Scheduling and Resource Optimization in Next Generation Heterogeneous Wireless Networks

A Dissertation Presented to the Graduate Faculty of the University of Louisiana at Lafayette In Partial Fulllment of the Requirements for the Degree Doctor of Philosophy

Haining Chen Spring 2008

c Haining Chen 2008 All Rights Reserved

Scheduling and Resource Optimization in Next Generation Heterogeneous Wireless Networks Haining Chen

APPROVED:

Hongyi Wu, Chair Associate Professor The Center for Advanced Computer Studies

Nian-Feng Tzeng Professor The Center for Advanced Computer Studies

Dmitri Perkins Assistant Professor The Center for Advanced Computer Studies

C. E. Palmer Dean of the Graduate School

This dissertation is dedicated to my wife Xiaojuan Xie, my daughter Julia Chen, my father Beixiang Chen, and my late mother Yili Yang.

ACKNOWLEDGMENTS

The work described here could not have been nished without help and support from others. The following words cannot express the depth of my appreciation. First, I want to thank my advisor, Dr. Hongyi Wu. Without his continuous support, advice, and encouragement throughout my Ph.D. study, I could not have nished this work. I also would like to thank Dr. Nian-Feng Tzeng and Dr. Dmitri Perkins for their valuable supervision of my academic study and research. I also appreciate the valuable feedback from Dr. Xiaoduan Sun, who served as the outside observer for my Ph.D. defense. Second, I would like to thank all CACS faculty for the wonderful knowledge they taught me and all CACS staff for their help during my time in CACS as a Ph.D. student. I also would like to thank all my friends and colleagues in the iWIN lab as well as in CACS for their unselsh help, especially Su Xia, Rajesh Prasad, Adam Lewis, Denvil Smith, Ha Dang, Lei Pan, Zhipeng Yang, Yan He, and Chong Wang. Third, I would like to pay respect to my father for his sacrices and encouragement, and to my late mother, who is always in my heart. I really have no words to express my love to my parents. Finally, I would like to thank my wife, Xiaojuan Xie, for her endless love, support, and encouragement at all times. She accompanies me in overcoming all difculties in life. Last but never the least, I would like to thank my 13-month-old daughter, Julia Chen, who cheers me up with her cute smile, her endless curiosity, her naughty requests, and sometimes even her cries.

TABLE OF CONTENTS

DEDICATION............................................................................................................. iv ACKNOWLEDGMENTS ............................................................................................ v

LIST OF TABLES....................................................................................................... ix LIST OF FIGURES ..................................................................................................... CHAPTER 1: Introduction .......................................................................................... 1.1 Wireless Networks and their Characteristics ..................................................... 1.1.1 Wireless Cellular Networks.................................................................. 1.1.2 Wireless Data Networks....................................................................... 1.2 Next Generation Wireless Networks................................................................. 1.3 Contributions and Dissertation Organizations ................................................... CHAPTER 2: System Overview and Challenges ........................................................... 2.1 A System Overview of Next Generation Wireless Networks .............................. 2.1.1 Major Components of Next Generation Wireless Networks .................... 2.1.2 Multiple Interfaces in Mobile Terminals ............................................... 2.2 Challenges in QoS Guarantee Problem ............................................................. 2.3 Challenges in Minimum Cost Problem ............................................................. 2.4 Challenges in System Maintenance Problem..................................................... 2.5 Challenges in Scheduling for Relay Networks .................................................. 2.6 Summary........................................................................................................ CHAPTER 3: Qos Guarantee for WiMAX Networks .................................................... 3.1 Introduction.................................................................................................... 3.2 WiMAX MAC Layer ...................................................................................... 3.2.1 WiMAX MAC Layer Structure ............................................................ 3.2.2 Uniqueness of Uplink Scheduling in WiMAX MAC Layer .................... 3.2.3 Related Scheduling Algorithms ............................................................ 3.3 Two-phase Fair and Efcient Queueing (FEQ) Scheduling Algorithm ................ 3.3.1 Overview of the Two-phase FEQ Scheduling Algorithm ........................ 3.3.2 Analysis of the Two-phase FEQ Scheduling Algorithm.......................... 3.3.3 Analysis of Other Approaches.............................................................. 3.4 Simulations and Discussion ............................................................................. 3.4.1 Poisson Trafc .................................................................................... x 1 1 1 3 4 6 8 8 9 11 12 15 17 19 21 22 22 24 24 27 29 33 33 35 42 42 47

3.5

3.4.2 Bursty Trafc ...................................................................................... 50 3.4.3 Further Discussion .............................................................................. 52 Summary........................................................................................................ 52 54 54 57 59 63 64 65 65 66 68 68 69 71 73 76 78 87

CHAPTER 4: Minimum-cost Data Delivery in Wireless Networks................................. 4.1 Introduction.................................................................................................... 4.2 Related Work.................................................................................................. 4.3 System Architecture........................................................................................ 4.4 Problem Formulation and NP-hard Proof.......................................................... 4.4.1 Problem Formulation ........................................................................... 4.4.2 NP-hard Proof..................................................................................... 4.4.2.1 CMCK problem.................................................................... 4.4.2.2 Polynomial Reduction to Cost Minimization Problem ............. 4.5 Proposed LP Algorithms for Cost Minimization................................................ 4.5.1 Single Route ....................................................................................... 4.5.1.1 Base Case ............................................................................ 4.5.1.2 Advanced Case..................................................................... 4.5.2 Multiple Possible Routes ..................................................................... 4.5.3 Variable Speed .................................................................................... 4.6 Simulations and Discussion ............................................................................. 4.7 Summary........................................................................................................

CHAPTER 5: Self-maintenance Scheduling for Wireless Networks ............................... 89 5.1 Introduction.................................................................................................... 89 5.2 Related Work.................................................................................................. 92 5.3 The Proposed Scheduling Algorithms .............................................................. 93 5.3.1 Problem formulation............................................................................ 93 5.3.2 An Integer Linear Programming (ILP) Model........................................ 94 5.3.3 RC-Cliques-RM Heuristic Algorithm ................................................... 97 5.3.4 RC-RM-Cliques Heuristic Algorithm ................................................... 100 5.4 Simulations and Discussion ............................................................................. 101 5.5 Summary........................................................................................................ 106 CHAPTER 6: Scheduling for Multi-hop Relay Wireless Cellular Networks .................... 109 6.1 Introduction.................................................................................................... 109 6.2 Related Work.................................................................................................. 112 6.3 Scheduling for Multi-hop Relay Cellular Networks ........................................... 115 6.3.1 Preferred Approach for Building Cellular Relay Networks ..................... 115 6.3.2 Linear Programming Model for the Scheduling Algorithm..................... 116 6.3.3 Finding Concurrent Transmission Scenarios.......................................... 119

vii

6.4 6.5

Simulations and Discussion ............................................................................. 121 Summary........................................................................................................ 125

CHAPTER 7: Conclusion............................................................................................ 130 BIBLIOGRAPHY........................................................................................................ 132 ABSTRACT ............................................................................................................... 142 BIOGRAPHICAL SKETCH ........................................................................................ 144

viii

LIST OF TABLES

Table 3.1: Simulation Parameters for WiMAX WirelessHUMAN-SCa...................... 43 Table 3.2: WirelessHUMAN-SCa Raw Data Rate. ................................................... 44 Table 3.3: MRR and Delay Budget of Different Connections.................................... 44 Table 3.4: Distribution of the 8 Connections in 4 SS. ............................................... 44 Table 4.1: Network Parameters for Simulation Setup. .............................................. 77 Table 4.2: Impact of System Overhead on Communication Cost. .............................. 77 Table 5.1: Impact of Increasing L and M . ............................................................... 102 Table 5.2: Impact of Increasing Lk and Mk . ............................................................ 103 Table 5.3: Impact of Increasing N . ......................................................................... 103 Table 5.4: Complexity and Optimality of the Proposed Algorithms. .......................... 106 Table 6.1: Simulation Network Parameters. ............................................................. 121

LIST OF FIGURES

Figure 2.1: Next Generation Wireless Networks. .....................................................

Figure 3.1: WiMAX MAC Frame Structure............................................................. 25 Figure 3.2: Analytical Model for the Two-phase FEQ Algorithm. ............................. 37 Figure 3.3: Markovian Model of M/D/1 + D Queue under EDF. ............................ 40 Figure 3.4: Performance under Poisson Trafc. ....................................................... 46 Figure 3.5: FEQ Algorithm Running Time with the Number of Connections. ............ 49 Figure 3.6: MMPP Model for Bursty Trafc............................................................ 50 Figure 3.7: Performance under Bursty Trafc. ......................................................... 51 Figure 3.8: Impact of Bandwidth Allocation between Two Phases (). ..................... 51 Figure 4.1: A Scenario of Heterogeneous Wireless Networks. .................................. 55 Figure 4.2: Signaling Protocols for Cost Minimization Problem. .............................. 59 Figure 4.3: Inputs and Outputs for the Cost Minimization Problem. .......................... 63 Figure 4.4: LP Model for Predetermined Route. ...................................................... 69 Figure 4.5: LP Model for Predetermined Route with Prefetching. ............................. 72 Figure 4.6: LP Model for Multiple Possible Routes.................................................. 74 Figure 4.7: Heterogeneous Networks with WiMAX, 3G, and WiFi cells. .................. 78 Figure 4.8: Performance Averaged over Random Routes. ......................................... 78 Figure 4.9: Performance Averaged over Random Maps. ........................................... 79 Figure 4.10: Impact of Delay Bounds on Downloaded Data. ...................................... 79

Figure 4.11: Impact of Route Probability on Communication Cost. ............................ 79 Figure 4.12: Impact of Route Probability on Prefetched Data. .................................... 79 Figure 4.13: Impact of Increasing Cell Dwelling Time............................................... 80 Figure 4.14: Impact of Decreasing Cell Dwelling Time.............................................. 80 Figure 4.15: Cost Comparison when Cell Dwelling Time Varies Randomly................. 80 Figure 5.1: (a) Gf Resource Conicting and (b) Gp Resource Compatible................. 97 Figure 5.2: Integer Linear Programming (ILP) Model. ............................................. 108 Figure 6.1: Architecture of Cellular Relay Networks................................................ 117 Figure 6.2: LP Model for Scheduling in Cellular Relay Networks............................. 127 Figure 6.3: Cellular Relay Network Topology for Simulation. .................................. 128 Figure 6.4: Throughput Comparison of Different Scheduling Algorithms.................. 128 Figure 6.5: Fairness Comparison of Different Scheduling Algorithms. ...................... 128 Figure 6.6: Impact of Distance from the BS on Network Throughput ........................ 129 Figure 6.7: Impact of Number of Scenarios on Network Throughput. ....................... 129

xi

CHAPTER 1 INTRODUCTION

Wireless networks have been making tremendous progress during the past two decades. From the rst generation wireless cellular networks to the second generation (2G) and the third generation (3G), from narrow-bandwidth systems to wide-bandwidth systems, from analog systems to digital systems, the pace of evolution never stops. Nevertheless, next generation wireless networks, known as Beyond 3G (B3G), are on the horizon. All of these various wireless networks possess different characteristics and provide different services to mobile users. In this chapter, I rst review wireless cellular networks from the rst generation to 2G and 3G, followed by an introduction of wireless data networks including Wireless Local Area Network (WLAN) and Wireless Personal Area Network (WPAN). Secondly, after briey describing the characteristics of next generation wireless networks, I summarize the contributions of the dissertation. 1.1 Wireless Networks and their Characteristics There are many types of wireless networks. The focus of my research is on wireless cellular networks and wireless data networks, and their evolution paths and characteristics are discussed in this section. 1.1.1 Wireless Cellular Networks Initially, wireless phone systems came as replacements for cable. In North America, the Advanced Mobile Phone System (AMPS) [1] belongs to the rst generation wireless cellular networks, which provides voice service to subscribers. The counterpart of AMPS in Europe is

the Total Access Communication System (TACS) [2]. Both AMPS and TACS appeared in the early 1980s. The idea of cellular was invented in AMPS because the covered area is divided into small hexagonal cells and frequency reuse is employed among cells so that those cells far enough apart to prevent mutual interference can be assigned the same frequency. This cellular concept together with handoff schemes between neighboring cells enable wireless cellular networks to support a large number of mobile users over a wide geographic area. Within the same cell, Frequency Division Multiple Access (FDMA) technology is used to achieve multiple access to mobile users. Each FDMA channel in AMPS has a bandwidth of 30 kHz, which only supports voice service. Both AMPS and TACS belong to analog systems, and they are susceptible to noise and interference and vulnerable to security threats. Besides, the capacity limitation of analog systems prevents AMPS and TACS from meeting the challenge of the growing numbers of mobile subscribers. The 2G systems represented by Global System for Mobile Communications (GSM) [3] and Interim Standard 95 (IS-95) [4] are digital systems, in contrast to the rst generation analog systems. The advantages of digital systems include higher bandwidth efciency and less susceptibility to security threats. Besides FDMA, other multiple access schemes such as Time Division Multiple Access (TDMA) and Code Division Multiple Access (CDMA) are used in GSM and IS-95, respectively. TDMA and CDMA further improve the bandwidth efciency of the 2G systems. The voice quality of the digital systems is better than that of the analog systems since digital systems have more tolerance for noise and interference. Besides voice service, data service is also supported in GSM and IS-95, although the available data rate is limited to 13 Kbps in GSM and 14.4 Kbps in IS-95. Both GSM and IS-95 have experienced huge success in the global mobile communication markets from the late 1980s 2

through the 1990s. Despite the huge success, 2G systems are tagged with limited data rates. Evolutions of 2G systems aim at enhancing the data rate, such as in 2.5G and 3G systems. In 2.5G systems, General Packet Radio Service (GPRS) and Enhanced Data Rates for GSM Evolution (EDGE) [5] are enhancements for GSM, while IS-95B [6] is an enhancement for IS-95. 2.5G systems enter the commercial stage after 1995, and their data rates can reach more than 100 Kbps due to the enhanced coding and modulation schemes that can put more data bits into one symbol transmitted over the air. The evolution of GSM, GPRS, and EDGE towards 3G is Wide-band CDMA (WCDMA) [7], and the 3G evolution of IS-95 and IS-95B is CDMA2000 [8]. Both WCDMA and CDMA2000 are wide bandwidth communication systems with CDMA technology in the physical layers. 3G systems can achieve a 2 Mbps data rate, and voice service and high-speed data service are supported in the 3G systems. 3G systems entered the phase of worldwide commercial deployment after 2000. 1.1.2 Wireless Data Networks From the evolution paths of the wireless cellular networks described above, it is clear that cellular networks start with voice service, and then data service is gradually incorporated into the systems capable of increased data rate. While the wireless cellular networks are making their ways to the hybrid service mode of both voice and data, another type of wireless network becomes more and more popular, which is widely known as WLAN or WiFi. WLAN targets data service from the very beginning. Since WLAN is developed under the IEEE 802.11 working groups, its standards are given names starting with 802.11. The rst popular WLAN product is 802.11b [9], whose standard

was released in 1999. 802.11b supports data rates up to 11 Mbps, and it uses the free Industrial, Scientic, and Medical (ISM) spectrum at 2.4 GHz. The physical technology used in 802.11b is Direct Sequence Spread Spectrum (DSSS). 802.11a [10] is released along with 802.11b, and it operates at the 5 GHz ISM spectrum with a maximum data rate of 54 Mbps. The physical layer technology of 802.11a is Orthogonal Frequency Division Multiplexing (OFDM). Due to increase in hardware complexity and a shorter coverage range, 802.11a is less popular than 802.11b. A breakthrough in WLAN occurred in 2003 when 802.11g [11] standard was released, which uses OFDM technology in the physical layer and operates at the 2.4 GHz ISM spectrum with a maximum data rate of 54 Mbps. 802.11b and 802.11g are the most popular WLAN technologies being deployed so far. The cell radius of a single WLAN cell is around 100 feet, and the handoff function between neighboring WLAN cells are usually not available unless specically supported. This means that the WLAN cannot provide seamless handover to mobile users moving across neighboring WLAN cells. With the maturity of Voice over IP (VoIP) technology and Video over IP technology, voice and video services are also available in WLAN. Besides WLAN, WPAN [12] (also known as bluetooth) belongs to wireless data networks, which uses short-range high-speed wireless channels to connect personal electronic devices, such as laptops, headsets, ash drives, and so on. WPAN is developed under the 802.15 working group, and current bluetooth devices can reach 3 Mbps. 1.2 Next Generation Wireless Networks The discussions in Section 1.1 reveal a trend of wireless networks evolving to support higher data rates with broader bandwidth, and the evolution of wireless cellular networks from

3G to B3G has testied to that. The B3G evolution along the GSM and WCDMA track is High Speed Packet Access (HSPA) [13], which uses a 5 MHz channel. HSPA in the downlink is known as High Speed Downlink Packet Access (HSDPA) and it can reach a data rate of 14.4 Mbps, while the counter part in the uplink is known as High Speed Uplink Packet Access (HSUPA) and it can reach 5.7 Mbps. The technology enhancements in HSPA include adaptive modulation and coding schemes, fast packet scheduling, and hybrid automatic repeat-request. Beyond HSPA is Long term Evolution (LTE) [14], which aims at achieving data rates more than 100 Mbps by applying OFDM and Multiple Input Multiple Output (MIMO) technologies. MIMO technology exploits the spatial diversity and can signicantly improve the data rate of wireless communication systems. In the CDMA2000 evolution path towards B3G, CDMA2000 EV-DV [15] stands for CDMA2000s evolution version with support for data and voice, and CDMA2000 EV-DO [16] represents CDMA2000 evolution for data only, with voice being supported through VoIP. Both EV-DV and EV-DO support a data rate of 3.7 Mbps in the downlink and 1.8 Mbps in the uplink, with a channel bandwidth of 1.25 MHz, which is comparable to the data rate per hertz in HSPA. CDMA2000 multi-carrier technology use three times more bandwidth than CDMA2000 single carrier technology and can achieve higher data rates. The counter part for LTE in the CDMA2000 family is Ultra Mobile Band (UMB) [17], which uses OFDM technology in the physical layer and aims at data rates of more than 100 Mbps. In the evolution path of WLAN, 802.11n [18] is proposed to achieve data rates of more than 200 Mbps with MIMO and OFDM technologies. With the increasing data rate of next generation wireless networks, it is possible to include services that require higher data rates, such as realtime video or high-speed data downloading. Also possible in next generation wireless networks is the integration of 5

different types of services, such as voice, video, and data. For example, wireless cellular networks offer data service, while WLAN offers voice service. Higher data rates enable wireless networks to integrate different types of service, and there are several approaches to achieve the goal of integrated services. One approach is to establish heterogeneous wireless networks, which effectively combine existing wireless networks such as wireless cellular networks and WLAN. An effective combination of different wireless networks should enable them to complement each other and provide voice and data service of high quality. The other approach is to develop new wireless standards, which provide full services with broadband wireless access from the very beginning. One such new standard is the Worldwide Interoperability for Microwave Access (WiMAX) standard, which is attracting attention from both academy and industry. Next generation wireless networks also have features of self-organization and self-maintenance, and it is desirable to have intelligent networks that can perform self-maintenance tasks. Characterized by features such as support for different services, heterogeneous networks, broadband access, and self-maintenance, next generation wireless networks should also be able to provide wide area coverage. One approach to achieve better coverage than the current wireless cellular networks is to introduce relay stations into the cellular networks and build multi-hop relay networks. 1.3 Contributions and Dissertation Organizations In this section, I present a summary of the contributions of my research work and state the organization of the dissertation. In my research work, I have conducted research on the following problems for next

generation wireless networks. The rst problem is the Quality of Service (QoS) guarantee issue in the WiMAX network. A scheduling algorithm in the WiMAX Medium Access Control (MAC) layer is proposed to address the QoS guarantee for different types of services in WiMAX network. The second problem studied is the minimum cost problem for data delivery in heterogeneous wireless networks. Several scenarios of the minimum cost data delivery problem are identied and linear programming models are constructed to solve the optimization problem. The third problem studied is the self-maintenance scheduling in next generation wireless networks. A self-maintenance scheduling algorithm is proposed to minimize the total maintenance time while meeting various resource conicting and resource maintenance constraints of next generation wireless networks. The fourth problem studied is the scheduling algorithm of multi-hop relay wireless cellular networks in order to maximize network throughput. The proposed multi-hop scheduling algorithm is frame-based and queue-aware, implemented by a linear programming model. Simulation results identify the effectiveness of the proposed scheduling algorithm. The organization of the dissertation is as follows. In Chapter 2, a system overview of next generation wireless networks is provided, followed by analysis of the challenges for each research problem. In Chapters 3, 4, 5, and 6, we discuss in detail the research work on QoS guarantee of WiMAX network, minimum cost data delivery problem for heterogeneous wireless networks, self-maintenance scheduling for next generation wireless networks, and scheduling algorithm of multi-hop relay cellular networks, respectively. For each research topic, rst the optimization problem is formulated, then solutions are proposed with analysis models built, and the proposed solutions are veried through simulations. Chapter 7 concludes the dissertation. 7

CHAPTER 2 SYSTEM OVERVIEW AND CHALLENGES

In this chapter, rst an overview of next generation wireless networks is presented, then four challenges are identied that correspond to the four research problems studied in the dissertation. For each research problem, we illustrate the challenges associated with it and review a list of related work. We briey describe how our approaches address the challenges and why our proposed approaches are unique compared to related work. 2.1 A System Overview of Next Generation Wireless Networks In Section 1.2 , we have discussed the characteristics and requirements of next generation wireless networks, such as high-speed data rate, integrated services, wide area coverage, and QoS guarantee. In order to meet these requirements, one approach is to integrate different wireless networks together and build heterogeneous networks with the intent of meeting the requirements of next generation wireless networks. The elements for building heterogeneous wireless networks include wireless cellular networks and wireless data networks. The other approach is to develop new wireless network standards, such as WiMAX, LTE [14], and UMB [17]. In the foreseeable future, these two approaches will coexist and thus allow a variety of wireless networks to participate in the construction of next generation wireless networks and form heterogeneous wireless networks. This picture of different wireless networks combined together, including new wireless technologies, is shown in Fig. 2.1 . As we can see from Fig. 2.1 , next generation wireless networks are comprised of different wireless networks, such as WiMAX, WLAN including

Figure 2.1: Next Generation Wireless Networks.

802.11b/g/n, and wireless cellular networks including WCDMA/HSPA and CDMA2000/EV-DO. The major components of next generation wireless networks and the characteristics of the mobile terminals are illustrated in this section. 2.1.1 Major Components of Next Generation Wireless Networks In next generation wireless networks, different types of networks coexist and together they provide integrated services to mobile users. One element is a WiMAX network, which includes Base Stations (BSs), Subscriber Stations (SSs), and Mobile Stations (MSs). In Fig. 2.1 , WiMAX BS1 provides backhaul connections to other stations, such as WiMAX BS2, WiMAX SS1, and WiMAX SS2. Designed to accommodate broadband wireless applications over long distance, a WiMAX link is suitable for trafc aggregation. Usually a SS is installed at a xed position, and it connects to the downstream subscribers with wired connections. For example, WiMAX SS1

in Fig. 2.1 connects to the base stations of cellular networks via a router, and WiMAX SS2 connects to WLAN Access Point (AP) and other wired subscribers also via a router. As indicated in Fig. 2.1 , mobile users such as MS3 and MS4 can connect directly to WiMAX BS2. The broadband capability and architecture exibility enables a WiMAX network to play an important role in next generation wireless networks. When traditional wireless cellular networks evolve to WCDMA/HSPA or CDMA2000 EV-DO, they also become an indispensable part of next generation wireless networks. Cellular networks consist of a wireless access network and a Core Network (CN). A wireless access network connects to a CN via router and gateway, and a CN further connects to internet, data network, or Public Switched Telephone Network (PSTN), as indicated in Fig. 2.1 . Wireless access networks include BS and MS, plus an optional Relay Station (RS). Traditionally, a BS connects to a CN through a wired or ber connection, and with the emergence of the WiMAX network, this wired or ber connection between BS and CN can be replaced by a WiMAX backhaul link, as shown in Fig. 2.1 . MS in wireless cellular networks could be cell phone, laptop, tablet PC, or Personal Digital Assistance (PDA), and all of them need data communication capability in order to enjoy the data service of cellular networks. Notice that MS in cellular networks can connect to a BS through an RS. Cellular networks are indispensable in next generation wireless networks, since they have the largest number of mobile subscribers over wide areas, and they can provide full mobility and handoff support. WLAN is characterized by a high data rate, a small cell radius of around 100m, and a low cost of deployment, and thus WLAN is suitable for providing hotspot and indoor coverage in next generation wireless networks. Several scenarios are shown in Fig. 2.1 . The rst scenario is the deployment of 802.11b/g/n WLAN AP1 and WLAN AP2 within the cells 10

of cellular networks for hotspot coverage, where trafc intensity is high. The second scenario is using WLAN for indoor coverage shown by WLAN AP from AP4 to AP8, where the signal of cellular networks does not penetrate well into the building due to signal attenuations. Besides, cellular networks cannot meet the high-speed data rate requirement as WLAN can. Also notice that, in the second scenario, WLAN AP can form multi-hop relay networks to relay trafc. The third scenario is shown by WLAN AP3 in Fig. 2.1 , where a handoff procedure is provided to support mobile users. WLAN belongs to wireless access networks, and thus it needs to connect to a CN through a router and a gateway, and then to the internet or PSTN. Both data and voice service are supported in WLAN, and the wireless subscribers of WLAN include laptop, cell phone, or PDA with data and/or VoIP features. 2.1.2 Multiple Interfaces in Mobile Terminals As we can see from Fig. 2.1 , some mobile terminals have multiple physical interfaces, such as MS1 and Laptop1 connecting to both a cellular network and a WiMAX network, MS5 and Laptop2 connecting to both a WiMAX network and a WLAN, and MS2 and Laptop1 connecting to both a cellular network and a WLAN. Having multiple physical interfaces gives the mobile devices more choices when facing the question of which network to connect to, and it also helps to diversify the applications that can be run concurrently on mobile devices. For example, a mobile terminal can use one physical interface to connect to a cellular network for voice service and at the same time uses a different physical interface to connect to a WiMAX network for high-speed data downloading. This scenario of multiple interfaces or multiple operating modes of the mobile terminal truly enables interactive and multimedia applications for the mobile subscribers. It also brings challenges to the network operators in

11

different aspects, such as network maintenance, network operating cost, and QoS guarantee, since different types of networks need to work together to achieve optimization goals related to system performance. In the following sections, the challenges of next generation wireless networks will be discussed and the research problems will be outlined. 2.2 Challenges in QoS Guarantee Problem Since providing different types of services is one characteristic of next generation wireless networks, how to guarantee the QoS of the services is of great concern to the network service provider. In WLAN, a new standard 802.11e [19] is proposed in 2005 to provide QoS guarantee. In wireless cellular networks, HSDPA/HSUPA and CDMA2000 EV-DO have entered the stage of commercial deployment, and they all use time division multiplexing to schedule different types of services. The scheduling algorithms in HSPA and CDMA2000 EV-DO have been studied intensively in [20], [21], and [22]. Compared to HSPA and CDMA2000 EV-DO, WiMAX has not entered the commercial stage yet, and the scheduling issue in WiMAX needs to be addressed before it can be widely deployed. Thus in our research, we focus on the scheduling algorithm of WiMAX. In a WiMAX network, trafc streams are classied into different categories, and each category has its own set of QoS parameters such as data rate and delay. The QoS guarantee issue in WiMAX is by nature how to allocate bandwidth resources among these trafc streams so that each stream can have its QoS requirements satised. A WiMAX network is designed to incorporate different types of data streams, and it aims at providing QoS guarantee for all the data streams being served by WiMAX. The WiMAX protocol covers physical layer and MAC layer, and there are several challenges for QoS

12

guarantee in WiMAX. In the physical layer, one challenge is the uncertainty of the wireless channel, which makes the guarantee of broadband wireless data service difcult and renders the static resource allocation scheme unsuitable. In the MAC layer, one challenge is the diversied service types, which requires the WiMAX scheduling scheme to be adaptive to the various QoS parameters of different service types. In order to address these challenges and achieve QoS guarantee, WiMAX has some unique features and mechanisms. The rst salient feature is that WiMAX follows the policy of centralized control in BS. Compared to the distributed access control scheme in WLAN, WiMAX adopts the centralized approach. In WLAN, Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) is used to resolve the coniction of medium access, and there is no guarantee for QoS parameters like data rate or delay. The second feature is that WiMAX BS has two mechanisms in MAC layer for QoS guarantee, i.e., the admission control unit and the bandwidth resource allocation unit. Before any service can run on WiMAX, it has to be admitted by WiMAX BS rst, and thus all WiMAX services are connection-oriented. After a connection is admitted, dynamic bandwidth allocation is deployed to meet the QoS requirement of this connection. Here dynamic bandwidth allocation is chosen instead of static bandwidth allocation because of the variable bit rate nature of connections and the uctuation of the wireless channel quality. In comparison, there is no admission control or dynamic resource allocation scheme in WLAN. The third feature is that WiMAX networks differentiate various service types and precisely dene a set of QoS parameters for each service type. This differentiation scheme allows BS to uniquely treat every service type according to its own characteristics and enables QoS guarantee for each service type. Other QoS guarantee mechanisms in WiMAX network include Automatic Repeat Request (ARQ), 13

adaptive modulation and coding, and power control. The key components in WiMAX QoS guarantee are the admission control and the bandwidth allocation in BS. WiMAX standard denes adequate signalling schemes to support admission control and bandwidth allocation, but does not dene the algorithms for them. This absence of denition allows more exibility in the implementation of admission control and bandwidth allocation. In this work, the focus is the bandwidth allocation, i.e., the scheduling algorithm in the WiMAX MAC layer. The admission control algorithm has been well studied in both wired and wireless networks, and one common admission control algorithm is to admit a connection if the remaining bandwidth is larger than the average bandwidth consumption of the connection candidate. In this dissertation, I adopt this common approach for admission control. The research problem being investigated here is, after connections are admitted into the WiMAX network, how to allocate bandwidth resources and perform scheduling services, so that the QoS requirements of the connections can be satised. Scheduling has also been studied intensively in many disciplines, such as CPU task scheduling in operating systems, service scheduling in a client-server model, and events scheduling in communication and computer networks. Thus a lot of scheduling algorithms have been developed. However, compared with the traditional scheduling problems, the WiMAX MAC layer scheduling problem is unique and worth study for the following reasons. First, the total bandwidth in a WiMAX network is adaptive since AMC is deployed in the physical layer and the number of bytes each time slot can carry depends on the coding and modulation scheme. Second, multiple service types have been dened and their QoS requirements need to be satised at the same time. How to satisfy various QoS requirements of different service types simultaneously has not been addressed by any other wireless access 14

standard before. Third, the time complexity of the WiMAX scheduling algorithm must be simple since realtime service demands a fast response from the central controller in BS. Fourth, the frame boundary in the WiMAX MAC layer also serves as the scheduling boundary, which makes the WiMAX scheduling problem different from the continuous-time scheduling problem. The above four characteristics make the resource allocation in the WiMAX MAC layer a challenging problem. There have been some studies of the WiMAX MAC scheduling problem [23], [24], [25], and [26]. Among these solutions, none of them can address the QoS requirements of various service types, and no analytical work has been developed to derive the theoretical boundary of the performance of the scheduling algorithms. The scheduling algorithm proposed in this work not only addresses the QoS requirements of various service types, but it also builds up an analysis model to derive its performance metrics in theory. The details of the proposed scheduling algorithm for the WiMAX MAC layer are discussed in Chapter 3. 2.3 Challenges in Minimum Cost Problem From Fig. 2.1 , it can be seen that next generation wireless networks integrate different types of wireless networks, e.g., wireless cellular networks, broadband access networks such as WiMAX, and wireless data networks such as WLAN. When a mobile terminal travels through different wireless networks, it will experience different service types. Even if the mobile terminal stays in one place, it still may be under the coverage of multiple networks, thus facing the choice of which network to connect to. The multiple interfaces in the physical layer enable a mobile terminal to communicate to different networks. From the point view of network operators, different network services incur different costs, and it is the network

15

operators concern to reduce the network communication cost. Similarly, from the point view of a mobile subscriber, it is also desirable to reduce the total communication cost during data download or upload. Every data download or upload service has its own QoS requirements to be satised in the data download/upload process, such as delay budget and data rate. The optimization problem studied here is how to minimize the communication cost in heterogeneous wireless networks if a mobile subscriber needs to perform data download/upload across various networks, while meeting the QoS requirement. Intuitively, always choosing the network with the lowest communication cost will result in the minimum cost, but it will not necessarily satisfy the QoS requirements of the data service. On the other hand, always choosing the network with the highest data rate may help to satisfy the QoS requirements, but it will increase the communication cost. Consequently, there must be an optimal point which reaches a balance between communication efciency and communication cost. Thus it is non-trivial to study this minimum cost problem to nd out the optimal solution. There has been some pioneering work regarding the choice of networks in integrated wireless networks. In [27], handoff among overlaid wireless networks is studied, and the concept of vertical handoff is introduced to describe the handoff process between overlapped wireless networks. In this approach, the mobile user always switches to the network with the highest data rate per unit coverage area. In [28], a cost function is dened to serve as the metrics for handoff, and the mobile subscriber tends to switch to the network with the least cost function. Although the cost function sheds some light on the selection of networks, it reects only the local optimization result since the mobile user stays in a specic location. The optimization problem studied in this dissertation envisions a global view of the networks, and the optimization process is performed with respect to the dynamic data 16

download process across different networks, instead of local optimization in a static place. The contributions of the approach in the dissertation include the NP-hard proof of the cost minimization problem and the development of several realistic scenarios, such as prefetched data, multiple routes, variable speeds, and random routes. Optimization problem is formulated by developing linear programming models for all these scenarios. Linear programming models are then solved by standard linear programming tools. The details of the optimization problem for minimum cost problem in data delivery are discussed in Chapter 4. 2.4 Challenges in System Maintenance Problem Besides integrating different service types, next generation wireless networks also incorporate wireless multi-hop architecture. As can be seen from Fig. 2.1 , MS3 and MS4 form multi-hop wireless networks, and so do WLAN AP5, AP6, AP7, and AP8. This multi-hop structure requires each wireless node to take certain responsibilities in the networks, such as trafc forwarding or data backup. After performing service in the networks for a certain amount of time, each wireless node may generate system maintenance requests. The reasons for system maintenance requests of wireless node can be limited amount of resources (such as battery power or storage capacity), or routine maintenance requirements (such as software update). When there is a need for system maintenance scheduling, we formulate two types of constrains that must be considered. One is Resource Conicting (RC) constraint, and the other is Resource Maintenance (RM) constraint. Examples are indicated in Fig. 2.1 . RC constraint applies when WLAN AP3, AP4, and AP5 share the same backup server, and these three APs can only be scheduled for maintenance at different time slots since the backup server can only serve one node at a time. An example for RM constraint is when WLAN AP6,

17

AP7, and AP8 serve as forwarding nodes. They cannot all be ofine for maintenance at the same time, since at least one AP is required to stay online to forward trafc. The scheduling problem that needs to be solved here is how to schedule the maintenance requests generated by wireless nodes so that the total maintenance time is minimized while satisfying all the RC and RM constraints. The maintenance scheduling problem is crucial for next generation wireless networks, because each wireless node need its maintenance service being performed smoothly. In addition, minimizing the total time used in scheduling can reduce the maintenance overhead and help to achieve the optimal performance of the networks. From another perspective, wireless networks with self-maintenance feature reects the idea that next generation wireless networks should be intelligent networks, and the wireless networks can heal itself automatically. There has been very little prior work in the constraint-based scheduling algorithm for system maintenance in wireless networks. [29] and [30] both aim at prolonging the lifetime of wireless nodes in ad hoc networks by switching some nodes off and saving their power consumptions, but only remaining power is considered in the scheduling decision instead of RC or RM constraints. In the proposed scheduling algorithm for maintenance, the optimization problem is dened rst, then a linear program model is proposed as the scheduling maintenance algorithm. In addition, two heuristic algorithms are developed for the scheduling problem. Our contributions in the maintenance scheduling problem include not only the linear programming model and the two heuristic approaches, but also the consideration of RC and RM constraints that have not been addressed before. The details of the scheduling problem for system maintenance are discussed in Chapter 5.

18

2.5

Challenges in Scheduling for Relay Networks In Fig.2.1 , besides the multi-hop transmission in ad hoc networks formed by WLAN

APs, multi-hop transmission in cellular networks is also shown by the relay station RS1 inside BS2 of cellular wireless networks. There are different approaches to build multi-hop relay networks under the architecture of wireless cellular networks, and in general we can divide the approaches into two categories, the ad hoc networks approach and the centralized approach. In an ad hoc network approach to build relay networks, mobile stations can also serve as relay stations, and they together form ad hoc networks to relay trafc. One advantage of this ad hoc approach is that the relay networks can be formed in a more arbitrary manner, and the other is that the mobile stations can have more route candidates to choose. Since mobile stations have to route packets for other destinations in the ad hoc approach, mobile stations will consume extra bandwidth and need software updates to route packets. In addition, it is difcult for the BS to provide QoS guarantee since the BS does not control the data delivery process in ad hoc networks. Compared to an ad hoc network approach, we are in favor of the centralized approach, where a base station is in control of all the associated relay stations and mobile stations, and all the data transmission schedules are determined by the base station. In this centralized approach, mobile stations do not need to relay trafc and thus no software change is needed. Besides, no extra bandwidth will be consumed in mobile stations to relay trafc for other mobile stations. Given this centralized approach of building relay networks in cellular networks, one important question is how the BS will schedule the concurrent transmissions so that the network throughput can be increased. On the one hand, relay stations can provide

19

wireless channels with higher data rate, and on the other hand, the increase of transmission hops will prolong the transmission time and can potentially harm the throughput. Thus a scheduling algorithm is needed to ensure that network throughput will not decrease due to multiple hops in transmission. Scheduling in cellular relay networks is challenging due to the following reasons. First, as the input for the scheduling algorithm, concurrent transmission scenarios need to be determined in an efciently way. When considering a data link candidate for a concurrent transmission scenario, it must be guaranteed that the adding of this link into the scenario will not decrease the total throughput of all the links in the scenario. It is not practical to traverse all the links to search for concurrent scenarios due to the non-linear growth of links with respect to the number of nodes. The second challenge is due to the frame-based nature of cellular networks. In each frame, the scheduling algorithm will aim at maximizing the throughput, thus for different concurrent scenarios, they have to share the frame duration. How to maintain fairness among various users in different scenarios while achieving max throughput is an important issue. The third challenge is how to consider the dynamic queue size in the relay nodes and let the scheduling algorithm adjust to the realtime change of the queues. There have been several approaches to address the scheduling in relay cellular networks. In [31], two concurrent transmission scenarios are enumerated in a simple network topology. This enumeration approach cannot scale with the growth of wireless nodes in the cellular networks. Besides, this approach is based on ad hoc networks. In [32], both the scheduling algorithm and concurrent scenarios are provided for a specic topology, and it does not address how to derive them in a general network topology. In [33], a linear programming 20

model is proposed to schedule the concurrent transmissions in relay cellular networks, but several practical factors of the cellular networks such as frame boundary and dynamic queue status are not considered in this scheduling algorithm. Nor does this approach try to balance throughput and fairness. We propose a linear programming model for the scheduling algorithm, and we aim at maximizing network throughput and achieving fairness. We also consider the frame-based nature of cellular networks and the dynamic queues of the relay stations. All these make our scheduling algorithm a practical and an effective approach to address the multi-hop transmissions in relay cellular networks. The details of the scheduling algorithm for multi-hop relay cellular networks are discussed in Chapter 6. 2.6 Summary In this chapter, an overview of next generation wireless networks is depicted in Fig. 2.1 . The characteristics of the various wireless access technologies in next generation wireless networks are discussed, and the conclusion reached is that next generation wireless networks should be heterogeneous networks with integrated services. The four research problems identied are the WiMAX MAC layer scheduling problem, the minimum cost data delivery problem among different networks, the self-maintenance scheduling problem in next generation wireless networks, and the multi-hop relay scheduling problem in cellular networks. For each research problem, we rst present its formulation, then illustrate why it is considered a challenging problem. After introducing the existing solutions for each of the problems, we highlight the unique contributions of our proposed solutions. In the following chapters, we will discuss the four scheduling and resource optimization problems in detail.

21

CHAPTER 3 QOS GUARANTEE FOR WIMAX NETWORKS

In this chapter, the QoS guarantee issue in a WiMAX network is discussed. 3.1 Introduction With the emergence of various innovative applications, business models, and low-cost but powerful mobile devices, the WiMAX network is prepared for unprecedented growth, especially after the release of the wireless broadband access standard IEEE 802.16 [34] in 2004 and its mobile version IEEE 802.16e [35] in 2005. One of the salient features of the WiMAX network is its ability to support a variety of applications of different types, including realtime broadband video conferencing, non-realtime large volume data transfer, traditional voice/data trafc throughpu E1/T1 connection, and web browsing. The need to support all of these different types of services brings new challenges in the design of the WiMAX network. In this work, we develop an efcient uplink scheduling algorithm for the WiMAX network, which is a key component for differentiating different trafcs and supporting their QoS. In the WiMAX network, a BS usually operates in point to multi-point mode, where it connects to multiple SS simultaneously. All data transmissions are connection-oriented and the connections are classied into four types, namely, Unsolicited Grant Service (UGS), also known as Constant Bit Rate (CBR), realtime Variable Bit Rate (rtVR), non-realtime Variable Bit Rate (nrtVR), and BE (Best Effort). For each established connection, the bandwidth requests are sent from the SS to the BS. Given the limited bandwidth of the uplink data channel shared by all connections, the BS needs to run an efcient scheduling algorithm to

allocate bandwidth to different connections in order to meet their QoS requirements. As to be elaborated in Section 3.2 , the WiMAX uplink scheduling is unique, due to its adaptive total bandwidth, QoS requirement, short scheduling delay budget, and frame-based data arrival and service. In this work, we introduce a two-phase Fair and Efcient Queueing (FEQ) algorithm tailored for uplink scheduling in the WiMAX network. It combines Weighted Round Robin (WRR) and Earliest Deadline First (EDF) algorithms, aiming to strike the balance between delay requirement and fair bandwidth allocation. In order to gain deep understanding of and insights into FEQ, we establish an elegant queueing model to derive in theory the performance metrics in terms of packet drop rate and throughput. The analytic model is veried by extensive simulations, carried out on the basis of a broad set of parameters according to the WiMAX physical layer standard WirelessHUMAN-SCa. In simulations, we consider both popular Poisson trafc and practical bursty trafc that is modelled by the Markov Modulated Poisson Process (MMPP). The simulation results verify the correctness of our analytical models and compare FEQ with other scheduling schemes. Both analytic and simulation results clearly show that FEQ effectively combines WRR and EDF. It achieves a low drop rate and high throughput while maintaining fairness among different connections at the same time. The rest of this chapter is organized as follows. Section 3.2 provides background for the WiMAX network, discusses the uniqueness of WiMAX uplink scheduling, and presents the related work. Section 3.3 elaborates the proposed two-phase FEQ algorithm, followed by an elegant analytic model for evaluating its performance and providing deep insights into its queuing behavior under QoS-oriented queuing disciplines. Simulation results are illustrated and discussed in Section 3.4 . Section 3.5 summarizes the chapter. 23

3.2

WiMAX MAC Layer In this section, we discuss the WiMAX MAC layer, the uniqueness of WiMAX uplink

scheduling, and the related scheduling algorithms. 3.2.1 WiMAX MAC Layer Structure WiMAX data transmissions include uplink (from multiple SSs to a BS) and downlink (from a BS to multiple SSs). To support the two-way communication, either frequency division duplex (FDD) or time division duplex (TDD) can be adopted. In the following discussion, we focus on the popular TDD, while our proposed approach applies to the FDD as well. In TDD mode, a WiMAX MAC frame consists of two subframes, DL-subframe for downlink transmission and UL-subframe for uplink transmission, as shown in Fig. 3.1. The DL-subframe comprises a burst preamble (BP) and a frame control header (FCH), followed by DL-MAP, UL-MAP, and a number of downlink payload bursts (DL-PL1 , , DL-PLn ). As will be discussed later, the UL-MAP contains the uplink scheduling results, i.e., the uplink bandwidth granted to each SS. The UL-subframe starts with inital ranging contention slots and bandwidth request contention slots, followed by a number of uplink payload bursts (UL-PL1 , , UL-PLn ). The uplink and downlink bursts are not necessarily equal and their length can be adjusted dynamically in order to adapt to the trafc variation. All data transmissions in the WiMAX network are connection-oriented. Each connection can be one of four possible types, UGS, rtVR, nrtVR, or BE, with different QoS requirements. In order to establish a connection, either a SS or a BS can initiate a connection request that contains necessary QoS parameters of the connection. For example, each connection may request a minimum reserved rate (MRR), a maximum delay, and/or a maximum tolerable

24

One Frame Burst FCH Preamble ( DL_MAP,

UL_MAP )

DL PL1

DL Initial Ranging BW request PL n contention contention

UL PL 1

UL PL n

Downlink Subframe Adaptive

Uplink Subframe

Figure 3.1: WiMAX MAC Frame Structure. jitter. The Connection Admission Control (CAC) is performed by the BS. The basic idea of CAC is to ensure the total reserved bandwidth (i.e., the sum of the MRR of each admitted connection) not exceeding a given threshold, which is usually the total available bandwidth. Once a connection is admitted, the QoS of this connection should be guaranteed by the BS through a proper scheduling algorithm. A unique connection ID (CID) is assigned to each admitted connection, which helps to identify the link. Like many other infrastructure-based wireless systems, downlink transmission is easier to manage, because the BS has all trafc information and corresponding QoS requirements. For uplink, however, the BS does not have the trafc information of each SS, and thus a request/grant mechanism is needed to gather the request information from each SS and then to schedule and grant the requests. Different types of connections may employ different request/grant mechanisms, as discussed below. Since UGS trafc requests a constant bit rate with stringent requirements for delay and jitter, the data grant is given out in a periodic manner without actual request being sent from the SS. Once a UGS connection is admitted, the periodic data grant interval and the data grant size are determined. The BS only needs to allocate bandwidth to the UGS connection accordingly. The rtVR connection has a variable data rate, and thus its trafc cannot be precisely predicted on a frame-by-frame basis. For

25

efcient bandwidth utilization, a polling scheme is employed. Upon being polled by the BS, the SS reports the number of bytes waiting for transmission in its sending queue. Since rtVR is sensitive to delay, unicast polling is directed to each individual rtVR connection at a xed polling interval, without contentions with any other connections. As a result, the SS can always send in its rtVR request timely. Clearly, the polling interval must be short enough to meet the delay requirement of rtVR. In nrtVR, the BS also gives unicast polling grant to each connection, as that in rtVR, but with a longer interval, because the trafc type is non-realtime. Besides the endowed unicast polling opportunity, each nrtVR connection can also send its bandwidth request during the bandwidth request contention slot, as shown in Fig. 3.1. Such requests are subject to possible collisions, and thus are not always received by the BS. The BE connection has no minimum reserved bandwidth. Thus the BE connections get whatever bandwidth remains after all UGS, rtVR, and nrtVR connections are served. Similar to nrtVR, each SS send BE bandwidth requests to the BS in the bandwidth request contention period. At the beginning of each frame, all uplink requests gathered by the BS in the previous frame are passed to the scheduling unit. Then, the scheduling result is broadcasted to all SS in the UL-MAP messages (see Fig. 3.1) of the current frame and/or the successive frames. By looking up UL-MAP, each SS understands when to start its transmission and how much data to transmit. This procedure repeats in every frame. The above bandwidth request/grant mechanism and the corresponding signalling messages have been dened in the WiMAX standard [34]. The scheduling policy, however, still remains open. In our work, we will focus on uplink scheduling.

26

3.2.2

Uniqueness of Uplink Scheduling in WiMAX MAC Layer Scheduling has been extensively researched in computer and communication systems. It

deals with the problem of sharing one output link among multiple input links. Compared with previous studies, the WiMAX MAC uplink channel has a few characteristics and requirements, as elaborated below, which together make the WiMAX uplink scheduling a unique and challenging problem. Adaptive Total Bandwidth: Each WiMAX physical frame is composed of a xed number of time slots, which are the minislots under single carrier systems such as WirelessMAN-SC and WirelessMAN-SCa or the modulation symbols in multiple carrier systems like WirelessMAN-OFDM and WirelessMAN-OFDMA [34]. Each time slot, however, can accommodate a variable number of payload bytes, depending on the adaptive modulation and coding (AMC) scheme used. AMC dynamically adjusts the data rate of each connection according to its channel condition, resulting in different bandwidth efciency, i.e., the number of bytes that can be carried per time slot. Consequently, the total bandwidth of the frame is adaptive. QoS Requirement: Four types of connections, UGS, rtVR, nrtVR and BE, are supported in the WiMAX network. In general, once a connection is admitted, the scheduling algorithm needs to satisfy its requested QoS parameters, such as MRR and delay. A delay budget is usually associated with the rtVR and the nrtVR connections. The data packets are dropped if they cannot be delivered within the delay budget. By considering this delay constraint, the scheduling algorithm gives higher priority to the data with tighter delay budget (i.e., schedules them rst). Meanwhile, MRR is 27

employed to ensure a minimum data rate for every connection, in order to avoid starvation of connections with loose or no delay requirement. It is non-trivial to satisfy both MRR and delay requirements. Short Scheduling Delay Budget: As can be seen in Fig. 3.1, the BS receives uplink data bursts from each SS one by one. Since the bandwidth request may be included in any data burst, the BS can accumulate the complete requests only after it receives the last burst. Meanwhile, the scheduling results need to be broadcasted in the UL-MAP message at the beginning of the next frame. Therefore, the turnaround time for scheduling is a short period from the end of current frame to the beginning of UL-MAP message of the next frame. Clearly, the uplink scheduling algorithm must be simple in time complexity. Frame-based Data Arrival and Service: While data arrival at the SS may be continuous, the bandwidth requests are frame-based from the perspective of the BS. More specically, the input for the scheduling algorithm includes the requests received by the BS in the current frame and those requests received in the previous frame that have not been scheduled. If a request from a SS arrives at the BS in the current frame successfully, it will be fed as input to the scheduler and may be scheduled in the next frame; if it fails to arrive in the current frame, it has to wait for the next polling opportunity or successful contention slot, which could be one or several frames later, depending on the polling interval or contention collision probability. An analogy of this scheduling is work-and-wait, in contrast to typical whenever come whenever serve scheduler. As a result, all requests received in one frame are assumed to be orderless,

28

(i.e., with the same arrival time). This frame-based polling structure renders the exact calculation of any sort of virtual nish time in the packet level unnecessary. 3.2.3 Related Scheduling Algorithms In general, the scheduling algorithms can be classied as frame-based scheduling and sorted-based scheduling. Frame-based scheduling algorithms include Weighted Round Robin (WRR) [36], Decit Round Robin (DRR) [37], etc. Sorted-based scheduling algorithms include Weighted Fair Queueing (WFQ) [38], also known as Packet-based Generalized Processor Sharing (PGPS) [39], and a number of variations of WFQ such as Worst Case Fair Queuing (WF2Q) [40], Self-Clock Faire Queuing (SCFQ) [41]. The advantage of frame-based scheduling algorithms is their low computing complexity, while the disadvantage is the signicant worst case delay. On the contrary, scheduling algorithm in the WFQ family has better performance in worst case delay, but the algorithm complexity is much higher than that of the frame-based scheduling algorithms. When it comes to scheduling tasks with delay budgets, EDF [42] has been proven to be optimal if the trafc load is appropriate. However, when the system is overloaded, the performance of EDF is largely unpredictable. Although these algorithms are effective for solving many scheduling problems, they do not address the unique scheduling requirements of the WiMAX network, as discussed in the previous subsection. Several WiMAX scheduling solutions have been proposed. WFQ is proposed and analyzed by an M/G/FQ queueing model in [23]. Although WFQ can guarantee the minimum data rate of the connections, it does not take into consideration the delay constraint. Besides, the time complexity of WFQ scheduling algorithm is high and thus becomes a potential

29

problem for its implementation in the WiMAX network. A two-tier hierarchical architecture is proposed in [24] for WiMAX uplink scheduling. In the higher hierarchy, strict prioritization is used to direct the trafc into the four queues, according to its type. Then, each queue is scheduled according to a particular algorithm, i.e., xed allocation for UGS, EDF for rtVR, WFQ for nrtVR, and equal division of remaining bandwidth for BE. Although EDF takes care of the delay requirement of the rtVR, grouping multiple rtVR connections into one queue fails to guarantee the minimum bandwidth requirement of each individual rtVR connection. For example, one rtVR connection with tight delay budget may dominate the bandwidth allocation, resulting in starvation of other rtVR connections. A similar approach is proposed in [43], and it replaces the strict priority algorithm in the higher hierarchy with Decit Fair Priority Queue (DFPQ). The basic idea of DFPQ is to use DRR to guarantee the bandwidth allocation of each of the four queues, thus preventing queues with higher priorities from depleting the bandwidth and causing starvation of queues with lower priorities. However, the same unfairness problem exists in each queue as discussed above, since multiple connections of the same type share one queue. In addition, since WFQ is used for non-realtime connections, the overall complexity of the algorithm is high as discussed earlier. A cross-layer approach is proposed in [25], where a priority function (PRF) is dened for every rtVR, nrtVR, and BE connection and is updated on a frame basis. The advantage of this cross-layer approach is to take into consideration the adaptive modulation and coding scheme of each SS to achieve high bandwidth efciency, while its disadvantage is the lack of support for the bandwidth requirement of rtVR. Furthermore, it requires keeping track of the average throughput of nrtVR connections, and thus increases algorithm complexity. A queue-based scheduling algorithm is proposed in [44]. In this approach, the bandwidth is divided between 30

realtime and non-realtime trafc according to a ratio, and this ratio is calculated in direct proportion to the number of packets waiting in the queue. This scheme cannot guarantee the QoS of each individual realtime or non-realtime connection since it consider all the realtime or non-realtime trafc as a group, neither can it guarantee the fairness among different connections. In [45], the authors propose an enhanced algorithm for Wireless Decit Round Robin (WDRR), called Uniformly-Fair WDRR (UF-DRR). In WDRR, connections that are in good channel state with packets to send can take the quantum of connections that are in bad channel state or do not have any packet to send. The authors in [45] argue that in WDRR, the search for good state connections always starting from connection number 1 can cause unfairness. In UF-DRR, the information of connections in good state and bad state are gathered rst, and the surplus resources are distributed among all the connections and it results in fairness. This approach does not address the individual needs of each connection and it cannot prevent greedy connections from occupying the entire bandwidth. In [46], a Service Criticality (SC) based scheduling is proposed for the WiMAX network, where an SC index is calculated in every SS for each connection and then sent to BS, and BS sorts the SC of all the connections and assigns bandwidth according to the descending order of SC. SC is derived according to the buffer occupancy and waiting time of each connection. If a malicious connection always reports a high SC, or a connection is generating excessive trafc to occupy its sending buffer, this connection will dominate the available bandwidth and affect other connections. In [47], each downlink connection is assigned an index value proportional to the its queue length. The larger the index value, the higher the probability the BS will schedule the connection. This approach apply the Littles theorem to derive the index, which implies that it tries to maintain the average delay of the connection, but not the delay of every packet. 31

Also this approach is not a fair algorithm, since it always favors connections with more backlogged trafc. In [48], the requests are classied into three priorities, P0 , P1 , and P2 , with decreasing priorities. P0 corresponds to the realtime trafc within the minimum reserve data rate and about to violate the delay budget. P1 corresponds to the realtime trafc within the minimum reserve data rate but not about to violate the delay budget, plus the non-realtime trafc within the minimum reserved rate. P2 corresponds to the realtime and non-realtime trafc above the minimum reserve rate, plus the best effort trafc. The scheduling algorithm satises the P0 request rst, and lets the P1 trafc share the remaining bandwidth according to the minimum reserved rate of each connection. After P0 and P1 are satised, the remaining bandwidth (if there is any) will be equally divided among the connections in P3 . This approach is not fair since it always favors connections with short delay budgets, and by equally dividing the P2 trafc, it fails to deal with burst trafc that deserves more bandwidth allocation when burst trafc comes. Analysis model is an important part of the research in scheduling algorithms. In [23], mean delay bound is derived by using the M/G/FQ queueing model, but it is based on WFQ algorithm and does not consider QoS parameters other than MRR. QoS parameters are not considered in [26], where an MMPP model is employed to derive the packet drop rate and characterize uplink rtVR and nrtVR trafc that share a single First-Come-First-Serve (FCFS) queue. For other scheduling algorithms [24], [43], [25], [44], [45], [46], [47], and [48], their performance and effectiveness are demonstrated and compared via simulations only, without in-depth theoretical analysis.

32

3.3

Two-phase Fair and Efcient Queueing (FEQ) Scheduling Algorithm From the above discussion, it is desirable to have a simple and effective scheduling

algorithm, which can take care of the QoS requirements of all connections. In this section, we rst present a two-phase Fair and Efcient Queueing (FEQ) algorithm that addresses the delay and bandwidth requirements while balancing the fairness and efciency among different connections. Then, an elegant queueing model is established to derive in theory its performance in terms of packet drop rate and throughput. 3.3.1 Overview of the Two-phase FEQ Scheduling Algorithm As we have discussed in Section 3.2 , the WiMAX scheduling algorithm is frame-based. At the beginning of each frame, all uplink requests gathered by the BS in the previous frame are passed to the scheduler as inputs, and the scheduling result is broadcasted to all SS in the UL-MAP messages of the current frame and/or the successive frames. This scheduling procedure repeats itself in every frame. Note that since AMC is employed in the WiMAX network, different SS may choose different modulation and coding schemes, resulting in different bandwidth efciency, i.e., the number of payload bytes per time slot. In order to accurately and efciently allocate the available bandwidth, the uplink bandwidth requests that are expressed in terms of bytes are all converted into time slots according to the specic AMC scheme used by the SS. Consequently, the grant to the request is also in terms of time slots. Similarly, the minimum reserved rate (MRR) is converted into time slots per frame, for scheduling and CAC. The conversion from bytes to time slots are discussed later in Section 3.4 . To facilitate and simplify our following discussion, we use bytes and time slots interchangeably in the rest

33

of this chapter without distinction. The scheduling algorithm aims to meet the QoS requirements of all types of connections. For UGS, the scheduling is straightforward, because a xed amount of bandwidth is always allocated to each UGS connection during each grant interval, which is determined according to the requested data rate. BE connection has no specic QoS requirement, thus it is not the interest of the study. BEs scheduling can be done via a simple scheme like equal division of the remaining bandwidth among all BE connections. In the following discussion, we focus on rtVR and nrtVR trafc that has specic QoS requirements on delay and bandwidth. To strike the balance between delay and bandwidth requirements, the proposed scheduling algorithm decouples them and addresses them separately in two phases. Phase 1 employs WRR to guarantee each connection with a reserved bandwidth share and to maintain fairness among all connections. WRR [36] has been well studied thus its details are omitted here. Briey, the bandwidth (i.e., the time slots) in a frame is allocated to each connection according to its weight, which is chosen to be the MRR of the connection. By applying WRR in Phase 1, every connection can enjoy a basic service rate according to its reserved bandwidth. At the same time, it can prevent any connection from depleting the entire bandwidth, especially when this connection generates extra amount of trafc over its MRR. In other words, Phase 1 tries to separate and protect each connection according to its MRR. Note that due to the random and bursty nature of incoming trafc, the requested bandwidth of a connection in a given frame may be higher or lower than its MRR. If it is higher, part of the request cannot be scheduled in Phase 1, and will be passed to Phase 2; otherwise, the entire request is served in Phase 1, and at the same time, the remaining bandwidth is left for use in Phase 2 by other connections that still have unscheduled requests. 34

The principle of Phase 2 is to let all connections compete with each other over the remaining bandwidth after Phase 1, with priority given to the trafc with smaller delay budget, which is more likely to be dropped if not scheduled in this frame. To this end, we apply the EDF algorithm [42] in Phase 2. Specically, all remaining requests are put into a single queue, sorted in an ascending order according to their delay budgets. The one with the smallest delay budget is scheduled rst, until all bandwidth in the frame is allocated or all requests are served. EDF has been proven to be optimal for scheduling packets with delays under appropriate trafc load. It can be seen that in the design of the scheduling algorithm in Phase 1, we emphasize fairness and try to provide every connection with a minimum amount of service guarantee, while in Phase 2, we emphasize efciency and try to minimize data dropping probability. Thus we name our proposed two-phase scheduling algorithm Fair and Efcient Queueing (FEQ). 3.3.2 Analysis of the Two-phase FEQ Scheduling Algorithm In this subsection, we analyze the performance of the proposed FEQ algorithm, in order to gain insight into it and to demonstrate in theory its performance. Similar to our earlier discussions, we focus on the QoS performance of rtVR and nrtVR connections only in our analysis. Note that, we do not explicitly distinguish rtVR and nrtVR connections, because they only differ in the amount of delay budget. As to be discussed next, our analytic model is generally applicable to any rtVR or nrtVR connections with given arrival rates, MRR, and delay bound. The analysis of the FEQ algorithm is non-trivial, due to the following reasons. Different queuing principles: FEQ involves two phases with different queuing

35

principles, i.e., FCFS in Phase 1 and EDF in Phase 2. Delay constraint: Each data packet in the queue is associated with a specic delay bound. A data packet is dropped if its waiting time is longer than its delay budget. The delay constraint dramatically increases the analytic complexity. Unknown input for Phase 2: The input for Phase 2 is the data not served in Phase 1, which is an unknown parameter to be derived in analysis. As a result, none of the available analytic models can be applied here. In this work, an elegant queueing model as shown in Fig. 3.2 is constructed to describe the two-phase FEQ algorithm. Our key idea is to creatively use the M/D/1+D (for Markov arrival, deterministic service time, one server, plus deterministic delay budget) queuing model with FCFS and EDF queuing disciplines. We consider a general scenario, where n rtVR and/or nrtVR connections are established. Connection i has an arrival rate of i , a delay budget of i , and a minimum reserved bandwidth of M RRi . For the sake of analytic tractability, we assume that the data arrival forms a Poisson process and all queues have innite size. Other types of trafc (such as the more practical bursty trafc) are studied through simulations, as to be discussed in Section 3.4 . Following the FEQ scheduling algorithm, the queuing model also consists of two phases. In Phase 1, since the connections are served according to WRR with their reserved bandwidth (i.e., MRR), each of them can be modelled as a separate and independent queue. Assuming the data packet size is xed, the service rate is a constant. More specically, the service rate of Queue i (for Connection i) is i = M RRi , l 36 (3.1)

Phase 1 1 2

1 Queue FCFS 1 2 Queue FCFS 2

P1 1
d

Phase 2

P2 2
d

Queue EDF

Pd

n Queue FCFS n

Pn n
d

Figure 3.2: Analytical Model for the Two-phase FEQ Algorithm. where l denotes the length of the data packet. Clearly, the service rate of Queue i is proportional to the MRR of Connection i. Note that, data packets that cannot be served in Phase 1 due to the limited MRR will all be passed to Phase 2. Therefore, we articially set a delay bound of T (which is the period of one frame) for the queues in Phase I, in order to track the dropped data packets of Phase 1, which will be the inputs for Phase 2. Based on the above discussion, we arrive at an M/D/1+D queue for each connection in Phase 1, with an arrival rate of i , a service rate of i , and a delay budget of T . The M/D/1+D queue under FCFS queuing discipline has been well studied. According to [49], the packet drop rate of Queue i is: Pid where i equals 1 i = 1 + [i + 2 ie i
r

(1)j
j =0

(i i j )j i j 1 e ] , j!

(3.2)

i , is the delay bound which is set to T , and r is the integer satisfying i

(i 1) < r i . The packets dropped from Phase 1 become the input of Phase 2. Because such packet dropping is random, the dropped packets of Queue i also form a Poisson process, with mean arrival rate of i Pid . Since the EDF scheduling scheme in Phase 2 utilizes all remaining bandwidth and does 37

not distinguish the packets according to which queues (in Phase 1) they come from, the Phase 2 results in a single queue model. As shown in Fig. 3.2, the input of the queue is the aggregation of the dropped packets from all queues in Phase 1. Since the dropped packets from each queue are Poisson, the aggregation of them are also Poisson, with a mean arrival rate of
n

=
i=1

i Pid .

(3.3)

The service rate in Phase 2 denoted by is estimated as follows. The amount of bandwidth available for Phase 2 is what has been left after Phase 1. In other words, if there is remaining bandwidth after the scheduler has performed the WRR scheduling algorithm in Phase 1, this amount of remaining bandwidth will be allocated for EDF scheduling in Phase 2. We assume that CAC ensures the total reserved bandwidth never greater the total available
n

bandwidth, i.e.,
i=1

M RRi W , where W denotes the total available bandwidth. Thus the

remaining bandwidth comes from:

n

the bandwidth that is not reserved by any connections, i.e., W

i=1

M RRi , and

the bandwidth that is reserved but not actually utilized in full by some connections. While the former is straightforward, the latter is worth elaboration. When a connection is idle, its reserved bandwidth is not used, and thus contributing to the bandwidth of Phase 2. To gure out this potential bandwidth, we rst calculate the fraction of time, during which a connection is idle. For Queue i in Phase 1, the fraction of time when packets are served is: (1 Pid ) t tbusy = t t
1 i

= (1 Pid )

1 , i

(3.4)

where t denotes any given time interval, and tbusy represents the busy period within t. In the 38

above equation, (1 Pid ) t is the actual number of arrivals in t (excluding those being dropped), and by Pio , is Pio = 1 i (1 Pid ) Accordingly, the overall service rate in Phase 2 is = W
n i=1

1 is the service time for each packet. Thus the fraction of idle time, denoted i 1 . i

(3.5)

M RRi

+
i=1

Pio i .

(3.6)

Actually, Equation (3.4) is a form of Littles theorem [50], and 1 P 0 is also known as the utilization factor. Equation (3.6) shows the average service rate in Phase 2. The actual available bandwidth in Phase 2 is clearly dynamic, depending on the remaining bandwidth in Phase 1. For analytic tractability, however, we argue that the variation of available bandwidth in Phase 2 is small, because it is the aggregation of many connections. When there are a large number of connections, the result converges to the average value given by Equation (3.6), according to the Law of Large Number. Therefore, we model Phase 2 as an M/D/1+D queue, with an arrival rate of and a service rate of . In addition, each packet in the queue is associated with a delay budget, e.g., i for the packet from Connection i. Note that, although we still arrive at an M/D/1+D queue here, Equation (3.2) cannot be employed to derive the dropping probability, because Phase 2 follows EDF discipline. The queuing analysis of M/D/1 + D under EDF discipline has been studied in [51]. Next, we discuss how to apply this method in our model to derive Pd . For simplifying our discussion, lets consider two types of connections, with delay L for Type 1 and L + M for Type 2, respectively. Both L and L + M are measured in terms of time 39

Plane A (1-P1)k+1-p(1-P2)k+1-q P1 p P2 q Spq(0) Spq(1)

(1-P1)k+1-p(1-P2)k+1-q P1 p P2 q

(1-P1)k+1-p(1-P2)k+1-q P1 p P2 q

Spq(k) (p,q) = {(0,1),(1,0),(1,1)} 0 [k [LN-1

(1-P1)LN(1-P2)(L+M)N

(1-P1)LN(1-P2)m P2

(1-P1)k(1-P2)k- max(0, j-N)

S0 Plane B (1-P1)i P1 (1-P1)LN(1-P2)(L+M)N max(0, j-N) S(0,1) S(0,2) S(0,3) S(0,4)

S(1,2)

S(1,3)

S(1,4)

S(i,j) 0 [i [LN-1, i+1 [j[(L+M)N

(1-P1) P1 (1-P1)N(1-P2)(L+M)N ( m-N+1)

(1-P1)

k-(L-1)N

(1-P2)

k-(m-N+1)

(1-P1)i-(L-1)NP1

(1-P1)LN(1-P2)m P2

Plane C (1-P1)N(1-P2)N-2 P2 S(2,2) (1-P1)N(1-P2)N P2 (1-P1)N(1-P2)N-1 P2 S(LN,m) LN[m [(L+M)N-1 (1-P1)N(1-P2)N-1 P2 S(2,3)

Figure 3.3: Markovian Model of M/D/1 + D Queue under EDF. unit. Each time unit is divided into N intervals, where N ranges from 45 to 70 as suggested in [51]. The Poisson arrival process is approximated by N Bernoulli processes, with parameter P = . By dividing the time unit into N discrete intervals, we arrive at a set of countable N

states and thus a nite state Markovian model can be constructed. More specically, a two-dimensional Markovian model is established to describe the scheduling scenario under EDF discipline. The size of the Markovian model is determined by N , L, and M . For illustration purpose, a diagram of the Markovian model is depicted in Fig. 3.3, with simple parameters of N = 2, L = 1 and M = 1. P1 and P2 in Fig. 3.3 stand for the packet arrival probability of Type 1 and Type 2 connections, respectively. A window of (L + M ) N intervals is considered in the Markovian model. The arriving packets are put into this window according to their remaining delay budget. An arrived packet is either served during its window or dropped when it moves out of the window. In Fig. 3.3,

40

the states are classied into four different groups, indicated by S 0 and three planes, i.e., Plane A, B , and C . S 0 represents those states in which all packets have been served during their stay in the window. States in Plane A are denoted by Spq (k ), where 0 k (LN 1) and (p, q ) {(0, 1), (1, 0), (1, 1)}. Spq (k ) represents those states with no unserved packets in the rst k intervals. The combination of (p, q ) indicates whether it applies to Type 1 connection or Type 2 connection or both. States in Plane B are denoted by S (i, j ), where 0 i (LN 1) and (i + 1) j (L + M )N . S (i, j ) represents those states in which there are no unserved Type 1 packets in the rst i intervals and no unserved Type 2 packet in the rst j intervals, but there is one Type 1 packet in the (i + 1)th interval. States in Plane C are denoted by S (LN, m), where LN m (L + M )N 1. S (LN, m) represents those states in which there are no unserved Type 1 packets and no unserved Type 2 packets in the rst m intervals, but there is one Type 2 packet in the (m + 1)th interval. The state transition probabilities of the Markov model are also depicted in Fig. 3.3. For example, the transition probability from state S (2, 3) to state S 0 is
N (1 P1 )(1 P2 (L+M )N (3N +1) 2 2 ) = (1 P1 )(1 P2 ), since in S (2, 3) there is only one Type

2 arrival in the 4th interval, and this packet will be served by the EDF policy since no packet is ahead of it. After one unit time has passed, S 0 can be achieved if no Type 1 and Type 2 arrivals occur in the last two intervals since the only packet in the previous two intervals has been served. Refer to [51] for details of the state transition probabilities. The calculation of packet drop rate is discussed below. When a state transition happens, a gain of 1 is associated with it if this state transition leads to the delivery of a packet; on the contrary, a gain of 0 is associated with a state transition if no packet is delivered. Based on the theorem in [52], it is proved that the gain of the per state transition of the Markovian model 41

converges [51]. As a result, a value iteration algorithm is devised to compute the gain, and accordingly the network throughput and packet drop rate Pd . The packet drop rate Pd obtained above is with respect to the input of Phase 2, i.e., the arrivals with a mean rate of . As shown in Fig 3.2, the actual inputs of the entire system are the arrivals with means of 1 , 2 , , and n in Phase 1. Therefore, the overall dropping probability is Pd = Pd and the throughput is 1 Pd . 3.3.3 Analysis of Other Approaches For comparison, we have also analyzed WRR and EDF scheduling algorithms. The analytical model of WRR is similar to that of Phase 1 in the two-phase FEQ algorithm. More specically, we establish a separate queue for each connection. The total packet drop rate of WRR is calculated as Pd =
n i=1

n i=1

(3.7)

i Pid , where i and Pid are the arrival rate and dropping n i=1 i

probability of Queue i, respectively. The analytical model of EDF is similar to that of Phase 2 in Fig. 3.2, and thus it is omitted here. 3.4 Simulations and Discussion We have carried out extensive simulations to verify the correctness of the proposed analytical model for the FEQ algorithm, and to compare its performance with other scheduling algorithms, under a broad set of simulation parameters. In particular, we consider several comparable scheduling algorithms, including WRR, EDF , and DFPQ [43] (which is a representative WiMAX scheduling algorithm and has been patented and well received). Besides packet drop rate and throughput that have been studied in analysis, we are also

42

Table 3.1: Simulation Parameters for WiMAX WirelessHUMAN-SCa. Bandwidth BSCa Roll-off factor Symbol rate Rsymb PS duration TP S Frame duration Tf r Number of PS in one frame Length of one minislot Number of minislots per frame Number of exible minislots in the uplink subframe Outer code rate Router 10 M Hz 0.25 7.9296 M Baud 0.5044 s 10 ms 19824 8 PS 2478 540 239 255

interested in the fairness performance, which is measured by Jains Fairness Index [53] dened as follows:

f (x1 , x2 , ..., xn ) =

( n

n 2 i=1 xi ) , n 2 i=1 xi

(3.8)

where xi is the normalized throughput of connection i, and n is the total number of connections. Here we use the normalized throughput of a connection, i.e., xi = T hi , with M RRi

T hi and M RRi stand for the connection is actual data rate and reserved data rate, respectively. The Jains Fairness Index ranges between 0 and 1. The higher the index , the better the fairness. If T hi = M RRi for all i, or in other words, every connection obtains its reserved data rate, then xi = 1 for all i, and Jains Fairness Index equals 1. All simulations and analytic calculations are done using Java and Matlab programs. For accurate performance evaluation, we adopt the WiMAX physical layer standard WirelessHUMAN-SCa [34] in our simulations. As shown in Table 3.1, the bandwidth for WirelessHUMAN-SCa is BSCa = 10M Hz , and the roll-off factor is = 0.25. Thus the

43

Table 3.2: WirelessHUMAN-SCa Raw Data Rate. SS ID Modulation type Inner code rate Rinner Bits / symbol Bpm Raw data rate Rb (Mbps) Bytes / minislot SS1 BPSK 1/2 1 3.7160 1.875 SS2 QPSK 2/3 2 9.9094 4.999 SS3 16QAM 3/4 4 22.2962 11.247 SS4 64QAM 5/6 6 37.1603 18.745

Table 3.3: MRR and Delay Budget of Different Connections. rtVR MRR (kbps) Delay (ms) nrtVR MRR (kbps) Delay (ms) rtVR1 19.2 10 rtVR2 64 20 rtVR3 384 30 nrtVR3 256 170 rtVR4 1024 40 nrtVR4 768 200

nrtVR1 nrtVR2 24 48 100 130

Table 3.4: Distribution of the 8 Connections in 4 SS. SS ID rtVR1 + nrtVR1 rtVR2 + nrtVR2 rtVR3 + nrtVR3 rtVR4 + nrtVR4 SS1 SS2 SS3 SS4

44

BSCa 0.088 = 7.9296M Baud [34]. Since one physical lot (PS) 1+ 4 equals four modulation symbols, the duration of one PS, TP S , is = 0.5044s, and the Rsymb Rsymb Tf r number of total PSs in a frame of duration Tf r = 10ms is = 19824. Let one 4 symbol rate Rsymb = minislot equal 8 PS, we have 2478 minislots in one frame. We are concerned with the TDD model only, where a frame is shared for downlink and uplink transmission. We assume that in one uplink subframe, the number of exible minislots (excluding those already allocated for UGS connections, initial ranging contention, and bandwidth request contention) is 540, which is the total bandwidth available for rtVR and nrtVR connections. Since AMC is employed in the WiMAX network, different SS may choose different modulation and coding schemes and thus have different data rate. For example, we consider four SS in our simulation, which employs BPSK, QPSK, 16QAM, and 64QAM, respectively, as shown in Table 3.2. The raw data rate is calculated as Rb = Rsymb Bpm Router Rinner , where Bpm is the number of bits per modulation symbol, Router is the outer Reed-Solomon (RS) code rate, and Rinner is the inner Trellis Code Modulation (TCM) code rate. In WirelessHUMAN-SCa, Router has a xed value of 239 , and 255

Rinner can be chosen from a number of values according to the modulation type. In Table 3.2, Bpm and Rinner are listed under the specic modulation type of each SS. Based on that, we calculate Rb and consequently the number of bytes per minislot. Each SS establishes a number of connections to the BS in our simulation. We consider four rtVR connections, named rtVR1, rtVR2, rtVR3, and rtVR4, and four nrtVR connections, named nrtVR1, nrtVR2, nrtVR3, and nrtVR4. As shown in Table 3.3, each type of connection is associated with an MRR and a delay budget.

45

0.14 0.12 0.1 Drop Rate 0.08 0.06 0.04 0.02 0 FEQ Simulation FEQ Analysis EDF Simulation EDF Analysis WRR Simulation WRR Analysis DFPQ Simulation

1 0.99 0.98 Fairness 0.97 0.96 0.95 0.94 0.93 FEQ EDF WRR DFPQ 0.8 1 1.2 1.4 1.6

0.8

1.2

1.4

1.6

(a) Drop rate.

(b) Fairness.

Figure 3.4: Performance under Poisson Trafc. Every SS has four connections as shown in Table 3.4. For example, SS1 has these four connections: rtVR1, nrtVR1, rtVR2, and nrtVR2. As a result, there are 16 connections, which request a total MRR of 537 minislots. This is almost the maximum number of connections that can be admitted by CAC, given a total available bandwidth of 540 minislots in uplink subframe for rtVR and nrtVR trafc. We use the WirelessHUMAN-SCa physical layer technology in our simulations, and the scheduling algorithm also applies to the WirelessMAN-OFDMA physical layer technology. Notice that OFDMA symbols are the basic resource allocation units in WirelessMAN-OFDMA as minislots in WirelessHUMAN-SCa, and thus we can perform simulation based on WirelessMAN-OFDMA, if we replace the minislots in WirelessHUMAN-SCa with the OFDMA symbols in WirelessMAN-OFDMA. We consider two types of trafc in our simulations, i.e., the popular Poisson trafc and the bursty trafc generated by the Markov Modulated Poisson Process (MMPP) model, as discussed below.

46

3.4.1

Poisson Trafc First, we use Poisson trafc, with packet size of l = 800 bytes. The arrival rates of all

connections except rtVR3 are xed and match their reserved data rates. For example, the packet arrival rate of Connection i equals M RRi /l. Meanwhile, we vary the arrival rate of rtVR3 (which corresponds to three connections established by SS2, SS3, and SS4, respectively, as shown in Table 3.4), in order to study the inuence on the scheduling results. Based on these system parameters, we have obtained analytic and simulation results for FEQ, as well as the simple WRR and EDF schemes. For DFPQ approach in [43], the results are obtained via simulation only. The analysis and simulations are compared in Fig. 3.4(a), where X-axis indicates the trafc load of rtVR3 (where equals arrival rate over M RRi /l) and Y-axis shows the overall packet drop rate of all connections. As can been seen, the analytical results of FEQ, WRR, and EDF all match well with their simulation results, verifying the correctness of our analytical models. The simulation result of DFPQ algorithm is also depicted in Fig. 3.4(a). We observe that EDF, FEQ, and DFPQ have similar packet drop rate, while the drop rate of WRR is much higher, because WRR does not consider the delay budget and is more likely to drop realtime data packets. Note that, with the increase of the trafc load of rtVR3, the total trafc load becomes signicantly higher than the total available bandwidth. Therefore, all four algorithms exhibit high drop rate. Simulation results of the Jains Fairness Index are shown in Fig. 3.4(b). As we can see, the fairness index of EDF drops dramatically when the trafc load of rtVR3 becomes higher than its MRR, because the additional realtime trafc (with tight delay budget) aggressively

47

takes bandwidth from the nrtVR connections under the EDF algorithm, and thus leading to low throughput and unfairness to non-realtime trafc. When of connection rtVR3 is greater than 1, the fairness index of both FEQ and DFPQ drops, but DFPQs fairness index drops more rapidly, exibiting worse fairness performance than FEQ. The reason can be two-folded. First, DFPQ always gives realtime trafc higher priority than non-realtime trafc, resulting in unfairness. The other reason is that DFPQ does not guarantee the minimum reserved rate of each realtime connection since EDF is deployed within all the realtime connections. Among the four scheduling algorithms, WRR s fairness index is the stablest when the input trafc varies. Notice that since WRR tends to drop packets with urgent delay budget, the fairness of WRR is not exactly equal to 1. In summary, the fairness index of FEQ is better than that of DFPQ and EDF and comparable to WRR. We have also studied the running time of the proposed FEQ algorithm. As we have mentioned in Section 3.2.2 , the WiMAX scheduling algorithm must have a short running time. The reason is that the scheduling result of each frame needs to be broadcasted at the beginning of the frame, while the inputs for the scheduling algorithm, i.e., the uplink requests, may arrive at the BS in the last burst of the previous frame. This leads to the fact that the available algorithm running time is the length of the frame header composing of the BP (Burst Preamble) and the FCH (Frame Control Header). Under the WiMAX physical layer technology WirelessHUMAN-SCa [34], the length of BP and FCH can vary from 60 s to 200 s, and a typical value is around 100 s. The other WiMAX physical layer technology WirelessHUMAN-OFDMA [54] also has a typical value of 100 s for BP and FCH duration. Thus we choose 100 s as the BP and FCH length of each frame header, denoted by Th , to evaluate the algorithm performance in terms of running time. In the simulation parameters of 48

Table 3.4, there are 16 uplink connections. The FEQ algorithm running time under these 16 connections is 8 s, which is within the range of Th . If the number of connections goes beyond that, we want to study its impact on the algorithm running time. Notice that from Table 3.1, there are total 540 minilots available for exible uplink bandwidth allocation. If one connection occupies one minislot in each frame, there could be 540 connections in the uplink. Let n denote for the number of connections in the uplink channel, and n can go beyond 540 since there could be connections with data rates less than one minislot per frame. We increase n and record down the FEQ algorithm running time. Fig. 3.5 shows the result of the algorithm scalability, and FEQ algorithm can support up to 1800 connections with a Th of 100 s, thus FEQ algorithm is scalable.
120 100 Run Time (us) 80 60 40 20 0 0 500 1000 1500 Connection Number (n) 2000

Figure 3.5: FEQ Algorithm Running Time with the Number of Connections.

The above analytic and simulation results clearly show that our proposed FEQ effectively combines WRR and EDF, achieving the low packet drop rate and high throughput features of EDF while maintaining the fairness feature of WRR. FEQ outperforms DFPQ algorithm in fairness and has a similar performance as DFPQ in packet drop rate. In addition, the implementation of FEQ is easier than DFPQ because it avoids the complexity of WFQ, which is a key component of DFPQ.

49

1 1 2 2

Figure 3.6: MMPP Model for Bursty Trafc. 3.4.2 Bursty Trafc To study the performances of the four scheduling algorithms under bursty trafc, we use Markov Modulated Poisson Process (MMPP) model to generate data packets. Fig. 3.6 illustrates the MMPP trafc model. There are two states. Each state is a Poisson process with arrival rate of 1 (or 2 ), and has a transition rate denoted by 1 (or 2 ) to transit to the other state. Without loss of generality, we assume max = 2 > 1 . The average data arrival rate of 1 2 + 2 1 . In our simulation, we let 1 = 0.01 1 + 2 max and 2 = 0.1. The burstiness of the trafc is dened as b = . The higher the value of b, avg MMPP process is dened as avg = the more bursty the trafc is. When b equals to 1, the MMPP model is equivalent to the Poisson model. To focus on the impact of bursty trafc, we vary b, and x the arrival rate of all connections, i.e., let i = 1.5M RRi /l for rtVR3 connections and i = M RRi /l for other connection. The packet drop rates of the four scheduling algorithms under bursty trafc are shown in Fig. 3.7(a). As can be seen, WRR is rather vulnerable to the bursty trafc. With the increase of b, its drop rate increases dramatically. On the other hand, EDF, FEQ, and DFPQ can maintain a reasonably low packet drop rate even when b is large. The fairness of the scheduling algorithms under bursty trafc is shown in Fig. 3.7(b). As we can see, FEQ always maintains a high fairness index, while the fairness of EDF algorithm is the worst among the four algorithms. Notice that when b increases, the fairness index of

50

0.35 0.3 0.25 Drop Rate 0.2 0.15 0.1 0.05 0 Fairness FEQ EDF WRR DFPQ

1.0

FEQ EDF WRR DFPQ

0.975

0.95

0.925

b=1

b=3

b=5

0.9

b=1

b=3

b=5

(a) Drop rate.

(b) Fairness.

Figure 3.7: Performance under Bursty Trafc. EDF becomes higher. This is due to the fact that some realtime packets of rtVR3 connection are dropped under high burstiness, and thus the throughput of rtVR3 decreases to a value more closer to its reserved throughput. Accordingly, the normalized throughput of rtVR3 becomes close to 1, which appears more fair. On the contrary, the fairness of WRR drops with the increase of trafc burstiness. The reason is that the normalized throughput of rtVR3 under WRR is less than 1, and it keeps decreasing when trafc burst increases, leading to a higher degree of unfairness. In summary, under bursty trafc, the FEQ can still achieve a balance between WRR and EDF, with a low packet drop rate and a good fairness performance.
0.99 0.118 0.116 0.114 0.98

0.97

Drop Rate

0.112 0.11 0.108 0.106 0.104 0.102 0.1 0 0.2 0.4 0.6 0.8 1

Fairness

0.96

0.95

0.94

0.93

0.2

0.4

0.6

0.8

(a) Drop rate.

(b) Fairness.

Figure 3.8: Impact of Bandwidth Allocation between Two Phases ().

51

3.4.3

Further Discussion In our discussion so far of FEQ, the bandwidth allocation in Phase 1 is based on MRR.

More specically, each connection is allocated a bandwidth that equals its MRR, unless its total request is lower than its MRR. The remaining bandwidth is then used in Phase 2. Apparently, this is not the only option. In theory, we can make any bandwidth allocation to the two phases. To study the impact of bandwidth allocation between two phases, we introduce a parameter , with 0 1, and allocate M RRi to Connection i in Phase 1. In general, there is a trade-off between the drop rate and fairness when varies from 0 to 1. For example, a small indicates more bandwidth allocated to Phase 2, which results in a smaller drop rate but higher unfairness. However, our simulation results show that the algorithm achieves higher overall performance under larger . As can be seen in Fig. 3.8, when increases from 0 to 1, the drop rate increases only marginally by about 0.1%, while the Jains Fairness Index increases from around 0.93 to 0.98, which is close to the perfect fairness. This result justies our choice of letting = 1 in FEQ. 3.5 Summary In this chapter, we have presented a two-phase Fair and Efcient Queueing (FEQ) algorithm for uplink scheduling in the WiMAX network. It combines Weighted Round Robin (WRR) and Earliest Deadline First (EDF) algorithms, aiming to strike the balance between fair bandwidth allocation and delay requirement. In order to gain deep understanding of and insights into FEQ algorithm, we have established an elegant queueing model to derive in theory the performance metrics in terms of packet drop rate and throughput. The analytic

52

model has been veried by extensive simulations, carried out on the basis of a broad set of parameters according to WiMAX physical layer standards. In simulations, we have considered both popular Poisson trafc and practical bursty trafc that is modelled by the Markov Modulated Poisson Process (MMPP). The simulation results have veried the correctness of our analytical models. We have also compared FEQ with WRR, EDF, and Decit Fair Priority Queue (DFPQ). Both analytic and simulation results have clearly shown that FEQ algorithm effectively combines WRR and EDF and achieves low packet drop rate and high throughput while maintaining the fairness among different connections.

53

CHAPTER 4 MINIMUM-COST DATA DELIVERY IN WIRELESS NETWORKS

In this chapter, the minimum cost problem for data delivery in the heterogeneous wireless networks is discussed. 4.1 Introduction With various network characteristics and commercial concerns, a number of wireless technologies have been developed over past years and they are likely to coexist for many years to come. For example, the cellular systems [55], [56], and [57] have evolved from the rst generation analog systems to the second generation digital systems, and they are presently entering the era of the third generation (3G) systems that support not only voice but also data trafc at a speed of up to 2 Mbps, while the fourth generation (4G) systems are under development for achieving the data rate of ten times higher. On the other hand, a series of complementary IEEE standards, including 802.20 [58], 802.16e [35], 802.16 [34], 802.11 [9], and 802.15 [12], have been developed or are currently under development for effecting data communication in mobile and xed broadband wireless access networks, local and metropolitan area networks, and personal area networks, respectively. In particular, 802.20 and 802.16e target at mobile broadband wireless access networks, providing users moving at vehicular speed with a data rate from 1 Mbps to 30 Mbps in a wide area. 802.16 offers xed broadband wireless access with data rate up to 75 Mbps, which can be allotted to T1-level connections for business customers and/or to the best effort DSL-speed connections for home customers. 802.11 supports low-mobility users in small cells, at the data rates varying from 1

Figure 4.1: A Scenario of Heterogeneous Wireless Networks.

to 54 Mbps. Recently, this cost-effective technology is being deployed aggressively for establishing metro-scale cellular WiFi networks [59] to support seamless Internet access. In addition to aforementioned terrestrial communication systems, the satellite [60] is a vital component in the wireless systems, providing global coverage and high-speed data transmissions. While most of these wireless technologies are deployed independently for now, the service providers have an interest to own and operate overlaid heterogeneous wireless systems, which integrate multiple wireless technologies with partially overlapped coverage areas and provide ubiquitous network service to mobile users. For example, several mobile carriers such as Verizon, Sprint PCS, and T-Mobile are anxious to include WLAN access among their service offerings. In order to access various wireless networks/technologies, the mobile host (MH) may be equipped with one or multiple programmable wireless interface card(s) (e.g., based on the programmable radio technology [61], [62] or an approach similar to mobile access router [63]), resulting in two-fold exibility that may enable the optimization of data delivery: (1) an MH may select one of multiple available wireless access technologies at

55

a particular location, because one area may be covered by multiple wireless networks with different costs, data rates, and mobility support capabilities; and (2) an MH may use different access technologies when it travels in the networks and arrives at different locations covered by various wireless networks. From the standpoint of the service provider, it is an important issue to minimize the overall communication cost by intelligently using the available heterogeneous wireless technologies. In this work, we consider a typical scenario where an MH X is involved in massive data transmission while travelling (or staying, as a special case). For example, MH X may participate in a large Peer-to-Peer (P2P) network, where the members share resources such as movie les [64], [65]. Given the large data volume and the limited link capacity, a long data transmission time (e.g., up to hours) may be expected, during which MH X needs to serve as either a receiver or a data source continuously upon its travel or stay. MH X accesses the Internet through available heterogeneous wireless links in order to communicate with its peers. In addition, certain QoS requirements (e.g., delay) may be called for, depending on the type of applications. For example, the user may require downloading movie le and playing at the same time, and thus delay requirement is associated with such an application. The objective is to minimize the communication costs of the data transmission to/from the MH, while meeting the QoS requirements. Intuitively, the least expensive technology available should always be employed for data transmission in order to minimize the communication cost. This naive approach, however, does not guarantee QoS requirement, since the low-cost networks may not provide sufcient bandwidth to achieve the required QoS. At the same time, the low-cost networks are not available anytime anywhere either. For example, when MH X in Fig. 4.1 is covered by 56

high-cost cells only, continuing aggressive data transmission will noticeably increase the total communication cost. Thus, MH X may defer its data transmission or decrease its data rate until it enters the coverage of high-speed, low-cost cells, as long as the delay budget can be met. Although a simple greedy algorithm can be derived based on intuition, it is non-trivial to develop effective schemes for minimizing the communication cost while simultaneously meeting QoS constraints. In this chapter, we propose a linear programming (LP) [66] algorithm that takes into consideration such constraints as channel bandwidth, link costs, delay budget, and user routes. In case of insufcient bandwidth for communication with the core network, prefetch is employed to fully utilize the wireless network capacity. When multiple routes are available, a probability-based approach is taken for cost minimization. Extensive simulations are carried out to evaluate the performance of the proposed approach. Our results show that the LP algorithm can effectively reduce the overall cost of data transmission. The rest of this chapter is organized as follows. Section 4.2 discusses background and related work. Section 4.3 introduces the system architecture and signaling protocols. Section 4.4 denes the cost minimization problem and proves it to be NP-hard. Section 4.5 presents our proposed LP algorithms. Simulation results are illustrated in Section 4.6 . Section 4.7 summarizes the chapter. 4.2 Related Work A series of efforts have been made so far to enable and enhance the integration of heterogeneous wireless networks, focusing on the system framework that supports inter-operable signaling, seamless roaming, Internet accessing, security, authentication,

57

authorization, and billing. For example, led by telecommunication industry, some initial steps have been taken for integrating the emerging wireless LANs and the well-established cellular systems [67], [68]. On the other hand, there are several proposals from the Internet society [69], [70], [71], aiming to efciently implement Internet protocols in wireless networks for connecting mobile ad hoc networks to the Internet. In the meantime, new techniques, such as iCAR [72], MACA [73], PARCELS [74], and Multi-hop Cellular [75], [76], [77], have been proposed for employing ad hoc technology to improve the performance of cellular systems, or vice versa. Cost minimization has been long recognized as a critical design issue in heterogeneous wireless networks. In [27], Katz et al. consider an overlaid structure of room-size (e.g., infrared LAN), building-size (e.g., IEEE 802.11b WLAN), and wide area data networks (e.g., the 2.5G or the 3G cellular systems), where the lower the level of overlay, the smaller area it covers, and the higher data rate per unit coverage area it has. Vertical handoff is employed to support mobile host roaming from one network to another. The mobile host always switches to the lowest reachable overlay to achieve the highest data rate per unit coverage area. In order to adapt to the system dynamics (e.g., the varying trafc load), a policy-enabled handoff scheme is proposed in [28] to take multiple factors into considerations. Specically, a cost function is dened as the weighted sum of network bandwidth, the MHs power consumption, and the network access cost. The network with the least cost is chosen for communication. The pioneering work of Katz et al. has motivated our research on cost minimization in heterogeneous wireless networks. The approaches considered in [27] and [28], however, only take into consideration the local networks being accessed, which may not result in optimized overall performance (see the example discussed in Section 4.1 ). Thus, a longer term 58

Figure 4.2: Signaling Protocols for Cost Minimization Problem.

planning strategy for the entire data transmission period (rather than only at a given snapshot) is needed to achieve global optimization. In addition, the QoS requirement of the MH is not considered in [27], [28]. In this research, we establish a more general cost-minimization model based on not only system information (e.g., available bandwidth, access costs, network congestion) but also mobile users status (e.g., trafc type/amounts and moving routes), in order to minimize the overall communication cost while simultaneously meeting the QoS requirement. 4.3 System Architecture In this section, we give an overview of the system architecture considered in our research. Five components are involved in cost minimization: (a) the MH, (b) the communication peer, (c) the AP, (d) the network map server (NMS), and (e) the mobility agent, as shown in Fig. 4.2. The MH receives data from, or sends data to, its communication peer (e.g., a node in a P2P network or a data server in the Internet). An AP is a unit that provides wireless access to the MH. It could be a wireless LAN access point, a cellular base station, an IEEE 802.16

59

base station, an IEEE 802.20 base station, or a satellite. The NMS maintains a database of every AP in the system. The mobility agent helps the MH obtain its mobility information. Before initiating data transmission, the MH rst negotiates with its communication peer to acquire data trafc information (e.g., the total amount of data to be transmitted and its delay budget). If the amount of transmitted data is less than a predened threshold, no attempt for minimizing the communication cost is made, due to overhead involved. Otherwise, cost minimization is initiated. More specically, the MH collects needed information, including the MH mobility and the network topology, from the mobility agent and the NMS, respectively, and runs the cost minimization algorithm (to be discussed in Section 4.5 ). The network topology and user mobility are important inputs for running our proposed cost minimization algorithm. In the rest of this section, we briey discuss the signaling protocols to maintain that information, although their design and implementation are not the focus of our work. Signaling Protocol for Network Map Maintenance. When an AP is established, it sends to its associated NMS a network update message, which includes its location (e.g., GPS coordinates), estimated coverage area, transmission technology, communication costs, and data rates. The coverage area could be estimated as an ideal circle or in any arbitrary shape depending on the environment and the estimation model. Upon receiving a network update message, the NMS creates an entry in its database, with a timer associated. The AP periodically sends update messages to refresh the timers. If no new update message is received when the timer expires, the entry will be removed from the database. Note that once a cell has been established, no signicant changes of the APs location, coverage, data rates,

60

and/or communication costs are expected in the near future. Thus, the timer can be set to a large value and the AP can send update messages to NMS infrequently, resulting in ignorable overhead. Based on information maintained in the NMS database, a 2-dimensional network map can be established. Upon receiving a request from the MH, the NMS replies with a (partial) network map and information associated with the AP. Such information can also be buffered by the MH for future use, where the MH only needs to contact the NMS to check if the buffered network map is out of date. Signaling Protocol for User Mobility. The mobile users moving pattern (i.e., the traveling route) could be predicted precisely, roughly, or may be unpredictable at all. Three typical types of user mobility are discussed as follows. 1. Predetermined Routes and Speeds. The mobile user may follow a predetermined route and move at predetermined speeds. A typical example is the railway system (which is a main carrier in many metropolitan areas), transporting multitudinous people everyday. For instance, 500,000 passengers take trains daily in the New York Pennsylvania Station; tens of millions of people travel by trains daily in other major cities around the world; and moreover, all of these passengers are current or potential mobile network users. The train follows predetermined routes and precise schedules, and may travel through many cells (in systems such as cellular, IEEE 802.20, or satellite), and stop at a number of stations, where for instance, wireless LANs are available. 2. Predetermined Routes but Variable Speeds. For example, buses follow regular routes, but depending on the trafc condition, they may move faster or slower than the predetermined schedule. Another example is self-guided touring. A tourist may carry

61

an electronic cicerone provided by the tourist agent and follow the recommended route to visit the scenic landmarks. The MH moving schedule can be estimated according to the average pacer speed. The actual moving speeds of the tourists, however, may vary, resulting in deviation from the estimated schedules. 3. Uncertain Routes. The moving route of a mobile user is sometimes undetermined. For example, there may be two (or more) alternative routes available from home to ofce, and the mobile user may choose either one of them with a certain probability. A self-learning model (to be discussed next) may be adopted to learn the possible routes of the mobile users and the probability that a route is taken. The MH needs to acquire route information as the input of the cost minimization algorithm. Two approaches for mobility management are considered in this research. 1. System Mobility Management. where mobility information is maintained by the system and provided to the MH via mobility agents that are installed on the carriers (e.g., trains, buses, or cars) to broadcast mobility information periodically. 2. Autonomous Mobility Management where mobility information is maintained by every MH via a self-learning model. More specically, the moving routes are mapped to a graph, where the edges represent routes traversed by the MH and the vertices represent the intersection points. A visiting counter and a stale timer are associated with each edge. The visiting counter is increased by one when the MH takes the corresponding route, and decreased by one whenever the stale timer expires. At any vertex, the probability of taking a particular edge for next movement is calculated by dividing the visiting counter of this edge by the sum of the counters of all edges connecting to this

62

INPUT: 1. N : number of cells; 2. K : number of data blocks to be transmitted; 3. : number of wireless interface cards per MH; 4. M : total amount of data (in terms of bits) to be transmitted; 5. Mk (1 k K ): size (in terms of bits) of data block k ; 6. dk (1 k K ): time delay bound of data block k ;
c 7. rij : data rate between the core network and Aij ; m 8. rij : data rate between the MH and Aij ;

9. pc ij : communication cost rate between the core network and Aij ; 10. pm ij : communication cost rate between the MH and Aij ; 11. i : dwelling time during which the MH travels in cell i; OUTPUT: 1. tij : communication time between the MH and Aij ; 2. Lij : amount of data downloaded between the MH and Aij ; 3. fij : amount of prefetched data at Aij ; 4. gij : cost associated with fij ; 5. C : minimized total communication cost.

Figure 4.3: Inputs and Outputs for the Cost Minimization Problem.

vertex. As a result, a route is associated with a higher probability if it was traversed more frequently in the recent past.

63

4.4

Problem Formulation and NP-hard Proof In this section, we formulate the cost minimization (CM) problem and prove its

NP-hardness. 4.4.1 Problem Formulation To facilitate our discussion, we consider an abstract model of the heterogeneous wireless system. Specically, we assume the system is divided into N cells, where each cell (e.g., cell i, 1 i N ) is served by and only by a set of APs, denoted by {Aij |1 j |Ai |}, where |Ai | is the number of APs serving cell i. Clearly, the cells discussed here are different from the nature cells formed by the coverage of APs. The former can be obtained by dividing the whole area according to the boundaries of the natural cells. An AP, Aij , connects to the core network through a dedicated communication link that may be either wired or wireless, with a
c cost rate of p c ij and an average data rate of r ij bps. At the same time, Aij provides wireless

access for the MH within its coverage, with a cost rate of p m ij and an average data rate of
m r ij bps. The cost rate is the weighted sum of several factors such as service charge, signaling

overhead and power consumption, measured by the cost per unit amount of data. We consider an MH with programmable wireless network interface cards, each of which can tune to different wireless technology. Note that one such wireless interface card can be used to access one AP at a given time only, although it can switch from one to another at different times. Assume the MH sends or receives a sequence of K data blocks with M bits of data in total, while travelling in the coverage area of the heterogeneous wireless system.
K

Let M k denote the size (in terms of bits) of data block k (1 k K ). Clearly, M =
k=1

Mk .

A data block k is assumed to have a delay bound d k , indicating the time by which data block

64

k should be received. We denote i as the dwelling time of the MH in each cell i. i is obtained from the calculation based on the network map and the user mobility. When an MH is in cell i, it may or may not communicate with Aij during the entire period of i , in order to minimize the communication cost. We denote tij to be the time that the MH connects to Aij . Clearly, tij i . The total amount of data downloaded during tij is denoted by Lij . We also introduce a prefetch methodology where a certain amount of data may be prefetched by APs in order to enhance the cost minimization process. We denote fij as the amount of prefetched data at Aij with a corresponding cost gij . The cost of data transmission/reception through Aij is denoted
c m m as Cij and calculated based on tij , rij , pc ij , rij , and pij . The goal of the cost minimization

algorithm is to determine the values of tij , Lij , fij , and gij , such that the total cost
N |Ai |

C=
i=1 j =1

Cij is minimum, with all delay constraints satised. The inputs and outputs of

the cost minimization algorithms are summarized in Fig. 4.3. 4.4.2 NP-hard Proof In this section, we prove that the cost minimization problem is NP-hard by providing a polynomial reduction from a known NP-hard problem, the Continuous Multiple Choice Knapsack (CMCK) problem, to the cost minimization problem. 4.4.2.1 CMCK problem

The CMCK problem was rst formulated by T. Ibaraki in 1978 and proven to be NP-hard [78]. The CMCK problem considers N groups, where each group i has Ki items. The item j in group i has a value of vij and a size of sij . The CMCK problem is to select a fraction, xij (0 xij 1), of at most one item from each of the N groups, in order to satisfy both the

65

value and the size constraints. More specically, the sum of the size is less than a size threshold S , and the sum of the value is greater than a value threshold V . The CMCK problem is dened as follows. Given V > 0 and S > 0, do a set of 0 xij 1 exist, such that
N Ki

vij xij > V,

i=1 j =1

(4.1)

and
N Ki

sij xij < S,

i=1 j =1

(4.2)

hold, where, at most, one {xij |1 j Ki } is none-zero, for 1 i N ? 4.4.2.2 Polynomial Reduction to Cost Minimization Problem

Now we consider a simplied scenario of the cost minimization problem, where only one delay bound needs to be satised for completing the transmission of all data (M ). The simplied cost minimization problem with an objective minimum cost value (denoted by C ) can be dened as follows: Given M > 0 and C > 0, do a set of 0 xij = tij /i 1 exist, such that
N |Ai | m i xij > M, rij i=1 j =1

(4.3)

and
N |Ai | m m (pc ij + pij )rij i xij < C, i=1 j =1 |Ai |

(4.4)

hold, where
j =1

xij 1, for 1 i N ?

m m , i , pc In the above cost minimization problem, the denition of rij ij , and pij can be found m bps. tij is represented as in Fig. 4.3. We assume that data is downloaded to the MH at rij

66

i xij , where xij is between 0 and 1. To facilitate the discussion, we reiterate equation 4.3 and
m m m 4.4 by letting V = M , S = C , xij = xij , Ki = |Ai |, vij = rij i , sij = (pc ij + pij )rij i , then

we can reformulate the minimum cost problem as follows. Given V > 0 and S > 0, do a set of 0 xij 1 exist, such that
N Ki

vij xij > V ,

i=1 j =1

(4.5)

and
N Ki

sij xij < S ,

i=1 j =1 Ki

(4.6)

hold, where
j =1

xij 1, for 1 i N ?

Clearly, the parameters of Equation 4.5 and Equation 4.6 can be mapped to those of Equation 4.1 and Equation 4.2, and we can see that the difference between the CMCK problem and the minimum cost problem is the chosen of xij or xij . In CMCK, at most one item in group i can be chosen, i.e., xij = 0 is true for at most one j in 1 j Ki ; in minimal cost problem, multiple items can be chosen in each group i, i.e., there can be more than one
Ki

xij = 0 for 1 j Ki . One addition constraint of the CMCK problem is that

j =1

xij 1.

Recall that a polynomial reduction from problem A to problem B requires mapping the YES instance of A to the YES instance of B , i.e., YES YES, and mapping the YES instance of B to the YES instance of A, i.e., YES YES. We rst prove YES YES : given a set of xij that satises inequations 4.1 and 4.2, can we construct an instance of the minimum cost problem that satises inequations 4.5 and 4.6? This can be easily achieved by matching the variables in inequations 4.1 and 4.2 to inequations 4.5 and 4.6, respectively. That is to say, we let V = V , S = S , xij = xij ,

67

Ki

Ki = Ki , vij = vij , sij = sij , then we can verify that

j =1

xij 1 holds since at most one

xij = 0 for 1 j Ki , and all xij is between 0 and 1. We then prove YES YES : given a set of xij that satises inequations 4.5 and 4.6, can we construct an instance of the CMCK problem that satises inequations 4.1 and 4.2? Starting from inequation 4.5, we rst nd the largest value of vij xij in each group i, and we denote this index j as pi . Then we can rewrite the left hand side of equation 4.5 as
N N Ki

LHSeq5 =
i=1

vipi xipi +
i=1 j =1,j =pi

vij xij . Notice that vipi xipi is the largest value of vij xij in

group i for 1 j Ki , we have:

N N Ki N N

LHSeq5 =
i=1 N

vipi xipi +
i=1 j =1,j =pi N

vij xij
i=1

vipi xipi +
i=1

(Ki 1)vipi xipi =

vipi xipi Ki Kimax

i=1 N i=1

vipi xipi , where Kimax = maxN i=1 Ki . Thus we have

N

Kimax
i=1

vipi xipi LHSeq5 > V

i=1 N

vipi xipi >

V = V > 0. Kimax
N N Ki

Also it is easy to verify that

i=1

sipi xipi < S , since

i=1

sipi xipi
i=1 j =1

sij xij < S .

Now we summarize the YES YES proof as follows: given a set of xij that satises inequations 4.5 and 4.6, we rst nd out the pi for each group i in polynomial time, such that vipi xipi is the largest among vij xij , for 1 j Ki . Then we construct the following inequations in polynomial time which match inequations 4.1 and 4.2 of CMCK problem: 1)
N

vipi xipi
i=1

V , Kimax = maxN > V > 0, where V = i=1 Ki ; and 2) Kimax

sipi xipi < S . We

i=1

have proved that the above two inequations hold. Thus we nish the YES YES proof. 4.5 Proposed LP Algorithms for Cost Minimization In this section, we examine several cases of the cost minimization problem with increasing levels of complexity and discuss the proposed LP algorithms for each of them.

68

OBJECTIVE: minimize C ; VARIABLES: tij , 1 i N and 1 j |Ai | ; CONSTRAINTS:

N |Ai |

1: C = 2: 3:

Cij ; Lij ;

i=1 j =1 c Cij = (pm ij + pij ) m Lij = rij tij ; N |Ai |

4: M =
i=1 j =1 |Au |

Lij ;
k

5:
u(dk ) j =1

Luj
v =1

Mv , for 1 k K ;

6: tij i ;
|Ai |

7:
j =1

tij i min(, |Ai |);

(x) is a function that returns the set of all cells that the MH has visited by time x.

Figure 4.4: LP Model for Predetermined Route.

4.5.1

Single Route We rst consider the scenario where an MH traverses a predetermined route and follows

predetermined speeds. Note that the MHs moving speed is not necessarily constant. As long as the speeds (which may vary with time) are predetermined, one can precisely calculate the dwelling time of the MH in each cell i, i.e., i . To facilitate our discussion, we perform a preprocessing on the network map. More specically, we divide the N cells into N small cells so that all delay bounds are due at the boundary of the small cells.

69

4.5.1.1

Base Case

We start with a base case, where the following assumption is made: the wireless link
m c rij ( 1 j | A i |, between AP and the MH is the performance bottleneck, i.e., rij m 1 i N ). Thus APij can always communicate with the MH at rij bps.

We have developed an LP model as shown in Fig. 4.4 for the base case. The primary objective of this LP model is to minimize the total communication cost, C , by choosing a proper value for each tij , such that all constraints 1 through 7 (given in Fig. 4.4) are satised. The rst three constraints dene the communication cost, which is calculated based on the unit communication cost between Aij and the core network (i.e., pc ij ), the unit communication
m cost between Aij and the MH (i.e., pm ij ), the data rate rij , and the communication time tij . The

fourth constraint ensures the total transmitted data to be M . The delay budget is specied in the fth constraint, which ensures the transmission of the k th data block to be nished no later than dk . The sixth constraint limits tij , the data transmission time between the MH and Aij , to be no longer than the dwelling time of the MH in cell i (i.e., i ). An AP is available for an MH, if and only if the MH is within the coverage of the AP and the MH has the suitable wireless interface to access the AP. Note that, even when multiple APs are available, hardware may limit the number of simultaneous wireless connections that can be established by an MH. More specically, with programmable wireless interface cards, the MH may communicate with at most min(, |Ai |) number of APs in cell i. With the seventh constraint, the total data transmission time between the MH and the access points in cell i should not exceed i multiplied by min(, |Ai |). Finally, (x) is a function that returns the set of all cells that the MH has visited by time x. The formulated LP model in Fig. 4.4 is

70

solved by using LP-SOLVE [79], yielding optimal values of {tij | 1 j |Ai | and 1 i N } that minimize the total communication cost. In FDD system where the uplink channel and downlink channel are separated, the optimization of uplink cost is just a dual problem of the downlink case. In TDD system where uplink and downlink share the same channel, if we try to optimize both uplink and downlink cost, then we need to let the uplink and downlink channel share the MHs cell dwelling time in each AP, which can be easily implemented in the LP model by adding the term for uplink dwelling time in constraint 7 of Fig. 4.4. Other modications of constrains in Fig. 4.4 are also straightforward, such as adding terms for data requirement and delay requirement of uplink, and the associated cost. Since the LP model for considering both uplink and downlink is similar to that of downlink only, we do not bother to list the solution for considering both uplink and downlink. Notice that, in TDD system, the individual optimization goal of uplink and downlink may conict with each other, and the yielding optimized solution of considering both uplink and downlink is a compromise of the two individual goals. 4.5.1.2 Advanced Case

m c In the base case, we have assumed rij rij . Now we nullify this assumption in order to m c establish a more realistic model. If rij > rij , the data rate limitation between Aij and the core

network may lead to inefciency, because Aij cannot draw down data blocks from the core
m , and thus limits the data transmission to the MH. Two network at a rate as high as rij

approaches have been considered to address this problem, as discussed below.

c m , Aij may lower the data rate at its wireless interface by > rij Reduced data rate: rij m c setting rij = rij . This approach is simple, but it sacrices the channel efciency and

71

OBJECTIVE: minimize C ; VARIABLES: tij and fij , 1 i N and 1 j |Ai | ; CONSTRAINTS:

N |Ai |

1: C =
i=1 j =1 Cij = (pm ij +

Cij ;

2: pc ij ) Lij ; m 3.1: for non-prefetch Aij : Lij = rij tij ; 3.2: for prefetch Aij : m c 3.2.1: Lij = rij tij 1 + rij tij 2 ; m c 3.2.2: fij = (rij rij ) tij 1 ;
i1

3.2.3: fij

c rij

l=1

l ;

3.2.4: tij = tij 1 + tij 2 ;

N |Ai |

4: M =
i=1 j =1 |Au |

Lij ;
k

5:
u(dk ) j =1

Luj
v =1

Mv , for 1 k K ;

6: tij i ;
|Ai |

7:
j =1

tij i min(, |Ai |);

(x) is a function that returns the set of all cells that the MH has visited by time x;

Figure 4.5: LP Model for Predetermined Route with Prefetching.

results in reduced throughput. Moreover, it may also increase the communication cost,
c because the data rate in the low-cost cell is limited by rij and thus the low-cost resource

cannot be utilized thoroughly.

c m , Aij may draw down a certain number of data blocks beforehand > rij Prefetching: rij

and store them in buffer so that Aij can transmit data to the MH at a high data rate when

72

the MH enters cell i. To enable efcient prefetching, we divide tij into two parts, tij 1 and tij 2 , with
m tij = tij 1 + tij 2 . tij 1 represents the time period when Aij transmits at data rate rij , with the c help of the prefetched data. tij 2 represents the time period when Aij transmits at data rate rij .

The optimal values of tij 1 and tij 2 will be determined by solving the LP model in Fig. 4.5.
m c ) tij 1 . If tij 1 = 0, no data is prefetched. rij The amount of prefetched data fij equals (rij

Prefetching should be completed before the MH enters the coverage area of Aij , as specied by Constraint 3.2 in Fig. 4.5. Given the assumption that the route and the speed of the MH are predetermined, the exact time to start prefetching is not a crucial issue, as long as prefetching can be nished before data transmission begins. 4.5.2 Multiple Possible Routes

We now consider the case where the mobile users moving route is undetermined when the data transmission is initiated. Assume that the mobile user may take one of W possible routes, with a probability of Pw associated with Route w (1 w W ) according to the mobility management schemes discussed in Section 4.3 . The multiple possible routes may have common or uncommon paths. Common paths refer to those shared routes, whereas uncommon paths are those completely disjoint from each other. There are two issues that need to be handled properly in multiple possible routes scenario. First, the delay bounds should be satised in the worst case. In other words, the data blocks are to be transmitted within their delay budgets even the route with the lowest capacity is eventually taken by the mobile user. Second, prefetch has two impacts on the overall communication cost. On one hand, prefetch may reduce the communication cost by enabling

73

Given W possible routes,

W

OBJECTIVE: minimize C =
w=1

Cw ;

VARIABLES: tij and fij , 1 i N 1 j |Ai |; For each route w, where 1 w W , LOOP: Compute Cw , with the following CONSTRAINTS:
N |Ai |

1: Cw =

[Cij Pw + gij (1 Pw )];

i=1 j =1 c [(pm ij + pij )

2: Cij = Lij ; 3.1: for non-prefetch Aij : m 3.1.1: Lij = rij tij ; 3.1.2: gij = 0, since there is no prefetch data; 3.2: for prefetch Aij : m c 3.2.1: Lij = rij tij 1 + rij tij 2 ; m c 3.2.2: fij = (rij rij ) tij 1 ;
i1 c 3.2.3: fij rij l=1

lw ;

3.2.4: tij = tij 1 + tij 2 ; 3.2.5: gij = pc ij fij ;

N |Ai |

4: M =
i=1 j =1 |Au |

Lij ;
k

5:
u (dk ) j =1

Luj
v =1

Mv , for 1 k K ;

6: tij iw ;
|Ai |

7:
j =1

tij iw min(, |Ai |);

END LOOP; (x) is a function that returns the set of all cells that the MH has visited by time x;

Figure 4.6: LP Model for Multiple Possible Routes.

74

the thorough use of low-cost resources. But on the other hand, the prefetched data in the untaken routes become wasteful, and thus may increase the communication cost. It is non-trivial to decide whether to employ prefetch or not and how to use it. If prefetching is disabled for simplicity, handling multiple possible routes is not much different from the single route case, as one can always optimize each route separately without considering other routes. In the presence of prefetching, however, it is inefcient to take the simple approach that considers each route separately by following the LP model presented in Fig. 4.5, since the cost of wasted prefetched data is not taken into account. Our proposed approach is summarized in Fig. 4.6, where all routes and their probabilities are considered in order to minimize the overall communication cost. We denote iw as the dwelling time of the MH in cell i if Route w is taken. The total communication cost is the weighted sum of the costs of all possible routes. The communicating cost of Route w (denoted by Cw ) is the sum of its communication cost if it is taken with a probability of Pw and the cost of wasted prefetched data if the route is not taken with a probability of (1 Pw ), as shown in Fig. 4.6. The cost of the wasted prefetched data is pc ij fij , since only the communication link to the core network is used for prefetching data. Based on the above cost calculation, a penalty is added to those routes with high prefetching costs. As a result, the route with lower probability will prefetch less data to contain the cost. For the common path of multiple routes, there is only one set of variables to represent its network parameters and prefetch data, and the yielding result of this common path will reect the compromises of multiple routes. The LP model aims to minimize the overall cost, based on the probability of each route when there are multiple possible routes, and it does not intend to decide for the MH which route to take. Once the MH 75

chooses any one of the possible routes, the LP model can be rerun for further optimization of the cost. However, there is no technical hurdle to let AP decide for an MH which route to take. For example, AP can suggest to an MH the route with the lowest cost. In the above discussion, all possible routes have been taken into consideration. To reduce computing complexity, one may consider the routes with high probabilities only. Specically, the possible routes are sorted via a decreasing order of their probabilities. The routes are chosen for consideration from the top until
w

Pw , where is a predened constant. The

LP model yields results based on currently available route information. While the MH moves, the number of possible routes may change, or one predetermined route may not be taken. The LP model is then run again based on updated route information. The probability-based approach for multiple routes scenario does not aim at dealing with very large (non-polynomial) number of routes. First, a non-polynomial number of routes is not reasonable in real scenarios. The number of routes for an MH is usually limited. Second, the probability of each route will drop with the increase of the number of possible routes, and the prefetched data in each route will decrease to near zero if the probability is low. We assume that the MHs dwelling time in each cell is known, and this information in needed as the input of the LP model. If the MHs actual dwelling time in any cell is different from the predicted value, a re-run of the LP model is necessary. 4.5.3 Variable Speed

In above discussion, we have assumed that the MHs moving speeds are predetermined. This assumption does not always hold in practical applications. In this subsection, we study possible speed variations and their impacts on cost minimization.

76

Note that not all speed variation affects cost minimization. For example, the MH may vary its speed in a cell i. As long as the average speed equals the expected value, i does not change. As a result, the same cost minimization schemes discussed in Section 4.5.1 and Section 4.5.2 can be employed to yield the same results as the case with a predetermined speed. If i changes, however, the cost minimization scheme will be different. The above observation motivates us to develop an approach for the scenario with variable moving speeds of the MH, centering at the variation of i . More specically, the MH maintains an expected moving speed (V ). Based on the moving speed, the moving route (either single route or multiple possible routes) and the network map, { i | 1 i N } can be calculated. Upon entering a cell (e.g., cell i), the MH keeps tracking its actual average speed in this cell (denoted by V ). The MH compares V with V at either of the following two occasions, whichever arrives earlier: (1) when the MH has spent i in cell i or (2) when the MH has nished traversing cell i. If V = V , i changes. As a result, the MH updates V with V and performs cost minimization based on the updated i (1 i N ) and the remaining data to be delivered. Note that since the MH already has the moving route and the network map, no extra signaling overhead is incurred. Although additional delay is expected for running the LP algorithm, such a computation delay is short as to be discussed in Section 4.6 . 4.6 Simulations and Discussion We have carried out extensive simulations to evaluate the effectiveness and efciency of our proposed minimum-cost data delivery algorithm. Three typical wireless technologies with different coverage areas, data transmission rates and costs are considered in our simulation. Specically, we simulate a number of IEEE 802.11b WiFi, 3G Cellular, and IEEE 802.16

77

Table 4.1: Network Parameters for Simulation Setup. Wireless networks Cell radius m ) Wireless access data rate (rij c Core network data rate (rij ) Core network data cost (pc ij ) m Total data cost (pij + pc ij ) 802.11 WiFi 200m 5Mbps 3Mbps 0.4/MBytes 1/MBytes 3G Cellular 1000m 1Mbps 3Mbps 0.7/MBytes 1.5/MBytes 802.16 WiMAX 3000m 6Mbps 9Mbps 1.9/MBytes 4/MBytes

Table 4.2: Impact of System Overhead on Communication Cost. M(MB) cost w/o overhead cost with overhead cost increased 190 285 288 1.1% 210 230 374 464 381 474 1.9% 2.2% 250 270 554 644 567 659 2.3% 2.3% 290 819 837 2.2%

1 S R2 A1 5 R1

4 3G A2

WiFi

WiMax 9 3km A3 13 14 15 A4 16 D 10 11 12

4707m

Figure 4.7: Heterogeneous Networks with WiMAX, 3G, and WiFi cells.

WiMAX cells. Their network parameters, including cell radii, data rates, and costs, are summarized in Table 4.1. The WiMAX network aims at providing wide area broadband access, with the longest cell radius and the highest wireless data rate among the three, as indicated in Table 4.1. We set the data rate of a WiMAX user to be 6 Mbps, although IEEE 802.16 can reach a peak rate of 78

1400 Greedy algorithm Prefetch off, one interface card Prefetch on, one interface card Prefetch off, two interface cards

1200

1000

Cost

800

600

400

200 180

200

220

240 M (MB)

260

280

300

320

Figure 4.8: Performance Averaged over Random Routes.

1400 Greedy algorithm Prefetch off, one interface card Prefetch on, one interface card Prefetch off, two interface cards

1200

1000

800 Cost 600 400 200 0 180

200

220

240 M (MB)

260

280

300

320

Figure 4.9: Performance Averaged over Random Maps.

240 One delay bound Three delay bounds 220 delay bound 3

200 Downloaded Data (MB)

180

delay bound 2 160

140

120

delay bound 1

100 150

200

250

300 350 Time (seconds)

400

450

500

Figure 4.10: Impact of Delay Bounds on Downloaded Data.

Route 1 Route 2 and Route 3 Average of all routes 700

650

Cost 600 550 500 0.1

0.2

0.3

0.4 0.5 Probability of Route 1

0.6

0.7

0.8

Figure 4.11: Impact of Route Probability on Communication Cost.

79

25 Route 1 Route 2 and Route 3

20

Cost of Prefetched Data

15

10

0 0.1

0.2

0.3

0.4 0.5 Probability of Route 1

0.6

0.7

0.8

Figure 4.12: Impact of Route Probability on Prefetched Data.

10% increase 20% increase 30% increase 40% increase 50% increase

580

560

540 Cost 520 500 480 0

6 8 Number of rerun

10

12

14

Figure 4.13: Impact of Increasing Cell Dwelling Time.

850 10% decrease 20% decrease 30% decrease 40% decrease 50% decrease

800

750

Cost

700

650

600

550

6 8 Number of rerun

10

12

14

Figure 4.14: Impact of Decreasing Cell Dwelling Time.

620 cost if cell dwelling time is known in advance cost under LP model rerun cost under LP model rerun plus system overhead 610

600

590 Cost 580 570 560 550 0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Figure 4.15: Cost Comparison when Cell Dwelling Time Varies Randomly.

80

75 Mbps. WiFi cells locate in the hot-spot areas with the smallest cell size. The average user data rate of WiFi is assumed to be 5Mbps, given its maximum raw data rate of 11 Mbps. The peak wireless data rate of the 3G network is 2 Mbps. We assume its average data rate to be 1
c m Mbps for mobile users. Also we notice that rij is usually higher than rij in the WiMAX and

the 3G systems, and as a result, prefetching is not necessary. The communication cost has no unit because it is a weighted sum of service charge and power consumption. It is a separate issue (outside the scope of our work) to establish efcient models for determining cost rates. The network topology and parameters in our simulations are shown in Fig. 4.7. It is a square area with diagonal length of 4 km. We divide this square area into 16 small square blocks, and each block has a diagonal length of 1 km. We number these 16 blocks from 1 to 16, as indicated in the upper left corner of each block. A WiMAX AP is placed at the center of the square, and covers the entire square area. Four 3G APs are placed at locations A1 , A2 , A3 , and A4 , respectively. Since the radius of a 3G AP is 1 km, each 3G AP covers its surrounding four blocks. For example, the 3G AP at location A1 covers blocks 1, 2, 5, and 6. There are 16 WiFi APs in total. Each of them is randomly and uniformly distributed inside a block. In Fig. 4.7, only three WiFi APs are drawn to avoid too many items included in the gure. Handoff procedure happens between two APs of different types, or of the same type. Without loss of generality, we set the start point of all the routes at the center of block 1, and the destination at the center of block 16, denoted by S and D in Fig. 4.7, respectively. We let the MH travel along the straight line connecting the centers of two adjacent blocks. When arriving at the center of a block, the MH can choose any of its neighboring blocks to be the next block to traverse, except the block from which this MH enters the current block. Two random routes are shown as examples in Fig. 4.7, marked as R1 in dash dot line and R2 in 81

solid line. The speed of the MH can be constant (at 10 m/s) or variable. The MHs dwelling time in each cell (i.e., i , 1 i N , where N is the number of cells in the route) can be calculated given the route and the speed information. The amount of data (M ) to be delivered and its associated delay bounds may vary in different simulation scenarios. We dene two thresholds Ml and Mh . Ml is the total amount of data that can be downloaded if the least expensive AP in each cell is always selected (i.e., by following a greedy algorithm that selects the AP with the lowest cost). Mh is the total amount of data that can be downloaded if the AP with the highest data rate in each cell is always chosen (i.e., by following a greedy algorithm for minimizing delay). Obviously, Ml Mh . If Ml = Mh , it implies that the least expensive AP is the AP with the highest data rata in every cell, and thus the optimization problem becomes trivial. If the total amount of data, M , falls below Ml , then a greedy algorithm choosing the least expensive AP in each cell yields the optimal solution. For M between Ml and Mh , the greedy algorithm selects the access technology with the highest data rate in each cell. If M is greater than Mh , then there is no feasible solution, because even data are downloaded at the highest data rate in each cell, the delay budget still cannot be met. Thus, for a non-trivial cost minimization problem, we let Ml < M < Mh . We study the performance of the proposed LP model under four different scenarios: random single route, random map, random multiple routes, and random distribution of cell dwelling time. For each scenario, we generate 20 different samples, and average their results. In these dynamic scenarios, the routes, the maps, or the MHs cell dwelling time change randomly. A random single route is created by randomly generating a path from S to D in Fig. 4.7. A random map is created by randomly distributing the location of each WiFi AP inside its block. A random multiple routes scenario is created by combining multiple single 82

routes together and then feeding them to the multiple routes LP model in Fig. 4.6. We have several patterns for the random distribution of the MHs cell dwelling time. One is to let the dwelling time conform to a certain distribution, e.g., a uniform distribution. Other patterns include increasing or decreasing the cell dwelling time by a certain amount. Due to the change in cell dwelling time, the MH will rerun the LP model for a new download scheme when it is necessary. Next we discuss the simulations in detail. In the random single route scenario, we keep the same map, but generate different route randomly for each scenario. Only one delay bound is specied at the end of route for delivering a total of M megabytes data, where M varies from 190 MB to 312 MB. Fig. 4.8 compares the performances of the LP model and the greedy algorithm. As can be seen, the LP Model without prefetch achieves a cost reduction of up to 32% (or 20% in average) compared to the greedy algorithm. Prefetch can also help to further reduce the communication cost in
m c any cell i with rij > rij . By employing prefetch, Aij may draw down a certain number of data

blocks beforehand and store them in buffer so that Aij can transmit to the MH at a high data rate when the MH enters cell i. In our simulation, prefetch is employed at the WiFi APs only. As can be seen from Fig. 4.8, prefetch reduces the overall communication cost by 27% in average, compared to the non-prefetching case. Though results are not shown here, the benet
c m . Prefetch incurs an extra and rij of prefetch increases with the increase of the gap between rij

buffering cost at the access point, which is assumed negligible. We also consider the scenario where the MH is equipped with more than one programmable wireless interface card. Fig. 4.8 shows the result of employing two interface cards, which lead to a drop of 50% in average in the communication cost, compared to the one interface card scenario. In the random maps simulation, we x one route and generate random maps. The result shown in Fig. 4.9 indicates 83

that the LP model achieves similar cost reduction as we discussed in Fig. 4.8. In the above discussion we have ignored the possible overhead that may result in additional communication cost. The proposed cost minimization approach introduces three types of overhead, i.e., signaling overhead, computing overhead, and handoff overhead. They are elaborated below. First, the signaling overhead is there since the MH needs to obtain such information as network map and MH mobility before initiating cost minimization. The impact of the signaling overhead on the communication cost includes the extra transmission time and the transmission cost associated with this signaling data. With proper design of data structure, the total amount of signaling data is far less than 1 KB, which takes less than 3 ms of transmission time given the core network data rate in Table 4.1. The actual transmission cost of the signaling data can be calculated according to the core network data cost in Table 4.1. The second type of overhead is introduced by the computing time needed to solve the LP model. Our simulation shows that the computation delay is usually between 10 ms to 20 ms on a Celeron 2 GHz CPU, by using the LP-SOLVE [79] tool. The third type of overhead is caused by handoff, which can be either intro-system or inter-system handoff. The exact handoff delay depends on the handoff strategy and the network trafc condition. Typical handoff delays among the three different types of APs are listed below: the handoff delay between the WiFi and the 3G is around 400 ms [80]; the handoff delay between the 3G cells is usually less than 100 ms [81]; the handoff delay between WiFi cells is less than 100 ms [82]. Without known experimental results, we assume that the handoff delay of the WiMAX is similar to that of the WiFi, i.e., 100 ms between the WiMAX cells, 100 ms between the WiMAX cell and the WiFi cell, and 300 ms between the WiMAX cell and the 3G cell. Now we want to estimate the delay by taking the above three different types of overhead into 84

consideration. When the MH is inside a cell, the longest handoff delay this MH can experience is to handoff from the WiFi to the 3G, and then to the WiMAX, which takes 700 ms. Considering the handoff delay of 100 ms to the next cell, we have 800 ms in total. Since the sum of the computing delay and the signaling delay is less than 100 ms, the total delay is less than 900 ms in a cell. For the worst-case estimation, we assume a delay of 1 s in each cell introduced by the three types of system overhead. Given this delay being subtracted from i in each cell i, and the transmission cost of signaling data counted into the total cost, we can compute the optimal data delivery scheme by running the LP model. We x one route and generate random maps. The averaged results are shown in Table 4.2,indicating the total communication cost increase due to system overhead is no more than 3%. Now we study the scenario with multiple delay bounds. We x the route and the map, and introduce three delay bounds, at t1 = 153 s for 114 MB, at t2 = 257 s for 163 MB, and t3 = 424 s for 230 MB. In Fig. 4.10, the downloaded data is compared with the scenario where only one delay bound is specied at time T = 424 s for M = 230 MB data. As can be seen in Fig. 4.10, extra data must be downloaded by time t1 and t2 in order to satisfy the corresponding delay bounds, thus the MH is not able to postpone data transmission to future low cost cells. Consequently, the overall communication cost rises in the multiple delay bounds scenario. We have also studied the scenario where multiple possible routes exist. In particular, we choose three candidate routes and keep them xed in the simulation. Associated with each route is its probability, denoted by P1 , P2 , and P3 , respectively. In our simulation, P1 increases from 0.1 to 0.8, while P2 = P3 = (1 P1 )/2. One delay bound at T = 424 s is set for M = 230 M B data, and one wireless interface card is used. It is indicated in Fig. 4.11 85

that, with the increase of P1 , the cost of route 1 decreases, since the optimization algorithm is more in favor of the route with higher probability. At the same time, the costs of route 2 and route 3 increase with P1 , and the overall cost decreases slightly. Two curves are shown in Fig. 4.12 for the cost of prefetched data of route 1 and the average cost of prefetched data of route 2 and route 3. As we can see from Fig. 4.12, the increase of P1 results in more prefetch in route 1 and less prefetch in routes 2 and 3. In general, a higher taken probability of a route leads to a more aggressive use of this route and a lower total communication cost in this route. We change the MHs cell dwelling time to study the impact of variable moving speed of the MH. Let i denote the cell dwelling time of cell i, and denote the percentage of variation of i . We create three different patterns for the variation of the cell dwelling time. The rst pattern is to increase i by ranging from 10% to 50%. The second pattern is to decrease i by ranging from 10% to 50%. The third pattern is to let the cell dwelling time uniformly distributed between (1 ) i and (1 + ) i . In our simulation, we x a route and let M = 230M B . When the MH detects a change in i , the LP model will be regenerated and rerun. The simulation results of pattern 1 and 2 are shown in Fig. 4.13 and Fig. 4.14, respectively. We show the communication cost versus the number of reruns in these two gures. As we can see from Fig. 4.13, when i increases by from 10% to 50%, the corresponding communication cost drops accordingly. The reason is that the MH has more time to nish the download task due to the increase of i in the cell, and thus the MH can resort to cheaper AP to download data instead of more expensive ones. Also is true that the communication cost reduces with the increase of rerun times for any value. On the contrary, as shown in Fig. 4.14, if i decreases by from 10% to 50%, the communication cost rises, or 86

the MH cannot even nish the downloading task because it becomes infeasible for the data left to be downloaded in the remaining time. Fig. 4.15 shows the results under pattern 3, which compares the actual communication cost with the ideal communication cost if we know the exact cell dwelling time in advance. The motivation for such comparison is to learn the impact of variable speed on the proposed LP model. As depicted in Fig. 4.15, the difference in communication cost between these two scenarios increases from 3% to 6%, when goes from 10% to 50%. The result shows that the proposed LP model works effectively by rerunning the LP model, even when the MH varies its speed signicantly. Fig. 4.15 also shows that the system overhead due to handoff, rerunning the LP model, and signaling overhead is around 3%. This observation conforms to the simulation results of system overhead in Table 4.2. 4.7 Summary In this chapter, we aim to minimize the communication cost in an overlay heterogeneous wireless network. We have investigated and formulated the cost minimization problem, which has been proven to be NP-hard. We have also proposed an efcient minimum-cost data delivery algorithm based on linear programming, with various constraints taken into consideration, such as channel bandwidth, link costs, delay budget, and user mobility. In case of insufcient bandwidth in the core network, we have proposed a prefetch scheme in order to fully utilize the bandwidth of wireless channels. When multiple routes are available, a probability-based approach is brought up to minimize cost. Extensive simulations have been carried out to evaluate the proposed cost minimization schemes. Our simulation results show that the proposed linear programming approach can effectively reduce the overall

87

communication cost in various application scenarios, with small overhead (less than 3%) in signaling, computing, and handoff. The proposed algorithm can be integrated into the heterogeneous wireless networks and the emerging 4G wireless networks for minimum-cost data delivery.

88

CHAPTER 5 SELF-MAINTENANCE SCHEDULING FOR WIRELESS NETWORKS

In this chapter, the self-maintenance scheduling problem in next generation wireless networks is discussed. 5.1 Introduction A ubiquitous and integrated architecture has been envisioned for future wireless communication. In particular, integrating multi-hop wireless networking technologies (such as ad hoc networks [83], [84], sensor networks [85], and/or Bluetooth [86]) with infrastructure-based systems (such as 3G cellular systems [7]) is becoming one of the most promising trends in building next generation wireless networks. The multi-hop networking technologies introduce unmatched exibility into wireless systems. At the same time, however, the mobile nodes in the multi-hop wireless networks need to take up more responsibilities (such as routing) and are inherently unreliable due to their limited computing, storage, transmission, and/or battery power, resulting in failures that may dramatically increase the system maintenance cost. To improve reliability, each wireless node may monitor its available resources, such as battery power and storage capacity, and generate a maintenance request whenever the available resource becomes lower than a predened threshold. The requests should be processed timely and properly such that the maintenance activities (either with or without human involvement) are performed without conicting each other or interfering normal network operations. In this research, we consider a network with a number of active wireless nodes. Each

node maintains certain resources to provide, e.g., database inquiring service, computing service, and/or data forwarding service. At the same time, each node has a limited online life time (e.g., due to the limited battery life or storage capacity), called nodal life cycle. Different types of nodes may have different life cycles. Carefully scheduled maintenance is needed before the end of the nodal life cycle, in order to avoid network failure due to, e.g., the loss of important data or the partition of the network. Two types of constraints need to be considered during scheduling, namely resource conicting (RC) constraints and resource maintaining (RM) constraints. An RC constraint indicates a set of nodes among which any two of them must not perform maintenance at the same time, because these nodes share, the same data backup server, the same communication route for data/program update, the same battery charger, and/or the same network operator/administrator if human assistance is needed for maintenance. An RM constraint indicates a set of nodes that cannot be all in maintenance mode simultaneously. For instance, at least one node in a minimum cut of the network should be online at any time to avoid network partition. For another example, multiple copies of important data (such as mobile commerce database) are stored by different nodes, and at least one of those nodes should be in working mode. In each network (or subnetwork), a static and computationally powerful station serves as the scheduler, which keeps a set of RC and RM constraints and determines the maintenance schedule. For example, in a large sensor network, the sensor nodes may store the collected data in local storage, instead of sending out the data continuously, in order to reduce energy consumption. When the local storage becomes full, the sensor should ofoad the data and release the storage for new data acquisition. Given a tremendous set of working sensors, a large number of ofoading requests may be generated and sent to the scheduler in a short 90

period of time. Upon receiving the maintenance requests, the scheduler runs the proposed scheduling algorithms with the consideration of various constraints (such as shared data server, routing path, and so on) to reach a schedule with minimum amount of total maintenance time. Finally the determined schedule is sent to the requesting nodes, indicating when they can perform their maintenance. In this chapter, we develop three approaches to solve the scheduling problem discussed above, i.e., an integer linear programming (ILP) approach and two heuristic approaches. In the ILP approach, RC and RM constraints are converted into a set of linear constraint functions, which can be solved by available tools. The ILP model minimizes the total time used for maintenance while satisfying all constraints. In the RC-Cliques-RM heuristic approach, we rst try to satisfy the RC constraints by generating a resource compatible graph and nding all maximal cliques on this graph, aiming at scheduling as many concurrent tasks as possible. Then the RM constraints are checked, and a set-covering algorithm is employed to select a subset of the cliques to cover all requesting nodes, resulting in the desired schedule with minimized total maintenance time. In the RC-RM-Cliques approach, the RM constraints are checked before applying the nding-cliques algorithm, which results in a signicant reduction in the algorithm running time. We implement the proposed scheduling schemes and evaluate their performance in terms of algorithm complexity and total maintenance time. Our results show that all schemes can effectively schedule the requests within a reasonable period of time, but with different applicable scenarios. ILP approach is effective when the number of requests is large, while RC-Cliques-RM heuristic approach performs well at the presence of many constraints. When both the number of requests and the number of constraints are large, the RC-RM-Cliques scheme can be used to balance the tradeoff between accuracy and time 91

complexity. It is anticipated that the proposed scheduling algorithms may be generally applicable to various mobile wireless networks where self-maintenance is needed. The rest of this chapter is organized as follows. Section 5.2 discusses the related work. Section 5.3 describes the proposed scheduling schemes. Section 5.4 presents the simulation results. Section 5.5 summarizes the chapter. 5.2 Related Work Self-maintenance scheduling in multi-hop wireless networks is a largely untouched area, and there is few prior work. A centralized scheduling algorithm is proposed in [29] for ad hoc networks, in order to determine the time interval during which a node with low energy can be powered off to recharge its battery, while keeping the whole network interconnected all the time. Whenever low energy is detected, the node sends out a request to the scheduler, which employs a model prediction to estimate the power consumption and the life cycle of the mobile node based on the nodal mobility, the transmitter/receiver distance, and the height of antennas. The maintenance period is carefully scheduled in order to maintain the connectivity of the network when those mobile nodes are powered off. This is actually a special example of our RM constraints. In addition, the RC constraints are not considered here. The Energy Dependent Participation (EDP) scheme is proposed in [30]. It is a distributed approach that aims at prolonging the life cycle of the mobile nodes in ad hoc networks by allowing some nodes to enter sleeping mode temporarily, and at the same time, preventing any unwanted partition of network due to the absence of those nodes. In this algorithm, the probability that a node enters the sleeping mode depends on the remaining power of the node. Accordingly, the participation of a node in the network is reduced gradually with the decrease

92

of its remaining power. The algorithm depends totally on the residue power, and thus is insensitive to other restrictions in the network as we have discussed in Section 5.1 . More discussions with regard to the withdrawing of nodes from the network can be found in [87], [88], and [89], and their goals are to achieve energy efciency. In [90], the authors evaluate different scheduling algorithms for queuing packets on the mobile ad hoc nodes. Although many scheduling algorithms have been proposed, none of them are suitable for maintenance scheduling because of the difference in the nature of problems involved. In addition, the readers are referred to [91], and [92] for more general scheduling algorithms. 5.3 The Proposed Scheduling Algorithms In this section, we introduce the proposed scheduling algorithms. We rst formulate the scheduling problem to be addressed in this research. Then we present an ILP model and two heuristic algorithms to solve the problem. 5.3.1 Problem formulation Assume the scheduler receives N maintenance requests generated by a set of nodes, {ni | 1 i N }, where ni is the node identication (ID) number. Each ni has an associated maintenance time di . The RC constraints are represented by a number of resource conicting sets, RCS = {RCk | 1 k |RCS |}. |RCS | denotes the total number of RC constraints (i.e., the number of resource conicting sets). RCk is a resource conicting set, in which any two nodes cannot perform maintenance at the same time. |RCk | denotes the number of requests (or requesting nodes) in RCk . RCkj (1 j |RCk |) is a request (or a requesting node) in RCk . The RM constraints are represented by a number of resource maintaining sets,

93

RM S = {RMm | 1 m |RM S |}). Not all nodes in a set RMm can perform maintenance simultaneously. |RM S | denotes the total number of RM constraints. |RMm | denotes the number of requests (or requesting nodes) in a resource maintaining set RMm . RMmn (1 n |RMm |) is a request (or a request node) in RMm . In the rest of this section, a simple network with seven nodes (n1 n7 ) as shown in is used as an example for illustration. The resource conicting sets and the resource maintaining sets of the network are assumed to be RC16 and RM16 , respectively, as shown below. RC1 RC2 RC3 RC4 RC5 RC6 = n2 , = n1 , = n1 , = n2 , = n3 , = n4 , n5 , n6 ; n3 ; n5 ; n3 ; n7 ; n5 ; RM1 RM2 RM3 RM4 RM5 RM6 = n2 , = n3 , = n2 , = n4 , = n2 , = n2 , n5 , n 7 ; n4 , n 7 ; n3 ; n5 ; n4 ; n6 ;

The goal of the scheduling algorithm is to nd the maintenance starting time of each task denoted as {ti |1 i N |}, in order to minimize the total maintenance time (T ). 5.3.2 An Integer Linear Programming (ILP) Model Linear programming [93] is widely used for maximizing or minimizing the objective functions. Standard tools (such as lpsolve [79]) are available to solve the linear programming problem. A special linear programming model called integer linear programming (ILP) is characterized by having all variables taking on integer values. In this research, we have developed an ILP model to solve the self-maintenance scheduling problem. The basic idea is to minimize the ending time of the last maintenance job, with the given RC and RM constraints. More specically, we minimize max1iN {ti + di } with these constraints: 1. ti + di tj , or tj + dj ti , for any i, j RCk , 1 k |RCS |; 2. tRMki + dRMki tRMkj , or tRMkj + dRMkj tRMki , for at least one pair of RMki and 94

RMkj , where 1 i, j |RMk |, i = j, 1 k |RM S |. The above non-linear functions/constraints can be linearized by standard methods. More specically, the non-linear maximum function is linearized by minimizing a variable T with the added constraints T ti + di (1 i N ), as shown in line 1 of Fig. 5.2. The upper bound of T is the sum of all the di , denoted by in line 2 of Fig. 5.2. For any two elements ni and nj in a resource conicting set RCk (1 k |RCS |), there is a constraint ti + di tj , or tj + dj ti , in order to guarantee no overlap between their maintenance period. To linearize this constraint, a 0 1 integer variable ij (that can take values of either 0 or 1) is introduced so that ti tj ij di , and tj ti (1 ij ) dj , (5.2) (5.1)

where ij = 1 indicates ti > tj , and ij = 0 reects ti < tj under the constraints of Equations 5.1 and 5.2. For a resource maintaining set RMk = {RMk1 , RMk2 , ..., RMk|RMk | } (1 k |RM S |), at least two nodes RMki and RMkj perform maintenance at different time, i.e., tRMki + dRMki tRMkj , or tRMkj + dRMkj tRMki . Similarly, we transform this constraint into the following inequations by introducing 0 1 variables xy , where x and y take appropriate values as below.

|RMk |1 |RMk |

[(tRMki tRMkj dj )RMki ,RMkj +

i=1 j =i+1

(tRMkj tRMki di )RMkj ,RMki ] 0, 95

(5.3)

and
|RMk |1 |RMk |

(RMki ,RMkj + RMkj ,RMki ) 1,

i=1 j =i+1

(5.4)

and 0 RMki ,RMkj + RMkj ,RMki 1, f or all 1 i (|RMk | 1) and (i + 1) j |RMk |. Inequation 5.3 is still non-linear because it includes the product of two variables. In order to linearize it, we let x,y,z = tx y,z . Thus Inequation 5.3 is transformed to
|RMk |1 |RMk |

(5.5)

(RMki ,RMki ,RMkj RMkj ,RMki ,RMkj

i=1 j =i+1

+ RMkj ,RMkj ,RMki RMki ,RMkj ,RMki RMki ,RMkj dj RMkj ,RMki di ) 0 with the constraints x,y,z ( dx ) y,z , x,y,z 0, x,y,z tx , x,y,z tx ( dx )(1 y,z ). All the inequations related to RM constraints are shown in lines 4.1 through 4.4 in Fig. 5.2. According to the above discussion (as summarized in Fig. 5.2), the linear inequations can be generated and solved by using lpsolve (or other tools), and consequently one can obtain the optimized maintenance schedule (i.e. {ti |1 i N }). There is not an effective method to predict the running time of the ILP approach. However, we can estimate the size of the problem in terms of the number of constraint inequations. We notice that there are N + 1 inequations generated in lines 1 and line 2 of Fig. 5.2, while the number of inequations generated by the RC constraints is O(L L2 k ), and that generated by the RM constraints is
2 O(M Mk ), where L = |RCS |, M = |RM S |, Lk = max1iL {|RCi |}, and

(5.6)

96

n2

n3

n2

n3

n1

n4

n1

n4

n7 n6
(a)

n5

n7 n6
(b)

n5

Figure 5.1: (a) Gf Resource Conicting and (b) Gp Resource Compatible.

Mk = max1iM {|RMi |}. Thus the total number of inequations is in the order of O(L L2 k)
2 + O(M Mk ) + O(N ).

5.3.3

RC-Cliques-RM Heuristic Algorithm The ILP model yields close-to-optimal results but with undetermined running time. In

this subsection, we present a heuristic algorithm called RC-Cliques-RM, whose basic idea is to nd the minimum number of compatible sets, within each of which the requests can be scheduled to execute simultaneously. Here are the major steps of this algorithm.

Step 1: Construct Resource Conicting and Resource Compatible Graph. We rst construct the resource conicting graph Gf where each vertex represents a maintenance request, and two vertices are connected by an edge if the corresponding nodes are in the same resource conicting set. For example, Fig. 5.1 (a) illustrates the resource conicting graph of the example network. Based on the resource conicting graph, the resource compatible graph Gp can be obtained easily by a complement operation, i.e., Gp = Gf . More specically, Gf contains edges not in Gp and excludes edges that are in Gp . The resource compatible graph of the above example is shown in Fig. 5.1 (b).

97

Step 2: Find All Maximal Cliques of Resource Compatible Graph. A clique is a complete subgraph, where any two nodes are connected by an edge. Clearly, the nodes in a clique of the compatible graph Gp can perform maintenance concurrently, if only RC constraints are considered. A maximal clique is a clique that cannot be enlarged by adding any additional vertex. For example, {n1 , n4 , n7 , n6 } in Fig. 5.1 (b) is a maximal clique. The problem of nding all maximal cliques has been studied extensively. An efcient approach proposed in [94] is adopted in our algorithm. The basic idea of this algorithm is to use backtracking together with branch and bound technique to reduce time complexity. While the details of the algorithm is skipped here, we can nd all maximal cliques of the compatible graph in Fig. 5.1 (b) to be: S1 = {n1 , n4 , n6 , n7 }, S2 = {n1 , n2 , n4 , n7 }, S3 = {n3 , n4 , n6 }, S4 = {n3 , n5 }, and S5 = {n5 , n7 }. Step 3: Check Resource Maintaining Constraints. Assume we have obtained q maximal cliques in Step 2. In this step, we consider the RM constraints. More specically, if a maximal clique contains a resource maintaining set, i.e., Si RMj (1 i q and 1 j |RM S |), then Si is broken down into smaller cliques in size, so that they do not contain any resource maintaining sets and are not subsets of any other cliques. This is accomplished via a recursive function call in our implementation. For example, in this step, the maximal cliques S1 to S5 obtained in Step 2 are updated to cliques: S1 = {n1 , n4 , n6 , n7 }, S2 = {n1 , n2 , n7 }, S3 = {n3 , n4 , n6 }, S4 = {n3 , n5 }, and S5 = {n5 , n7 }. Step 4: Select Sets with Minimum Total Maintenance Time. We now select sets with minimum total maintenance time among the sets obtained in Step 3 to cover all the requesting nodes. This is a typical set-covering problem, and can be solved by an existing simple approximation algorithm [93] with modications in order to take different maintenance time 98

of the nodes into consideration. First, the algorithm picks, at each stage, the set covering the greatest number of vertices that have not been selected, with a polynomial time complexity. If there is a tie, the set with the greatest weight is selected (the weight of each set is calculated by adding the maintenance time of all the nodes in this set). This is reasonable because if the set with the greatest weight were not selected, then the nodes in this set would perform maintenance separately, which may cause possible longer maintenance time. Applying this set-covering algorithm to the sets in Step 3, and assuming the maintenance time for each node ni to be i (just for simplicity), we select three sets: S1 = {n1 , n4 , n6 , n7 }, S2 = {n3 , n5 }, and S3 = {n2 }. (Note that a node is eliminated from a set Si if it has been covered by sets S1 to Si1 ). Clearly, the nodes in the selected set Si meet both the RC constraints and the RM constraints, and thus can perform maintenance simultaneously. For example, nodes n1 , n4 , n6 , and n7 are scheduled for maintenance starting at the same time, say t1 = t4 = t6 = t7 = To . It will be To + 7 when this set is nished. Similarly, n3 and n5 are scheduled for maintenance at time t3 = t5 = To + 7, and will nish at To + 7 + 5, while n2 is scheduled to start at t2 = To + 7 + 5, and will terminate at To + 7 + 5 + 2 = To + 14. So the total maintenance time is T = 14. The time complexity of this heuristic approach is analyzed as follows. The complexity of Step 1 is O(N 2 ) because there are at most N (N 1) edges. In Step 2, the problem of 2

nding all cliques is NP-hard and we implement the approximation algorithm proposed in [94], whose time complexity is O(3.14N ). In Step 3, the set breaking problem is solved by a recursive method, with a complexity of O((q M )N ), where q is the number of cliques generated with upper bound 3(N/3) [95], and M stands for the number of RM constraints groups. The time complexity in this step is O(3N ). Let p be the number of sets after set 99
2

breaking, which is comparable with q , and thus p is also in the order of O(3(N/3) ). In Step 4, set-covering problem takes p and N as input, and is solved by a greedy algorithm [93], with time complexity of O(N 2 p). In summary, the time complexity of the heuristic approach is in exponential order of N 2 . Note that, although the complexity of the heuristic algorithm is in very high order of N , it is unrelated to the number and size of the RC and RM constraints, thus being effective in some scenarios, as to be discussed in Section 5.4 . 5.3.4 RC-RM-Cliques Heuristic Algorithm

We have noticed that the third step of the RC-Cliques-RM heuristic algorithm dominates the time complexity. This is because that a large number of cliques are generated in Step 2 of the RC-Cliques-RM algorithm, accordingly resulting in high complexity in Step 3 of the RC-Cliques-RM, where every clique is checked with the RM constraints. In order to reduce the time complexity, we address the RM constraints before generating cliques. More specically, we propose an RC-RM-Cliques heuristic with the following steps.

Step 1: Construct Resource Compatible Graph. This step is the same as step 1 in RC-Cliques-RM algorithm. Step 2: Remove One Edge According to Each RM Set. We rst construct a complete graph of each RM set, then check if this complete graph is a subgraph of the compatible graph. If it is, we randomly select a pair of nodes in the RM set, and remove the corresponding edge in the compatible graph, which connects these two nodes. Step 3: Find All Maximal Cliques in the Graph Generated by Step 2. The same algorithm used in Step 2 of the RC-Cliques-RM heuristic is proposed in [94]. Note that, all maximal cliques generated in this step satisfy the RM constraints, which have already been checked in

100

the previous step. Step 4: Select Sets with the Minimum Total Maintenance Time. This step is the same as Step 4 in the RC-Cliques-RM algorithm. Applying this algorithm to the example network, three compatible sets are obtained: S1 = {n1 , n4 , n6 , n7 }, S2 = {n3 , n5 }, and S3 = {n2 }, which are the same results we have obtained by using the RC-Cliques-RM algorithm. The time complexity of Step 2 is O(N 3 ), and similar to what we analyzed before, the time complexity of Steps 1, 3, and 4 are polynomial, exponential, and polynomial, respectively. Clearly, the total time complexity of this algorithm is still exponential (O(N 2 3N )), but much lower than that of the RC-Cliques-RM algorithm (O(3N )), and the bottleneck is in the third step. As we will show in our simulation results later, this algorithm dramatically reduces the running time compared with that of the RC-Cliques-RM heuristic. Meanwhile, its optimality is less than but still close to that of RC-Cliques-RM approach. As we can see, the RC-RM-Cliques heuristic trades the accuracy for algorithm running time. In particular, the random selection of the removed edge in Step 2 may lead to non-optimal results, although it reduces the number of cliques generated, and thus resulting in lower time complexity. 5.4 Simulations and Discussion In this section, we present our simulation results of the proposed maintenance scheduling algorithms. The simulation program is written in C language and runs on a SUN workstation with one 333 MHz CPU. Due to the lack of benchmark data, we randomly generate RC and RM constraints for given values of L, Lk , M, and Mk . The total number of requests, N , is also
2

101

Table 5.1: Impact of Increasing L and M . L, Lk , M, Mk N =15 ILP RC-Clq-RM RC-RM-Clq Greedy N =15 ILP RC-Clq-RM RC-RM-Clq Greedy 11,2,4,3 10 0.1 0.1 0.1 12,2,5,3 13,2,6,3 14,2,7,3 algorithm running time 6.6 110 0.2 0.2 0.2 0.1 0.1 0.1 0.1 0.1 0.1

lines/ time for ILP, sets/ time for others 192/ 20 220/ 14 289/ 17 358/ 3/ 22 3/ 23 3/ 22 3/ 28 3/ 22 3/ 23 3/ 22 3/ 28 4/ 37 4/ 33 4/ 33 4/ 39

randomly selected. The variable maintenance time of each requesting task, which is assumed to range from 1 to 10 time units, is also randomly generated. To evaluate the effectiveness of the proposed approaches, a greedy algorithm is used as baseline for our comparison. The greedy algorithm works in a very simple and straightforward manner. It picks up the tasks in a random order, and schedules each task without violating RC and RM constrains. If a task cannot be executed in parallel with existing tasks due to the constraints, it is put at the end of the schedule. This greedy process repeats until all tasks have been scheduled. Three experiments are conducted to evaluate the performance of the proposed algorithms in terms of the algorithm running time (i.e., complexity), and the total maintenance time (i.e., optimality), and scalability. Notice that in Table 5.1, 5.2, and 5.3, the number of equations involved by the ILP approach is recorded to show the problem complexity, and the number of sets calculated by the heuristics are also recorded, along with the total maintenance time (T ). In the rst experiment, we increase {L, M } from {11, 4} to {14, 7}, while keeping

102

Table 5.2: Impact of Increasing Lk and Mk . L, Lk , M, Mk N =15 ILP RC-Clq-RM RC-RM-Clq Greedy N =15 ILP RC-Clq-RM RC-RM-Clq Greedy 11,2,4,2 1.7 0.1 0.1 0.1 11, 2, 4, 3 11,3,4,3 11,4,4,4 algorithm running time 10 45 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1

lines/ time for ILP, sets/ time for others 151/ 24 193/ 20 201/ 17 326/ 3/ 24 3/ 22 3/ 26 4/ 32 3/ 24 3/ 22 3/ 26 4/ 32 4/ 38 4/ 37 4/ 35 5/ 41

Table 5.3: Impact of Increasing N . N 14 24 25 26 L = 10, Lk = 2, M = 3, Mk = 3, algorithm running time ILP 0.3 3.5 3.1 1.1 RC-Clq-RM 0.2 548 2204 RC-RM-Clq 0.1 0.1 0.1 0.2 Greedy 0.1 0.2 0.1 0.1 ILP RC-Clq-RM RC-RM-Clq Greedy lines/ time for ILP, sets/ time for others 165/ 18 176/ 18 177/ 17 178/ 17 2/ 18 2/ 18 2/ 20 2/ 18 2/ 18 2/ 20 2/ 18 3/ 28 2/ 20 3/ 24 2/ 20

103

Lk = 2, Mk = 3, and N = 15 as constants (See Table 5.1). At each step of increment, new RC sets and RM sets are added to the existing RC and RM sets, respectively, so that we can observe the effect of L and M on the time complexity and optimality of the algorithms while maintaining other factors unchanged. As we can see from Table 5.1, the running time of all heuristic approaches are less than 1 second, and that of the ILP approach increases signicantly in respect to L and M . When L and M are large (e.g., L = 14, M = 7), the ILP approach cannot yield converged results within one hour. The ILP model yields close to optimal result, and the two heuristic algorithms are second best in terms of optimality, while the greedy algorithm results in the longest maintenance time. In the second experiment, we increase {Lk , Mk } from {2, 2} to {4, 4}, while keeping L = 11, M = 4, and N = 15 as constants (See Table 5.2). Similar to the rst experiment, the ILP approach consumes longer time than the heuristic approaches, but results in better maintenance schedules (i.e., shorter total maintenance time). When Lk = 4 and Mk = 4, no converged results can be obtained in ILP, which indicates that the complexity of ILP model is more sensitive to Lk and Mk than to L and M . This is reasonable because the number of
2 constraints in ILP model is proportional to O(L2 k ) and O (Mk ), a higher order compared with

O(L) and O(M ), as discussed in Section 5.3.2 . On the other hand, the running time of the three heuristic approaches are kept short. Note that the increase of the number and the size of the constraints does not necessarily increase the running time of the heuristic approaches (see Table 5.1 and Table 5.2). The complexity of the heuristic algorithm is largely determined by the number and size of the cliques, which may or may not increase monotonically with more constraints in RC and RM. As an extreme case, when every pair of nodes appears in the RC sets, the resource conicting 104

graph is fully connected, and there is no edge in the resource compatible graph. Thus the clique size will be one, resulting in very short running time for the heuristic algorithms. A similar phenomenon can be observed for ILP approach, where the increase of the number of constraint functions does not necessarily increase the time complexity. This is because the actual time needed for ILP to converge depends on the geometrical shape of each ILP problem in the multi-dimensional space. From both Tables 5.1 and Table 5.2, we observe that when the number of equations in ILP exceeds 300, the ILP model is unable to converge because of the increased complexity of the problem. In the third experiment, we increase N from 14 to 26, while maintaining other parameters unchanged, i.e., L = 10, Lk = 2, M = 3, and Mk = 3 (See Table 5.3). From the results, we notice that the running time of RC-Cliques-RM algorithm increases exponentially with regards to N , while the complexity of ILP approach is rather insensitive to N . This also veries our discussions in Section 5.3.2 and Section 5.3.3 . When N is 26 or above, the time needed for the RC-Cliques-RM approach is more than 1 hour, which makes it impractical to handle a large number of requests because of the excessive time needed for completing this scheduling algorithm. The RC-RM-Cliques algorithm and greedy algorithm have very short running time. Similar to experiments 1 and 2, whenever there is a difference in terms of total maintenance time, the ILP approach achieves better result than the heuristic algorithms. Thought the results are not shown here, we have further increased the size of the problem, i.e., the value of N , L, Lk , M , and Mk . We observe that both RC-RM-Cliques and greedy algorithm can adapt to problems of large size, while in terms of optimality, RC-RM-Cliques algorithm outperforms the greedy algorithm. Another interesting observation is that in some cases not shown in the above tables, the greedy algorithm yields less number of sets but 105

Table 5.4: Complexity and Optimality of the Proposed Algorithms. Algorithm Complexity Optimality Scalability ILP highest very good low RC-Clq-RM exponential good medium RC-RM-Clq exponential medium good

longer maintenance time than the RC-RM-Cliques algorithm does. This veries the effectiveness of the RC-RM-Cliques algorithm, which aims at nding the minimum maintenance time, not necessary the minimum number of sets. From the above three experiments, we observe that the ILP model and the two heuristic approaches are suitable for different scenarios. When N is large, e.g., greater than 25, the time complexity of RC-Cliques-RM algorithm is unaffordable. When L, Lk , M , and Mk are large, the ILP approach is unable to solve the scheduling problem within a reasonable amount of time. The RC-RM-Cliques balances the tradeoff between optimality and complexity. Thus upon receiving maintaining requests, the scheduler needs to rst analyze the parameters, and choose an appropriate approach for this particular problem. A comparison between these algorithms in term of complexity, optimality and scalability is given in Table 5.4. 5.5 Summary In this chapter, we have studied the self-maintenance scheduling problem in multi-hop wireless networks. We have considered a general model with resource conicting (RC) constraints and resource maintaining (RM) constraints, which broadens previous research that solely focuses on power reservation. An integer linear programming (ILP) approach and two heuristic algorithms have been developed to solve the maintenance scheduling problem. Their performance have been evaluated and compared through analysis and extensive simulations. It

106

is shown that all schemes can effectively schedule the maintenance requests, but with different running time at different scenarios. The ILP approach performs well with a small number of constraints, while the RC-Cliques-RM heuristic approach is more sensitive to the number of requests, and both of them yield accurate results. The RC-RM-Cliques algorithm is a tradeoff between optimality and complexity, and its algorithm running time is reduced at the expense of increasing total maintenance time.

107

Minimize T, subject to 1. T ti + di , 1 i N .
N

2. =
i=1

di .

3. For any two elements i and j in RCk , 1 k |RCS |, 3.1. ti tj ij di . 3.2. tj ti (1 ij ) dj . 3.3. int ij , 0 ij 1. 4. For RMk = {RMk1 , RMk2 , ..., RMk|RMk | }, 1 k |RM S |,
|RMk |1 |RMk |

4.1.
i=1 j =i+1

(RMki ,RMki ,RMkj RMkj ,RMki ,RMkj + RMkj ,RMkj ,RMki RMki ,RMkj ,RMki RMki ,RMkj dj RMkj ,RMki di ) 0.
|RMk |1 |RMk |

4.2.
i=1 j =i+1

(RMki ,RMkj + RMkj ,RMki ) 1.

4.3. For each x,y,z , 4.3.1. x,y,z ( dx ) y,z . 4.3.2. x,y,z 0. 4.3.3. x,y,z tx . 4.3.4. x,y,z tx ( dx )(1 y,z ). 4.4. For each set RMk , int RMki ,RMkj , 0 RMki ,RMkj 1, and for each pair of RMki ,RMkj and RMkj ,RMki , let RMki ,RMkj + RMkj ,RMki 1.

Figure 5.2: Integer Linear Programming (ILP) Model.

108

CHAPTER 6 SCHEDULING FOR MULTI-HOP RELAY WIRELESS CELLULAR NETWORKS

In this chapter, the scheduling problem for multi-hop relay wireless cellular networks is discussed. 6.1 Introduction Wireless cellular networks have undergone rapid development over the past several decades. Wireless technologies and application scenarios have changed a lot during the evolution of wireless cellular networks, and a variety of wireless technologies currently exists. However, there are some common goals that different wireless networks want to achieve when they aim to meet the requirements of future wireless networks. These common goals include higher network throughput, wider coverage area, larger capacity, and better quality. Over the years, wireless cellular networks have moved from the rst generation [1], [2], to the second generation [3], [4], then to the third generation [7], [8], and even to the future wireless networks [14], [17]. The growth of wireless cellular networks shows that the quest for higher network throughput, wider network coverage, larger network capacity, and better network quality never stops. There has been a variety of approaches to address these important requirements of wireless cellular networks. For example, people have designed modulation and coding scheme with more bandwidth-efciency in order to carry data streams at higher speed with the same bandwidth. Larger cell size, seamless handoff algorithms, and efcient frequency reuse schemes have been developed to achieve wider coverage. Network capacity can be increased

by avoiding network congestion with careful trafc planning. We can also add more spectrum into the wireless networks to increase network capacity. Scheduling algorithms are designed to improve network quality by providing QoS guarantee, such as network throughput and delay. In addition, more sensitive receivers are built to retrieve wireless signals under unfavorable conditions. Besides the aforementioned approaches, an architectural enhancement called relay network have been applied to the current wireless cellular networks to achieve better throughput, coverage, capacity, and quality. In the relay network, one or multiple RS are placed between a BS and an MS for data relaying, thus creating a multi-hop wireless network. The number of hops between the BS and the MS can be more than two. The effect of using an RS in the wireless network is multifold. First, the RS can expand the coverage area of wireless cellular networks. It is obvious that the radius of the cellular cell can be extended since the RS adds its own radius into the reachable range of the cellular cell. Besides, the RS can help to reach areas where there is no direct link available, for example, some shadow areas behind high and dense buildings are difcult for wireless signal to reach with direct link, but with the RS placed at suitable position, these shadow areas can be reached easily. Second, relay networks can also increase the capacity of wireless cellular networks. It is indicated in [72] that the RS enables neighboring cells to borrow channels from each other so that when one cell is congested with high volume of trafc, the borrowed channel from its neighboring cell can help relieve the congestion and thus increase network capacity and reduce network packet dropped rate. Third, the RS can improve the signal-to-noise ratio (SNR) of each relay. Compared to 110

the direct link transmission, the distance of each relay link is shortened and better SNR is expected due to less attenuation in the transmission path. Moreover, the reduction in transmission power in the relay link also alleviates the interference in the mobile receivers. Besides the above two factors, relay link can eliminate possible blind spots hence further contributes to the better SNR of the network. Therefore, the quality of the wireless cellular network can be improved by relay architecture. Relay links have positive effects for network throughput. First, since relay links have better SNR, the throughput of each relay link is higher than that of a direct link. Second, since a direct link is divided into several relay links, concurrent transmissions can happen among relay links, resulting in spatial reuse of wireless channels and an increase of network throughput. However, there exists negative effects on the throughput because of relay links. Notice that multiple hops caused by relay links force any packet travelling from source to destination to experience multiple transmission sections, which could potentially decrease the network throughput due to the prolonged transmission time. Consequently, it will be desirable to have a scheduling algorithm that can take into consideration both positive and negative factors and let positive effects outweigh the negative effects. The problem studied here is to design a scheduling algorithm for multi-hop relay wireless cellular networks, so that it can properly arrange the concurrent transmission scenarios for multi-hop relay links, and the overall network throughput can be improved. The rest of this chapter is organized as follows. Section 6.2 discusses the related work on cellular networks with multi-hop relays. Section 6.3 describes the proposed scheduling algorithm for multi-hop relay networks. Section 6.4 presents the simulation result. Section 6.5 summarizes the chapter. 111

6.2

Related Work The study of combining relay networks with cellular networks has been going on for

quite a while. In [72], an RS is equipped with both WLAN and celluar network interfaces to assist in cellular trafc forwarding with WLAN links, aiming to achieve throughput improvement and load balancing. In [96], the MS is assumed to have both WLAN and 3G interfaces for trafc relaying. On-demand routing protocols are developed to nd relay paths, and incentive schemes are designed to encourage the MS to relay trafc for other MS. In [97], a clique nding algorithm is proposed to locate the concurrent transmissions in a multi-hop relay cellular network, and relay nodes have WLAN or WPAN interface in addition to cellular network interface. The authors assume that the MS is available everywhere in the relay path and can serve as a relay node. Different from the above approaches where a separate relay interface is needed, other approaches described below let the relay path share the same physical interface of the cellular network. In [75], multiple mobile stations form multi-hop ad hoc network within cellular network, and ad hoc routing is exploited to improve network throughput. In [31], the MS is used to relay trafc, while concurrent transmissions are investigated with a simple network topology, although the derivation is based on enumeration and hence is not suitable for practical purpose. In [33], a distributed scheduling algorithm is developed to schedule concurrent transmissions in a cellular relay network aiming at maximizing throughput while avoiding data collision in the relay links. As a result, each mobile node has to pay for the signaling and bandwidth overhead for the distributed algorithm. The above approaches have one thing in common, which is the ad hoc implementation of relay network within cellular network architecture. One advantage of the ad hoc

112

implementation is the abundance of ad hoc routing protocols that can serve in relay routing. Besides, the ad hoc routing protocol results in exibility in the selection of relay routes. However, the ad hoc implementation has some drawbacks. The rst drawback is that ad hoc routing needs each node in the relay network to participate in the route nding process that involves broadcasting, feedback, and forwarding of routing message, and this participation requires signicant modications of the signaling protocols of wireless cellular networks. These signicant modications in the signalling protocols of the BS and the MS make the current cellular network operator reluctant to deploy relay network. Second, the algorithm for nding ad hoc routing occupies a certain amount of bandwidth from each mobile node, and more bandwidth consumption is anticipated in order to keep the routing table up to date, since nodes are mobile in the network. Third, the frequent changes in the routing table reduce the reliability of the data delivery and affects the total network throughput. Therefore, the ad hoc implementation is unattractive to the network operator in industry, since the BS, the RS, and the MS all have to adjust to the architecture changes of the cellular network, and the extra signaling overhead and bandwidth consumption incurred for each MS is signicant. Observing the drawbacks of the ad hoc implementation of a relay network, we prefer an approach that involves minimum network architecture changes and still can enjoy the advantages of multi-hop concurrent transmissions. A desirable solution should be able to incorporate a relay network into the current cellular network with insignicant architecture modication and small implementation overhead, and the new network architecture is still controllable from the point view of the network operator. When designing a relay network inside the cellular network, there are many optional factors, such as whether or not the relay network adopts the ad hoc implementation as 113

discussed above, and whether or not the RS should use the same cellular spectrum to relay trafc, referred to as in-band relay. Other choices include whether or not the MS can serve as the RS and whether or not the scheduling algorithm for concurrent transmissions should be centralized. How to wisely choose among these options to form an adaptive relay network is a challenging task. On such attempt is the development of a WiMAX draft standard, 802.16j [98], which is a revision of WiMAX with the intend of incorporating relay network into WiMAX network. This 802.16j draft standard has the following characteristics. The RS in 802.16j is for relay trafc only, and the scheduling algorithm is centralized and run in the BS. The RS uses the same spectrum as the BS and the MS, and no ad hoc routing is allowed in the relay paths. The approach in 802.16j allows WiMAX network to incorporate relay network without sacricing WiMAX network architecture, and the WiMAX BS is still able to manage the RS. One research work [32] study the scheduling algorithm in 802.16j, and it uses directional antennas to exploit the space reuse of frequency resources in multi-hop relay cellular network under a Manhattan-like environment. Although scheduling algorithm is provided to enhance network throughput, but how to determine the concurrent transmissions in the relay network is not addressed. What motivates our research is the work in [99], which assumes a multi-hop wireless mesh backhaul network with in-band relay. Each RS aggregates the trafc from the nearby MS. A linear programming model is then developed to calculate the minimum time to transmit a xed data load from the BS to every RS over possible multiple hops. Since the transmission time is minimized for the xed data load, throughput is maximized. However, it does not taken into consideration the varying queue size of each relay station, neither does it apply the frame boundary of cellular network into transmission. Since wireless cellular 114

networks are predominantly frame-based, it is non-trivial to study the scheduling algorithm for multi-hop relay network under frame-based assumptions. 6.3 Scheduling for Multi-hop Relay Cellular Networks In this section, we rst illustrate our approach to build a cellular relay network with centralized control, then we present a linear programming model developed for multi-hop relay scheduling, followed by a discussion of how to nd concurrent transmission scenarios in multi-hop relay networks. 6.3.1 Preferred Approach for Building Cellular Relay Networks Following the discussions in Section 6.2 , we study the scheduling algorithm for a wireless multi-hop relay network under the architecture of wireless cellular network. The characteristics of the cellular relay network are described below. It is a cellular network with frame-based transmission scenarios, with the BS connecting to the RS and/or the MS, and each RS can connect further to other RS and/or MS. The RS only forwards trafc to/from the MS, and the RS does not generate its own trafc. In this cellular relay architecture, the MS does not need to know whether it is communicating with a BS or an RS, and no software change is needed in the MS. Also due to the fact that the RS is solely responsible for routing, there is no need for the MS to route packets to other destinations. The BS, the RS, and the MS all share the same spectrum to communicate with each other, thus no hardware change such as a second physical interface is needed for the MS or the RS besides the physical interface of cellular networks. In the uplink frame, the BS gathers the queue information of its associated RS, and the resulting change in signalling protocol is not signicant since a small portion of uplink channel bandwidth is enough to carried realtime queueing status. Then the BS makes

115

the scheduling decision, which the RS and the MS obey, and the BS will broadcast the scheduling result to the RS and the MS in the downlink frame. Thus this is a centralized scheduling algorithm. This centralized approach of building a cellular relay network instead of the ad hoc approach enables the BS to be in full control of its associated RS and MS, and the network operator can rest assured that the relay network is still manageable. The architecture change in the cellular network is reduced to a scale that the current network operator will not consider the introduction of the RS a burden. Because there is no hardware and software change in the MS and no signicant signalling protocol change in the system, the centralized approach to incorporate relay network into the current cellular network is a feasible solution. The relay network draft standard 802.16j [98] conforms with our assumption of the relay network, but we do not constraint our study on 802.16j technology. Our research result can also be applied to other types of cellular networks if the relay network is constructed in a way that matches the characteristics described above. 6.3.2 Linear Programming Model for the Scheduling Algorithm Now we present a linear programming model to implement the scheduling algorithm for cellular multi-hop relay network. An example of the cellular relay network is shown in Fig. 6.1, which has M mobile stations and R relay stations under the control of 1 base station. The downlink frame size is T seconds, and the frames are indexed by t in the time domain. We consider the one-to-M downlink data transmission, with the BS serving as the single source, and the M MS serving as destinations. The transmission can be accomplished by a direct link from the BS to the MS,

116

Figure 6.1: Architecture of Cellular Relay Networks.

or by relay links via the RS. In each RS, a designated queue is maintained for every MS that can be reached by this RS. At the beginning of each frame, the queue information of each RS is transmitted to the BS. Based on the queue information, the BS makes the scheduling decision. Each node (the BS, the MS, or the RS) is given a unique node ID and each link is given a unique link ID, as indicated in Fig. 6.1. We dene a concurrent transmission scenario as a set of all links that can transmit at the same time, for example, l1 , l8 and l12 forms a concurrent transmission scenario since these three links can happen simultaneously. Similarly, l4 and l7 forms another concurrent transmission scenario, while l4 and l8 cannot form a transmission scenario. Let the k th transmission scenario denoted by Sk , and k ranges from 1 to K , where K is the total number of concurrent transmission scenarios. We assume all transmission scenarios share each frame duration T in Time Division Multiplexing (TDM) manner, and the time portion occupied by

117

scenario Sk at frame t is denoted by Tk (t). Let xm ij (k, t) denote the packet transferred from node i to node j destined for MS node m in scenario k at frame t, given that there is a link lij from node i to node j . We name node i the upstream node of node j , and node j the downstream node of node i. Let rij (k, t) denote the actual data rate of link lij , and rij (k, t) can be calculated according to Shannons Theorem. The actual packets sent in lij is
m

xm ij (k, t), and it should be less than the capacity of lij ,

represented by rij (k, t) Tk (t). This capacity constraint indicates that the data load of a link is subject to the actual data rate and the scheduled time portion of this link. As mentioned earlier in this section, each relay node i has a queue for every reachable MS node m, and we let Qm i (t) denote the size of this queue at the beginning of frame t. Since the MS consumes the packets destined for itself and the MS does not relay packet, the MS does not need to maintain any queue. If we add Qm i (t) and the packets arrived from node is upstream nodes destined for MS node m, the sum should be equal to the packets forwarded to node is downstream nodes destined for MS node m, plus the remaining packets in this queue, which becomes the outstanding packets at frame t + 1. This ow constraint reects the dynamic change of queues in each RS. Let am (t) denote the packets arrived in MS node m at frame t. Apparently,
K

am (t) =
k=1 n

xnm (k, t), where node n and m has a link lnm between them. From the

expression of am (t), we can see that the number of packets arriving in MS node m at frame t is the sum of all the packets arriving through its neighboring nodes over all the transmission scenarios scheduled at this frame t. In order to maximize each MS node ms throughput am (t) at each frame, the scheduling algorithm need to determine how long each scenario Sk should be activated at frame t, 118

indicated Tk (t), and how much data destined for m should be transmitted at every link lij of each active scenario Sk , indicated by xm ij (k, t). The input of the scheduling algorithm include the size of all the queues in the RS and the transmission power of each link. A linear programming (LP) model is developed as the scheduling algorithm, as shown in Fig. 6.2. The constraints of the linear programming model in Fig. 6.2 highlight the characteristics of the scheduling algorithm. Constraint 1 calculates the throughput of MS node m. Constraint 2 shows that the proposed scheduling algorithm is a queue-aware algorithm, since the trafc ow is monitored and the RS queue status is maintained accordingly. This queue-aware feature is not possessed by the related work discussed in Section 6.2 . Constraint 3 shows that the actual data transmitted of a link in a scenario is subjected to the link data rate, and the total scheduled time of this link in this scenario. Constraint 4 applies Shannons Theorem to calculate link data rate with consideration of the interference caused by concurrent transmissions. Constraint 5 says that the sum of the total scheduled time of all the scenarios is no larger than one frames duration T . It shows that this scheduling algorithm is a frame-based approach considering frame boundary, which is neglected by the approach in [99]. 6.3.3 Finding Concurrent Transmission Scenarios

Notice that in Fig. 6.2, input variable 5 is a set of concurrent transmission scenarios Sk , 1 k K . Each Sk needs to be derived based on the current network topology. Since the number of links grows non-linearly with the growth of nodes in the network, it is unpractical to use an exhaust algorithm to search every possible scenarios. Thus a greedy algorithm is usually used to derive a subset of scenarios. In [99], [97], the greedy algorithm rst randomly selects a link from the set of all links, then the next link candidate is added into the scenario

119

only if the adding of this link candidate can maximize an object function. In [99], the object function is the sum of capacity for all the links already in the scenario plus the candidate link, while in [97], the objective function is the sum of the product of capacity and distance. This greedy algorithm adds one concurrent link to the scenario at a time, until no more concurrent link is available. Different scenarios can be derived when the initial link of the scenario is selected randomly. We follow the greedy algorithm used in [99] and [97], but we select a different object function, since we want to address both throughput and fairness. The research result in [100] states that that in order to maximize the end-to-end throughput in multi-hop wireless network, the selection of concurrent transmissions should be able to maximize the object function of wij Rij with respect to S , where (i, j ) are the pair of nodes that form a link in concurrent
(i,j )S

scenario S , Rij is the channel capacity of this link, and wij is the weight of this link, and
m m m wij = max(Qm i Qj ), where Qi and Qj are the queue length for destination node m in m

node i and node j , respectively. The weight denition for wij is called back-pressure. The objection function
(i,j )S

wij Rij always favors those links that can reduce the back pressure in

the network. Intuitively, this back pressure ow control has positive effect on the network throughput since it help to relieve the bottleneck trafc in the network, observing the fact that bottleneck trafc happens where back pressure is high. Nevertheless, back pressure ow control scheme also help to achieve fairness among links. Once the back pressure of one link is reduced, other links with high back pressure will be selected for transmission in the next round. Since it help to reduce the back pressure of all the links, not just favoring any specic link, it is fair scheduling algorithm. This back pressure ow control promotes both throughput and fairness, we apply it to derive concurrent scenarios Sk for the proposed scheduling 120

Table 6.1: Simulation Network Parameters. node type node id r (km) (degree) K 20 BS 0 0 0 T 1S RS 1 0.7 90 w 10MHz 2 0.6 210 N0 -75dBm 3 0.3 330 PBS 40dBm 4 0.78 330 PRS 37dBm 5 0.75 85 6 0.65 200 MS 7 8 0.65 0.33 230 320

9 0.9 325

algorithm. Thus in the greedy algorithm of nding Sk , when we try to add a link candidate into scenario Sk , we want to select the link that can maximize the object function, i.e, lij = arg(i,j ) max
(i,j )candidates

wij (t)rij (k, t), where wij is the back pressure weight of lij ,

and rij (k, t) is the data rate of lij . In summary, our proposed scheduling algorithm can be divided into two steps. In step 1, we apply the back pressure ow control scheme described above to derive a set of concurrent transmission scenarios; in step 2, we build a linear programming model stated in Fig. 6.2 to calculate the transmission schedules among different scenarios aiming at maximizing the throughput of each frame. 6.4 Simulations and Discussion We carry out extensive simulations to evaluate the proposed scheduling algorithm and study the effectiveness of deploying relay station in the cellular network. We also compared the performance of our approach with other scheduling algorithm. Fig. 6.3 shows the network topology used for simulation. It is a network with 1 BS, 4 RS, and 5 MS. The number in the parenthesis under each node in Fig. 6.3 is the node index. The parameters of the network are summarized in Table 6.1.

121

In Table 6.1, the (r, ) pair is the coordinate of each node in the polar coordinate system centered at the BS, measured in terms of km and degree, respectively. The number of concurrent scenarios K is set to 20. The greater the value of K , the more complete the concurrent scenario, at the cost of more computing overhead. Thus we select the value of K aiming at achieving enough accuracy without incurring too much computing overhead. We assume that the duration of one frame is 1 second, denoted by T , and the bandwidth for the downlink channel is 10 MHz, denoted by w. We set the BS power, the RS power, and the noise power to 40 dBm, 37 dBm, and -75 dBm, respectively. We compare our proposed scheduling algorithm with direct transmission. In direct transmission, we still run the linear programming model in Fig. 6.2, and the corresponding concurrent transmission scenarios only include direct links. We also compare our approach with a modied version of the linear programming model in Fig. 6.2. In the modied version, no buffer is provided in each RS, and all the packets arriving in each RS at a frame have to be delivered or dropped. Since there is no buffer in each RS, the concurrent scenario is calculated with no weight factor for each link throughput, and thus a link candidate is added into a concurrent transmission scenario if this link can maximize the total throughput of the scenario. This approach with no buffer in the RS is similar to the assumption of the approach in [99]. Trafc is generated in a BS on a per frame basis with Poisson distribution. The throughput of our proposed scheduling algorithm are the throughput value averaged over multiple frames. The comparison result of throughput is shown in Fig. 6.4. From Fig. 6.4, we can see that when the trafc request per MS node increase from 0.5 Mb per frame to 1.5 Mb per frame, all the three approaches can satisfy the input, and the average throughput per MS node can still catch up with the request. When the request goes between 2 Mb and 2.5 122

Mb per frame for each node, the direct transmission cannot satisfy the request anymore, while our approach can still satisfy the request. Thus it veries the effectiveness of the proposed scheduling algorithm. When the request per MS node exceeds 3 Mb, our scheduling algorithm cannot satisfy the requests either. The reason is that the average data rate of all the ve MS in one frame is around 9 Mbps in direct link, thus it is difcult for ve MS to reach a total data rate of 3 5 = 15 Mb in relay link given the request of 3 Mb per frame per MS. We can see from Fig. 6.4 that the total data rate of the ve MS is around 13.5 Mb (i.e., 2.7Mb per MS) when the input request is 15 Mb for all MS nodes (i.e., 3 Mb per MS). When the request reach 4 Mb in Fig. 6.4, we can see the scheduling result of linear programming can reach 3 Mb per MS node, i.e., a total of 15 Mb for ve MS, which is better than the total data rate of 9 Mb for ve MS in direct link. This also proves the effectiveness of the relay links. One more thing to mention is that the linear programming approach with queue constraint has better performance than the one without queue constraint. The reason is that with queue constraint, each relay node can store the remaining packets of this frame, and they are ready for delivery in the next frame. In comparison, when there is no queue available in the RS, no such reservation in queue is allowed and no data are in a ready-to-be-delivered state. For any data that need to be sent out, extra bandwidth has to be consumed to send them from upstream nodes rst. Besides network throughput, we are also interested in fairness performance, which is measured by Jains Fairness Index [53], dened in Equation 3.8 in Chapter 3. The fairness comparison result is shown in Fig. 6.5, and our approach has the best fairness performance. One reason is that in the nding of concurrent scenarios, we apply the back pressure ow control theorem and it helps to remove trafc bottleneck in the network, hence achieving fairness, while in the approach with no queue constraint, back pressure ow control cannot be 123

applied and it causes more unfairness than our approach. The other reason is that a relay channel can improve the throughput of the MS whose direct link is not in good state, and throughput of this MS is comparable to those MS whose direct link is in good state. Since these two different types of MS have similar throughput with the help of relay, the whole network seems more fair. Also as can be seen from Fig. 6.5, direct transmission is the worst among the three scheduling algorithms since the link with better quality always receives favor from the scheduler. For any scheduling algorithm, Fig. 6.4 shows that the throughput of scheduled data cannot catch up with the throughput of the input when the data rate of the input reaches a certain threshold value. If we examine the corresponding fairness performance of this scheduling algorithm at the input threshold value in Fig. 6.5, we can see that the fairness also begins to drop. One interesting question of the relay network is whether using relay links is always benecial. We study this issue by changing the distance between the BS and the MS/RS, and investigate its impact on the network throughput. The result is shown in Fig. 6.6. The x-axis of Fig. 6.6 is the distance scale factor. If we multiply the distance scale factor with the r values in Table 6.1, we get the real distances from the BS to each RS or MS. The y-axis in Fig. 6.6 is the average throughput per MS node per frame. We can see from Fig. 6.6 that when the distance scale reduces from 1 to 0.4, the throughput gain of our scheduling algorithm over the direct transmission decreases gradually. This is because when the MS/RS is close enough to the BS, the drawback of the multi-hop transmission outweighs the gain, and thus direct transmission becomes a better choice than multi-hop relay. The conclusion we can draw is that relay links and direct links apply to different applicable scenarios. A relay link is more effective when the MS is far away from the BS and suffers from bad channel quality, while 124

direct link is more effective when the MS is close to the RS and enjoys good channel quality. Notice that the channel quality of each link in the relay cellular network is not only related to the length of the link, but also related to the how many concurrent transmissions are going on. The concurrent transmissions will generate interferences and therefore affect the data rate of every link in the scenario. The number of concurrent transmission scenarios, K , also affects the performance of the scheduling algorithm. Fig. 6.7 shows the throughput performance of our proposed scheduling algorithm under different K . Intuitively, when K is large, the linear programming yields result more close to the optimum solution, as indicated in Fig. 6.7. But when K reaches a certain value, the improvement in throughput is marginal. The reason is that when K reaches a certain value, the concurrent scenarios that can benet the network throughput could have all been added into the set, and the including of more scenarios will not boost the network throughput signicantly. We want to apply a reasonably large value to K to achieve near optimal result, but we do not want go to the other extreme of using very large K , which implies high computation overhead. 6.5 Summary We have presented a scheduling algorithm for multi-hop relay wireless cellular networks. Through our analysis, we argue that following a centralized approach for building cellular relay networks best reects the interest of the cellular networks. This centralized approach implies that relay stations and mobile stations do not form ad hoc networks and they are under the control of base station. Other choices of building relay cellular networks we follow include using in-band spectrum for relay stations, not allowing mobile stations to serve as

125

relay stations, and applying centralized scheduling algorithm. A centralized scheduling algorithm is developed and base stations will run this scheduling algorithm. In the scheduling algorithm, rst a set of concurrent transmission scenarios is derived and then it is used as input for a linear programming model that determines the transmission schedules for the multi-hop relay network. The linear programming model aims at maximizing the overall throughput of the all the mobile stations, while taking into consideration the frame-based nature of cellular network and the dynamic queue change in the relay stations. The features of frame-based and queue-awareness of the scheduling algorithm are the unique contributions that have not been addressed by previous work. Simulations measure performance metrics such as throughput and fairness of the proposed scheduling algorithm. Two other scheduling algorithms are compared with our approach via simulations. One is scheduling for direct transmission only, and the other is scheduling without buffer in the relay nodes. The effectiveness of our approach is validated by the simulation results.

126

OBJECTIVE: maximize
m

am (t);

INPUT VARIABLES: 1: MS index m; 2: frame index t; 3: frame duration T ; 4: RS node is queue status Qm i (t); 5: a set of concurrent transmission scenarios Sk , 1 k K ; 6: power used from node i to j , Pij ; 7: distance between node i to j , dij ; OUTPUT VARIABLES: 1: xm ij (k, t), scheduled packets transmitted from node i to node j in Sk at frame t, which are destined for MS node m; 2: Tk (t), scheduled time portion for scenario Sk ; CONSTRAINTS:
K

1: am (t) =
k=1 K s

xsm (k, t), where s is MS node ms upstream node index;

K

2:

Qm i (t)

+
k=1 s

xm si (k, t)

=
k=1 r

m xm ir (k, t) + Qi (t + 1), where i is RS index,

and s and r stands for node is upstream and downstream node, respectively; 3: xm ij (k, t) rij (k, t) Tk (t);
m

4: rij (k, t) = w log2 (1 +

K

), N0 + (x,y)Sk ,(x,y)=(i,j ) Pxy /d xj where is the path loss exponent, and N0 is noise power; Tk (t) = T ;
k=1

Pij /d ij

5:

Figure 6.2: LP Model for Scheduling in Cellular Relay Networks.

127

Figure 6.3: Cellular Relay Network Topology for Simulation.

Avg throughput per node(Mbits/frame)

3.5 3 2.5 2 1.5 1 0.5 0 1 2 3 Request per node(Mbits/frame) 4 With Queue Constraint W/O Queue Constraint Direct Transmission

Figure 6.4: Throughput Comparison of Different Scheduling Algorithms.

1 0.95 Fairness 0.9 0.85 0.8 0.75 0 With Queue Constraint W/O Queue Constraint Direct Transmission 1 2 3 Request per node(Mbits/frame) 4

Figure 6.5: Fairness Comparison of Different Scheduling Algorithms.

128

Avg throughput per node(Mbits/frame)

Req = 10 Mbits/frame 10

With Queue Constraint W/O Queue Constraint Direct Transmission 0.2 0.4 0.6 distance scale 0.8 1

2 0

Figure 6.6: Impact of Distance from the BS on Network Throughput

Avg throughput per node(Mbits/frame)

3.5 3 2.5 2 1.5 1 0.5 0 1 2 3 Request per node(Mbits/frame) 4 K = 10 K = 15 K = 20 K = 30 K = 40

Figure 6.7: Impact of Number of Scenarios on Network Throughput.

129

CHAPTER 7 CONCLUSION

The focus of the dissertation is the scheduling and optimization issues in next generation heterogeneous wireless networks. Among the many features of next generation wireless networks are heterogeneous networks, broadband, integrated services of voice and data, Quality of Service (QoS) guarantee, self-maintenance scheduling requirements, and the integration of relay networks with cellular networks. Based on these features, challenges for building next generation wireless networks are identied and several optimization problems are formulated accordingly. The scheduling and resource optimization problems studied in the dissertation include the QoS guarantee issue in the Worldwide Interoperability for Microwave Access (WiMAX) network, the minimum cost data delivery issue with QoS guarantee in heterogeneous wireless networks, the self-maintenance scheduling issue in next generation wireless networks, and the scheduling issue in multi-hop relay wireless cellular networks. Scheduling solutions are proposed to address these problems and analysis models are built for the proposed solutions. Then we carry out extensive simulations to verify the correctness of the proposed solution and analysis models. In the WiMAX QoS guarantee problem, our scheduling algorithm can effectively provide QoS guarantee to various types of connections, which is crucial to the system requirements of the WiMAX. Besides, the proposed analysis model precisely describes the dynamic behavior of the WiMAX scheduling algorithm, behavior that has not been shown in the previous related work. Thus our scheduling scheme is efcient and the analysis model is

novel. In the minimum cost data delivery problem, our proposed scheduling scheme can effectively reduce the total system cost of data delivery, and several scenarios are studied to provide a complete picture of the data delivery process, including multiple physical interfaces, prefetched data, multiple routes, dynamic speeds, and system overhead. Our approach is an inventive approach since no such attempt has been made before to address the cost issue under dynamic data delivery. Nevertheless, the QoS requirements associated with data delivery are observed in the cost minimization process, and the proposed scheduling algorithm can serve as a guideline for cost minimization with QoS guarantee. In the maintenance scheduling problem, the three scheduling algorithms we proposed can t into practical applications of minimizing the total schedule maintenance time while obeying various resource conicting and resource maintenance constraints. The considerations of these constraints make our approach unique from other approaches, and they reect the concrete maintenance needs of next generation wireless networks. In the scheduling of cellular relay networks, we are the rst to consider the dynamic queue status of the relay nodes. We also creatively apply back pressure ow control to the searching of concurrent transmission scenarios, resulting in network throughput improvement and fairness enhancement. In addition, with the concurrent transmission interference taken into account, the proposed scheduling algorithm for cellular multi-hop networks can be applied to practical scenarios and can successfully increase the network throughput. To summarize, we conclude that this dissertation addresses important scheduling and resource optimization problems for next generation wireless networks, and the proposed solutions provide deeper understandings not only of the scheduling algorithms but also of the features and characteristics of next generation heterogeneous wireless networks. 131

BIBLIOGRAPHY

[1] F. H. Blecher, Advanced Mobile Phone Service, IEEE Trans. Veh. Tech., vol. 29, no. 2, pp. 238244, Jul. 1980. [2] L. C. Godara, Handbook of Antennas in Wireless Communications. Boca Raton, FL: CRC, 2002. [3] M. Mouly and M. B. Pautet, The GSM System for Mobile Communications. Paris, France: Telecom, 1992. [4] L. Harte, M. Hoenig, D. McLaughlin, and R. Kikta, CDMA IS-95 for Cellular and PCS: Technology, Applications, and Resource Guide. Chicago, IL: McGraw-Hill Professional, 1999. [5] A. Fares, GPRS and EDGE Engineering. Charleston, SC: BookSurge, 2006. [6] Mobile Station-Base Station Compatibility Standard for Dual-Mode Spread Spectrum Systems, 1998, TIA. [Online]. Available: http://www.tiaonline.org/standards/. [7] H. Holma and A. Toskala, WCDMA for UMTS. Hoboken, NJ: John Wiley and Sons, 2002. [8] Y. S. Rao and A. Kripalani, cdma2000 mobile radio access for IMT 2000, in Proc. IEEE Int. Conf. PWC, 1999, pp. 615. [9] IEEE Standard for Local And Metropolitan Area NetworksPart 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specications: Higher-speed Physical Layer Extension in the 2.4 GHz Band, 1999, IEEE802.11b. [Online]. Available: http://www.ieee802.org/11/. [10] IEEE Standard for Local and Metropolitan Area NetworksPart 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)Amendment: High-speed Physical Layer in the 5 GHz band, 1996, IEEE802.11a. [Online]. Available: http://www.ieee802.org/11/.

[11] IEEE Standard for Local and metropolitan area networksPart 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)Amendment 4: Further Higher Data Rate Extension in the 2.4 GHz Band, 2003, IEEE802.11g. [Online]. Available: http://www.ieee802.org/11/. [12] IEEE Standard for Local and Metropolitan Area NetworksPart 15.1: Wireless Medium Access Control (MAC) and Physical Layer (PHY) specications for wireless personal area networks (WPANs), 2005, IEEE802.15.1. [Online]. Available: http://www.ieee802.org/15/. [13] D. Mulvey, HSPA, IET Commun. Eng., vol. 5, no. 1, pp. 3841, Jan. 2007. [14] UTRA-UTRAN Long Term Evolution (LTE) and 3GPP System Architecture Evolution (SAE), 2006, 3GPP. [Online]. Available: http://www.3gpp.org. [15] CDMA2000 Standard for Spread Spectrum Systems, Revision D, C.S0002-D, 2004, 3GPP2. [Online]. Available: http://www.3gpp2.org. [16] CDMA2000 Standard for Spread Spectrum Systems, Revision E, C.S0002-E, 2005, 3GPP2. [Online]. Available: http://www.3gpp2.org. [17] Ultra Mobile Broadband (UMB) Air Interface Specication, C.S0084, 2007, 3GPP2. [Online]. Available: http://www.3gpp2.org. [18] IEEE Draft Standard for Local and Metropolitan Area NetworksPart 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)Amendment: Enhancements for Higher Throughput, 2007, IEEE802.11n. [Online]. Available: http://www.ieee802.org/11/. [19] IEEE Standard for Local and metropolitan area networksPart 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)Amendment 8: Medium Access Control (MAC) Quality of Service Enhancements, 2005, IEEE802.11e. [Online]. Available: http://www.ieee802.org/11/.

133

[20] J. S. Gomes, M. Yun, H. A. Choi, J. H. Kim, J. K. Sohn, and H. I. Choi, Scheduling Algorithms For Policy Driven QoS Support in HSDPA Networks, in Proc. 65th IEEE Veh. Tech. Conf., 2007, pp. 799803. [21] Y. J. Kang, J. Kim, D. K. Sung, and S. Lee, Hybrid Scheduling Algorithm for Guaranteeing QoS of Real-Time Trafc in High Speed Uplink Packet Access (HSUPA), in Proc. IEEE Int. Symp. PIMRC, 2007, pp. 15. [22] Y. Ou, C. Wang, and X. Wang, Scheduling for multi-user packet in CDMA2000 1x EV-DO, in Proc. 2nd IEEE Int. Conf. MobiSys, 2005, pp. 510. [23] M. Hawa, Stochastic Evaluation of Fair Scheduling with Applications to Quality-of-Service in Broadband Wireless Access Networks, 2003, Ph.D. dissertation, University of Kansas. [24] K. Wongthavarawat and A. Ganz, IEEE 802.16 Based Last Mile Broadband Wireless Military Networks with Quality of Service Support, in Proc. Mil. Commun. Conf., 2003, pp. 779784. [25] Q. Liu, X. Wang, and G. Giannakis, A Cross-Layer Scheduling Algorithm with QoS Support in Wireless Networks, IEEE Trans. Veh. Tech., vol. 55, no. 3, pp. 839847, May 2006. [26] D. Niyato and E. Hossain, Queue-aware Uplink Banwidth Allocation and Rate Control for Polling Service in IEEE 802.16 Broadband Wireless Networks, IEEE Trans. Mobile Comp., vol. 5, no. 8, pp. 668679, Dec. 2006. [27] M. Stemm and R. H. Katz, Vertical Handoffs in Wireless Overlay Networks, ACM Mobile Netw. Special Issue Mobile Netw. Internet, vol. 3, no. 4, pp. 335350, Sep. 1998. [28] H. J. Wang, R. H. Katz, and J. Giese, Policy-Enabled Handoffs Across Heterogenous Wireless Networks, in Proc. 2nd IEEE WMCSA, 1999, pp. 5160.

134

[29] S. Tragoudas, Power dissipation component of a management protocol for ad-hoc networks, in Proc. IEEE WCNC, 2000, pp. 555559. [30] M. R. Pearlman, D. Jing, B. Liang, and Z. J. Haas, Elective participation in ad hoc networks based on energy consumption, in Proc. IEEE GLobal Telecommun. Conf., 2002, pp. 2631. [31] J. Cho and Z. J. Hass, On the Throughput Enhancement of the Downstream Channel in Cellular Radio Networks Through Multihop Relaying, IEEE J. Sel. Areas Commun., vol. 22, no. 7, pp. 12061219, Sept. 2004. [32] S.-J. Lin, W. H. Sheen, I.-K. Fu, and C.-C. Huang, Resource Schduling with Directional Antennas for Multihop Relay Networks in Manhattan-linke Environment, in Proc. IEEE Mobile WiMAX Symp., 2007, pp. 108113. [33] Y.-N. Lee, J.-C. Chen, Y.-C. Wang, and J.-T. Chen, A Novel Distributed Scheduling Alorithm for Downlink Relay Networks, IEEE Trans. Wireless Commun., vol. 6, no. 6, pp. 19851991, Jun. 2007. [34] IEEE Standard for Local and Metropolitan Area NetworksPart 16: Air Interface for Fixed Broadband Wireless Access Systems, 2004, IEEE802.16. [Online]. Available: http://www.ieee802.org/16/. [35] IEEE Standard for Local and Metropolitan Area NetworksPart 16: Air Interface for Fixed Broadband Wireless Access SystemsAmendment 2: Physical and Medium Access Control Layers for Combined Fixed and Mobile Operation in Licensed Bands, 2005, IEEE802.16e. [Online]. Available: http://www.ieee802.org/16/. [36] M. Katavenis, S. Sidiropoulos, and C. Courcoubetis, Weighted Round-Robin Cell Multiplexing in A General-Purpose ATM Switch Chip, IEEE J. Sel. Areas Commun., vol. 9, no. 8, pp. 12651279, Jan. 1991. [37] M. Shreedhar and G. Varghese, Efcient Fair Queueing Using Decit Round Robin, in Proc. IEEE SIGCOMM, 1995, pp. 231242.

135

[38] A. Demeres, S. Keshav, and S. Shenker, Analysis and Simulation of A Fair Queueing Algorithm, in Proc. IEEE SIGCOMM, 1989, pp. 112. [39] A. Parekh and R. Gallager, A Qeneralized Processor Sharing Approach to Flow Control: The Single Node Case, in Proc. IEEE INFOCOM, 1992, pp. 915924. [40] J. Bennet and H. Zhang, WF2Q: Worst-Case Fair Weighted Fair Queueing, in Proc. IEEE INFOCOM, 1996, pp. 120128. [41] S. Golestani, A Self-Clocked Fair Queueing Scheme for Broadband Applications, in Proc. IEEE INFOCOM, 1994, pp. 636646. [42] C. Liu and J. Layland, Scheduling Algorithms for Multiprogramming in A Hard-Real-Time Environment, J. ACM, vol. 20, no. 1, pp. 116, Jan. 1973. [43] J. Chen, W. Jiao, and H. Wang, A Service Flow Management Strategy for 802.16 Broadband Wireless Access System in TDD Mode, in Proc. IEEE ICC, 2005, pp. 34223426. [44] K. R. Raghu, S. K. Bose, and M. Ma, Queue Based Scheduling for IEEE 802.16 Wireless Broadband, in Proc. 6th IEEE Int. Conf. ICICS, 2007, pp. 15. [45] V. Sharma and N. Vamaney, The Uniformly-Fair Decit Round-Robin (UF-DRR) Scheduler for Improved QoS Guarantees in IEEE 802.16 WiMAX Networks, in Proc. Mil. Commun. Conf., 2007, pp. 17. [46] A. Shejwal and A. Parhar, Service Criticality Based Scheduling for IEEE 802.16 WirelessMAN, in Proc. 2nd IEEE Int. Conf. AusWireless, 2007, pp. 1218. [47] N. Salodjar and A. Karandikar, An Indexing Scheduler for Delay Constrained Scheduling with Applications to IEEE 802.16, in Proc. IEEE WCNC, 2008, pp. 14711476. [48] X. Bai, A. Shami, and Y. Ye, Robust QoS Control for Single Carrier PMP Mode IEEE 802.16 Systems, IEEE Trans. Mobile Comp., vol. 7, no. 4, pp. 416429, May 2008.

136

[49] A. G. D. Kok and H. C. Tijms, A Queueing System with Impatient Customers, J. Appl. Prob., vol. 22, no. 3, pp. 688696, Nov. 1985. [50] L. Kleinrock, Queueing Systems, Volume 1: Theory. Hoboken, NJ: John Wiley and Sons, 1975. [51] S. S. Panwar, Time Constrained and Multiaccess Communications, 1986, Ph.D. dissertation, University of Massachusetts. [52] A. R. Odoni, On Finding The Maximal Gain for Markov Decision Processes, J. Oper. Res., vol. 17, no. 5, pp. 857860, Oct. 1969. [53] D. M. C. R. Jain and W. Hawe, A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Systems, 1984, dEC Research Report, TR-301. [54] F. Wang, A. Ghosh, C. Sankaran, and P. Fleming, WiMAX Overview and System Performance, in Proc. 64th IEEE Veh. Tech. Conf., 2006, pp. 15. [55] V. H. MacDonald, The Cellular Concept, Bell Syst. Tech. J., vol. 58, no. 1, pp. 1543, Feb. 1978. [56] T. Rappaport, Wireless Communications: Principles and Practice. Englewood Cliffs, NJ: Prentice Hall, 1996. [57] R. Prasad, W. Konhauser, and W. Mohr, Third Generation Mobile Radio Systems. Norwood, MA: Artech House, 2000. [58] IEEE Standard for Local and Metropolitan Area NetworksStandard Air Interface for Mobile Broadband Wireless Access Systems Supporting Vehicular MobilityPhysical and Media Access Control Layer Specication, 2002, IEEE802.20. [Online]. Available: http://www.ieee802.org/20/. [59] Cellular WiFi. [Online]. Available: http://www.troposnetworks.com/. [60] J. Gibson, The Mobile Communications Handbook. Boca Raton, FL: CRC, 1996.

137

[61] J. Mitola III, The Software Radio Architecture, IEEE Commun. Mag., vol. 33, no. 5, pp. 2638, May 1995. [62] A. Munro, Mobile Middleware for the Recongurable Software Radio, IEEE Commun. Mag., vol. 38, no. 8, pp. 152161, Aug. 2000. [63] P. Rodriguez, R. Chakravorty, J. Chestereld, I. Pratt, and S. Banerjee, MAR: A Commuter Router Infrastructure for the Mobile Internet, in Proc. 2nd Int. Conf. MobiSys, 2004, pp. 217230. [64] R. Schollmeier, A Denition of Peer-to-Peer Networking for The Classication of Peer-to-Peer Architectures and Applications, in Proc. 1st Int. Conf. Peer-to-Peer Comp., 2001, pp. 101102. [65] G. Fox, Peer-to-Peer Networks, IEEE Comp. Sci. Eng., vol. 3, no. 3, pp. 7577, May/Jun. 2001. [66] W. Winston and M. Venkataramanan, Introduction to Mathematical Programming. Pacic Grove, CA: Brooks/Cole, 2003. [67] J. Blau, Wi-Fi Hotspot Networks Sprout Like Mushrooms, IEEE Spectr., vol. 39, no. 9, pp. 1820, Sept. 2002. [68] X. Gao, G. Wu, and T. Miki, QoS Framework for Mobile Heterogeneous Networks, in Proc. IEEE Int. Conf. Commun., 2003, pp. 933937. [69] J. J. Garcia-Luna-Aceves, C. Fullmer, E. Madruga, D. Beyer, and T. Frivold, Wireless Internet Gateway (WINGS), in Proc. Mil. Commun. Conf., 1997, pp. 12711276. [70] A. T. Campbell, J. Gomez, and A. G. Valko, An Overview of Cellular IP, in Proc. IEEE WCNC, 1999, pp. 606610. [71] S. Bae, S. Lee, and M. Gerla, Unicast Performance Analysis of Extended ODMRP in a Wired-to-Wireless Hybrid Ad-Hoc Network, in Proc. Mil. Commun. Conf., 2002, pp. 12281232.

138

[72] H. Wu, C. Qiao, S. De, and O. Tonguz, Integrated Cellular and Ad-hoc Relay Systems: iCAR, IEEE J. Sel. Areas Commun., vol. 19, no. 10, pp. 21052115, Oct. 2001. [73] X. Wu, B. Mukerjee, and S. G. Chan, MACAAn Efcient Channel Allocation Scheme in Cellular Networks, in Proc. IEEE GLobal Telecommun. Conf., 2000, pp. 13851389. [74] J. Zhou and Y. Yang, Parcels: Pervasive Ad-hoc Relaying for Cellular Systems, in Proc. 1st Annu. Mediterranean Ad Hoc Netw. Workshop, 2002, pp. 5160. [75] Y. D. Lin and Y. C. Hsu, Multihop Cellular: A New Architecture for Wireless Communication, in Proc. IEEE INFOCOM, 2000, pp. 12731282. [76] H. Hsieh and R. Sivakumar, On Using the Ad-hoc Network Model in Cellular Packet Data Networks, in Proc. ACM MobiHOC, 2002, pp. 3647. [77] B. Bhargava, X. Wu, Y. Lu, and W. Wang, Integrating Heterogeneous Wireless Technologies: A Cellular Aided Mobile Ad hoc Network (CAMA), ACM Mobile Netw. Appl., vol. 9, no. 5, pp. 393408, Aug. 2004. [78] T. Ibaraki, T. Hasegawa, K. Teranaka, and J. Iwase, The Multiple Choice Knapsack Problem, J. Oper. Res. Soc. Jpn., vol. 21, no. 3, pp. 5993, Jan. 1978. [79] LP-SOLVE. [Online]. Available: http://sourceforge.net/projects/lpsolve. [80] R. Rajavelsamy, V. Jeedigunta, B. Holur, M. Choudhary, and O. Song, Performance evaluation of VoIP over 3G-WLAN interworking system, in Proc. IEEE WCNC, 2005, pp. 23122317. [81] J. Wang, J. C. L. Liu, and Y. Cen, Handoff algorithms in dynamic spreading WCDMA system supporting multimedia trafc, IEEE J. Sel. Areas Commun., vol. 21, no. 10, pp. 16521662, Dec. 2003. [82] I. Samprakou, C. Bouras, and T. Karoubalis, Fast IP handoff support for VoIP and multimedia applications in 802.11 WLANs, in Proc. 6th IEEE Int. Symp. WoWMoM, 2005, pp. 332337. 139

[83] C. E. Perkins, Ad Hoc Networking. Reading, MA: Addison-Wesley, 2001. [84] C.-K. Toh, Ad Hoc Mobile Wireless Networks: Protocols and Systems. Englewood Cliffs, NJ: Prentice Hall, 2001. [85] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, A Survey On Sensor Networks, IEEE Commun. Mag., vol. 40, no. 5, pp. 102116, Aug. 2002. [86] K. Sairam, N. Gunasekaran, and S. R. Redd, Bluetooth in wireless communication, IEEE Commun. Mag., vol. 40, no. 6, pp. 9096, Jun. 2002. [87] Y. Xu, J. Heidemann, and D. Estrin, Geography-informed energy conservation for ad hoc routing, in Proc. ACM Mobilcom, 2001, pp. 115. [88] B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris, Span: an eneyrgy-efcienct coordination algorithm for topology maintenance in ad hoc wireless network, in Proc. ACM Mobilcom, 2001, pp. 3142. [89] T. Imielinski, S. Viswanathan, and B. R. Badrinath, Energy efciency indexing on air, in Proc. ACM Mobilcom, 2001, pp. 8697. [90] B. G. Chun and M. Baker, Evaluation of Packet Scheduling Algorithms in Mobile Ad Hoc Networks, in Proc. ACM MC2R, 2002, pp. 2631. [91] W. Stallings, Operating systems: Internals and Design Principles. Englewood Cliffs, NJ: Prentice Hall, 1998. [92] M. Singhal and N. G. Shivaratri, Advanced concepts in operating systems. Cambridge, MA: MIT Press, 2001. [93] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to Algorithms. Cambridge, MA: MIT Press, 2001. [94] C. Bron and J. Kerbosch, Finding All Cliques in An Undirected Graph, in Proc. ACM Commun., 1973, pp. 575577.

140

[95] J. W. Moon and L. Moser, On Cliques in Graphs, Israel J. Math., vol. 2, no. 3, pp. 2328, Feb. 1965. [96] H. Luo, X. Meng, R. Ramjee, P. Sinha, and L. Li, The Design and Evaluation of Unied Cellular and Ad Hoc Networks, IEEE Trans. Mobile Comp., vol. 6, no. 9, pp. 10601074, Jun. 2007. [97] B. Radunovic and A. Proutiere, On Downlink Capacity of Cellular Data Networks with WLAN/WPAN Relays, in Proc. 3rd IEEE RAWNET, 2007, pp. 17. [98] IEEE Standard for Local and Metropolitan Area NetworksPart 16: Air Interface for Fixed Broadband Wireless Access SystemsAmendment: Multihop Relay Specication, 2007, IEEE802.16j. [Online]. Available: http://www.ieee802.org/16/. [99] H. Viswanathan and S. Mukherjee, Throughput-Range Tradeoff of Wireless Mesh Backhaul Networks, IEEE J. Sel. Areas Commun., vol. 24, no. 3, pp. 593602, Mar. 2006. [100] L. Tassiulas and A. Ephremides, Statibility Properites of Constrained Queueing Systems and Scheduling Policies for Maximum Throughput in multihop Radio Networks, IEEE Trans. Auto. Ctrl., vol. 37, no. 12, pp. 19361948, Jan. 1992.

141

ABSTRACT Next generation wireless networks have many promising features, such as high data rates, hybrid voice and data services, integration of different wireless access technologies, and intelligent networks with self-maintenance. With these features, several scheduling and resource optimization issues arise. In this dissertation, I aim to address these issues by rst formulating the problems and then proposing solutions accordingly. These challenging research problems include the resource allocation problem for the Worldwide Microwave Interoperability for Mobile Access (WiMAX) networks, the minimum-cost data delivery problem in heterogeneous wireless networks, the self-maintenance scheduling problem for next generation wireless networks, and the scheduling problem for multi-hop cellular relay networks. In the WiMAX scheduling, I propose a two-phase Fair and Efcient Queueing (FEQ) scheduling algorithm, which effectively combines Weighted Round Robin (WRR) and Earliest Deadline First (EDF) algorithms. I also build an elegant queueing model to derive in theory the performance metrics. In minimum-cost data delivery, I rst prove the cost minimization problem to be NP-hard, and then present an efcient minimum-cost data delivery algorithm based on linear programming, with various constraints and scenarios taken into consideration. In self-maintenance scheduling, I consider resource maintenance and resource conicting constraints and propose a linear programming model and two heuristic algorithms to schedule the maintenance requests, aiming at minimizing the total maintenance time. In scheduling for multi-hop cellular relay networks, I develop a scheduling algorithm based on linear programming to maximize system throughput while maintaining system

fairness. This scheduling algorithm takes into account the dynamic queue change in the relay stations, the frame boundary of the cellular networks, and fairness among different concurrent transmission scenarios. Extensive simulations have been conducted to verify the effectiveness of the scheduling algorithms proposed above. In summary, we have proposed several scheduling algorithms to address various scheduling and optimization issues for next generation heterogeneous wireless networks, and we anticipate that the proposed scheduling algorithms will become necessary components of next generation heterogeneous wireless networks.

143

BIOGRAPHICAL SKETCH Haining Chen received a Bachelor of Science in Electrical Engineering from Tsinghua University, Beijing, China, in 1996, and a Master of Science in Electrical Engineering from the Chinese Academy of Sciences, Beijing, China, in 1999. He has been working toward a Ph.D. in Computer Engineering with The Center for Advanced Computer Studies (CACS) at the University of Louisiana at Lafayette since 2003. From 1999 to 2002, he was a networking engineer and a software engineer in China. His current research interests include QoS guarantee for broadband wireless access networks, resource optimization for next generation wireless networks, and integration of heterogeneous wireless networks.

Chen, Haining. Bachelor of Science, Tsinghua University, Beijing, China, Spring 1996; Master of Science, Chinese Academy of Science, China, Spring 1999; Master of Engineering, University of Louisiana at Lafayette, Spring 2004; Master of Science, University of Louisiana at Lafayette, Winter 2007; Doctor of Philosophy, University of Louisiana at Lafayette, Spring 2008 Major: Computer Engineering Title of Dissertation: Scheduling and Resource Optimization in Next Generation Heterogeneous Wireless Networks Dissertation Director: Dr. Hongyi Wu Pages in Dissertation: 155; Words in Abstract: 323 ABSTRACT Next generation wireless networks have many promising features, such as high data rates, hybrid voice and data services, integration of different wireless access technologies, and intelligent networks with self-maintenance. With these features, several scheduling and resource optimization issues arise. In this dissertation, I aim to address these issues by rst formulating the problems and then proposing solutions accordingly. These challenging research problems include the resource allocation problem for the Worldwide Microwave Interoperability for Mobile Access (WiMAX) networks, the minimum-cost data delivery problem in heterogeneous wireless networks, the self-maintenance scheduling problem for next generation wireless networks, and the scheduling problem for multi-hop cellular relay networks. In the WiMAX scheduling, I propose a two-phase Fair and Efcient Queueing (FEQ) scheduling algorithm, which effectively combines Weighted Round Robin (WRR) and Earliest Deadline First (EDF) algorithms. I also build an elegant queueing model to derive in theory the performance metrics. In minimum-cost data delivery, I rst prove the cost minimization problem to be NP-hard, and then present an efcient minimum-cost data delivery algorithm based on linear programming, with various constraints and scenarios taken

into consideration. In self-maintenance scheduling, I consider resource maintenance and resource conicting constraints and propose a linear programming model and two heuristic algorithms to schedule the maintenance requests, aiming at minimizing the total maintenance time. In scheduling for multi-hop cellular relay networks, I develop a scheduling algorithm based on linear programming to maximize system throughput while maintaining system fairness. This scheduling algorithm takes into account the dynamic queue change in the relay stations, the frame boundary of the cellular networks, and fairness among different concurrent transmission scenarios. Extensive simulations have been conducted to verify the effectiveness of the scheduling algorithms proposed above. In summary, we have proposed several scheduling algorithms to address various scheduling and optimization issues for next generation heterogeneous wireless networks, and we anticipate that the proposed scheduling algorithms will become necessary components of next generation heterogeneous wireless networks.