku, and F. Andersen. Mapping of Filesharing onto Mobile Environments: Enhancement by UMTS. In Proc. of IEEE Pervasive Computing and Communications (PerCom), Kauai Island, Hawaii, Mar. 2005.

[33] T. Hossfeld, K. Tutschku, F. Andersen, H. de Meer, and J. Oberender. Simulative Performance Evaluation of a Mobile Peer-to-Peer File-Sharing System. In Proc. of IEEE Next Generation Internet Networks (NGI), Rome, Italy, Apr. 2005.

[34] L. Huang and C.-C. J. Kuo. Joint Connection-Level and Packet-Level Quality-of- Service Support for VBR Trafﬁc in Wireless Multimedia Networks. IEEE Journal on Selected Areas in Communications, 23(6), June 2005.

[35] ITU-T Recommendation. Theory for Full Availability Group, Delay System, 1982. Tetrapro, edited by H. Leijon, ITU, 2008.

[36] V.B. Iversen. Modelling Restricted Accessibility for Wireless Multi-Service System. In Proc. of the 2rd Workshop on Wireless and Mobility, Euro-NGI, Springer, July 2005.

[37] V. B. Iversen. Teletrafﬁc Engineering Handbook, ITU-D Study Group 2, Question 16/2. http://www.com.dtu.dk/teletraffic, 2005.

[38] V.B. Iversen. Performance measures for multi-rate loss systems with limited accessibility and their evaluation. In Proc. of the 3rd Workshop on Wireless and Mobility, Euro-NGI, Springer, Apr. 2006.

[39] V.B. Iversen, V.Benetis, N. Ha, and S. Stepanov. Evaluation of Multi-service CDMA Networks with Soft Blocking. In Proc. of ITC 16, Spec. Sem., Antwerp. Belgium, Sept. 2004. 174 Bibliography

[40] C. Jedrzycki and C. M. Leung. Probability distribution of channel holding time in cellular telephony system. In Proc. of IEEE 44th VTC, Sweden, June 1994.

[41] S. Katti, D. Katabi, W. H. H. Rahul, and M. Medard. The importance of being opportunistic: Practical network coding for wireless environments. In Proc. of 43rd Annual Allerton Conference on Communication, Control and Computing, Mon- ticello, IL, USA, Sept. 2005.

[42] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Medard, and J. Crowcroft. Xors in the air: Practical wireless network coding. In Proc. of Conference on Applications, technologies, architectures, and protocols for computer communications, (SIGCOM), New York, NY, USA, 2006.

[43] J. Kaufman. Blocking in a Shared Resource Environment. IEEE Trans.Commun., 29(10):1474–1481, 1981.

[44] D. G. Kendall. Stochastic processes occurring in the theory of queues and their analysis by the method of the imbedded markov chain. The Annals of Mathemati- cal Statistics, 24(3):338–354, 1953.

[45] R. Koetter and M. Medard. Beyond Routing: An Algebraic Approach to Network Coding. In Proc. of IEEE INFOCOM, New York, USA, 2002.

[46] J. Laiho, A. Wacker, and T. Novosad. Radio Network Planning and Optimization for UMTS, Nokia. Wiley Interscience Publication, 2002.

[47] P. E. Lassila and J. T. Virtamo. Nearly optimal importance sampling for Monte Carlo simulation of loss systems. ACM Trans. on Modeling and Computer Simula- tion, 4:326–347, Oct. 2000.

[48] W. C. Lee. Mobile Cellular Telecommunications System. McGraw Hill International Editions, 1989.

[49] Y. R. Li, R. W. Yeung, and N. Cai. Linear Network Coding. IEEE Trans. of Informa- tion Theory, 46, July 2000.

[50] J. Lloret, J. R. Diaz, J. M. Jiménez, and M. Esteve. Public Domain P2P File Sharing Networks Content and Their Evolution. In Proc. of IEEE Communication Systems and Networks (CSN), Benidorm, Spain, Sept. 2005.

[51] H. Luo, R. Ramjee, P. Sinha, L. Li, and S. Lu. UCAN: A Uniﬁed Cellular and Ad- Hoc Network Architecture. In Proc. of ACM MOBICOM, San Diego, California, USA, Sept. 2003.

[52] I. Moscholios, M. Logothetis, and G. Kokkinakis. Connection-dependent threshold model. Performance Evaluation, Elsevier Publisher, 48, 2002. Bibliography 175

[53] I. D. Moscholios and M. Logothetis. Call-Level Blocking of ON-OFF Trafﬁc Sources in a Shared Resource Environment with Batched Poisson Arrival Processes. In Proc. International Teletrafﬁc Congress, ITC-19, pages 143–152, Beijing, China, September 2005.

[54] M. Naghshineh and S. Schwartz. Distributed Call Admission Control in Mo- bile/Wireless Systems. IEEE Journal on Selected Areas in Communications, 14(4):711–717, May 1996.

[55] J. M. Peha and A. Sutivong. Admission Control Algorithms for Cellular Systems. Wireless Networks, 7(2):117–125, Apr. 2001.

[56] L. Popova, T. Herpel, W. Gerstacker, and W. Koch. Cooperative Mobile-to-Mobile File Dissemination in Cellular Networks within a Uniﬁed Radio Interface. Regular Issue of Computer Networks Journal, Elsevier Publisher, 2008.

[57] L. Popova, T. Herpel, and W. Koch. Efﬁciency and Dependability of Direct Mobile- to-Mobile Data Transfer for UMTS Downlink in Multi-Service Networks. In Proc. of IEEE Wireless Communications & Networking Conference (WCNC), Hong Kong, Mar. 2007.

[58] L. Popova, T. Herpel, and W. Koch. Enhanced Downlink Capacity in UMTS supported by Mobile-to-Mobile Data Transfer. In Proc. of 6th International IFIP-TC6 Networking Conference, Springer, LNCS 4479, Atlanta, GA, USA, May 2007.

[59] L. Popova, T. Herpel, and W. Koch. Improving Downlink UMTS Capacity by Ex- ploiting Direct Mobile-to-Mobile Data Transfer. In Proc. of the 5th International Symposium on Modelling and Optimization in Mobile, Ad Hoc, and Wireless Net- works (OPNET), Limassol, Cyprus, Apr. 2007.

[60] L. Popova and V. B. Iversen. Analysis of Two-Layer Performance Models by Using Generalized Approaches from Extended Teletrafﬁc Theory. In 4rd Workshop on Wireless and Mobility, Euro-NGI, Jan. 2008.

[61] L. Popova, A. Schmidt, W. Gerstacker, and W. Koch. Network Coding Assisted Mobile-to-Mobile File Transfer. In Proc. of IEEE Australasian Telecommunication Networks and Applications Conference (ATNAC), Christchurch, New Zealand, Dec. 2007.

[62] L. N. Popova, W. H. Gerstacker, and W. Koch. Efﬁcient Distribution of Large Files in UMTS supported by Network Coded M2M Data Transfer with Multiple Gener- ations. In Proc. of the 1st International Conference on Ad Hoc Networks, Springer, LNICST 28, Niagara Falls, Ontario, Canada, Sept. 2009.

[63] L. N. Popova and W. Koch. Analytical Performance Evaluation of Mixed Services with Variable Data Rates for the Uplink of UMTS. In Proc. of IEEE VTC Spring, Bermingham, Sept. 2006. 176 Bibliography

[64] R. Ramjee, D. Tomwley, and R. Nagarajan. On Optimal Call Admission Control in Cellular Systems. Wireless Networks, 3(1):29–41, May 1997.

[65] J. W.Robert. A service system with heterogeneous user requirements - application to multi-service telecommunication systems. In Proc. of Performance of Data Com- munication Systems and their Applications, Amsterdam, The Netherlands: North- Holland, 1981.

[66] J. W. Robert. Teletrafﬁc models for the Telcom 1 integrated services network. In Proc. of ITC 10, Monreal, Canada, Sept. 1983.

[67] P. Sander, S. Egner, and L. Tolhuizen. Polynomial time algorithms for network informaton ﬂow. In Proc. of Symp. Parallel Algorithms and Archtectures, (ACM), San Diego, CA, June 2003.

[68] D. Staehle, K. Leibnitz, K. Heck, B. Schroeder, A. Weller, and P. Tran-Gia. Ap- proximating the Othercell Interference Distribution in Inhomogeneous UMTS Net- works. In Proc. of IEEE ISWCS, Valencia, Spain, May 2002.

[69] D. Staehle and A. Maeder. An analytic approximation of the uplink capacity in a UMTS network with heterogeneous trafﬁc. In Proc. of ITC 18, Berlin, Sept. 2003.

[70] M. Stasiak. An approximate model of a switching network carrying mixture of different multichannel trafﬁc stream. IEEE Trans. on Communications, 41(6):836– 840, 1993.

[71] M. Stasiak. Blocking probability in a limited-availability group carrying mixture of different multichannel trafﬁc stream. Annales des Telecommunications, 48(1- 2):71–76, 1993.

[72] M. Stasiak, A. Wisniewski, and P.Zwierzykowski. Blocking probability calculation in the uplink direction for cellular systems with WCDMA radio interface. In Proc. of 12 GI/ITG Conference on Measuring, Modelling and Evaluation of Computer and Communication Systems (MMB), Dresden, Sept. 2004.

[73] M. Stasiak and P. Zwierzykowski. Analytical model of ATM node with multicast switching. In Proc. th Mediterranean Electrotechnical Conference MELECON 98, volume 2, pages 683–687, 18–20 May 1998.

[74] S. N. Stepanov and V.S. Lagutin. Teletrafﬁc of Multiservice Systems. Radio & Svjas, Moscow, 2000.

[75] G. Swedberg. Ericsson’s mobile location solution. Ericsson Review, 4, 1999.

[76] D. Tacconi, C. Saraydar, and S. Tekinay. Ad Hoc Enhanced Routing in UMTS for Increased Packet Delivery Rates. In Proc. of IEEE WCNC, New Orleans, USA, Mar. 2005. Bibliography 177

[77] A. M. Viterbi and A. J. Viterbi. Erlang capacity of a power controlled CDMA system. IEEE Selected Areas in Commun., 11(6):892–893, Aug. 1993.

[78] A. M. Viterbi, A. J. Viterbi, and E. Zehavi. Other-cell interference in cellular power- controlled CDMA. IEEE Trans.Commun., 42(2):1501–1504, Feb. 1994.

[79] J. Walﬁsch and H. L. Bertoni. A Theoretical model of UHF propagation in urban environments. IEEE Trans. on Antennas and Propagation, 36:1788–1796, 1988.

[80] Z. Wang, E. Tameh, and A. Nix. Statistical Peer-to-Peer Channel Models for Out- door Urban Environments at 2GHz and 5GHz. In Proc. of IEEE VTC Fall, page 5101, Los Angeles, USA, Sept. 2004.

[81] J. Widmer and J. Y. L. Boudec. Network coding for efﬁcient communication in extreme networks. In Proc. of ACM SIGCOMM Workshop on Delay-tolerant networking, pages 284–291, Philadelphia, Pennsylvania, USA, Aug. 2005.

[82] H. Wu, C. Qiao, S. De, and O. Tonguz. Integrated Cellular and Ad Hoc Relaying System: iCAR. IEEE Selected Areas in Commun., 19(10):2105–2115, Oct. 2001.

[83] S. Wu, K. Y. M. Wong, and B. Li. A New Distributed and Dynamic Call Admission Policy for Mobile Wireless Networks with QoS Guarantee. In Proc. of IEEE PIMRC, Boston, MA, Sept. 1998.

[84] Y. Wu, P. A. Chou, and S. Y. Kung. Information Exchange in Wireless Networks with Network Coding and Physical-layer Broadcast. Technical report, MSR-TR- 2004-78, 2004.