QUALITY OF SERVICE MANAGEMENT TECHNIQUES FOR IP BASED HETEROGENEOUS … · 2014-10-29 · universita degli studi di pavia` facolta di ingegneria` tesi di dottorato di ricerca matteo

UNIVERSITA DEGLI STUDI DI PAVIAFACOLTA DI INGEGNERIA

TESI DI DOTTORATO DI RICERCAMatteo Lanati

XX CICLO

QUALITY OF SERVICE MANAGEMENTTECHNIQUES FOR IP BASED

HETEROGENEOUS NETWORKS

Tutor:Chiar.mo Prof. Lorenzo Favalli

Ai miei genitori.

Contents

Introduction 1

1 Topics on Quality of Service 41.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2 Protocol behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Applications performance analysis . . . . . . . . . . . . . . . . . . . . . . . 71.4 Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.5 Resource allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Protocol behaviour 142.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2 Transmission Control Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2.1 TCP Tahoe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2.2 TCP Reno . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.2.3 TCP New Reno . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.2.4 TCP Selective Acknowledgements . . . . . . . . . . . . . . . . . . . 18

2.3 UMTS overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.3.1 UTRAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.3.2 Core Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.4 Simulation environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.5 Simulation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.5.1 Hidden Markov Model . . . . . . . . . . . . . . . . . . . . . . . . . 282.5.2 UMTS TDD module . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.6 Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

iii

Contents

3 Application awareness and performance assurance 393.1 The video coding case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.2 Video coding principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.3 Buffer management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.4 Simulation scenarios and results . . . . . . . . . . . . . . . . . . . . . . . . 423.5 More general approach using DiffServ . . . . . . . . . . . . . . . . . . . . . 443.6 DiffServ architecture principles . . . . . . . . . . . . . . . . . . . . . . . . . 443.7 Interactions between DiffServ and UMTS . . . . . . . . . . . . . . . . . . . 463.8 Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

4 Impact of resource allocation on QoS provisioning 514.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514.2 Granularity and efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534.3 Diversity and adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5 Efficient retransmission resource allocation: Hybrid ARQ case 565.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565.2 UMTS Long Term Evolution: system overview . . . . . . . . . . . . . . . . 57

5.2.1 Physical layer and OFDM . . . . . . . . . . . . . . . . . . . . . . . . 575.2.2 MAC layer and Hybrid ARQ . . . . . . . . . . . . . . . . . . . . . . 60

5.3 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625.4 Simulations and results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6 Example of a complete resource allocation framework 716.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716.2 Technology overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716.3 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 766.4 Allocation algorithm performances evaluation . . . . . . . . . . . . . . . . 826.5 Impact of Frequency Reuse Factor . . . . . . . . . . . . . . . . . . . . . . . 846.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Conclusions 88

Bibliography 94

List of Figures 100

List of Tables 101

iv

List of Acronyms

2G 2nd Generation cellular system3G 3rd Generation cellular system3GPP 3rd Generation Partnership Project4G 4th Generation cellular systemAAL ATM Adaptation LayerACK AcknowledgementACM Adaptive Coding and ModulationARQ Automatic Repeat reQuestATM Asynchronous Transfer ModeBER Bit Error RateBTS Base Transceiver StationCAC Call Admission ControlCBR Constant Bit RateCC Chase CombiningCDMA Code Division Multiple AccessCID Connection IDentifierCS Circuit SwitchedDiffServ Differentiated ServicesFC Fountain CodesFDD Frequency Division DuplexFDMA Frequency Division Multiple AccessFEC Forward Error CorrectionFRF Frequency Reuse Factor

v

List of Acronyms

FTP File Transfer ProtocolGPRS General Packet Radio ServiceGSM Global System for Mobile communicationsHARQ Hybrid ARQHSDPA High Speed Downlink Packet AccesIETF Internet Engineering Task ForceIntServ Integrated ServicesIP Internet ProtocolIR Incremental RedundancyISO International Standard OrganisationLOS Line of SightMAC Medium Access ControlMIP Mobile IPMPLS Multi Protocol Label SwitchingMUD Multi User DiversityNLOS Non-Line of SightNRT Non Real TimeOFDM Orthogonal Frequency Division MultiplexOFDMA Orthogonal Frequency Division Multiple AccessOSI Open Standard InitiativePDU Packet Data UnitPS Packet SwitchedQoS Quality of ServiceRAT Radio Access TechnologyRLC Radio Link ControlRNC Radio Network ControllerRNS Radio Network SubsystemRRC Radio Resource ControlRSS Received Signal StrengthRSVP Resource reSerVation ProtocolRT Real TimeRTP Real Time ProtocolRTT Round Trip TimeSNR Signal to Noise RatioSOFDMA Slotted Orthogonal Frequency Division Multiple Access

vi

List of Acronyms

SR Selective RepeatSSCS Service Specific Convergence SublayerTOS Type Of ServiceTCP Transmission Control ProtocolTDD Time Division DuplexTDMA Time Division Multiple AccessUDP User Datagram ProtocolUMTS Universal Mobile Telephone SystemUTRAN UMTS Terrestrial Radio Access NetworkVoIP Voice over IPWLAN Wireless Local Area Network

vii

Introduction

Speed up of data connection is the most noticeable aspect of technology evolution indata oriented networks and, at the same time, it is also the most promoted feature by

service providers. Users are often requesting more bandwidth to run simultaneously anincreasing number of applications, but the simple approach consisting in enlargementof the infrastructure is not always the right solution. Referring to a wired environment,it is known that telecommunication companies can rely on the so called “dark fibers”,that is unused cables already deployed, preserved for redundancy purpose and part ofthe strategic business plan. If we think to the wireless case, where there are fixed legalboundaries, usually an authorisation is needed for transmission and the radio channelis unique for both customers and operators, installation of new network components iseven more difficult.Another significant trend on which we can reflect is the change of role played by packetnetworks, paying particular attention to the widest example, Internet. Packet switch-ing was born to provide a flexible and reconfigurable framework for communicationsbetween computers thanks to a layered, adaptable and open structure. Nowadays, dueto these peculiarities, Internet is the hearth of services with features completely differ-ent from the original project, being in use for audio-video content distribution, voicetransport, information broadcasting, ...

The examples just mentioned stress that the need for efficiency in resource manage-ment and the demand for service differentiation and performance assurance are twoof the reasons leading to the introduction of the concept of Quality of Service (QoS). Itis quite difficult to give an exact and universally accepted definition because there are

Introduction 2

multiple aspects to take into account, moreover QoS is involved in very heterogeneousscenarios. However, in both the situations described, we want that users are satisfiedwith their application from the qualitative point of view, but, for this purpose, we haveto interact with the system in a quantitative way. The customer is more concerned withthe qualitative experience he can obtain, while a provider is interested in identifyingfew quantitative parameters strongly charaterising the services delivered and that canbe handled to differentiate the offer.

In our work we deal with QoS in a mixed wired and wireless scenario because wethink that convergence is the challenge for next years. QoS is implicit in circuit switch-ing networks, such as Public Switched Telephone Network (PSTN) or first generationcellular network, because all the resources needed for a communication are reservedfor the entire duration of the conversation. An important step in this research field wasthe introduction of ISDN (Integrated Services Digital Network) due to the possibility ofusing at the same time voice and data services. Researcher and operators started to con-sider a new problem, that is how to manage various types of traffic, with different needsin term of error rate and latency. Contributions given in this field were then extendedto the wireless domain when second generation cellular network was designed. In fact,ISDN solutions found a large application and some of its protocols were adapted towork in the new environment. Despite this wireless technology was strongly orientedto voice communications, a packet switched domain was added and the criteria usedfor QoS classes definition are still now a reference point. Third generation inherited thisstructure increasing the variety of services and mixing together Internet, broadband (or,at least, not too narrow) access and mobility. In fact, the recently released radio interfaceaffords a data rate very near to those offered by a wired technology. Despite there is alot of work to do, in particular to reach a deeper integration, standardisation entities,manufacturers and researcher are already working on advanced third and fourth gen-eration. High bit rate in both downlink and uplink directions, strong IP orientation andtight interoperability for a seamless Internet experience are the goals of this workinggroup.

We tried to recognise and to analyse main points of this process, starting from trans-port protocols and applications. They occupy the upper part of the layered protocolmodel, have an end-to-end validity, from the server to the mobile station and vicev-ersa, and consequently represent the common part between wired and wireless domain.Then we considered the impact of resource allocation because it is not possible to dis-

Introduction 3

regard all the issues related to the radio channel. The bandwidth is scarce and linkconditions vary quickly, especially in relation with user’s mobility. We did not limit ourcontribution to allocation algorithms based on known approach like water filling, butwe went deeper, focusing on new techniques like hybrid ARQ error correction.

Chapter 1

Topics on Quality of Service

1.1 Introduction

The term Quality of Service (QoS) is referred to service performance and satisfactionlevel experienced by the user. On the other side, from the provider point of view,

QoS identifies a certain number of techniques to manage available resources to achieveoptimal throughput, reduce congestion effects and satisfy application’s needs. Basic ser-vice offered by datagram networks is “best effort”, that is, nodes try to deliver error-freeand ordered packets, but without guarantees. This paradigm was enough for the initialpurpose of building a decentralised and redundant network for military use. QoS wasborn later, when Internet became a wide used media and Real Time (RT) applicationsstarted to run on the network. The parameters affecting QoS may also differ, depend-ing on the type of application so, for example, for what multimedia content delivery isconcerned, it is more important to keep delay under control rather than obtain an in-credibly low and constant error rate. If a speech or video frame is corrupted, a smartdecoder can try to recover it, but if the delay is too high, the information is completelyunuseful.

Packets flowing in a network can encounter various problems, such as:

• packets can be dropped, because routers’ queues are full, that is there is conges-tion;

• packets can be delayed because of congestion or routing strategies (a change inthe path to avoid congested links);

• packets are out-of-order, because they followed different paths;

• packets contain errors.

1.2. Protocol behaviour 5

Due to the nature of the applications that were popular when Internet was born, one ofthe main problems to be faced was the guarantee of error-free packet delivery, eventu-ally avoiding traffic congestion. These requirements led to the definition of Transmis-sion Control Protocol (TCP), characterised by the presence of tools for error correctionand rate control. At the same time, User Datagram Protocol (UDP) was developed withthe goal to provide only a simple multiplexing and demultiplexing service. Thanks tothis feature, it was immediately clear that UDP was suitable for delay-intolerant ser-vices. However, multimedia streams need source and synchronisation management, sodedicated headers were introduced for this purpose with the Real Time Protocol (RTP).

In order to satisfy bandwidth requests, a qualitative definition of “applications be-having well” is no more usefull, on the contrary, monitoring of some quantities such asthroughput, delay, maximum tolerable burst is needed. For all of these reasons, onlyworking on the protocol stack may not lead to encouraging results, so IETF (InternetEngineering Task Force) started to think to new architectures. Recently this scenario be-came more complicated because the integration between Internet and cellular networksis taking place. This attempt focused limitations of the two systems. Wired networkscan rely on a large amount of bandwidth, while telephone systems have a huge poten-tial market and a strong QoS infrastructure, but optimal radio channel exploitation isstill in development

We think that an analysis taking into account various aspects can be useful, in par-ticular we have individuated four distinguishing points for end to end QoS:

• Protocol behaviour

• Performance analysis of applications to be run

• Mobility

• Resource allocation.

Each item will be briefly explained in next sections.

1.2 Protocol behaviour

TCP/IP protocol stack gained a relevant role in providing connectivity and a wide rangeof communication services, thank to flexibility and capability of supporting variousphysical networks. This trend is consolidated in traditional telephone network, while in

1.2. Protocol behaviour 6

mobile radio networks, IP support was taken into account at standardisation time. Forexample, GSM (Global System for Mobile communications), the 2nd generation packetswitched cellular system, was made IP aware introducing GPRS (General Packet RadioService), that is, new nodes providing access to Internet. 3GPP UMTS (3rd GenerationPartnership Project Universal Mobile Telephone System) planned a transitional archi-tecture [1][2], called UMTS 2000 Release 4, to maintain compatibility with GSM/GPRS.In fact the structure is the same and it is made up of a Circuit Switching (CS) domainand a Packet Switching (PS) domain. The first domain has to support voice services,the second one is dedicated to data services. However, UMTS 2000 Release 5 [3] is al-ready defined, adding an IP multimedia subsystem on top of PS domain, in order tocomplete the transition towards an all-IP network. Systems beyond 3G do not supporta CS domain any more, all services are delivered by IP protocol.

As we can see, telecommunications are moving to a unique platform for delivery ofhigh speed data and multimedia services, however IP cannot tackle two fundamentalissues of wireless networks: mobility and variable radio channel conditions. Mobilitywill be dealt in next section, while here an introduction to error correction will be given.

IETF already defined TCP (Transmission Control Protocol), a transport layer proto-col responsible for error free communication between end to end entities. Since its firstappearance, many features were improved:

• Selective Repeat (SR) scheme substituted Go-back-N ARQ (Automatic Repeat re-Quest) strategy;

• retransmission timer was refined, introducing better estimation;

• enhanced mechanisms to detect packet loss were added.

However, TCP was designed for wired networks, where congestion is the most frequentproblem and the physical medium employed (copper, optical fibber, ...) has a low prob-ability of error. Radio channel behaves in the opposite way, that is, data loss is usuallydue to change in propagation conditions. These considerations do not encourage torelay on TCP to guarantee correct data reception in a wireless environment. The ma-jority of packets are lost over the air interface between base station and mobile station,so waiting for a retransmission from the source leads to a waste of time and resources.Waste of time is due to the fact that TCP is not optimised for a wireless environment,despite some algorithms are under development, and the reaction is not immediate. Itis possible that a time out is triggered, yielding a bit rate reduction. In other words,

1.3. Applications performance analysis 7

it could be difficult to distinguish congestion and a burst error resulting from fading.Moreover, the mobile user could be in communication with an Internet node (i.e. awebsite, a mail server, a streaming server ...), so a retransmission from the source is notefficient, involving various nodes and wasting intermediate resources.

The key point is to note that radio channel error is a local problem, limited to theair link between base station and mobile user, so the best solution should be local too.In fact, 3GPP inserted in wireless protocol stack an RLC (Radio Link Control) sublayerto manage local retransmissions. RLC entity has to segment a packet in PDUs (PacketData Unit), data blocks of smaller size to maximise efficiency in channel use, and thenapply a SR retransmission scheme. Retransmissions, however, are not always the bestsolution: the number of attempts could be high, delaying data delivery. For this reasonit is important to consider FEC (Forward Error Correction) coding techniques appliedat the physical layer. These problems will be explained in Chapter 5.

1.3 Applications performance analysis

With 3G cellular system, multimedia contents broke into nomadic user experience thankto a great variety of applications spanning from interactive services (conversational, pre-recorded or live streaming, messaging) to distributive ones, with different constraintson network and end-to-end performances. Conversational services are those which im-pose the strictest transmission requirements, in order to guarantee a continuous and”real-time” perception of a telephone call or video sequence: the strongest constraintsconcern the maximum acceptable end-to-end delay and delay jitter.

Theoretically, such constraints could be fulfilled using a CBR (Constant Bit Rate)bearer service and a CBR encoding, however this is difficult to carry out in practice.CBR encoding itself is utopian, especially if the source is changing. Moreover, the radiochannel imposes variable and harsh transmission conditions. From the network side,ARQ and smart buffering can help to reduce the error rate seen by an application, whilea resilient encoding technique can recover residual error on the basis of video-level in-formations, eventually exploiting benefits of a cross layer approach.

1.4. Mobility 8

1.4 Mobility

Cellular networks, in particular 3G or more advanced systems, and local wireless net-works (WLANs) are two access technologies with different features. The formers havea clear QoS framework, usually operating in a nationwide context and using a licensedspectrum. The latters can provide large bandwidth but in a restricted area, usually re-ferred to as “hot spot”, guaranteeing a fast and simple deployment because operatingfrequencies are in an unlicensed band. These technologies have been developed al-most in the same period, so many operators believe them to be competitors in the samemarket. This is partially true, but some efforts have been made towards cooperation,in order to provide best access opportunities to mobile users. In particular, recent de-velopments [4] in both 3GPP UMTS standardisation process and the IEEE 802 group,consider the interworking procedures between IEEE 802.11 WLAN and third genera-tion mobile networks. The objective is combining the complementary aspects of thetwo different technologies, leading to development of recongurable multi-RAT (RadioAccess Technology) terminals.

Two reference scenarios are possible [5]:

• an improved WLAN with interworking capabilities interacts with 3GPP systemsproviding enhanced services to legacy single-mode terminals;

• a multimodal user equipment can handle all interworking services through 3GPPsystem or interworked WLAN.

In the first approach, design efforts are concentrated on the network side, assuring max-imum compatibility. In the second one, interoperability can be accomplished either us-ing independent interfaces or reconfigurable hardware (see, for example [6]), involvingtechnologies such as vertical handover and Mobile IP (MIP) to ensure service continuity.

Cellular systems, starting from their first appearance, support handover operation,that is the capability for a mobile terminal to switch from a cell to an adjacent one ina transparent way for the user. Usually, the first implementation is called “hard han-dover”, in the sense that all the radio links with the original cell are removed, withoutcompromising service provisioning. UMTS introduced the concept of “soft handover”,in the sense that a user equipment can maintain service connection with more than onebase station. Vertical handover further develops this concept, involving upper layers:in a certain sense applications are notified that an handover will take place. To per-form these operations a certain amount of time is needed, so it is not convenient to start

1.4. Mobility 9

handover when users are at the cell edge. Since user’s position is quite difficult to cal-culate, the natural parameter to monitor is received signal strength (ROSS). A powerthreshold, higher than minimum receivable power, is needed to start routing opera-tions and authorisation requests. While performing association with new Access Pointor Base station, the mobile equipment can still receive packets from the serving node, asin a soft handover scenario. Summarising, vertical handover between UMTS and IEEE802.11 has three main features

• Base Station and Access Point covering areas should partially overlap: this is fun-damental for service continuity;

• the two technologies support different data-rates, so an adaptation strategy inneeded together with different handover priorities (i.e.from WLAN to 3G shouldbe higher);

• different thresholds are needed because two technologies have different level ofpower transmission (due to different coverage area extension).

It is important to set a proper value for vertical handover threshold because real timeapplications have strict temporal constraints. For example, the maximum transfer delaysuggested [6] for UMTS Real Time services varies between 20 ms and 300 ms whileprocessing time in downlink or uplink direction are heavily implementation dependent.

As we said, vertical handover involves also upper layers, e.g. network layer, inmanagement of mobility problem, so MIP can be successfully applied. Flexibility ofthe IP protocol is well known, but in an inter-system handover the user is changingaccess technology, and this fact should not influence data packets delivery. MIP allowa mobile host to change serving node, keeping the same IP address. For this purpose,both network and host IP agent should be modified. As we can see in Fig 1.1 there arethree elements: home agent, foreign agent and mobile host. Mobile host is the nomadicdevice, home agent belongs to home network (the network where mobile host usuallycan be found), while foreign agent belongs to foreign network (the network from whichmobile host is accessing Internet), and, finally, corresponding node is a node exchangingdata with the mobile host. So, an home agent allows home network’s nodes to moveaway, while a foreign agent allows mobile hosts to visit a foreign network.

The mobile host has two addresses: the home address, always identifying this partic-ular node, and a care-of-address to exchange data when the mobile host is in a foreign

1.5. Resource allocation 10

Figure 1.1: Mobile IP.

network. This address can be the foreign agent’s address or it can be temporarly as-signed to the mobile host. Assuming that the mobile host starts from his home network,a corresponding node can reach the mobile node through the home agent. If the mobilehost moves towards the foreign network, a care-of-address is assigned by the foreignagent by means of agent advertisement messages. The mobile node informs his homeagent that it can be reached through the foreign agent, so that communication with thecorresponding node is kept alive. In particular, if the corresponding node wants to senda packet to the mobile host, it keeps sending data to the home agent. Packets are thenforwarded to the foreign agent and, finally they can reach the mobile host. In the op-posite direction, from mobile host to corresponding node, communications are easier,because only the foreign agent is involved, acting as an ordinary router. Usually, a tun-nel is established between home and foreign agent, that is packets flowing along thisdirection are encapsulated. This sort of triangulation minimises routing management,at cost of a little overhead. In particular, only nodes at the network edge have to changetheir routing tables.

1.5 Resource allocation

Evolution from 2G to “near 4g” cellular systems extends the concept of broadband ac-cess has been extended to the wireless world. However, bandwidth still remains a scarceresource since newer services requesting more and more spectrum, are offered. Thegreat variety of services arises another problem, that is variability of traffic character-istics and requirements. IP, due to its flexibility, can manage this situation, but cannot


guarantee any performance, a key point for RT (real time) applications. In order to takeinto account applications’ nature, a resource allocation and management is needed. Theeasiest way is to introduce a call admission control (CAC), for example ordering userson the basis of experienced channel quality, but it is not enough. A CAC can avoid per-formance degradation, but the allocation is neither optimal nor fair. So, two solutionsare investigated, one at network layer, the other at a lower level.

In the IP header there are two fields dedicated to QoS, namely Type of Service (TOS)and precedence, respectively four and three bits. TOS was intended to be a commoncode to identify the type of traffic carried by the datagram, while precedence fieldshould have informed routers about stream priority. These fields did not found ap-plication because, at the begininning, most popular applications were e-mails, remotesessions using telnet, exchange of file by FTP (File Transfer Protocol) between few uni-versities. Only in early 90’s web pages appeared and VoIP (Voice over IP) or streamingvideos were far away. In other words there was no need to distinguish traffic and itwas more important to keep a router as simple, fast and reliable as possible. A simpleand cheap way to do this was ignoring TOS and precedence. When traffic compositionchanged, network engineers came up with some interesting solutions such as MPLS(Multi Protocol Label Switching), IntServ (Integrated Services) and DiffServ (Differenti-ated Services).

In MPLS [7], packets are forwarded on the basis of a label inserted in the header: thegoal is to add support to QoS, not to substitute routing infrastructure. The label is usedby a Label Switched Router (LSR) to identify a well defined path, called Label SwitchedPath (LSP), implementing a virtual circuit transmission over an IP network. A LSP canbe established using two approaches:

• control driven: the path is initialised by control messages, carried by routing up-date or managed by a label distribution protocol;

• data driven: the path is established when a long life streaming is detected (thenumber of packets is greater than a certain threshold or the stream is directed tosome particular ports).

In the first case, when data arrive, the path is already established and it is highly recon-figurable, but scalability is reduced and the network load is increased. In the secondcase performances are traffic dependent and an high number of short connections rep-resents a problem (i.e. a request of a HTML page with a lot of pictures).


A label represents an index of the switching table and it has local meaning, that is itis changed by every LSR. It is important to note that the information contained in the la-bel is analysed only the first time encountered, to fill in the forwarding table, while in IPforwarding the packet header has to be processed every time a datagram is received. Inother words, MPLS can speed up forwarding process, reducing classification to a sim-ple label identification, avoiding analysis of parameters such as source and destinationaddresses and ports.

IntServ [8] is another architecture proposed by IETF to introduce a QoS aware frame-work. Basically it can classify traffic in three categories:

• Best effort: standard service offered by IP, without any guarantee;

• Guaranteed service: an upper bound for packet delay is fixed;

• Controlled Load: a statistical characterisation of delay and packet loss is given.

IntServ strongly relies on RSVP (Resource reSerVation Protocol) [9], a signalling proto-col allowing

• an application to request a specific QoS;

• a router to select among different forwarding paths.

RSVP has to manage two types of informations, Filter Spec, which identifies IP pack-ets having allocated resources, and Flow Spec, which concerns resources allocated for aflow. It is quite difficult to deploy this protocol because it has to manage a lot of infor-mations and it introduces inter level dependency, in fact, every router has to maintainflow informations belonging to different layers.

As we can see from the brief explanation of MPLS and IntServ features, it is clear thatscalability is a limiting drawback of both architectures. For this reason a more flexibleproposal was made: DiffServ [10]. In this case classification is performed only in routersplaced at the border of the domain, while internal routers have only to apply knownrules. DiffServ was indicated by 3GPP for transition towards all-IP architecture and itwill be exposed in Chapter3.

Despite valuable efforts to define a general QoS architecture, suitable for any phys-ical carrier IP can support, we think that it is useful to look at the specific nature ofthe radio channel and try to work at a layer lower than network level. QoS satisfac-tion cannot leave available resources apart, in other words it is important to remem-ber that some points became milestones in the definition of 4G radio access systems


[11], and OFDM (Orthogonal Frequency Division Multiplex) is one of them. From theresource management point of view, OFDM is very powerfull because bandwidth isdivided in sub-carriers or in sub-bands allowing fine granularity allocation schemes.Power management and adaptive coding and modulation (ACM) help to follow radiolink variations in order to maximise throughput (rate adaptive system) or to minimisepower (margin adaptive system). Another important issue is the impact of the new Hy-brid ARQ technique on error correction, mixing efficiency of FEC coding and effective-ness of retransmission. In particular, we proposed a scheme that benefits from resiliencyof Fountain Codes (FC). The best way to analyse allocation and optimisation problem isthrough a crosslayer approach, as stated in Chapter 4.

Chapter 2

Protocol behaviour

2.1 Introduction

TcP role in provisioning reliable end-to-end transfers has been a de-facto standard intraditional Internet architecture for many years. TCP was born to ensure error-free

data delivery on a wired medium by means of a stop-and-wait retransmission algorithmand a transmission timer set to two times the round trip time (RTT). Complexity andeffectiveness of the protocol grew with Internet, in a process involving some decadesand still in development. Nowadays efforts are addressed to adapt TCP to a wirelessenvironment, where distinction between congestion and errors is fundamental. Radiochannel has a bursty error distribution which can create several problems to end-to-end correction schemes, for this reason 3GPP introduced a local recovery mechanism,included in RLC sublayer. In this chapter we propose a study on interaction of TCP andRLC implementation in UMTS (the first cellular system strongly data oriented, also formultimedia applications) after a short introduction to TCP, UMTS and the simulationtool used.

2.2 Transmission Control Protocol

TCP belongs to the 4th layer of the ISO/OSI stack, the transport level, which has tomake available a logic communication between two processes placed on distinct host,independently of the number of hops or the physical medium in use. In particular TCPprovides:

full-duplex transmissions: packets can be sent and received simultaneously;

point-to-point connections: there is always one sender and one receiver;

2.2. Transmission Control Protocol 15

connection-oriented services: before communication can take place, there is a param-eter negotiation phase in which the two entities exchange three packets (three-wayhandshake);

reliable communications: it is possible to recover erroneous segments through retrans-mission;

congestion control and flow control algorithms: the sender try to avoid saturation ofboth intermediate nodes and receiver.

In the rest of the paragraph a brief overview of some TCP implementation used in sim-ulations is proposed.

2.2.1 TCP Tahoe

This protocol version has all basics instruments that will be developed in recent imple-mentations.

Reliable data transfer is obtained using acknowledgement mechanism, that is everytime the receiver gets a packet, it has to send back to the transmitter a packet called ACK(acknowledgement). An ACK has the homonymous field set to one and its dimensionis 40 byte (the sum of TCP and IP headers). If the communication is bidirectional, ACKscan transport data flowing in the opposite direction, from receiver to sender, adoptinga piggyback scheme. ACKs are only positive and are cumulative, that is, if a ACKfor packet n is lost, but ACK for packet n+1 is correct, it is implicit that packet n wassuccessfully received.

Error correction is committed to a Selective Repeat ARQ scheme, in other words onlycorrupted packets are retransmitted. This means that the transmitter has to keep trackof packets sent consecutively without waiting for respective ACKs, according to the di-mension of the transmitting window, while the receiver has to store received segments ina buffer, with dimension equal to the receiving window. If the transmitter window staysbelow the receiving window, flow control is successfull, because the receiver buffer isnot saturated. Since an ACK is only positive, two others mechanisms are used for errordetection:

retransmission timer: it is associated to the oldest segment not yet acknowledged. Ifthe timer expires, TCP supposes that there is congestion, so it retransmits thatsegment. In the other case, if the segment is acknowledged before time out, the


timer is associated to the least recent segment still on the channel. It is clear thattime out value is fundamental, to avoid undue retransmissions or to avoid waitingfor a long period. So, there are two components, estimated round trip time, RTTE ,and its deviation, DevRTT .

TimeOut = RTTE + 4 ∗DevRTT

RTT is estimated using an exponential weighted moving average, taking a sample,SRTT every round trip time

RTTE = (1− α) ∗RTTE + α ∗ SRTT

where α is 0.125 and RTTE appears two times because the definition is recursive,using old values of estimated RTT. Deviation is calculated averaging the differencethe sampled value and its estimated value

DevRTT = (1− β) ∗DevRTT + β ∗ |SRTT −RTTE|

where β is 0.25, or, in general, 1− α. Every time a retransmission occurs, time outis doubled, but retransmitted segments are not used in this estimation.

duplicated ACKs: if a segment is lost, ACKs for other segments signal this situation.So, when the transmitter receives a certain number of this duplicated acknowl-edgements, say 3, it is clear that a segment was not received. It is a waste of timeto wait for the timer to time out, it is more efficient retransmit immediately. Thisalgorithm is called fast retransmit.

Error detection and correction is strictly related to congestion avoidance and to con-gestion window. First of all, TCP can control the bit rate of the application reducing it inorder to avoid saturation of routers. This mechanism is called self clocking and consists


in sending a packet only when an ACK is received. It is important to note that TCPdoes not use a Stop and Wait protocol, so it has not to wait for the ACK referred to thelast packet sent, but an ACK for a packet in the transmitting window is enough. In thisway it is not possible to send a new segment before an old one has left the network,according to packet conservation principle. However, the fundamental tool for conges-tion avoidance is the congestion window, representing the number of packets sent but notyet acknowledged. The congestion window works together with the receiving windowin the sense that the smallest one becomes the transmitting window. Of course, if thenumber of packets sent in the network doesn’t saturate the receiver or the network, bothcongestion and flow control are obtained. The evolution of the congestion windows isinfluenced by errors, in particular there are two algorithms:

slow start: congestion window is below a certain threshold and every ACK receivedaugments window’s dimension by one unit. In other words, in a round trip timethe window is doubled and, despite the name, it is growing quite fast.

congestion avoidance: congestion window is above a certain threshold and a quantityequal to the inverse of its dimension is added for every ACK received. In a roundtrip time, the window is increased by one unit and the growing rate is reducedwith respect to the previous case.

Using the first algorithm, TCP starts to send data at a low rate, increasing it in a quitefast way, until the threshold is reached. In this phase, the protocol is probing capabilitiesand level of congestion of the network, avoiding to start transmission with a rate tooaggressive. Beyond the threshold, rate is increased again, in order to exploit availableresources, but using small steps. In case of error, due to both retransmission timeoutor duplicate ACKs, congestion window is shrunk to one and the threshold is set tohalf the dimension before detection, triggering a slow start. If sketched, the congestionwindow has a saw-tooth shape and the rules (slow start, congestion avoidance, errormanagement) regulating its behaviour are usually boiled down in a single AdditiveIncrease Multiplicative Decrease (AIMD) algorithm.

2.2.2 TCP Reno

Considering more carefully error detection and its impact on congestion window, it isclear that it is quite penalising for the achievable average throughput. Retransmission


timeout represent, for sure, high congestion situation because no packets arrived to thereceiver. Since acknowledgements are cumulative, if only one of the packets sent laterthan the one associated to the timer had arrived, it could have confirmed it, avoidingtimeout. On the contrary, duplicated ACKs represents light congestion, because somepackets are arrived and it is confirmed because at least three ACKs are sent back to thetransmitter.

According to these considerations, TCP Reno treats in a different way timeouts andduplicated ACKs. In particular, TCP Reno behaviour is the same as TCP Tahoe in case ofthe timer is expired, that is a slow start is triggered. On the contrary, if a light congestiondue to triple ACKs is detected, a new algorithm is used, called fast recovery. When thethird ACK is received, the threshold is set to half the dimension the congestion windowhad before detection, while the congestion window itself becomes equal to the thresholdplus three (all the packets sent plus the three acknowledged). Slow start is avoided andTCP enters fast recovery mode. Every other duplicated ACK enlarges the congestionwindow by one unit, in order to increase rate and to speed up error recovery, until thelost packet is acknowledged. From this moment, TCP returns to congestion avoidancealgorithm because the congestion window is set equal to the threshold.

2.2.3 TCP New Reno

Fast recovery behaves correctly in case only one packet is lost for each congestion win-dow. In all others cases, i.e. two or more packets are erroneous, TCP Reno can retransmitquickly only the first segment, then it exits fast recovery mode. For all others segmentsit is necessary to wait for timer expiration. When IETF realized this odd behaviour in-troduced TCP New Reno, which keep fast recovery algorithm active until all packets inthe congestion window are error-free.

2.2.4 TCP Selective Acknowledgements

Selective ACKs [12] are particularly useful in case error free segments are received in anon continuous order. The sequence number is always the first byte needed to completethe stream, so it is possible to use fast recovery and fast retransmit, but the ACK is notcumulative. This means that, when a packet is acknowledged, it is not implicit that allthe previous ones are received too. On the contrary, the Option field in TCP header isused to give detailed informations about data blocks correctly received. In particular,

2.3. UMTS overview 19

Frequency Band [MHz] Usage1900 - 1920 UMTS TDD1920 - 1980 UMTS FDD uplink1980 - 2010 UMTS satellite uplink2010 - 2025 UMTS TDD2110 - 2170 UMTS FDD downlink2170 - 2220 UMTS satellite downlink

Table 2.1: Frequencies assigned to UMTS.

each block is described using two 32 bit values, left edge and right edge, so bytes lowerthan left edge - 1 and greater than right edge are not received. Since the Option field is 44bytes long and each block needs 8 bytes, up to four blocks are described in a single ACKpacket. The remaining space is used for the block that generated the selective ACK andto indicate the length of the Option field.

2.3 UMTS overview

TCP performances over radio channel will be simulated using the UMTS radio interface,so a brief overview of system nodes and protocol stack will follow.

UMTS is a third generation cellular system adopted in European countries, stronglyoriented to data communications. It can provide classical point-to-point services to-gether with new point-to-multipoint ones, on the basis of a detailed QoS framework.As shown in Table 2.1, there are symmetric bands for Frequency Division Duplex usingterrestrial or satellite communication and two non symmetric band for Time DivisionDuplex. Total terrestrial band is 155 MHz, while satellite band is 75 MHz, channels are5 MHz wide with 200 KHz spacing.

Since terrestrial and satellite coverage is envisaged, a new feature introduced byUMTS is macrodiversity, that is cells have a hierarchical superposition (satellite cell,macrocell, microcell and picocell) not mutually exclusive, corresponding to differentachievable data rates (for example 64 Kb/s, 144 Kb/s, 384 Kb/s, 2 Mb/s, respectively.)A mobile terminal can choose among various cells, making up an Active set, on the basisof the desired QoS and data rate.

UMTS differs from 2G systems also for the multiple access technique chosen, CDMA(Code Division Multiple Access). User data, in form of strings of bits, are multiplied in


the time domain by a unique code, a sequence of ’chips’, at a higher data rate (the num-ber of chips used per each bit represents the spreading factor). In this way signal bandis enlarged and users can be multiplexed simply using orthogonal codes. So, all ter-minals are sharing the same bandwidth, 5 MHz, at the same time, allowing gracefulldegradation performances with respect to other techniques such as Frequency DivisionMultiple access (FDMA) or Time Division Multiple Access (TDMA). Wideband signalis then modulated and protected using FEC codes (convolutional or turbo codes) whilemultiplication by the same code gives back the original data. Codes are not perfectlyorthogonal because they have finite length and their properties are degraded by prop-agation and distorsions, so there is a residual interference during reception. Moreover,power control is much more important than in previous systems, not only for batterysaving: if a terminal is transmitting using too much power, it can “hide” other ones.Power control and code orthogonality are the two factors that limit the number of usersin a CDMA system.

UMTS is made up of two functional elements, as shown in Figure 2.1, in order toobtain a separation between the radio interface Uu and the other parts of the system.The first functional component is the Access Stratum which includes radio protocols andfunctions related to resource and mobility control. The second one, called Non AccessStratum, has to provide connectivity to other networks independently of the access tech-nology. From the implementation point of view, these two blocks correspond to UTRAN(UMTS Terrestrial Radio Access Network) and Core Network, respectively.

Figure 2.1: Access Stratum and Non Access Stratum.


2.3.1 UTRAN

UTRAN is made up of Radio Network Subsystems (RNS), as we can see in Figure 2.2,which communicate with the Core Network by means of the interface Iu and exchang-ing information each other in a direct way employing the Iur interface. This featurerepresents an innovation helping to reduce the Core Network load. In each RNS there

Figure 2.2: UTRAN and Core Network.

is a Radio Network Controller (RNC) supervising the activity of one or more Node Bthrough Iub interface. A Node B controls some Base Transceiver Stations (BTS), whichare cells seen by mobile users. RNC block allocates radio resources and manages han-dover between Node Bs under its supervision, while a Node B has the same tasks withrespect to its BTSs. Macrodiversity and handover are bordered to UTRAN domain andCore Network is involved only if Iu interface is not present.

The radio interface between the mobile station and UTRAN is called Uu and is basedon the stack protocol in Figure 2.3. At the bottom there is the physical layer, whichhas to transmit informations given by upper layers and the services it provides are thetransport channels. They can be divided in two categories: common channels, for trans-missions towards all terminals and dedicated channels, for communications with a welldefined mobile station. There are more transport channel than GSM because UMTS isnot oriented only to voice services. Physical layer also defines low level transmissionsparameters, shown in Table 2.2.

The tuple made up of frequency, code and time slot identifies a physical channel,used to carry one or more transport channels.

In the second level of the stack there are two sublayers, MAC (Medium Access Con-trol) and RLC (Radio Link Control). MAC sublayer has to regulate access to the shared


Figure 2.3: UMTS radio protocol stack.

Duplexing FDD TDDAccess Technique Wideband CDMA Time Division CDMA

Chip rate 3.84 Mchip/sData rate Variable, according to the spreading factor usedChannel 5 MHz

Frame duration 10 msNumber of slots in a frame 15

Modulation QPSK in DL, BPSK in UL QPSKDemodulation Coherent

Table 2.2: Physical parameters for UMTS.

radio channel. For this purpose it takes some measurements, then passed to the controlentity at layer three, which can decide a reconfiguration of resource allocation. MACdoes not implement acknowledgement nor segmentation, while the services providedare called logical channel: on the basis of the information carried, there are control andtraffic logical channels. MAC can multiplex logical channels into transport channelsand it can also manage flow priority to obtain QoS.

The other sublayer, RLC, can operate in three different ways:

transparent mode: no header is added, usefull for multimedia content delivery;

unacknowledged mode: a header is added, but no guarantees on correct reception are


given;

acknowledged mode: a selective repeat protocol ensure error correction by retransmis-sion.

The third mode is the most interesting from our point of view, for interactions originatedwith upper layers, while the other two are transparent to the application level. Figure2.4 shows that RLC provides padding (that is zero filling), concatenation or segmenta-tion in order to maximise flexibility in usage of its data structure, called PDU (ProtocolData Unit). If needed, data can be encrypted and header suppression contributes to savebandwidth. Moreover, a poll mechanism [13][14] is the basis of acknowledged mode op-

Figure 2.4: RLC acknowledged mode.

eration. The sender asks the receiver about correctness of last PDUs arrived by means ofa POLL PDU (a data block with a flag enabled). The receiver answers sending a reportcalled STATUS PDU. Polling frequency, consequently the number of blocks a STATUSPDU covers, is negotiated, and influences the performance of the protocol. However,piggybacking is envisaged to maintain high efficiency. The same mechanism can also bereceiver initiated, enabling the POLL flag to request a STATUS PDU. Starting from thereceiver side can be useful because the direct information about missing blocks is avail-able here, however, the sender initiated version is usually adopted because it preventsdeadlocks and stall conditions [15].


The third layer is occupied by the Radio Resource Control (RRC) block, performingmanagement functions involving all resources in the cell. First of all, it has to evaluatethe service requirements from upper layers and available resources, assigning them onthe basis of QoS and bit rate needed. RRC remains active for all the communications,collecting measurements from MAC and physical layer in order to reallocate resources,if useful. Services provided to Non Access Stratum includes notification, paging, con-nection establishment and release.

2.3.2 Core Network

The first implementation proposed for UMTS Core Network aims to maintain high com-patibility with GSM and GPRS infrastructure, so there is a Circuit Switched Domain anda Packet Switched Domain. ATM (Asynchronous Transfer Mode) plays a central role,in particular ATM Adaptation Layers (AAL), because it operates using virtual circuits,so it can adapt to both domains and it is also able to emulate telephone signalling. Datapackets are transported the GPRS Tunnelling Protocol (GTP) while mobility can be ef-fectively managed introducing Mobile IP (see Chapter 1, Section 1.4). There is only anew node, UMTS Mobile Switching Centre, which implements UTRAN protocols andstrictly interacts with legacy nodes.

The all-IP Core Network implementation, shown in Figure 2.5, is much more inter-esting and it will be considered in next paragraph and chapters. The architecture shouldbe independent of the first two layers in the ISO/OSI stack, providing voice and dataservices with QoS levels comparable with the previous solution. IP protocol has to beextended to the mobile equipment and should be employed also for signalling traffic.

Briefly, the functions of each node are:

Call State Control Function (CSCF): it is the supervisor of all calling functions, such asdatabase queries, localisation, resource assignment, service negotiation, addresstranslation.

Media Gateway Control Function (MGCF): this node handles IP based and circuit switchedsignalling.

Media GateWay (MGW): this is the interface between IP based Core Network and tra-ditional networks for voice traffic.

2.4. Simulation environment 25

Figure 2.5: Core Network all-IP architecture.

Multimedia Resource Function (MRF): it is used in case multiuser or conference callsare performed.

Signalling Gateway Function (SGW): this node allows interaction between commonchannel signalling system SS7 and IP based signalling solution.

Home Subscriber Server (HSS): it is a database containing the profiles of all users.

GGSN (Gateway GPRS Support Node) and SGSN (Serving GPRS Support Node): thegateway and switching node introduced by GPRS architecture.

As we can see, the circuit switching part of the network is limited to interaction withthe traditional telephone infrastructure. On the other side, IP protocol allows to exploita great variety of available applications without any standardisation process, but thereare open issues on implementation of a QoS framework suitable for multimedia andcontrol traffic, exposed in Chapter 3.

2.4 Simulation environment

Analysis of TCP performances in a wireless scenario with a RLC correction scheme atlayer two has been carried out using a simulative approach. We chose the well knownpacket level simulator Network Simulator 2 (NS2), available at [16], because it is widelysupported and tested, especially for TCP algorithm implementation. It was originally


developed at the Berkeley University, using two object oriented languages, C++ andOTcl. C++, due to fast execution time, is employed in the core part of the simulator,especially transport protocol routines, topology elements and scheduling functions. Onthe opposite, OTcl code is easy to modify (in fact it is executed by an interpreter), so thislanguage is mainly used for topology description, assuming a role of a user interface.The interoperability between the simulator internal functions and the OTcl interface isassured by an interpreter, which maps some C++ object classes in an OTcl hierarchy.

The reference design for a simulator taking into account a mixed scenario involvingboth wired and wireless nodes is depicted in Figure 2.6. Unfortunately, NS2 does not

Figure 2.6: Simulator’s protocol stack.

support UMTS protocol stack in the original distribution, so we adopted two differentsolutions. In the first case we developed a simple radio agent implementing a basic RLCprotocol and using a Markov model for error transmissions over the radio channel. Inthe second one, we referred to the UMTS TDD module [17] developed at the Universityof Rome “La Sapienza”, calculating probability of error using analytical formulas.

Despite different approaches, there are some common points that should be dis-cussed. First of all, it is necessary to add a retransmission protocol at layer two of theradio protocol stack. It is fundamental that errors originated by the bursty channel na-


ture are corrected locally, avoiding waste of distributed resources. 3GPP suggests to useSelective Repeat (SR), the most efficient of ARQ protocols, limiting retransmissions toerroneous blocks, as in Figure 2.7 In order to reduce the number of bytes used to assign

Figure 2.7: Selective Repeat scheme.

a sequence number to each RLC PDU (data structure in which is divided the IP protocolflowing from the third to the second layer), a sliding window is introduced. Once allthe blocks in the window are sent, the window can move ahead and sequence numbersbegin from the start. Erroneous PDUs are repeated, occupying a position in the newwindow. If all blocks are not yet acknowledged, there is a mismatch because a sequencenumber is no more unique. Consequently, a window should be large enough (depend-ing on the round trip time to obtain the first acknowledgement) to avoid this situation.However, if TCP, or other protocols performing retransmissions at transport layer, areused, SR could arise problems. Retransmission of a corrupted PDU takes place in thenext window, so if the window is too large, error correction is delayed. Consequently,also generation of TCP acknowledgements are delayed, triggering a retransmission timeout and lowering the bitrate. Moreover, there are other problems due to the “self clock-ing” mechanism of TCP (see Subsection 2.2.1). TCP sends a new packet when receivesan ACK, so the number of packets in the network remains under control. If RLC needsa large time interval to correct errors, generation of ACKs is delayed too and TCP couldbe unable to send new packets because it is not receiving any acknowledgemet This isa strange situation: maybe the channel error rate is low, but the bit rate experienced ispoor because TCP is slowing down the source due to the lack of acknowledgements,delayed by RLC sliding window. Usually, a PDU is 40 bytes (38 bytes of data and 2bytes for header), while a TCP segment is in the order of hundreds of bytes, so we useda sliding window containing few packets. For example, whit a TCP segment size of 576

2.5. Simulation and Results 28

bytes, a suitable sliding window can be made up of 15 PDUs. With this empirycal rule,obtained with some simulation employing out modified agent, risk of slow start or stallsiyuation is reduced.

Another important issue is the dimension of the RLC buffer, where packets wait-ing for transmission over radio interface are stored. Buffer occupancy is very similarto TCP congestion window [18], except for fluctuations during fading periods in thechannel. So, the minimum buffer should not fall below the congestion window in or-der to avoid packet dropping because TCP is not aware of the situation at level two (inother words, a cross layer approach is not adopted here). As RLC selective repeat hideschannel burstiness from upper layers, the buffer limits influence of flow and congestioncontrol algorithms on air interface. In Section 3.3 a detailed buffer discipline for videoapplication will be exposed.

2.5 Simulation and Results

When a wireless channel is involved in a simulation, it is important to carefully modelburstiness and phenomena characterising radio interface. The easiest way is to usea Binary Symmetric Channel at packet level [19], using a random variable to decidewhether the block is correct or not. This is true only for very low error rates or for longinterleaving lengths, which may not be compatible with the time outs of the flow controlmechanisms. In other works a Gilbert [20] channel is used with generic assumptions onthe state probabilities [21]. In next sections we discuss two alternative solutions, morecomplete and closer to real behaviour.

2.5.1 Hidden Markov Model

In order to take into account variability of conditions, but aiming to keep the model assimple as possible, we used a hidden Markov model (HMM). Starting from physicalmeasurements in form of a string of bits, where 1 stands for erroneous transmission,a new binary sequence was created grouping 320 bits in a 40 byte PDU. Next, the wellknown Baum-Welch algorithm was n to statistically characterise the retransmission pro-cess [22]. It has been found that at the target BER of 10−6 the transmission chain is rathergood in randomizing the residual error process at the PDU level [23], so a simple twostates Markov model is enough for our purposes. Note that this does not hold for the


error process at the bit level and also may not be considered a valid assumption if thebit error rate approaches 10−4 at which the number of states dramatically increases andat least 8 states are necessary in agreement with [24].

The network topology, see Figure 2.8, used is very simple, involving only few nodes,in particular node 0 is the source in the wired domain, node 1 is Node B and node 3 isthe mobile station, while the intermediate node 2 is needed to add a fixed delay takinginto account encoding, decoding and interleaving operations. The link between the

Figure 2.8: Network topology.

last two nodes represents the wireless channel, its capacity is fixed in accordance toUMTS supported rate (namely 64, 144 or 384 kbit/s) and it is provided with the HMMmodel previously introduced. In Table 2.3 main radio channel parameters used in thesimulation are listed,

Channel rate [Kbit/s] 64 144 384PDUs per block 4 9 12Block size [bytes] 160 360 480Transmission Time Interval [ms] 20 20 10Spreading Factor 16 8 4

Table 2.3: Wireless channel simulation parameters.

while RLC and TCP parameter are given in Table 2.4.

PDU size[bytes] 40RLC window size 15Polling bit every 5 PDUsTCP segment size [bytes] 576TCP flavours Tahoe, Reno, New Reno, Seal. ACKTotal traffic [Mbytes] 2Simulation time [s] 300

Table 2.4: RLC and TCP simulation parameters.


We can start our comparison [25] between the four implementations of TCP madeavailable by NS2 from Figure 2.9, noting the different amount of data exchanged dur-ing simulation time, 300 seconds, using a sequence whose Eb/N0 is equal to 3 dB. TCP

Figure 2.9: TCP sequence number over wireless link. Solid line: Sack; dotted line: New Reno;dashed-dotted: Tahoe; dashed line: Reno. Sack and New Reno are overlapped.

Tahoe, New Reno and Sack have sent about 2 Megabytes of data, while TCP Reno hasbeen able to send only the half.

These behaviour can be explained looking carefully at the congestion window evo-lution in Figure 2.10. Here, a detail of the congestion window is shown: it representsthe first 125 seconds of simulation, but it is able to justify different performances of thefour TCP versions. In the first 38 seconds they have a similar performance because ofscattered errors and absence of timeout events; after that a three duplicate ACK eventis generated, and different solutions are shown. Tahoe behaves as usual, putting thecongestion window equal to 1, then increases it firstly in slow start (exponentially), andfinally switches to congestion avoidance after reaching its threshold. New Reno andSack succeed in avoiding slow start phase and, after halving their threshold, adopt con-gestion avoidance algorithm (with linear increments). The worse performance, again,is shown by TCP Reno: the agent, at the receiving of 3rd duplicated acknowledgement,switches to fast recovery and increases congestion window by one each further du-plicated ACK reaching the sender. Window exceeds the limit of 30 packets set in thesimulation parameters due to the presence of previous packets stored in the buffer. This


Figure 2.10: Congestion window. Legend as in Figure 2.9.

is not a real problem because the transmission window is chosen as the minimum be-tween congestion and advertised window. TCP Reno supposes that only a packet islost, so, when retransmission of the datagram is confirmed, congestion avoidance algo-rithm is used. In case of a second corrupted packet, retransmission cannot be triggeredimmediately, it is necessary to wait for the timer expiration. During this period sourcestops sending data because of absence of ACKs (see ”self clocking” mechanism), more-over time out forces the use of slow start algorithm. Tahoe can better manage multipleerrors, even if less efficiently than evoluted versions of TCP: it always reacts to missingdata in flow using slow start, so time out is no more needed for retransmitting packets.Congestion window remains 1 for the time necessary to serve packets arrived before theduplicate one, as seen for Reno.

In Figure 2.11, the cumulative distribution functions of the acknowledgements inter-arrival times are shown. Note that all four TCP versions have a similar behaviour. Thiscould lead to suppose that they all have the same performance, but consider that theduplicate ACKs are included too. All acknowledgement packets are close each other,in all simulations and, considering that the return path is lossless and error free, thisbehaviour should not be surprising [26]. In the same figure there is also a comparisonusing a model for a channel with Eb/N0 equal to 2 dB. Poor channel conditions reflectsin an increased number of retransmissions, so more acknowledgements experience a


Figure 2.11: Cumulative distribution functions of interarrival time for 2 and 3 dB channels witha 64 Kbit/s source.

greater delay. Once again, TCP Reno avarage data rate is lower than other flavours.The same comparison is proposed in Figure 2.12 for a 144 Kbit/s source.

(a) 2 dB channel (b) 3 dB channel

Figure 2.12: Cumulative distribution function of interarrival time with a 144 Kbit/s source.

The conclusions that can be drawn are the same already discussed. Results for 384Kbit/s source are omitted because they are very similar to the previous ones.


2.5.2 UMTS TDD module

TCP performance analysis continues employing UMTS TDD module extension for NS2[17] developed at the University of Rome “La Sapienza”. The authors exploit wirelesssupport provided by the simulator, so it is no more necessary to use a statistical model.Channel behaviour depends on

• SIR (Signal to Interference Ratio), measured in dBm;

• fading;

• loss of orthogonality.

So, it is possible to calculate the transmission power using the following:

P = α · pl + (1− α) · plmean + I + SIR

where

• α is the weight parameter for orthogonality loss;

• pl is path loss, measured on the primary common control physical channel (P-CCPCH), a physical channel transmitted at a defined power level;

• plmean is the average path loss, calculated with a moving average;

• I is the interference experienced during the timeslot.

To calculate the SIR, the following two formulas are used, respectively for downlinkand uplink direction:

SIRDL =PTi · Li ·Gi · f 2

i ·m2i

PN + IEXTi + α · IINTRAi

SIRUL =PTi · Li ·Gi · f 2

i ·m2i

PN + IEXT + (1− β) · IINTRAi

where PTi is the transmitted power, PN is the noise power, Li is the path loss, Gi is theprocessing gain (depending on the spreading factor), fi is the fast fading, mi is the slowfading, IINTRA is the intra-cell interference, IEXT is the inter-cell interference and β isthe joint detector receiver gain. For slow fading generation, the Gudmunson model isused, while fast fading is modelized as a low pass filtered Gaussian white noise process.


Mobile speed 2 m/sCode orthogonality loss 0.06Multi-user detection gain 1Simulation time 200 sMax. number of PDU retransmissions 4

Table 2.5: UMTS TDD module parameters.

Intra-cell interference depends on the codes used by other users in the cell and loss oforthogonality, while inter-cell interference is a log-normal random variable.

We started with a set of simulations with a 4 dBm SIR channel [27], other significantparameter can be found in Table 2.5 and we plotted some statistics about number ofPDUs retransmissions, Figure 2.13, and distribution of ACKs interarrival times, Figure2.14, for the four TCP implementations provided by the simulator.


(a) New Reno (b) Tahoe

(c) Reno (d) Sack

Figure 2.13: PDU transmission statistics for 4 dB channel.



(c) Reno (d) Sack

Figure 2.14: Cumulative distribution functions of ACKs interarrival times for 4 dB channel.

As we can see, results are very similar to those proposed in the previous section us-ing HMM model: in some cases TCP Reno fails to recover multiple errors. Total numberof PDUs sent using TCP Reno is lower than in other cases, moreover the maximum de-lay is higher, because of time out needed to recover the second lost packet in a windowWe repeated the simulations worsening channel condition, that is lowering the SIR to 2dBm. Results are shown in Figures 2.15 and 2.16.



(c) Reno (d) Sack

Figure 2.15: PDU transmission statistics for 2 dB channel.

The number of PDUs sent is very close in all four cases, moreover maximum de-lays experienced are completely unsatisfactory from the user’s point of view. Difficultchannel conditions put all TCP flavours considered on the same level, so error correc-tion mechanisms at the transport level are too far from the origin of the problem, losingtheir effectiveness.

2.6. Considerations 38

Figure 2.16: Cumulative distribution functions of ACKs interarrival times for 2 dB channel.

2.6 Considerations

In this chapter we evaluated the avarage behaviour of four popular TCP implemen-tations using two different approaches to evaluate the impact of channel variability.Usually, if channel error rate is moderate, TCP with selective acknowledgements out-performs other versions, but when channel conditions get worse, maximum latency andvariance are remarkable in all cases. In this prohibitive conditions it is necessary to re-fer to the avarage behaviour because, after the first error, each TCP flavour adopts adifferent strategy, occupying the channel with different bit rates and in distinct instantsof time. So, it is possible that TCP Sack experiences poor performances simply becauseother versions use the channel in a “favourable” moment, when no errors occur. More-over selective acknowledgement mechanism is not suitable for high error rates, it hasbeen designed for wired networks and it is not optimized for air interface. It is notunlike that all the three slots available in the option field to indicate non-continuoussegments correctly received are full. If it is necessary to indicate another segment [28],the first one, that is the oldest, is overwritten. There is no more track of this block ofdata and data recovery is possible only by retransmission time out.

Chapter 3

Application awareness and performanceassurance

3.1 The video coding case

Channel reliability is fundamental for a successfull communication, but not all ap-plications only rely on error free information transfer. As already said, multimedia

applications need to provide an information flow with continuity and within strict timeconstraints. Quality of service provisioning cannot disregard the nature of the applica-tion requested, so resiliency tools should be added to traditional error correction tools.In other words, ARQ techniques are suitable to save bandwith and to recover punctualerrors, regardless the delay introduced. Higher layers (transport and application) arenot likely to be involved in ARQ solutions for real time services, essentially becauseof violation of time constraints. FEC techniques, on the contrary, are best candidateto protect those contents which are critical for the retrieval of a multimedia flow. It isalso intuitive that some kind of cross-layer interaction between the network and the up-per levels could provide a more flexible resource management and a higher resiliency.In this section ARQ, resilient encoding and smart buffering will be considered in anUTRAN scenario where a user is requesting a real time video service.

RLC AM was already reviewed in section 2.3.1 and it is particularly indicated forinteractive and background traffic thank to guaranteed delivery of PDUs. Despite thisconsideration, in this paragraph it is treated as a resiliency tool and a key component torealize a differentiated treatment of stored traffic on the basis of buffer level occupancy.In order not to introduce excessive delays, ARQ protocol is managed within a very shortwindos of 15 PDUs and only one retransmission is possible. The ARQ policy on packetsis selective, when a PDU is not received, even after a second attempt, the whole packetis discarded.

3.2. Video coding principles 40

3.2 Video coding principles

In this section a brief overview of video coding principles is given, in order to introducesome usefull terminology without any seek of completeness.

Video compression is fundamental to fit related information flow to wired and wire-less channel capacity now available and this operation is based on the fact that humaneye is unable to perceive high frequency variations in signals and colours. So it is pos-sible to reduce the quantity of information carried by the signal removing spatial, tem-poral and statistical redundancy.

Starting from the YUV signal, produced by the digitalisation process and describedin terms of luminance and chrominance components, each frame is divided in squareblocks made up of n x n pixels (with n varying usually from 2 to 16), to reduce compu-tation time. Using a Discrete Cosine Transform (DCT) is possible to extract from eachblock a matrix of n2 coefficients representing spatial frequency components. The originis in the upper left corner and coefficients rapidly decrease approaching to the bottomright corner, so a quantisation matrix allows to reduce spatial redundancy. In the quan-tised matrix, a great number of coefficients is zero, consequently a reduction of statisticalredundancy is achieved employing a Run Length Coding (substituting a group of zeroswith the number of elements in the groups) followed by Huffman coding (substitutingfrequent coefficients with short code words). The frame obtained after these operationsis called I (Intra) frame, and is a basic self-decodable block. All the information neededis embedded, I frames are used as reference and they should be present periodically, atmaximum every 12 frames, to avoid propagation of errors. Finally, time redundancy islimited using motion compensation techniques, that is using a motion vector to recordthe movement of a block between two frames, originating two other type of frames,called P and B. Predicted, or P, frames are coded starting from the previous one, eitherI or P. Bidirectional, or B, frames are coded interpolating two frames, the next and theprevious one, either I or P, obtaining maximum compression.

For our tests, we suppose that mobile user is requesting to a server a H.264/AVC(Advanced Video Coding) [29] video stream. 3GPP is considering H.264/AVC for mostrecent versions of its recommendation because of coding efficiency, network adapta-tion capability and effectiveness of CBR (Constant Bit Rate) performances. H.264/AVCincludes several resiliency tools [30], particularly useful in case of losses:

• prediction can be restricted within groups of interleaved “slices”, worsening the

3.3. Buffer management 41

compression rate but limiting intra-frame error propagation; flexible macroblockordering generalises this concept with flexible block-to-slice mapping;

• data partitioning and FEC-like methods (either syntactic or semantic) on the pic-ture (including redundant slices);

• a non-normative error concealment feature in the spatial and in the temporal do-mains.

Moreover, another feature is added: a small and controlled redundant informationwith FEC significance is inserted in I frames due to their critical role for sequence syn-chronization, as explained in [31]. The authors propose a method called ”Edge andResiduum Decomposition” (ERD), providing an edge-based sketch-description of thewhole frame and allowing a better spatial-domain error concealment. This approachhas the benefit of protecting the whole frame and not only a region of a frame, as mostof current methods do. The ERD components are wrapped up in an additional “pseudo-frame“ which will be called E frame. Experiments show that E frames strongly dependon the image characteristics: their percentage occupancy, referred to the correspondent Iframes, is always less than 20%, while they represent less than 5% overhead with respectto the whole sequence. Benefits of E frame are not limited to I frames and also reduceerror propagation; additional advantages are likely to come from ”smart” network de-cisions, such as differentiated policing, scheduling and retransmission according to thetype of the frame.

3.3 Buffer management

In previous paragraphs benefits of cross-layer approach where envisaged: use of lowlevel information to manage resources, that is retransmission opportunities, in order tostrengthen, by introducing redundancy, crucial parts of a multimedia flow. We thinkthat RLC buffer can play a central role in applying these concepts, thanks to a smartmanagement algorithm. In figure 3.1 it is possible to note that three threshold valuesare indicated, namely Thr 0, 1 and 2. If buffer occupancy is under the smallest threshold,Thr 2, all frames are stored and retransmitted one time. Between threshold 2 and 1, onlyP and E frames are accepted and eventually retransmitted. In the next portion, betweenthreshold 1 and 0, the only difference is that only E frames can be retransmitted, while

3.4. Simulation scenarios and results 42

Figure 3.1: Smart RLC buffer management.

P are discarded after first transmission attempt. Finally, when buffer is almost full andalso threshold 0 is exceeded, only E frames are stored and retransmitted.

In accordance to this criteria, when channel behaviour is getting worse, the sourcetries to maintain a minimum video quality and a human readable image, giving morerelevance to essential parts of the scene, the edges contained in E frames.

3.4 Simulation scenarios and results

These features will be integrated within a Node B (see Figure 3.2) where smart cross-layer decisions can be taken, such as selective RLC retransmission and video- and frame-aware network policies.

We used our modified NS2 version with HMM channel model to simulate trans-mission of a H.264/AVC coded sequence over a 144 Kbit/s UMTS channel [32]. Otherinteresting parameters are specified in Table 3.1

In our tests we considered four possible scenarios. The first one is used as referencebecause only the original H.264/AVC coder is used, while RLC is working in transpar-ent mode, so no retransmissions are performed. The second and the third scenario arecomplementary. In the first case RLC AM is used, attempting only one retransmissionof I and P frames. In the other case, E frames are added to the video stream, but noretransmissions are envisaged. Finally, the fourth scenario is a combination of the pre-vious ones, that is both RLC and video resiliency tools are introduced. PSNR of the re-ceived sequence can be found in Figure 3.3 We can see that the unmodified coder (blackline) is not enough to face channel errors, so the features added can improve perfor-

3.4. Simulation scenarios and results 43

Figure 3.2: Network environment, with Node B acting like a proxy.

Number of frames 2880Target bit rate 116 Kbit/sBuffer dimension 8.0 KbytesThreshold 2 5.5 KbytesThreshold 1 6.5 KbytesThreshold 0 7.0 KbytesRLC window 15 PDUsPolling bit Every 5 PDUsChannel SNR 3 dB

Table 3.1: Simulation parameters.

mances. Introduction of RLC ARQ (red dotted line) allows to recover a deep burst errorbetween 650th and 700th frame because PDU size, 40 bytes, is small, consequently pro-tocol reaction is high. Video layer FEC (red dotted line) behaves a little worse than theprevious solution, but seems to be better when burst of errors affect I-frames or whenthe queue tends to fill up. In fact redundant information of E frames is associated to Iframes, while P are unprotected. In contrary the ARQ mechanism in scenario 2 concernsall PDUs, regardless of the content. Finally, scenario 4 (green dotted line) seems to take

3.5. More general approach using DiffServ 44

Figure 3.3: PSNR of H.264 decoded video. Light blue line: correctly decoded sequence; Blackline: standard H.264 error resilience; Dark blue line: RLC-ARQ; Red line: redundant informationin video layer; Green line: RLC-ARQ and redundant information at video layer.

the best of scenarios 2 and 3, both for punctual and burst errors. Further improvementscould be achieved introducing more sophisticated policies to discriminate acceptance,discarding and retransmission of frames. Other simulations, however, demonstratedthat keeping the queue length fixed such policies do not improve the overall behaviour:the thresholds were either too close one to the other, or too low.

3.5 More general approach using DiffServ

In previous paragraphs we discussed and compared the benefits of RLC-ARQ mecha-nisms in a video coded signal transmission. We proposed a complete framework withsmart buffering management specific for video streaming, but in the remaining part ofthe chapter we would like to discuss a more general approach to the problem of per-formace assurance for any kind of multimedia application (voice, video, streaming andmultimedia in wide sense). 3GPP already identified the best candidate in DiffServ archi-tecture [10]: a brief overview of the system and problems related to UMTS integrationwill be given.

3.6 DiffServ architecture principles

DiffServ was proposed in 1997 and is based on flow identification to apply a differenttreatment in the DiffServ domain (the set of nodes implementing the architecture). IP

3.6. DiffServ architecture principles 45

protocol is not connection oriented, so flow identification can be obtained through clas-sification using some header fields such as source and destination IP address, sourceand destination port, transport protocol. All packets with same values belongs to thesame micro flow, but the key idea is the aggregation in macro flow, in order to preservean high level of scalability. For this purpose the DiffServ IETF Working Group strictlyfocuses on the difference between forwarding and routing. Forwarding involves eachflowing packet , taking a small amount of time and resources, since it consist of a sim-ple table lookup. Routing operations are more complex and, in this context, they dealwith the decision of which rules should be applied to an aggregated flow. Nowadaysthese methodologies are still in development, however an effective service differentia-tion can be obtained using simplified strategies and static configurations.

The forwarding operation is related to Per-Hop Behaviour (PHB), defined as ”a de-scription of the externally observable forwarding treatment applied at a differentiatedservices-compliant node to a behaviour aggregate“ [33]. In other words, a PHB putsinto practice the differentiated treatment a packet receives in a node in terms of priority,queue management and buffer allocation. Defined PHBs includes

Default PHB: it is the best effort behaviour characterising IP networks. It is used forcompatibility with nodes not DiffServ compliant or when another PHB is not spec-ified.

Expedited Forwarding (EF) PHB group: it is used to implement premium services withlow delay, low jitter, guaranteed band and low loss probability. This service offersdeterministic guarantees, provided that the input rate lies below a threshold fixedin the service level agreement between the user and the provider. In this way it ispossible to fix a service rate greater than the input rate, so EF queues are almostempty, limiting experienced delay. Out of profile traffic is redirected to anotherclass, or it can be discarded.

Assured Forwarding (AF) PHB group: it offers statistichal guarantees in terms, for ex-ample, of average throughput. A user has a minimum level service, but it alsocan temporarily exceed the rate fixed in the agreement, while the provider can ex-ploit low traffic periods. Four classes of services are defined, AFx, with x varyingfrom 1 to 4, corresponding to different performance levels in terms of bandwidthand buffer allocation. Moreover, there are three discarding priority for each class,AFxy, with y varying from 1 to 3, on the basis of the packet ”importance“ among

3.7. Interactions between DiffServ and UMTS 46

the class. AF PHB group can tolerate short period congestion to maximise band-width usage and its reaction is quite smooth

Each PHB is associated to a DiffServ Code Point (DSCP), corresponding to the six mostsignificant bits in the Type of Service field of IP header, while the other two are unused.This code acts as a table index to decide how to manage that particular packet.

In the DiffServ architecture there are two type of nodes, namely edge and core routers.Edge routers are placed at the boundary of the DiffServ domain, representing an inter-face for clients requesting service differentiation. This role entails traffic classification(using a multi filed classifier, analysing various field of IP header) to recognise microflow and their aggregation in macro flow or behaviour aggregate. Other blocks belong-ing to an edge routers include a meter, a marker, a shaper and a dropper. The metermeasures temporal properties of incoming flow, to decide if it compliant with a profileor if it out of profile. The most common meter is the Token Bucket, fixing average rateand burstiness of a source. The marker assigns the DSCP, according to the decision ofthe previous block, eventually acting as a remarker if the packet is coming from anotherDiffServ domain. The shaper can modify temporal characteristic of a flow, delayingpackets to make them compliant with a profile. Finally, the dropper can discard packetsto fit a certain profile, as the shaper.Core routers are inside the domain, linking both edge or core routers. They main func-tional block is a behaviour aggregate classifier, selecting packets only on the basis of theDSCP. In other words, they have simply to apply PHB assigned to the service class ofthe packet, obtaining a substantial reduction of load and improving scalability.

3.7 Interactions between DiffServ and UMTS

DiffServ is expected to have a central role in transition toward an all-IP UMTS CoreNetwork, as stated in Subsection 2.3.2. Some studies, for example [34], considered im-plementation issues such as mapping of UMTS QoS classes in DiffServ PHBs, yieldingto a plausible convergence scheme shown in Table 3.2 A potential performance-limitingproblem concerns packet management once left the DiffServ domain, when the RLCbuffer is reached. Refering to the simple scenario in Figure 3.4, we can see that theradio interface is outside the DiffServ domain, so, if the buffer RLC is a simple FIFO(First in, First out) queue, all the differentiation discipline in the Core Network is lostand all types of traffic are merged together, regardless the priority. In Figure 3.5 (a) we


Traff. Class DS PHB Max. Delay Max. Jitter Packet Loss Application ExampleConversational EF 20 ms 5 ms 0.5% VoIP, videoconferencing

Streaming AF41 40 ms 5 ms 0.5% Audio/video streaming

InteractiveAF31 250 ms - 0.1% Transactional servicesAF21 300 ms - 0.1% Web browsingAF11 350 ms - 0.1% Telnet

Background BE 400 ms - 0.1% E-mail download

Table 3.2: UMTS QoS classes and DiffServ PHB mapping.

Figure 3.4: A simple scenario with DiffServ Core Network.

can see an example, where high priority signaling traffic (gray packet) is in the samebuffer together with ordinary data traffic (black packet). We studied this problem us-

(a) Single RLC buffer (b) Multiple RLC buffer

Figure 3.5: Modification added to the simulator.

ing the DiffServ module for NS2 in conjunction with the UMTS TDD module already


introduced. The main change proposed concerns the RLC buffer: the simpliest solu-tion to the problem previously stated consists in increasing the number of buffers toaccomodate packets with different priorities, as in Figure 3.5 (b). Each buffer is asigneda different weight and this solution can be obtained in practice subdividing a uniquephysical queue in various virtual queues. Packets are stored in a queue on the basis oftheir priority and this decision can be taken reading the DSCP of the packet, supposingthat RRC can provide this information to the RLC through a control plane procedure.

In our simulations, we considered two traffic classes, that is we split RLC buffer intwo queues, one for high priority traffic, inluding signalling packets origineted at theRLC layer, the ohter one for low priority flows. Despite the scenario is quite simple,some interesting conclusions can be drawn.

Figures 3.6 (a) and (b) show, respectively, low and high priority buffer evolution intime.

(a) Low priority buffer (b) High priority buffer

Figure 3.6: Multiple RLC buffer split in two queues [Horizontal axis: time in seconds; veritcalaxis: queue length in bytes].

Note that the slope of the graph depends on source rate and segmentation: in ourcase the maximum packet size is 1024 bytes, the maximum segments size is 576 bytesand the maximum source rate is 384 Kbit/s. The low priority buffer is diverging be-cause the scheduler picks up packets from the other queue as they are received. In fact,this one has a saw-tooth shape, with peak values relatively small: when the schedulerfinishes to serve a low priority packet, it immediately checks the situation in the highpriority queue and serves this traffic. This trend is confirmed also by delay analysisin Figures 3.7 (a) and (b) where, except some peaks, high priority traffing interarrivaltime is quite constant, while, in the other case, interarrival time is growing. Starvationproblem can be reduced introducing a weight coefficient for each buffer. In this way,


(a) Low priority traffic (b) High priority traffic

Figure 3.7: Delay analysis [Horizontal axis: time in seconds; vertical axis: delay in seconds].

a minimum service level is assured also for low priority traffic, at the cost of reducinghigh priority traffic performances. Figures 3.8 (a) and (b) compare different behaviourobtained introducing a weight coefficient (green line) with respect to the previous situ-ation (red line), for low and high priority queues respectively.

(a) Low priority traffic (b) High priority traffic

Figure 3.8: Comparison between previous situation and the introduction of a weight factor[Horizontal axis: time in seconds; vertical axis: queue in length].

3.8 Considerations

In this chapter we discussed the fundamental role of application’s nature in satisfyinguser’s expectations. It is not possible to give parameters suitable for all cases, in con-trary a delay and error sensitivity analysis should be taken into account. In particularfor video streaming, a service that telecommunications companies consider strategicto attract and to keep customers, we considered the impact of specific resilience tech-niques. This solution was compared with low persistent ARQ at RLC layer, however


smart buffer management enhances flexibility in both cases. In order to generalize theapproach we moved to a DiffServ aware Core Network, introducing macro flow dif-ferent treatement. Once again RLC buffer management should be introduced because,in the other case, all the efforts to classify and to separate macro flows traversing thenetwork is unuseful. Despite preliminary results, it is clear that an efficient approachshould be aware of what is happening in the first two layers of the ISO/OSI stack. Someissues regarding cross layer oriented resource allocation schemes will be treated in nextchapters.

Chapter 4

Impact of resource allocation on QoSprovisioning

4.1 Introduction

In this chapter we will introduce the resource allocation problem, explaining issues re-lated to QoS provisioning. Nowadays, users are always requesting more bandwidth,

low error rates and also low latency times because, sitting in front of a desktop devicethey are able to (or will be, in a very short period):

• chat with friends;

• buy goods and services;

• query a video-on-demand service for a short clip or even a whole film;

• make a phone call or a video call;

• share/browse profiles in a social network;

• update their blog;

• many other upcoming activities in development in various startup companies.

The easiest way to meet these growing expectations is to enlarge network capabilitiesand supported traffic, investing more resources to increase the number and the exten-sion of the links. Enhancement of wireless networks capabilities, together with avail-ability of handled devices more and more alike small personal computers in terms ofmemory capabilities and processing power, are shifting previous demands in a newenvironment.

A wireless scenario is much more complicated because it is not possible to add basestations or access points in an arbitrary way for at least two reasons. First of all, the

4.1. Introduction 52

interference increases and propagation conditions deteriorate signal quality, introduc-ing a loss of orthogonality, so the hypothesis of perfect isolation between channels is nomore valid. The second issue is related to electro-magnetic pollution, in fact the Euro-pean Union, together with national and local administrations, fixes accurate limitationsin maximum transmitted power, out-of-band emissions, maximum field intensity andmany other parameters. Consequently, a provider cannot enlarge coverage area boost-ing power transmission to obtain better signal quality or go through obstacles, more-over, the nature of the physical carrier implies mobility of the users, another criticalpoint to guarantee quality and continuity of service. Evolution of radio access networksin recent years has shown more and more awareness of these problems, accompaniedby a tremendous improvement toward the goal of making broadband access availableto a variety of bandwidth hungry services. Despite the huge progress in enhancingradio-link capacity, bandwidth is still a scarce resource to be used carefully. Anothergreat challenge comes from the varying requirements and traffic characteristics of thedifferent services that must be provided in interaction with the Internet. In this sense,the IP protocol seems to be the best candidate to naturally support the heterogeneousmix of networks and services, as already discussed. In this very fluid scenario, somepoints appear to be widely accepted as both state of the art and milestones for the defi-nition of a ”real 4G” radio access system. These include the fact that OFDM (OrthogonalFrequency Division Multiplexing) is used as transmission scheme due to its spectral ef-ficiency, robustness and adaptability to user needs. The main feature is parallelisationof transmission, splitting the spectrum in sub-channels, each one centred about a sub-carrier. In case sub-carriers or groups of sub-carriers be assigned to a connection withproper power and modulation parameters, the concept of orthogonal frequency divi-sion multiple access (OFDMA) is introduced.

Most relevant issues affecting radio interface and related performances include:

• fading;

• multipath;

• interference due to other users in the cell;

• interference due to other cells;

• distance from the serving cell.

4.2. Granularity and efficiency 53

These points represent a limiting factor for broadband single-carrier based systems, buttheir effect can be reduced thanks to OFDM advantages, also offering opportunitiesfor allocation and adaptation at link layer [35]. Radio resources are defined with highgranularity, so it is possible to improve allocation efficiency reserving the needed band-width without waste of spectrum. Moreover, OFDM can exploit frequency, space andtime diversity, matching link layer request by means of transmission parameters adap-tation. The given diversity is best exploited by dynamic sub-carrier allocation in anFDM (Frequency Division Multiplexing) fashion and can be accompanied by dynamicpower allocation. Using either or both of these dynamic mechanisms, the system can bemade rate adaptive in order to either maximise the wireless channel’s capacity or marginadaptive to minimise the transmission power for a given rate per terminal.

These scheme cannot disregard cooperation between the first two adjacent levels,the well-known layered network structure is no more suitable for our purposes becauseof lack of flexibility. Isolated entities lead to inefficient use of resources, so we thinkthat a cross layer approach is more useful, with particular attention to channel awareallocation. This technique consists in a continuous flow of informations regarding ra-dio channel conditions (signal to noise ratio or probability of error, for example) calledChannel State Information, CSI, influencing sender’s decisions also at higher layers.

4.2 Granularity and efficiency

As already said, OFDM makes resource definition and allocation simpler, limiting inef-ficient use of bandwidth. Despite this contribution, on the basis of the communicationsystem implementation (WiMAX, WiFi, UMTS, ...) some other optimisation issues canbe discussed. In order to limit the data exchanged in signalling procedures, a subchan-nel is usually assigned to only one user. The number of modulation schemes used isfinite and limited to two or three schemes, and also in this case it seems unreasonableto add more constellations. It is much more promising a study on error protection cod-ing due to the emerging role of Hybrid ARQ techniques. In particular, Chapter 5 dealswith the introduction of a rateless code approach in a UMTS Release 7 environment,providing a simple method to simulate global system performances.

4.3. Diversity and adaptation 54

4.3 Diversity and adaptation

Since channel response varies with frequencies and number of users, algorithms fac-ing dynamic conditions are needed. Most promising ones are based on water fillingapproach: according to an initial budget, power is added to each subchannel, startingfrom frequencies with low loss. If we think that the power can be seen as water, revers-ing the channel response function we obtain a bin that should be filled until reachinga constant level of liquid, as in Figure 4.1. Adaptation of water filling to OFDM needs

Figure 4.1: Water filling.

some modifications because spectrum is not treated in a continuous way, in contrary itis subdivided in sub-bands. Moreover, the allocation process can exploit only a finiteand discrete set of Adaptive Coding and Modulation (ACM) schemes, so proposed solu-tions should be approximated, usually diverging from optimum. For all these reasons,the problem should be reformulated in the discrete case, originating the so-called FiniteTones water filling case [36]. In our work, we try to find a solution applying greedyheuristic algorithms, which are proved to be fast, but affected by some problems [37].In the common procedure, a subchannel is assigned to the user exploiting it in the bestway. However, maybe we are affecting system performances or damaging another userwhich would select the same subchannel because it represents the best choice, or eventhe only choice, among his set. In other words, it is not sure that the solution is optimal,but the allocation is given in polynomial time.

Water filling is suitable in the single user case or in a multiuser scenario with fixed

4.3. Diversity and adaptation 55

resource assignment. In the dynamic case we can exploit multi user diversity, that isthe variability of channel condition with frequency and user, serving mobile stationsalternatively.

Multiuser detection, water filling and allocation algorithm will be explained in Chap-ter 6 using a WiMAX technology reference scenario.

Chapter 5

Efficient retransmission resource allocation:Hybrid ARQ case

5.1 Introduction

Wireless networks are requested to provide high bandwidth services showing lit-tle delay tolerance to a growing number of users. The total amount of traf-

fic served by a cell entails an efficient management system and an allocation strategyavoiding waste of resource with the final objective of maximising the number of satis-fied users. In this chapter we go deep in the protocol stack focusing on the Hybrid ARQstrategies implemented in layer two. Despite the original idea is not so recent, HARQmechanisms have been successfully applied to data communications recently becausethey appear to be very promising from the efficiency point of view. The idea is to evalu-ate by means of simulations the impact of a retransmission strategy, improved with anefficient coding scheme, on system performances. Since error recovery is fundamentaldue to radio channel nature, smart allocation of resources dedicated to this task mayhave a great influence on perceived QoS. Moreover, HARQ is performed near the airinterface, so it is easy to think to an interchange of information between layer one andtwo, in a promising cross layer approach. For this reason we propose some interestingsolutions to simplify simulation tools, in particular regarding feedback information andprobability of error computation. We considered the new Release 7 of UMTS, usuallyalso referred to as Long Term Evolution (LTE), whose technical documentation is still indevelopment, but some basic issues have been already established, such as a new radiointerface and the central role of HARQ.

5.2. UMTS Long Term Evolution: system overview 57

5.2 UMTS Long Term Evolution: system overview

When UMTS broke into the market, a new communication paradigm was introduced,greatly different from GSM/GPRS and, in general, second generation cellular systems.Wideband CDMA (WCDMA) access technology can guarantee soft degradation perfor-mances, so there is a considerable flexibility gain with respect to FDMA/TDMA previ-ous solutions. A mobile user, under certain conditions, can afford to be connected upto 384 Kb/s, comparable to a wired home access to the Internet. Optimisation of mod-ulation and coding schemes and extension of MAC functionalities provided by HSDPA(High Speed Downlink Packet Access) [38] and HSUPA (High Speed Uplink PacketAccess) [39] move ahead maximum connection speed for data transmissions. However,new radio access technologies emerged, in particular OFDM (Orthogonal Frequency Di-vision Multiplexing), so 3GPP is defining a new UMTS release called LTE (Long TermEvolution), to assure competitiveness for at least 10 years [40]. The main objectives are:

• data rate increased to 100 Mb/s in downlink direction and 50 Mb/s in uplinkdirection;

• enhanced total cell throughput;

• increase spectrum efficiency;

• reduce round trip time in access network to a maximum value of 10 ms;

• maximise flexibility in spectrum allocation.

Most relevant changes are introduced at the physical layer, while radio protocol stackin UTRAN is basicly the same.

5.2.1 Physical layer and OFDM

OFDM consists in splitting the available bandwith in orthogonal subcarriers, transmit-ting simultaneously N data symbols rather than transmitting one symbol using thewhole spectrum. Orthogonality is important also to increase the number of subcar-riers in a given interval because they can be overlapped without loss of information.Supposing that the transmission occupies a time interval equal to 4t, the subcarrier


spacing should be 4f = 1/Tu = 1/4 t. In other words, subcarrier frequencies are har-monics and they can be written as fi = f0 + i4 f . Usually subcarriers are grouped insubchannels, however the channel response can be considered flat.

Theoretically, this system can be built with a serial to parallel converter subdividingthe high data rate flow in N low data rate flows, each one transmitted with a differentorthogonal subcarrier, as in Figure 5.1. It is easy to see that such a system is difficult

Figure 5.1: Ideal realization of an OFDM transmission.

to realize and to maintain, in particular for the trimming of local oscillators. The sameresult can be obtained substituting the bank of multipliers with an Inverse Fast FourierTransform block, as in Figure 5.2. This block has been optimised from the software pointof view and it is also available as a stand alone electronic component. A modulator

Figure 5.2: OFDM using IFFT.

produces a stream of modulation symbols with rate equal to N/Tu, then shaped into Nparallel symbol streams with a rate of 1/Tu each, where Tu is the useful OFDM symbolduration, defined later. The N point IFFT block acts as the N parallel multipliers seenin the previous realization and it creates discrete-time complex values, called OFDMsamples, treating each input line as the weight associated to spectral components of


a signal. Samples are then serialised, making up the useful OFDM symbol, the timedomain signal corresponding to the IFFT window [41], whose duration is the quantityTu, already used in this paragraph.

The OFDM symbol duration Tu is greater than the modulation symbol period, Tu/N ,so multipath fading effects are reduced, also simplifying receivers and equalisers. How-ever there is an upper limit to Tu: it should be smaller than the coherence channel time,in contrary, the channel varies heavily during the symbol transmission. The orthogonal-ity of the subcarriers reduces the Inter Carrier Interference (ICI), but the time dispersivenature of the channel yields to Inter Symbol Interference (ISI). A valid countermeasureto this problem consists in adding a guard interval Tg, also measured in terms of num-ber of samples, to the OFDM symbol, so the total duration is Ts = Tu + Tg, as we cansee in Figure 5.3. In LTE the guard interval is a cyclic prefix added at the beginning

Figure 5.3: Generation of cyclic prefix in an OFDM symbol.

of the symbol, replicating the ending portion of the symbol itself. However, the maindrawback of OFDM is sensitivity to frequency and phase errors, in particular the fre-quency offset between the transmitter and the local oscillator at the receiver bringingthe subcarriers down to base band. In fact, due to this offset, subcarriers are no moreorthogonal and there are multiple contributions to the baseband signal, despite greatpart of the energy comes from the desired frequency. Another important issue is themisalignment in clocks used at the transmitter and receiver side, varying the durationof the OFDM symbol, resulting in a phase rotation.

Most important changes involve the physical layer [42], in particular:

• the time transmission interval, TTI, is reduced to 0.5 ms;

• the subcarrier spacing, 4f , is 15 kHz and the inverse value is the useful symbolduration, Tu;

• the total bandwidth varies from 1.25 to 20 MHz, with intermediate solutions in-cluding 2.5, 5, 10 and 15 MHz. In order to use the highest values it is necessary toconsider overlapping frequencies with actual technologies.


• according to the bandwidth, also the IFFT block size varies: possible values are128, 256, 512, 1024, 1536 and 2048. The corresponding number of occupied sub-carriers is 76, 151, 301, 601, 901 and 1201, respectively. Lacking subcarriers areemployed for frequency spacing or as pilot tones;

• subcarriers cannot be assigned singularly, in contrary they are grouped in resourceunits. A resource unit (RU) is made up of 25 subcarriers and it is the minimumallocation unit that can be assigned to a user for a TTI. A graphical representationin the time-frequency domain can be found in Figure 5.4: this simple definitionallows a refined resource management.

Figure 5.4: Time-frequency representation of a Resource Unit.

5.2.2 MAC layer and Hybrid ARQ

The protocol structure is basically the same proposed in previous versions, but the NodeB has enhanced features. It is provided with retransmission capabilities (see Figure 5.5)located in the MAC High Speed (MAC-HS) entities, whose number is equal to the num-ber of users served by the cell. A MAC-HS block performs HARQ, so it is possible tolimit the round trip time in the access network to a maximum value of 10 ms. In fact, intraditional ARQ techniques, PDUs are generated, stored and retransmitted at the RLCblock, which is located in the Radio Network Controller, so quite far from the radiochannel. However, if the MAC-HS entity fails the retransmission for a certain num-ber of attempts, usually 3, the PDU is sent again at the RLC layer. HARQ is obtainedmixing ARQ mechanism with forward error correction (FEC) protection coding in or-der to avoid waste of resources during retransmissions. As a matter of fact, protectinguser information with a FEC code allows to minimise the number of retransmissionswhile an ARQ strategy permits to limit the redundancy introduced with respect to aFEC-only solution. This approach is usually known as Type I ARQ [43], while Type IIHARQ can be furtherly subdivided in Chase Combining (CC), or diversity-combining,


Figure 5.5: Protocol structure for LTE environment.

and Incremental Redundancy (IR), or code-combining. In CC, the sender receiving anegative acknowledgement retransmits the whole packet, applying the same FEC pro-tection (rate and coding scheme) such as in a traditional ARQ approach. The differencewith respect to Type I consists in the fact that the receiver does not discard erroneousblocks, they are rather stored in a buffer to be compared with replicated ones. In thisway it is possible to take advantage of energy and diversity gain. In IR, each retrans-mission is different from the previous one because additional redundant information isgiven, for example, simply varying the puncturing scheme. In this case the block is self-decodable because systematic bits are included, however it is possible also to add onlyredundancy bits to save much more resources, adopting a so-called Type III scheme [44].The main HARQ advantage is the possibility to allow a higher block error using highmodulation order and code rate, reflecting in a relevant throughput gain if the channelconditions are good. The counterpart is the need for a reliable feedback channel andvariable delays in case errors occur.

The retransmission scheme adopted in HARQ version for LTE is different from a Se-lective Repeat one, but it guarantees the same level of resiliency through multiple Stopand Wait (SaW) instances. As we can see from Figure 5.6, each time a user is given thepossibility to transmit, a new SaW process is started. If the same user can transmit alsoin the following TTI, the previous SaW instance is not completed because the acknowl-edgement has not come back yet, so a new process has to be run, in order to avoid waste

5.3. Problem formulation 62

Figure 5.6: Hybrid ARQ retransmission scheme.

of resources. The maximum number of parallel instances should be enough to allow acontinuous transmission, in other words it is necessary to guarantee, at least, the returnof one acknowledgement, to reuse that process for new data, without arising ambiguityin attribution of feedback messages. 3GPP estimated that the medium round trip timefor an acknowledgement is covered by 6 TTI, so the maximum number of instancesis 8, allowing a security margin. In case a positive acknowledgement is received (theblock marked with “+” in the receiving side of Figure 5.6), the associated SaW processbecomes free and it can be used for a new transmission. If the acknowledgement is neg-ative (the gray block marked with “-” in Figure 5.6), the erroneous PDU is retransmittedby the same process, the first time the user has the possibility to use the channel. Theseschemes can work thanks to a control channel transmitting two informations: the iden-tifier of the users that are going to use the air interface in the next TTI and the identifierof the SaW process involved for each user.

5.3 Problem formulation

For error correction coding, usually convolutional or turbo codes are used, but, unfor-tunately, only a discrete set of coding rates can be used, so the efficiency is a staircasefunction of the channel quality. From this point of view, efficiency can be further in-creased replacing FEC with fountain codes [45]. In this type of codes, the source splitsdata to be sent in packets which are coded and then transmitted in a continuous way,from which the analogy with a water fountain. The receiver wishing to correctly decodethe stream has to collect a certain number of packets, like collecting drops with a bucket.The exact number of packets needed depends on the channel conditions and it can notbe calculated at the beginning of the process, in contrary it is determined on the fly.Fountain codes are rateless in the sense that the number of encoded packets generatedby the source message is potentially limitless.


In this work we focus on a generic fountain code, adapted to work in an UMTS LTEsingle cell environment where the digital fountain source is the Node B, since we areinterested in downlink traffic. Fountain coding is applied to a Type III IR scheme in asimple and traditional way, protecting PDUs only with the needed redundancy, accord-ing to the channel state. In case of unsuccessful decoding, in order to reduce the codingrate, only a relative small quantity of redundancy bits is sent, calculated exploiting thefeedback information. Since we are interested to a simulative approach, the encodingand decoding chain has not been implemented, in contrary these two equations [46] areused to calculate the channel capacity and the probability of error:

C(h) =1

n

n∑i=1

log2(1 + γ|hi|2), where h = (h1, h2, ..., hn)

Pout = P

(k

n> C(h)

)

The first formula represents the channel capacity, that is the maximum supported rate,and it depends on the Signal to Noise (SNR) ratio, γ, and the channel state information,the vector h. So, when a user transmits a code word of n bits, containing k informationbits, if the coding rate k/n is higher than the channel capacity, the decoding operationfails and the theoretical lower limit of outage probability is given by the second equa-tion. The main idea to simplify simulations is to introduce the outage probability oferror at the receiver side to decide wether the block is correct or not. Then the channelcapacity is used as a feedback information provided by the receiver to the sender inorder to

• calculate the coding rate of the new block, if the acknowledgement is positive;

• determine the redundancy bits needed to push the coding rate below the channelcapacity, if the acknowledgement is negative.

Note that the resulting coding rate of a block should be kept as close as possible to thechannel capacity for efficiency reasons. In facts, with traditional FEC techniques, it isnecessary a tradeoff between desired rate and outage probability: to increase reliabilitya rate reduction is requested, in contrary to improve the rate a larger outage probabilityis needed.


Great flexibility is given by the fact that PDUs have variable size, so they can be adaptedto user’s needs and to channel conditions. In particular, as we can see in Figure 5.7, if

Figure 5.7: PDU generation using Fountain Code approach.

the channel conditions are good all the available bits in a RU are employed to transportuser data, without protection. Since the subcarrier spacing is 15 kHz and 4f = 1/Tu,after the multiplexer, but before the IFFT block in Figure 5.2, there are N parallel streamscarrying 15,000 symbols of modulation per second. These streams are then mapped insubcarriers by the IFFT block: since a RU is made up of 25 subcarriers, each user cantransmit in a TTI 187.5 symbols of modulation. According to the three possible modu-lation techniques, QPSK, 16 QAM and 64 QAM, the number of bit assigned to a RU is375, 750 and 1125, respectively. If the channel conditions get worst, the channel capac-ity is used to calculate the number of redundancy bits, however the allocation can bechanged every TTI, that is every time the feedback information is provided. If the de-coding operation is unsuccessful even in this case, only a portion of the new RU is usedto add redundancy to lower the coding rate of the erroneous block. Remaining bits areemployed for a new transmission, protected according to the feedback information. Aswe can see in Figure 5.7 when calculating the total coding rate of a block it is necessaryto take into account all the redundancy contributions, also those added in TTIs after thetransmission of systematic bits.

This new approach is compared to a traditional Chase Combining one, where PDUshave fixed size, selecting one of the modulation and coding scheme listed in Table 5.1.Starting from the proposed number of bits in a RU for the fountain code case, the refer-ence coding rate is 2/3 because the other code rate available, 1/3, penalises the trans-mission from the efficiency point of view. For example, a QPSK modulation carries 375bits but only 125 bits, that is one third, contain useful informations, the others are usedfor protection. Moreover we have also to take into account various headers (2 bytes


Eb/N0 Coding rate Res. Unit PDU per TB Modulation-0.52 1/3 2 1 QPSK3.15 2/3 1 1 QPSK4.7 1/3 2 2 16 QAM6.7 2/3 1 2 16 QAM

7.35 1/3 2 3 64 QAM8.65 2/3 1 3 64 QAM

Table 5.1: Modulation and coding scheme for traditional Chase Combining.

for RLC, 1 byte for MAC and 21 bits for MAC-HS), so the payload is only 10 bytes. Incontrary, with the coding rate set to 2/3 and considering again a QPSK modulation, thepayload is 25 bytes. These considerations justifies the choice to fit a 2/3 coded PDU ina RU, consequently a 1/3 coded block is transmitted using 2 RUs because the numberof redundancy bits is doubled, as we can see in the third column of Table 5.1. This stillremains valid also for higher order modulation (16 and 64 QAM), however the numberof RUs composing a transport block (the minimum quantity of information protected bythe turbo encoder) is increased, simply because the channel is behaving in a better way.Finally, the most important information is in the first column, where we can find theminimum SNR requested to obtain a BER equal to 10−6 for the given modulation andcoding scheme. The transmission performance is calculated for every transport block,however it is necessary to introduce a corrective factor for the coding gain. For example,the probability of error in case of QPSK modulation is

BERQPSK =1

2erfc

(√2 · Eb

N0

)

The feedback information is the SNR value, then used to choose the modulation andcoding scheme.

The simulation tool developed consists of the four blocks represented in Figure 5.8The first one is a traffic generator, generating packets according to some common pro-files like CBR (Constant Bit Rate), voice activity, with exponential distribution, or webtraffic, with pareto distribution. The second one is the scheduler which has to select theuser to be served using round robin or proportional fair algorithm. Moreover, at thispoint packets are split in PDUs for transmission over a fading channel. For each active


Figure 5.8: LTE single cell simulator: main components.

user in a TTI the SNR is calculated considering received power, noise power, inter andintra cell interference, as in the following equation:

Eb

N0

=Prx

PNoise + IIntra + IInter

In case of traditional HARQ, from the SNR it is derived the probability of error, whilein the other case the channel capacity is computed. The main contribution to the atten-uation of the signal during propagation is given by pathloss, with a mean value givenby:

mean value[dB] = 128.1 + 37.6 · log10(d)

where d is the distance, in kilometers, from the base station to the user. Deviation fromthe mean value is due to fading, modeled as a lognormal random variable with meanvalue equal to 0 dB and variance from 4 to 8 dB. Usually shadowing is evaluated startingfrom a normal distribution N with 0 mean and variance σ, because this kind of randomvariable is easier to manipulate and to implement, according to

S = C(d) · S +√

1− [C(d)]2 ·N(0, σ)

5.4. Simulations and results 67

where S is the value of the shadowing in the previous TTI and C(d) is the correlationfunction. Since the variance is quite high and we are considering pedestrian users,shadowing changes in a steep way from one TTI to the adjacent, so it is necessary touse a correlation function for smooth transitions:

C(d) = 2d

dcorr

where dcorr is the correlation distanche. The inter-cell interference is represented as anormal random variable with mean value equal to -75 dBm, variance 0.333 dBm and acorrelation coefficient of 0.5, used to calculate effective variance

σeff =√

σ2 · (1− ρ2)

Finally, the intra-cell interference is considered negligible because there is not loss oforthogonality.The last block has to decide whether the PDUs are correctly received or not, providingthe feedback information, namely channel capacity or SNR.

5.4 Simulations and results

Starting from the simulation tool described in the previous section, we propose someresults [47] showing advantages of fountain codes. Simulations parameters are sum-marised in Table 5.2 where slow fading variance has been fixed to 5 dB, supposing amoderate dense urban scenario for pedestrian users moving at a speed of 3 Km/h. Thetotal number of RUs, that is the total number oftransmitting simultaneously in a TTI is48, corresponding to 20 MHz total bandwidth. The total transmitting power availableat the base station is 25 W, equally distributed among the users in the downlink direc-tion. This solution is far from optimality, however this is out of the scope of this work.The total number of users in the cell is 150, each one transmitting at a rate of 350 Kb/s:the cell is heavily loaded and the traffic conditions are not very realistic, however thissituation allows to achieve maximum performances.

First of all, in Figure 5.9 we propose a statistics on PDUs coding rate As expected, the

5.4. Simulations and results 68

Slow fading, mean value 0 dBSlow fading, variance 5 dBMobile speed 3 Km/hNumber of subcarriers per RU 25Number of RUs per TTI 48Total TX power 25 WNumber of users 150Bit rate per user 350 Kb/s

Table 5.2: Some interesting simulation parameters.

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

% o

f PD

Us

Coding rate

Rateless coding

Chase combinig

Figure 5.9: Cumulative distribution function of coding rate.

traditional HARQ scheme (red line) can exploit only two values, 1/3 and 2/3, while, inthe other case, the coding rate varies continuously following in a close way the channelcapacity.Figure 5.10 is much more interesting because it shows the total throughput in the twocases. At the beginning, with low loaded system, the two approaches have same per-formances, but, after all users have entered the system, fountain code outperform thetraditional coding techniques. This is due to efficient use of bits available in a RU andto the introduction of the needed redundancy. Finally, same conclusions can be drawnlooking at the statistics of transmission delays in Figure 5.11. Using fountain codes agreat percentage of PDUs, more than 70 % compared to 40 % in the other case, bene-

5.5. Conclusions 69

0 2 4 6 8 10 12 14 16 18 20−1

0

1

2

3

4

5

6

7x 10

7

time (s)

Goo

dput

(bi

t/s)

Rateless codingChase combining

Figure 5.10: Base station goodput versus time.

fit from the minimum delay simply because no time is wasted in adding unnecessaryprotection.

5.5 Conclusions

In this chapter we have considered benefits of a more efficient coding techniques in theHARQ scheme of UTRAN LTE. In particular, we suggested to use the outage probabilityto evaluate the probability of error in a simulation environment. Using a simulationapproach, we have seen that, despite the heavy traffic conditions supposed, it is possibleto keep under a reasonable threshold the transmission delay for the majority of PDUs.

Moreover, we stressed the role of the channel capacity as feedback information, car-rying informations from the physical layer to the link layer, where HARQ is performed.It is important to note that HARQ techniques are gaining more and more relevance, infact they are planned not only in next UMTS release, but also in other recent commu-

5.5. Conclusions 70

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.020

0.2

0.4

0.6

0.8

1

delay [s]

% o

f PD

Us

Chase CombiningFountain Code

Figure 5.11: Cumulative distribution function of transmission delays.

nication systems, such as WiMAX or HSDPA. So, beyond these particular results, wecan say that a cross layer approach, referred to as exchange of informations betweenadjacent layers, can play a relevant role in efficiency improving and, consequently, inperformance enhancement.

Chapter 6

Example of a complete resource allocationframework

6.1 Introduction

Inthe previous chapter we analysed a particular aspect related to resource allocation,in particular we dealt with performances and simulation introduced using an efficient

protection coding scheme. In this chapter we extend our point of view to a whole cell,taking into account also the surrounding conditions. The base station and the mobilestations are equipped with WiMAX: our choice is due to the fact that it is a promisingtechnology relying on OFDM, so we can benefits from all the advantages described inChapters 4 and 5. Moreover, the standard describes only the functions of schedulingand allocation blocks, so manufacturers has a high grade of freedom and researchersare encouraged to suggest innovative solutions. We proposed and compared four al-gorithms, two of them being based on water filling theory, taking into account also theinfluence of some deployment factors such as frequency reuse and antenna directivity.

6.2 Technology overview

WiMAX means Worldwide Interoperability for Microwave Access and it is the commer-cial name of IEEE 802.16 compliant products. WiMAX manufacturers gathered in anon-profit organisation, WiMAX Forum [48], to promote and to assure interoperabilitybetween devices. The standard proposes a wireless wide band access technology formetropolitan area networks with high coverage and high datarate, so it can play var-ious roles thanks to its flexibility. In Figure 6.1 we can see one of the most promisingapplications of this technology, that is how WiMAX can provide wireless broad bandaccess, acting as wireless service provider backhaul, in alternative to wired solutions

6.2. Technology overview 72

Figure 6.1: WiMAX scenario in a metropolitan area.

or to face the digital divide problem. In some rural and low density populated areas itis not convenient for an ISP to place fixed infrastructures, the investment return timewould be very long. The service access can be direct, through WiMAX capable devicesinstalled in buildings, or mediated by WiFi access points. Moreover, IEEE developed amobile extension, so also handheld device will be able to benefits from this technology.

The standardisation activity started in late nineties and the first release goes back to2001, operating in line of sight conditions at a maximum distance of 5 Km, as we can seefrom Table 6.1. Next one, made available in 2003, operates at lower frequencies but in

802.16 - 2001 802.16a - 2003 802.16e - 2006Frequencies [GHz] 10 - 66 2 - 11 2 - 6Conditions LOS NLOS NLOSMobility No No YesBandwidth [MHz] 20, 25, 28 1.75 - 20 1.75 - 20Modulation QPSK,16QAM,64QAM QPSK,16QAM,64QAM QPSK,16QAM,64QAMSupported rate [Mbit/s] 134, 28 MHz chan. 75, 20 MHz chan. 15, 5 MHz chan.Maximum distance [Km] 5 50 5

Table 6.1: Summary of WiMAX evolution.

non-line of sight conditions, the rate is reduced, covering a wider are and, finally, thereis the introduction of OFDM. All the aspects considered so far were finally merged in aunique release, 802.16d or 802.16-2004 [49], nowadays the reference for WiMAX prod-ucts placed on the market and certified by WiMAX Forum. In 2006, with 802.16e [50],also mobility aspects are taken into account, such as handover and roaming manage-ment. Note that WiMAX Forum selected only twelve profiles for both fixed and mobileservices, employing 256 subcarriers OFDM solution, TDD or FDD duplexing and TimeDivision Multiple Access (TDMA). In 3.4-3.6 and 3.6-3.8 GHz band there are two possi-


ble channels, 3.5 and 7 MHz, while in 5.4-5.7 and 5.7-5.8 GHz band channel width is 5or 10 MHz. 802.16e mobile profiles are suitable also for 2.3 and 2.5 GHz band, howeverSlotted Orthogonal Frequency Division Multiplex Access (SOFDMA) replaces TDMA.

Like other IEEE 802 standards, also in this case there is only the definition of theair interface, in fact, in Figure 6.2 we can see that the protocol stack is limited to thefirst two ISO/OSI layers. Starting from the top, the first block is the Service Specific

Figure 6.2: WiMAX protocol stack.

Convergence Sublayer (SSCS), providing mapping and adaptation for data related toexternal networks. In particular, two convergence sublayers (CS) are defined, one forATM communications, the other for packet services such as IPv4, IPv6, Ethernet and vir-tual LANs. In both cases the sequence of operations required is very simple, consistingin classification, header suppression (if needed) and forwarding to lower levels. Clas-sification in ATM CS is based on virtual channel and virtual path indicators which aremapped using connection identifiers (CID) associated to wireless connections. How-ever, it is not possible to use all the ATM virtual channel and path combinations dueto the limited size of the connection identifier. The presence of header suppression isindicated by MAC management procedures and it allows to save bandwidth keepingfew bits of the original 5 bytes ATM header.As for the ATM case, packet CS classification allows to map an incoming packet in a ser-vice flow, that is a connection between MAC entities, where pairing criteria depends on


higher level protocols, scheduling priority and connection identifier. In case of headersuppression, fields to be removed are decided by means of a negotiation procedure andthen associated to a profile, with the main task of restoring the missing informations.It is obvious that there is a strong connection between this profile and the data channelcarrying traffic and it is not possible to change this situation dynamically, in contraryanother negotiation procedure should be started.

Next block encountered in the downlink direction is MAC Common Part Sublayer(CPS), covering the classical MAC function of regulating the access of the users to the ra-dio channel. Two topologies are supported by WiMAX, mesh and point-to-point, how-ever we are interested only in the second one, involving the base station (BS) also forcommunications between subscriber stations (SS). Usually, BS transmits in broadcast,so a SS receives all PDUs unless a portion of the subframe is explicitly reserved and the16 bit CID is used to identify data blocks of interest, while the other ones are discarded.According to the previous explanation, WiMAX MAC layer is connection oriented, evenwhen transferring non-oriented traffic, like IP datagrams. All the stations, BS and SS,are supposed to be capable of performing some basics operations such as concatenationand fragmentation to create fixed or variable length PDUs, adding a packing subheaderin case small packets are merged in a single PDU. Error detection is committed to a 32bit CRC code appended to the end of the PDU, while ARQ correction is optional. ARQis negotiated during the connection set up, in other words it is not possible to have pro-tected and unprotected flows sharing the same CID.MAC CPS has also scheduling functions to support different QoS parameters for eachdata flow assigned to a connection. There are management messages to dynamicallyrenegotiate parameters during the transmission and four possible scheduling services,namely Unsolicited Grant Service (UGS), Real-Time Polling Service (rtPS), Non Real-Time Polling Service (nrtPS) and Best Effort (BE). The first one, UGS, is designed to sup-port real time services with packets having constant length and periodic arrival time,for example uncompressed VoIP. If packet size is not fixed, but periodical arrivals areexpected, such as in a MPEG stream, the most appropriate class is rtPS. Best support fordelay-tolerant applications requesting a minimum rate is granted by nrtPS class whileBE services are employed in all other cases, when performance assurance is not needed.Finally, 802.16e assigns to MAC CPS mobility support, defining handoff procedures andpower saving strategies. In particular, a SS can operate in regular mode, idle mode orsleep mode. In idle mode a device is not associated to any BS but can receive broadcast


downlink messages, without performing handoff operations (frequency scanning is ahigh energy consuming operation). During sleep mode the SS is not available for anytype of traffic, uplink or downlink, and the duration of this period is negotiated withthe BS.

Before the physical medium, data flowing down encounters the Security Sublayer,added to face privacy and authentication problems. Since the air interface is sharedamong all users, it is important to avoid unwanted data access by other mobile equip-ments and to assure that only authenticated users can join the base station. Informationsto be transmitted are protected using a digital X.509 certificate and the RSA public keyalgorithm while traffic keying material exchange is committed to a dedicated key man-agement protocol. All the informations employed to secure the communication betweena SS and the BS are indicated as Security Associations: they can be static, for exampleduring initialisation, or dynamic, being a part of the service provided, but their validityperiod is time limited. The SS is supposed to ask for a renewal, in contrary another asso-ciation should be requested, howevere the BS periodically refresh security informationsto preclude resource misuse.

Finally, at the bottom of the stack there is the physical layer: depending on the op-eration condition (LOS or NLOS) and the frequency band, basically two versions canbe implemented. The first one exploits a single carrier transmission and is employedespecially in fixed applications, while we are interested in the other version, for mo-bile purposes, based on OFDM. In Chapter 5, Section 5.2.1 OFDM main concept, ad-vantages and drawbacks were already explained, so only some peculiarities are shownhere. We already said that OFDM is multicarrier transmission technique, but, to sim-plify allocation management, subcarriers are grouped in subchannels. Depending onthe permutation scheme, the tones making up a subchannel are adjacent or sparse. InDiversity Permutation, subcarriers assigned to a subchannel for downlink transmissionare selected in a pseudo random way, for example according to the scheme in Figure6.3. A subchannel is composed of two clusters, defined as 24 data subcarriers plus 4pilot tones. The same scheme is applied in the uplink direction, but the cluster is sub-stituted by the tile, equal to 8 data subcarriers and 4 pilot tones. Tiles are then groupedsix by six to form a slot. In a Continuous Permutation scheme all the subcarriers in a sub-channel are adjacent and the reference unit is the bin, equal to 8 data subcarriers plusone pilot tone. Bins are arranged in frequency and time, or symbol, domain to build upa slot using different combinations, some of the shown in Figure 6.4, so, subchannel is

6.3. System model 76

Figure 6.3: Example of a diversity permutation scheme in downlink.

Figure 6.4: Some examples of Continuous Permutation slots.

defined as the frequency occupation of the slot. We can see that it is possible to reducethe frequency extension of a subchannel spreading the slot over a higher number ofsymbols. Diversity permutation is suitable for high mobility environment because it al-lows to mitigate intercell interference exploiting frequency diversity, while continuouspermutation is applied to fixed and low mobility scenarios.

6.3 System model

Once again we consider traffic problems in the downlink direction with a simulative ap-proach and in this section a description of the base station model will be given. First ofall, the BS has a crucial role to exploit Multi User Diversity (MUD), a fundamental gaincontribution in modern wireless multiuser networks. Due to shadowing and multipatheffects, users experience different channel quality during a certain observation period,moreover it is reasonable to suppose that frequency responses are independent, so effi-ciency can be increased giving the possibility to transmit to the stations with best radiolink conditions. In Figure 6.5 (a) and (b) [51] we can see the channel quality of 2 userswhile in Figure 6.5 (c) it is summarised the allocation made by the BS: a subchannel isassigned to the user exploiting it in the best way. Multidiversity gain is proportional to


(a) Frequency response of user 1 (b) Frequency response of user 2

(c) Allocation exploiting MUD

Figure 6.5: Example of Multi User Diversity gain. Dotted lines are SINR, histograms are theaverage values for subchannels.

the number of users in the system because there is a greater probability to find at leastone station achieving good performances, increasing the set of channel combinationsamong which the BS can choose. A Continuous Permutation scheme is needed to ob-tain homogeneous values on a subchannel, on the contrary, with the other scheme, allbenefits are overridden. If subcarriers are chosen in a pseudo random way, all subchan-nels are very similar from the quality point of view [52] and there is no diversity.

The base station model is shown in Figure 6.6, where we can see that the core ofthe simulator is the closed loop between the Allocator and the Scheduler. The schedulerdistributes transmission resources between streams while the allocator assigns radioresources (power, modulation, coding and slot) to the users, similarly to a model de-scribed by [53] and [54]. In particular, the allocator receives bandwidth requests fromthe scheduler and SINR (signal to interference and noise ratio) from the channel estima-tion block and it uses these informations to control Adaptive Coding and Modulation(ACM) block. The allocator needs some informations from the physical layer to accom-plish its task, but, in the same time, it is located at the MAC level, so we are proposinga cross layer approach to this topic.

In general, the scheduler is aware of all the traffic flowing through the cell and it has


Figure 6.6: Base station model.

the possibility to mitigate congested situations, however we are not interested in eval-uating different scheduling disciplines. We do not implement a full-capability block,we prefer to simplify the simulation using a scheduler emulator. Supposing that thebuffers are full and each user has always a packet ready to be sent when he is giventhe possibility to transmit, the scheduler has only to generate traffic requests. These re-quest are made in term of minimum and maximum bandwidth, as shown in Figure 6.7.Minimum band request is the sum of delay sensitive traffic, UGS and rtPS, while max-imum band request is calculated as minimum request plus non-delay sensitive trafficand allocation residue. Delay insensitive traffic includes best effort and nrtPS, whereasthe allocation residue is the delay insensitive traffic not sent during the previous frame.If the scheduler is unable to satisfy all delay tolerant traffic request it is possible to use abuffer and try to transmit during next frames. This solution cannot be applied to delaysensitive traffic, so it is discarded if it not transmitted immediately.

Scheduler’s requests represent, together with channel state information, the inputparameters for the allocator which decide how to distribute radio resources amongusers, indicating for each of them the modulation and coding scheme to employ. Avail-able modulations include QPSK, 16 or 64 QAM and the coding rate varies from 1/2 to3/4. It is also possible to fall back upon repetition coding, that is filling a frame withmultiple copies of encoded data, in case the noise is heavily affecting the transmission.


Figure 6.7: Minimum and maximum bandwidth requests generation.

The allocator is supposed to solve an integer value optimisation problem consistingin assigning the slots of a Ns by Nt matrix to a subset of the Nu users connected to theBS, with Ns indicating the number of subchannels and Nt the number of time symbolsin a frame. Defining χk,s,t as a binary matrix where a value equal to 1 means that slot (s,t) is assigned to user k, the allocation maximising the throughput is

χ = arg maxχ

Nu∑

k=1

Ns∑s=1

Nt∑t=1

χk,s,t · bk,s,t

with the following limits:

1 ≤ k ≤ Nu

1 ≤ s ≤ Ns

1 ≤ t ≤ Nt

The number of bits the user k can transmit in the slot (s, t) is bk,s,t and it depends on thetransmission power pk,s,t, that is bk,s,t = f(pk,s,t). Referring to boundary conditions, first


of all, we have to impose that it is not possible to assign multiple slots to a user

Nu∑

k=1

χk,s,t ≤ 1 ∀(s, t)

then we have to take into account bandwidth and power constraints. Allocated band-width for each user k is delimited by requests coming from the scheduler in terms ofminimum and maximum values, Bmin

k and Bmaxk , respectively

Ns∑s=1

Nt∑t=1

χk,s,t · bk,s,t ≥ Bmink ∀k

Ns∑s=1

Nt∑t=1

χk,s,t · bk,s,t ≤ Bmaxk ∀k

Power constraint can be written as

Nu∑

k=1

Ns∑s=1

χk,s,t · pk,s,t ≤ Pmax ∀t

where Pmax is the maximum power allowd at the BS.As we said, the allocation problem requires integer solutions, but they are difficult

to find due to complexity of integer programming algorithms. On the other side, lin-ear programming algorithms are simpler and they provide continuous solutions, but itshould be difficult to bring back these results in the integer domain. A simple approxi-mation gives only an upper bound because the resource allocated are grater than thosereally needed. For these reasons, the four algorithms proposed [55] work over integervalues.

Algorithm A works at a constant transmission power, equal to the maximum pro-vided by the BS, evaluating the quality of each subchannel on the basis of the maximumbitload calculated by the ACM block. The algorithm, first of all, tries to satisfy min-imum bandwidth constraints, then it considers exceeding requests, till the maximumband value. The main procedure consists in using SINR, that is the achievable bitrate,


to rank all the subchannels, then each subchannel is assigned to the user realizing thebest performance on it in terms of supported traffic. It is possible to obtain fast execu-tion with maximum throughput because the best matching between user, channel andbitrate is pursued. However, the allocation is unfair because users with poor channelconditions are placed at the bottom of the ranking, so their situation cannot improve.

Algorithm B is very similar to the previous one, the only difference is that subchan-nel ranking is not generated using the maximum achievable bitrate, but an efficiencyfunction is introduced

ηs =bk,s∑Nu

k=1 bk,s

1 ≤ s ≤ Ns

The equation describes the efficieny obtained allocating bk,s bits to user k in subchannels. This value tends to zero if a lot of users have good performances on this channel whileit tends to one if only few users can succesfully transmit on that particular frequency.

Algorithm C introduces the possibility to vary the transmission power for each sub-channel, so classification relies upon SINR rather than on bitrate (remeber, however,that the two values are related). In particular, users are arranged in a table on the ba-sis of the avarage SINR they experience over all subchannels and this will be the orderfollowed to assign resources. In other words, best users are served first, to guaranteethroughput maximization. To calculate the avarage SINR, a SS evaluates the quality ofeach subchannel: this list is not discarded, in contrary it is employed by the algorithmduring the analysis of user’s situation. In fact, tha allocator assigns to a SS the bestsubchannel it has reported, if still unused; in contrary the list is scrolled down until anavailabe band is found. If the operation is unsuccessfull the algorithm switches to thenext user, checking if there are again free resources. Finally, power control is applied,but with a limited step of plus or minus 3 dB. The negative step is employed if a usercan transmit the same amount of bits and the power saved is redistributed to users withbad conditions to improve their performances. The objective is to keep the algorithm assimple as possible, but, in the same time, maximize system throughput without wasteof resources due to repetition coding.

Algorithm D adopts a power control scheme inspired by water-filling approach withfixed 3 dB step, ranging from -9 dB to +12 dB. The process described in the previous caseis repeated more times and during each iteration the algorithm tries a higher power

6.4. Allocation algorithm performances evaluation 82

profile. If a user reaches the maximum bandwidth constraint or the maximum ratesupported by the ACM scheme, its configuration is frozen, while the whole algorithmends when the BS is transmitting at the maximum power.

6.4 Allocation algorithm performances evaluation

We have evaluated the performances of these four algorithms by means of a series ofsimulations running for 250 frames and making the following assumptions:

• IEEE 802.16e version with 7 MHz channel profile;

• pedestrian radio channels;

• a traffic scenario with the presence of all QoS categories;

• the packet size is based on the Pareto distribution, except for UGS;

• a random bandwidth ranging from a minimum to a maximum value, according toTable 6.2;

• traffic generation only if the avarage SINR is at least 0 dB;

We also classify users in two categories, premium and ordinary subscribers, respectively20% and 80% of the population. The difference is that an ordinary subscriber can gener-ate up to 2 Mbit/s of traffic, while this value is raised to 4 Mbit/s for a premium user. In

QoS type UGS rtPS nrtPS BEPercentage 10 30 30 30Minimum bandwidth [kbit/s] 32 5 1024 10Maximum bandwidth [kbit/s] 64 2048 4096 2048

Table 6.2: Simulated traffic characteristics.

Figure 6.8 (a) and (b) we can find simulation results for 10 and 20 users with the trafficcomposition proposed in Table 6.2. Algorithms C and D outperform the other solutionswhile algorithms A and B are substantially overlapped. Moreover, we can note that theclosed loop between the scheduler and the allocator is very important, in fact algorithmD without this interaction provides very poor results. Simulations are repeated witha different mix of real time and delay tolerant network traffic with respect to Table 6.2

6.4. Allocation algorithm performances evaluation 83

(a) 10 users, 40% real time subscribers (b) 20 users, 40% real time subscribers

Figure 6.8: Allocation algorithms performances with 40% real time traffic [Horizontal axis: gen-erated traffic in Mb/s; vertical axis: allocated traffic in Mb/s].

(10% UGS, 10% rtPS, 30% nrtPS and 50% BE) and the results are shown in Figure 6.9 (a)and (b). We can see that higher performances are obtainable increasing delay toleranttraffic, as expected. In particular, with 20 users we can reach a throughput of 10 Mbit/s,very near to the theoretical throughput of a 7 MHz profile (13 Mbit/ss).

(a) 10 users, 20% real time subscribers (b) 20 users, 20% real time subscribers

Figure 6.9: Allocation algorithms performances with 20% real time traffic [Horizontal axis: gen-erated traffic in Mb/s; vertical axis: allocated traffic in Mb/s].

6.5. Impact of Frequency Reuse Factor 84

6.5 Impact of Frequency Reuse Factor

A simple way to deploy a cellular network is to assign to all cells the same frequencyband, but this yields to co-channel interference due to the poor distance. The obvioussolution is to subdivide the allocated spectrum in sub-bands, so adjacent cells work withdifferent frequenceis and co-channel interference is reduced thanks to a better isolation.It is equivalent to give each cell a color code such that neighbours cannot have the samecolor. The set of cells, with different colors, making up the original band represents acluster and the number of elements in this set is the frequency reuse factor (FRF). In Fig-ure 6.10 we can see some examples illustrating the distance between cells with the samecolor code an the FRF. It is trivial that co-channel interference is mitigated, but each cellcan use a smaller portion of spectrum, reducing the numbers of stations served [56].

Figure 6.10: Variation of distance between clusters with frequency reuse factor, indicated as N.

In our analysis on influence of network configuration we also considered the role ofdirectional antennas. In particular we compared a standard BS transmitting 10 W bymeans of an omni-directional 9 dBi gain antenna with a sectorialized cell, equippedwith a directional 16 dBi gain antenna supporting 2 W per sector. In the first case thechannel is 10 MHz, in the second one is 3.5 MHz per sector, that is about one third of theinitial value, so the total band allocated is the same. Figure 6.11 shows the interactionbetween FRF and number of sectors: using the same number of sectors and increasingthe FRF, the SINR is lowered and the same results can be achieved fixing the FRF andintroducing more sectors.

6.5. Impact of Frequency Reuse Factor 85

Figure 6.11: Impact of FRF on SINR and distance from the BS.

In our simulation scenario we suppose that the BS observed is hexagonal and it is sur-rounded by other six interferring cells, transmitting at a constant power. All the usersconnected are pedestrians, moving at a maximum speed of 3 Km/h and they are ran-domly assigned for all the simulation time to an indoor or outdoor environment. Char-acteristics of the generated traffic are the same as those explained in the previous sec-tion, however real time traffic is 60% of the total (30% UGS and 30% rtPS).

Proposed results [57] start with a comparison of the BS throughput in one of the threesectors served by a directional antenna. The simulated channel has a 3.5 MHz profileand it is shared among 16 users with different FRFs, namely 1, 3 and 4, as shown in Fig-ure 6.12. Increasing FRF yields to a better SINR, reflecting positively in sector through-put and both algorithm C and D can exploit in a more efficient way radio resources.We can also see that the slope of the allocated traffic changes when the generated traf-fic reaches 4 Mb/s: if available resources cannot satisfy all request, users experiencingbad channel are penalized and throughput keeps growing but with a reduced rate. Inthis case the scheduler decides to store delay tolerant requests and tries to satisfy moreusers, limiting to minimum band requests. To build the graphic in Figure 6.13 (a) weused only one sector and a 10 Mhz channel, that is we multiplied by 3 the bandwidth,however, we can see that the total throughput is multiplied only by a factor of 2, result-ing in a inefficient spectrum utilization. We also tried to modify the composition of the

6.6. Conclusions 86

(a) FRF equal to 1 (b) FRF equal to 3

(c) FRF equal to 4

Figure 6.12: Total throughput for 3.5 MHz channel and 60% real time traffic [Horizontal axis:generated traffic in Mb/s; vertical axis: allocated traffic in Mb/s].

traffic, reducing real time portion to 20% and the result is reported in Figure 6.13 (b):algorithms with power control perform much better than the other ones.

6.6 Conclusions

In this chapter we did not focused on a particular aspect, as in the previous case, butwe considered a framework involving an entire cell, together with its surrounding en-vironment, using WiMAX technology. We proposed four allocation algorithms tryingto solve an integer programming problem to select users to fill each frame, paying par-ticular attention to water filling approach. In our analysis we took into account direc-

6.6. Conclusions 87

(a) 10 MHz channel with one sectorand 60% of real time traffic

(b) 3.5 MHz channel with three sectorsand 20% of real time traffic

Figure 6.13: Total throughput for FRF equal to 4 [Horizontal axis: generated traffic in Mb/s;vertical axis: allocated traffic in Mb/s].

tional antennas, different transmission conditions (such as indoor/outdoor subscriberstations or different frequency reuse factors) and various input traffic combinations. Wewere interested in total throughput performances, however it is clear that a QoS awareconnection management is fundamental to satisfy multiuser wireless network needs,especially to allocate in an efficient way frequency and power resources. The block sup-posed to accomplish this task is the allocator, placed in the MAC layer despite it hasto interact with the scheduler and it also gathers informations from the physical layer.There are a certain number of issues that should be put together to allocate resources:the feedback channel information, the frequency assignment problem, the modulationand coding selection, the optimum power distribution and the packet scheduling. Allthese elements should cooperate to converge to optimum, or near optimum, allocationand the most promising approach is probably the cross layer one. This is a relativelyrecent research field and there is space for other exhaustive analysis.

Conclusions

In this work we tried to characterise QoS by means of few typical aspects and thissection is devoted to give a brief overview of the results making up our analysis.We started evaluating the performances of transport layer protocols on radio links

because they span both the wireless and the wired domain. Error correction mecha-nisms based on retransmission are replicated by TCP and by link layer entities, so it isinteresting to study how they interact. Despite the same concept, the objective is differ-ent, in fact low-level retransmissions are fundamental to face error distribution on thephysical medium. We considered four TCP flavours, from Tahoe to Selective Acknowl-edgement, with different robustness against bursty errors. Avarage performances ad-vantage Selective Acknowledgement TCP, but variation of typical parameters such ascongestion window size or sequence number is high. Even if the link layer tries to maskthe air interface, the decoupling is not perfect and the instant channel conditions com-promise the transport layer correction strategy.

Starting from this intermediate point, we then explored interaction of the heteroge-neous network with the application and with link layer management techniques. Wetook into account the application requested by the user, focusing on the special case ofmultimedia content delivery in downlink direction. We considered various solutionsto guarantee integrity of video frames, putting together an enhanced coding technique,a limited retransmission scheme and a smart buffer management. We have seen that,mixing all these tools, it is possible to maximise resiliency because ARQ can recovererrors affecting any component of the clip while frame aware protection can smooththe degradation. To optimise this process, a buffer with multiple discarding thresholds

Conclusions 89

is very usefull. We also tried to enlarge our analysis to other applications, evaluat-ing the introduction of a DiffServ domain in the UMTS Core Network. According tothe performances shown, it is clear that the application of the DiffServ framework is notimmediate, but its fixed allocation scheme should be made compliant with variable QoSsupport of UMTS. Due to the particularity of this task, the best approach is probably across layer one.

Cross layer is a key word also for resource allocation issues, in fact the block decidinghow to assign frequencies and power is placed in the link layer, but it needs informa-tions coming from the physical level. OFDM makes this problem a bit more complicatedbecause of the presence of discrete frequencies, so integer solutions are needed, but theyare not trivial since not obtainable by a simple approximation. We think that greedy al-gorithms, such as water filling, are the best compromise between sub-optimal resultsand low computation time. We have proposed four allocation algorithms, giving spe-cial attention to the distribution of power boost to improve transmission quality andto increase the number of served users. We tested different boundary conditions, suchas frequency reuse factor or cell sectorsation, to see how they reflect on global systemperformances. Efficiency can be pursued also implementing in a more effective waysome functions, such as error correction. ARQ techniques can potentially recover anyerror, but they can be successly employed if delay is not a limit. In this case, error pro-tection coding is more suitable, but there is a consumption of computation power anda waste of transmission resources, due to the presence of redundancy. A good compro-mise is the coexistence of the two solutions in a so-called Hybrid ARQ scheme, a verypromising approach implemented by modern wireless broadband networks (WiMAX,UMTS). We suggest to use a rateless coding based HARQ in a UMTS LTE environmentand our contribution is in performance simulation, exploiting the outage probability tocalculate the channel behaveour. Results obtained, compared with traditional HARQ,are very encouraging and show a good efficiency gain, with the advantage of using asimple simulation tool.

In this work we demonstrated that global system performances depend on a vari-ety of factors, distributed among the layered structure of the network. QoS involvesdifferent aspects of networking, merging various areas of knowledge and this is evenmore true since the development of data-oriented wireless communications. All themechanisms we used for parameters handling refer to well-known standards, so thenew challenge is to go beyond a network model made up of isolated levels, adopting a

Conclusions 90

cross layer approach. In fact, application and link layers can benefit from informationsprovided by the physical interface in order to enhance and to speed up the adaptationto a lossy environment. Analysis can be also extended taking into account multiserviceand multiprotocol nature of the configuration, maybe not disregarding related designissues, for example the balancing between centralised or distributed algorithms.

In summary, QoS is a fundamental part of network deployment and all the relatedissues should be treated in an harmonic way to contribute to improve user’s satisfac-tion together with efficient resource management. In our opinion, cross layer designoffers an innovative point of view together with promising capabilities that can be fullyexploited only by a deeper research.

List of Acronyms

2G 2nd Generation cellular system3G 3rd Generation cellular system3GPP 3rd Generation Partnership Project4G 4th Generation cellular systemAAL ATM Adaptation LayerACK AcknowledgementACM Adaptive Coding and ModulationARQ Automatic Repeat reQuestATM Asynchronous Transfer ModeBER Bit Error RateBTS Base Transceiver StationCAC Call Admission ControlCBR Constant Bit RateCC Chase CombiningCDMA Code Division Multiple AccessCID Connection IDentifierCS Circuit SwitchedDiffServ Differentiated ServicesFC Fountain CodesFDD Frequency Division Duplex

List of Acronyms 92

FDMA Frequency Division Multiple AccessFEC Forward Error CorrectionFRF Frequency Reuse FactorFTP File Transfer ProtocolGPRS General Packet Radio ServiceGSM Global System for Mobile communicationsHARQ Hybrid ARQHSDPA High Speed Downlink Packet AccesIETF Internet Engineering Task ForceIntServ Integrated ServicesIP Internet ProtocolIR Incremental RedundancyISO International Standard OrganisationLOS Line of SightMAC Medium Access ControlMIP Mobile IPMPLS Multi Protocol Label SwitchingMUD Multi User DiversityNLOS Non-Line of SightNRT Non Real TimeOFDM Orthogonal Frequency Division MultiplexOFDMA Orthogonal Frequency Division Multiple AccessOSI Open Standard InitiativePDU Packet Data UnitPS Packet SwitchedQoS Quality of ServiceRAT Radio Access TechnologyRLC Radio Link ControlRNC Radio Network ControllerRNS Radio Network SubsystemRRC Radio Resource Control

List of Acronyms 93

RSS Received Signal StrengthRSVP Resource reSerVation ProtocolRT Real TimeRTP Real Time ProtocolRTT Round Trip TimeSNR Signal to Noise RatioSOFDMA Slotted Orthogonal Frequency Division Multiple AccessSR Selective RepeatSSCS Service Specific Convergence SublayerTOS Type Of ServiceTCP Transmission Control ProtocolTDD Time Division DuplexTDMA Time Division Multiple AccessUDP User Datagram ProtocolUMTS Universal Mobile Telephone SystemUTRAN UMTS Terrestrial Radio Access NetworkVoIP Voice over IPWLAN Wireless Local Area Network

Bibliography

[1] B. Walke, Mobile Radio Networks, Second Edition, J. Wiley and Sons Ltd., Chichester, GB.

[2] H. Holma, A. Toskala, WCDMA for UMTS: Radio Access for Third Generation Mobile Commu-nications, J. Wiley and Sons Ltd. Chichester, GB.

[3] Y.-B. Lin, A.-C. Pang, Y.-R. Haung and I. Chlamtac, ”An All-IP Approach for UMTS Third

Generation Mobile Networks,” IEEE Network, Sept.-Oct. 2002, pp. 8-19.

[4] 3GPP, “Technical Specication Group Radio Access Network; Feasibility study on 3GPP

system to Wireless Local Area Network (WLAN) interworking (Release 6),” TR 22.934, v.

6.2.0, Sep. 2003.

[5] 3GPP, “3GPP system to Wireless Local Area Network (WLAN) interworking; System de-

scription (Release 6),” ITS 23.234, v. 6.0.0, Mar. 2004.

[6] http://ims.unipv.it/firb/

[7] E. Rosen, A. Viswanathan and R. Callon, “Multiprotocol Label Switching Architecture,“

RFC 3031, Jan. 2001.

[8] R. Braden, D. Clark and S. Shenker, “Integrated Services in the Internet Architecture: an

Overview,“ RFC 1633, Jul. 1994.

[9] P. P. White, ”RSVP and Integrated Services in the Internet: A Tutorial,“ IEEE Communica-tions Magazine, May 1997, Vol. 45, Issue 5, pp. 100-106.

[10] S. Blake, M. Carlson, E. Davies, Z. Wang and W. Weiss, ”An Architecture for Differentiated

Services,” RFC 2475, Dec. 1998.

BIBLIOGRAPHY 95

[11] A. Jamalipour, T. Wada and T. Yamazato, ”A Tutorial on Multiple Access Technologies for

Beyond 3G Mobile Networks,” IEEE Communications Magazine, Feb. 2005, Vol. 43, Issue 2,

pp. 100-117.

[12] M. Mathis, J. Mahdavi, S. Floyd and A. Romanow, “TCP Selective Acknowledgment Op-

tions,” RFC 2018, Oct. 1996.

[13] 3GPP, “Radio Link Control (RLC) Protocol Specification (Release 5),” TS 25.322, v. 5.11.0,

Jun. 2005.

[14] Zhang Q. and H.-J. Su, “Performance of UMTSRadio Link Control,” Proc. of IEEE Inter-

national Conference on Communications (ICC 2002), New York City, NY, USA, 28 Apr.- 2

May 2002, Vol. 5, pp. 3346-3350.

[15] A. Lo, G. Heijenk and C. Bruma, “Performance of TCP over UMTS Common and Dedicated

Channels,“ IST Mobile & Wireless Communications Summit 2003, Aveiro, P, 15-18 Jun.

2003.

[16] http://isi.edu/nsnam/ns

[17] A. Todini and F. Vacirca, UMTS module for ns, available via http://net.infocom.uniroma1.it.

[18] R. Bestak, P. Godlewski and P. Martins, ”RLC Buffer Occupancy when Using a TCP Con-

nection Over UMTS,“ Proc. of 13th IEEE International Symposium on Personal, Indoor

and Mobile Radio Communications (PIMRC 2002), Lisboa, P, 15-18 Sep. 2002, Vol. 3, pp.

1161-1165.

[19] S. Heier, D. Heindrichs and A. Kemper, ”Performance evaluation of Internet applications

over the UMTS radio interface,“ IEEE 55th Vehicular Technology Conference, 2002, Vol. 4 ,

2002, pp. 1834-1838.

[20] E.N. Gilbert, ”Capacity of a burst-noise channel,“ The Bell System Tech. J., Vol. 39, Sep.

1960, pp. 1253-1256.

[21] L.A. Grieco and S. Mascolo, ”Performance evaluation and comparison of Westwood+, New

Reno, Vegas TCP congestion control,“ ACM Sigcomm Computer Communications Review, Vol.

34, No. 2, Apr. 2004, pp. 25-38.

[22] L.R. Rabiner, ”A tutorial on hidden Markov models and selected applications in speech

recognition,“ Proc. of the IEEE, Vol. 77, Issue 2, Feb. 1989, pp. 257286.

BIBLIOGRAPHY 96

[23] E. Costamagna, L. Favalli, P. Savazzi and F. Tarantola, ”Performance analysis of TCP traffic

over UTRA-FDD channels,“ Proc. The 6th International Symposium on Wireless Personal

Multimedia Communications (WPMC 2003), Yokosuka, JP, 19-22 Oct. 2003.

[24] A. Umbert and P. Diaz, ”A radio channel emulator for WCDMA, based on the Hidden

Markov Model (HMM),“ Proc. of 52nd IEEE Vehicular Technology Conference (VTC-Fall

2000), Boston, MA, USA, 24-28 Sept. 2000, Vol. 5, pp. 2174-2179.

[25] E. Costamagna, L. Favalli, F. Tarantola and M. Lanati, ”An End to End Simulator for Per-

formance Comparison of Different TCP Schemes Over UTRA Channels,“ IST Mobile &

Wireless Communications Summit 2005, Dresden, D, 19-23 Jun. 2005.

[26] C. Barakat, E. Altman and W. Dabbous, ”On TCP Performance in a Heterogeneous Net-

work: A Survey“, IEEE Communications Magazine,Vol. 38, Issue 1, Jan. 2000, pp. 4046.

[27] L. Favalli, P. Savazzi, F. Tarantola and M. Lanati, ”Performance Comparison of Different

TCP Schemes Over UTRA-TDD Channels,“ Proc. of 62nd IEEE Vehicular Technology Con-

ference (VTC-Fall 2005), Dallas, TX, USA, 25-28 Sep. 2005, Vol. 4, pp. 2680 - 2683.

[28] K.N. Srijith, Lillykutty Jacob and A.L. Ananda, “Worst-case Performance Limitation of TCP

Sack and a Feasible Solution,” Proc. of 8th IEEE International Conference on Communica-

tions Systems (ICCS 2002), Singapore, SGP, 25-28 Nov. 2002, Vol. 2, pp. 1152-1156.

[29] ITU-T, ”H.264 - Advanced video coding for generic audiovisual services,“ 2003.

[30] T.Stockhammer, M. Hannuksela and T.Wiegand, ”H.264/AVC in wireless environments,“

IEEE Transactions on Circuits and Systems for Video Technology, Vol. 13, No. 7, Jul. 2003.

[31] R.Scopigno and S.Belfiore, ”Image decomposition for selective encryption and flexible net-

work services,“ Proc. of IEEE Global Telecommunications Conference 2004 (GLOBECOM

’04), Dallas, TX, USA, 29 Nov.-3 Dec. 2004, Vol. 4, pp. 2302 - 2307.

[32] S. Belfiore, R. Scopigno, F. Tarantola and M. Lanati, ”Evaluation of resiliency solutions for

real-time multimedia streams in a UTRAN scenario,“ Proc. of 5th International Conference

on Information, Communications and Signal Processings (ICICS 2005), Bangkok, T, 6-9 Dec.

2005, pp. 513-517.

[33] K. Nichols, S. Blake, F. Baker and D. Black, ”Definition of the Differentiated Services Field

(DS Field) in the IPv4 and IPv6 Headers,“ RFC 2474, Dec. 1998.

[34] Cisco Systems, ”IP class of service for mobile networks“, Cisco Systems Inc., 2006.

BIBLIOGRAPHY 97

[35] M. Bohge, J. Gross, A. Wolisz and M. Meyer, ”Dynamic Resource Allocation in OFDM

Systems: An Overview of Cross-Layer Optimization Principles and Techniques,“ IEEENetwork, Vol. 21, No. 1, Jan.-Feb. 2007, pp. 53-59.

[36] I. Kalet, ”The Multitone Channel,“ IEEE Transactions on Communications, Vol. 37, No. 2, Feb.

1989, pp. 11924.

[37] C. Y. Wong, R. S. Cheng, K. B. Letaief and R. D. Murch, ”Multiuser OFDM with Adaptive

Subcarrier, Bit and Power Allocation,“ IEEE Journal on Selected Areas in Communications, Vol.

17, No. 10, Oct. 1999, pp. 1747-1758.

[38] 3GPP, ”High Speed Downlink Packet Access (HSPDA); Overall Description , (Release 5),“

TS 25.308, Mar. 2002.

[39] 3GPP, ”FDD enhanced uplink; Overall description, (Stage 2),“ TS 25.309, Oct. 2004.

[40] H. Ekstrm, A. Furuskr, J. Karlsson, M. Meyer, S. Parkvall, J. Torsner and M. Wahlqvist,

”Technical Solutions for the 3G Long-Term Evolution,“ IEEE Communications Magazine,

Vol. 20, No. 20, Mar. 2006.

[41] 3GPP, ”Feasibility Study for Orthogonal Frequency Division Multiplexing (OFDM) for

UTRAN enhancement (Release 6),“ TR 25.892, Jun. 2004.

[42] 3GPP, ”Physical Layer Aspects for Evolved UTRA (Release 7)“ TR 25.814, Jun. 2006.

[43] H. Zheng and H. Viswanathan, ”Optimizing the ARQ performance in downlink packet

data systems with scheduling,“ IEEE Transactions on Wireless Communications, Vol. 4, Issue

2, Mar. 2005, pp. 495-506.

[44] Y. Wang, L. Zhang, and D. Yang, ”Performance analysis of type III HARQ with turbo

codes,“ Proc. of 57th IEEE Vehicular Technology Conference (VTC-Spring 2003), Seoul,

ROK, 22-25 Apr. 2003, Vol. 4, pp. 2740-2744.

[45] D.J.C. MacKay, ”Fountain codes,“ IEEE Proc.-Commun., Vol. 152, No. 6, Dec. 2005, pp.

1062-1068.

[46] J. Castura and Y. Mao, ”Rateless coding over fading channels,“ IEEE Communications Letters,

Vol. 10, No. 1, Jan. 2006, pp. 46-48.

[47] L. Favalli, M. Lanati and P. Savazzi, ”Hybrid ARQ Based on Rateless Coding for UTRAN

LTE Wireless Systems,“ Tyrrhenian International Workshop on Digital Communication

2007 (TIWDC), Ischia Island, IT, 9-12 Sep. 2007.

BIBLIOGRAPHY 98

[48] http://www.wimaxforum.org

[49] IEEE, ”IEEE standard for local and metropolitan area networks - Part 16: air interface for

fixed broadband wireless access system,“ IEEE 802.16-2004 standard.

[50] IEEE, ”Amendament to IEEE standard for local and metropolitan area networks - Part

16: air interface for fixed broadband wireless access system - Physical and medium access

control layers for combined fixed and mobile operation in licensed bands,“ IEEE 802.16e-

2005 standard.

[51] H. Liu and G. Li, ”OFDM-based broadband wireless networks,“ Wiley, 2005.

[52] WiMAX Forum, ”Mobile WiMAX - Part I: A Technical Overview and Performance Evalua-

tion, “ Aug. 2006.

[53] D. Messina, L. Scalia, I. Tinnirello, S. Merlin, A. Zanella and M. Moretti, ”Allocation Al-

gorithms for PRIMO system,“ Wireless Reconfigurable Terminals and Platforms (WiRTeP)

workshop, Rome, IT, 10-12 Apr. 2006.

[54] L. Badia, A. Baiocchi, S. Merlin, S. Pupolin, A. Todini and M. Zorzi, ”On the impact of

physical layer awareness on scheduling and resource allocation in broadband multi-cellular

IEEE 802.16 systems,“ IEEE Wireless Communications, Vol. 14, No. 1, Feb. 2007, pp. 36-43.

[55] L. Favalli, M. Lanati, S. L. Monighini and P. Savazzi, ”Resource Assigment in Multiser-

vice 802.16e,“ 18th IEEE International Symposium on Personal, Indoor and Mobile Radio

Communications (PIMRC 2007), Athens, GR, 3-7 Sep. 2007.

[56] O. Bertazioli and L. Favalli, ”GSM - GPRS Seconda Edizione,“ Hoepli, 2002

[57] M. Lanati, S. L. Monighini and P. Savazzi, ”Impact of Network Configuration on Resource

Assignment in Multiservice 802.16e,“ GTTI (Gruppo nazionale Telecomunicazioni e Teoria

dell’Informazione) annual meeting, Rome, IT, 18-20 Jun. 2007.

List of Figures

1.1 Mobile IP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.1 Access Stratum and Non Access Stratum. . . . . . . . . . . . . . . . . . . . . . . . . 20

2.2 UTRAN and Core Network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.3 UMTS radio protocol stack. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.4 RLC acknowledged mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.5 Core Network all-IP architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.6 Simulator’s protocol stack. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.7 Selective Repeat scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.8 Network topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.9 TCP sequence number over wireless link. . . . . . . . . . . . . . . . . . . . . . . . . 30

2.10 Congestion window. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.11 Cumulative distribution functions of interarrival time for 2 and 3 dB channels

with a 64 Kbit/s source. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.12 Cumulative distribution function of interarrival time with a 144 Kbit/s source. . . 32

2.13 PDU transmission statistics for 4 dB channel. . . . . . . . . . . . . . . . . . . . . . . 35

2.14 Cumulative distribution functions of ACKs interarrival times for 4 dB channel. . . 36

2.15 PDU transmission statistics for 2 dB channel. . . . . . . . . . . . . . . . . . . . . . . 37

2.16 Cumulative distribution functions of ACKs interarrival times for 2 dB channel. . . 38

3.1 Smart RLC buffer management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.2 Network environment, with Node B acting like a proxy. . . . . . . . . . . . . . . . . 43

3.3 PSNR of H.264 decoded video. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.4 A simple scenario with DiffServ Core Network. . . . . . . . . . . . . . . . . . . . . . 47

List of Figures 100

3.5 Modification added to the simulator. . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.6 Multiple RLC buffer split in two queues. . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.7 Delay analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.8 Comparison between previous situation and the introduction of a weight factor. . 49

4.1 Water filling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5.1 Ideal realization of an OFDM transmission. . . . . . . . . . . . . . . . . . . . . . . . 58

5.2 OFDM using IFFT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.3 Generation of cyclic prefix in an OFDM symbol. . . . . . . . . . . . . . . . . . . . . 59

5.4 Time-frequency representation of a Resource Unit. . . . . . . . . . . . . . . . . . . . 60

5.5 Protocol structure for LTE environment. . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.6 Hybrid ARQ retransmission scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5.7 PDU generation using Fountain Code approach. . . . . . . . . . . . . . . . . . . . . 64

5.8 LTE single cell simulator: main components. . . . . . . . . . . . . . . . . . . . . . . 66

5.9 Cumulative distribution function of coding rate. . . . . . . . . . . . . . . . . . . . . 68

5.10 Base station goodput versus time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5.11 Cumulative distribution function of transmission delays. . . . . . . . . . . . . . . . 70

6.1 WiMAX scenario in a metropolitan area. . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.2 WiMAX protocol stack. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

6.3 Example of a diversity permutation scheme in downlink. . . . . . . . . . . . . . . . 76

6.4 Some examples of Continuous Permutation slots. . . . . . . . . . . . . . . . . . . . 76

6.5 Example of Multi User Diversity gain. . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.6 Base station model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

6.7 Minimum and maximum bandwidth requests generation. . . . . . . . . . . . . . . 79

6.8 Allocation algorithms performances with 40% real time traffic. . . . . . . . . . . . . 83

6.9 Allocation algorithms performances with 20% real time traffic. . . . . . . . . . . . . 83

6.10 Variation of distance between clusters with frequency reuse factor, indicated as N. 84

6.11 Impact of FRF on SINR and distance from the BS. . . . . . . . . . . . . . . . . . . . 85

6.12 Total throughput for 3.5 MHz channel and 60% real time traffic. . . . . . . . . . . . 86

6.13 Total throughput for FRF equal to 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

List of Tables

2.1 Frequencies assigned to UMTS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2 Physical parameters for UMTS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.3 Wireless channel simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.4 RLC and TCP simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.5 UMTS TDD module parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.1 Simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.2 UMTS QoS classes and DiffServ PHB mapping. . . . . . . . . . . . . . . . . . . . . . 47

5.1 Modulation and coding scheme for traditional Chase Combining. . . . . . . . . . . 65

5.2 Some interesting simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.1 Summary of WiMAX evolution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.2 Simulated traffic characteristics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

Ringraziamenti

Vorrei ringraziare il prof. Favalli per i validi spunti di ricerca e per i numerosi suggerimenti

che mi hanno permesso di affrontare questi tre anni di dottorato e il prof. Costamagna per

avermi permesso di condividere l’esperienza dell’insegnamento. Un doveroso ringraziamento

a tutte le persone che lavorano o hanno lavorato nel laboratorio di Comunicazioni Elettriche

e hanno avuto un ruolo fondamentale per la creazione di un ambiente sereno e costruttivo: il

prof. Gamba, Fabio, Pietro, i colleghi di dottorato (Giovanna, Gianni, Mattia, Francesco e i nuovi

arrivati Marco, Andrea, Anna) e i tesisti che si sono susseguiti in questi anni.

Questa tesi e dedicata ai miei genitori, valido sostegno a tutte le mie attivit e iniziative, ma

un ricordo particolare va anche ai parenti piu stretti, ai due nonni recentemente scomparsi e al

nuovo arrivato, Daniele.

Durante questi tre anni sono entrato in contatto con tante nuove persone che hanno allargato

il campo dei miei interessi. Un ringraziamento ai membri del Centro Medico Polispecialistico,

per avermi accolto con cortesia e fiducia e ai volontari della Biblioteca Comunale di Pietra de’

Giorgi “R. Moretti” per l’aiuto alle tante iniziative create insieme, premiate con tanto entusiasmo

e soddisfazione. In particolare, una dedica a Marco e Paola per avere dimostrato che credere

fermamente e con entusiasmo in un progetto porta sempre a buoni risultati.

Infine, un ringraziamento agli amici che hanno riempito i pochi momenti liberi e ai compagni

incontrati durante un indimenticabile viaggio.

Documents

QUALITY OF SERVICE MANAGEMENT TECHNIQUES FOR IP BASED HETEROGENEOUS … · 2014-10-29 · universita degli studi di pavia` facolta di ingegneria` tesi di dottorato di ricerca matteo