Institutionen för systemteknik

Transcription

1 Institutionen för systemteknik Department of Electrical Engineering Examensarbete Traffic Scheduling for LTE Advanced Examensarbete utfört i Communication Systems vid Tekniska högskolan i Linköping av Zhiqiang Tang LiTH-ISY-EX--10/4413--SE Linköping 2010 Department of Electrical Engineering Linköpings universitet SE Linköping, Sweden Linköpings tekniska högskola Linköpings universitet Linköping

2

3 Traffic Scheduling for LTE Advanced Examensarbete utfört i Communication Systems vid Tekniska högskolan i Linköping av Zhiqiang Tang LiTH-ISY-EX--10/4413--SE Handledare: Examinator: Yi Wu isy, Linköpings universitet Eleftherios Karipidis isy, Linköpings universitet Linköping, 22 October, 2010

4

5 Avdelning, Institution Division, Department Division of Communication Systems Department of Electrical Engineering Linköpings universitet SE Linköping, Sweden Datum Date Språk Language Svenska/Swedish Engelska/English Rapporttyp Report category Licentiatavhandling Examensarbete C-uppsats D-uppsats Övrig rapport ISBN ISRN LiTH-ISY-EX--10/4413--SE Serietitel och serienummer Title of series, numbering ISSN URL för elektronisk version Titel Title Trafikskedulering för LTE Advanced Traffic Scheduling for LTE Advanced Författare Author Zhiqiang Tang Sammanfattning Abstract Nyckelord Keywords Long Term Evolution (LTE) is becoming the first choice of operators when constructing the new network infrastructure, because of its high throughput and low latency. Although the LTE can offer high speed data service as a benefit of wideband application, the large bandwidth also results in huge control signaling cost. This thesis studies how the available radio resources should be allocated to different users for the particular purpose. Three proposals are presented in this thesis. The first two aim at maximizing the overall net capacity. The factors of channel conditions and control signaling cost are considered in the first proposal whereas power control is supplemented as an additional factor in the second one. The third proposal aims at achieving a tradeoff between subframe efficiency and average data rate. Channel conditions and control signaling cost are taken into consideration. Resource blocks scheduling under a control signaling cost constraint is feasible with the knowledge of the channel condition of users. This is obtained from channel state information directly. The first proposal studies how this scheduling is done. The second proposal takes power allocation scheme into account. In terms of computational complexity, the methods of internal water-filling and external water-filling are described in the second proposal. The simulation results illustrate that the net capacity in the first proposal can be enhanced by about 5% - 60% in the second proposal. The exact percentage of increase depends on different control signaling penalty. The third proposal discusses how to implement flexible subframe length under the Rayleigh fading channel condition in the LTE system. The objective is to achieve subframe efficiency as high as possible, while maintaining the maximum average data rate. The simulation results illustrate that the optimal subframe length depends on control signaling cost penalty as well as on Doppler frequency. Finally, the fairness of the first proposal is compared with the improved versions of Round Robin, Dynamic Allocation and Proportional Fair scheduling algorithms. The simulation results show that Round Robin and Dynamic Allocation outperform the first proposal under a certain condition. LTE, LTE Advanced, Scheduling, Control Signalling Cost

6

7 Abstract Long Term Evolution (LTE) is becoming the first choice of operators when constructing the new network infrastructure, because of its high throughput and low latency. Although the LTE can offer high speed data service as a benefit of wideband application, the large bandwidth also results in huge control signaling cost. This thesis studies how the available radio resources should be allocated to different users for the particular purpose. Three proposals are presented in this thesis. The first two aim at maximizing the overall net capacity. The factors of channel conditions and control signaling cost are considered in the first proposal whereas power control is supplemented as an additional factor in the second one. The third proposal aims at achieving a tradeoff between subframe efficiency and average data rate. Channel conditions and control signaling cost are taken into consideration. Resource blocks scheduling under a control signaling cost constraint is feasible with the knowledge of the channel condition of users. This is obtained from channel state information directly. The first proposal studies how this scheduling is done. The second proposal takes power allocation scheme into account. In terms of computational complexity, the methods of internal water-filling and external water-filling are described in the second proposal. The simulation results illustrate that the net capacity in the first proposal can be enhanced by about 5% - 60% in the second proposal. The exact percentage of increase depends on different control signaling penalty. The third proposal discusses how to implement flexible subframe length under the Rayleigh fading channel condition in the LTE system. The objective is to achieve subframe efficiency as high as possible, while maintaining the maximum average data rate. The simulation results illustrate that the optimal subframe length depends on control signaling cost penalty as well as on Doppler frequency. Finally, the fairness of the first proposal is compared with the improved versions of Round Robin, Dynamic Allocation and Proportional Fair scheduling algorithms. The simulation results show that Round Robin and Dynamic Allocation outperform the first proposal under a certain condition. v

8

9 Acknowledgments First I would like to thank all at Communication Systems division of Linkoping University, especially my supervisor Yi Wu and examiner Eleftherios Karipidis. I also would like to thank all my friends in Sweden. Last but not least, I would like to thank my wonderful family, for all their enduring support and always believing in me. Zhiqiang Tang Linkoping, Sweden, 2010 vii

10

11 Contents 1 Introduction LOLA Project Problem Statement Thesis Scope Thesis Layout LTE and LTE Advanced Background LTE and LTE Advanced LTE Protocol Architecture LTE Protocol Architecture for the Downlink Scheduler LTE Physical Layer Frame Structure Resource Allocation for Signals Multiple Access Technique for LTE/LTE Adavanced OFDM OFDMA Scheduling Algorithms Scheduling Algorithms Classification Scheduling Algorithms for Non-Real-Time Flows Max Rate Round Robin Proportional Fair Scheduling Algorithms for Real-Time Flows Largest Delay First Modified Largest Weighted Delay First Scheduling Algorithms for Mixture of Real-Time and Non-Real- Time Flows Exponential Rule Utility Function ix

12 x Contents 4 Scheduling Under a Control Signaling Cost Constraint Background Proposed Scheduling Algorithm Under a Control Signaling Cost Constraint for LTE/LTE Advanced Notation Individual Transmission of Scheduling Maps{M k } Broadcasting of Joint Scheduling Map M Optimal Assignment S Given the User Set U Selecting the Set of Active Users U Power Optimization Water-filling External Water-filling Internal Water-filling Simulation Results Individual and Broadcast Scheduling Map Transmission External Water-filling Internal Water-filling Adaptive Length Subframe Rayleigh Fading Channel Proposed Scheduling Algorithm of Adaptive Length Subframe for LTE Advanced Motivation Subframe Length Optimization Algorithm Simulation Results Optimal Subframe Length for Individual and Broadcast Scheduling Map Transmission Modes Influence of Different Control Signaling Cost Penalties to Adaptive Subframe Length Influence of Doppler Frequency to Adaptive Subframe Length 55 6 Fairness Evaluation Proportional Fair Notation Proportional Fair Under a Control Signaling Cost Constraint Dynamic Allocation Round Robin Simulation Results Conclusion 65 8 Future Work 67 Bibliography 69

13 Abbreviations 2G Second Generation Mobile Telecommunication 3G Third Generation Mobile Telecommunication 3GPP 3rd Generation Partnership Project ARQ Automatic Repeat-reQuest AWGN Additive White Gaussian Noise CA Carrier Aggregation CP Cyclic Prefix CSI Channel State Information DCI Downlink Control Information DSCH Downlink Shard Channel DSL Digital Subscriber Line DTCH Dedicated Traffic Channel ETSI European Telecommunications Standards Institute FDMA Frequency Division Multiple Access i.i.d. Independent and Identically Distributed LOLA Low Latency in Wireless Communication LTE Long Term Evolution MAC Medium Access Control MBSFN Multicast/Broadcast over Single Frequency Network M2M Machine to Machine NRT Non-Real-Time OFDM Orthogonal Frequency Division Multiplexing OFDMA Orthogonal Frequency Division Multiple Access PCFICH Physical Control Format Indicator Channel PDCCH Physical Downlink Control Channel PDU Protocol Data Unit PHICH Physical Hybrid ARQ Indicator Channel PHY Physical Layer PRB Physical Resource Block QoS Quality of Service RLC Radio Link Control RNTI Radio Network Temporary Identifier RRC Radio Resource Control RT Real-Time SAE System Architecture Evolution SB Schedule Block SDU Service Data Unit TTI Transmission Time Interval WiMAX Worldwide Interoperability for Microwave Access

14

15 Chapter 1 Introduction Due to the development of electronic and information technology, telecommunication industry has been undergoing rapid growth. In 1991, the Finnish operator Radiolinjia launched the first 2G network in the world, which was built for voice service and low data transmission [3]. Ten years later, the first pre-commercial 3G network was launched by the Japanese operator NTT DoCoMo [4]. Besides low data transmission services, such as instance fax and conventional voice services, video and multimedia services are also supported by 3G network, as the result of its data transfer rate which can be as high as 2Mbit. In recent years, some data services, such as mobile TV, file sharing, location based services and social networking have grown very fast. Faster upload and download speeds are expected by the end users. As a 3.9G technology, Long Term Evolution (LTE) is proposed by 3rd Generation Partnership Project (3GPP) to meet these requirements. On December 14, 2009, the first commercial LTE was deployed in Scandinavian countries by the Swedish-Finnish operator TeliaSonera. The Stockholm survey shows that the throughput is up to 42.8Mbit per second for downlink and 5.3 Mbit per second for uplink with 10 MHz spectral bandwidth [5]. According to a report of the US research organization Maravedis, the investment of the top 25 operators which have committed to LTE for LTE infrastructure could amount to 14 billion dollars by At that time, these operators will provide commercial LTE service for more than 226 million subscribers [23]. There are several reasons why operators prefer LTE. One is that the higher bandwidth speed and the higher system capacity of LTE attract operators. Bandwidth is not the bottleneck for high data rate services any more. Furthermore, the latency of end to end in LTE is possible to be kept within 50 ms and it makes wireless access comparable to Digital Subscriber Line (DSL). From operators perspective, it means that there are several strategic low-latency application areas supported by LTE with respect to revenue potential [6]. Take Machine to Machine (M2M) applications for example. Comparing to 2.7 billion mobile phone users out of 6 billion people, there are 50 billion machines in the world, and only 50 million of them are connected to networks with cellular technology, as illus- 3

16 4 Introduction Figure 1.1. Market potential for M2M over LTE/LTE-Advanced [2] trated in Figure 1.1. LTE Advanced is the first wireless system to standardize the constraint of latency and traffic policies for M2M applications. Huge potential market of M2M applications on top of cellular network will be exploited soon, after LTE techniques become mature. That is another important reason why operators prefer LTE. 1.1 LOLA Project LOLA (Achieving Low Latency in Wireless Communication) is funded by EU FP7 (European Community s Seventh Framework Programme). It is a research project that is carried out by EURECOM, THALES COMMUNICATIONS SA, TECHNISCHE UNIVERSITAET WIEN, LINKOPINGS UNIVERSITET, AT4 WIRELESS, ERICSSON, TELEKOM SERBIA. LOLA is focused on access layer technologies targeting low latency, robust and spectrally efficient transmission in several emergency scenarios. Current research covers two basic types of wireless networks, namely long-range LTE-Advanced cellular networks and medium-range rapidly-deployable mesh networks. This thesis is one part of this project and it is carried out on LTE/LTE- Advanced based cellular networks. 1.2 Problem Statement In wireless communication systems, terminals must know exactly which subcarrier of certain particular time slots conveys their data, before decoding. This

17 1.3 Thesis Scope 5 information is packaged into scheduling maps. The scheduling maps are sent to the terminals by control signaling. The overhead caused by control signaling is significant in LTE/LTE Advanced due to the huge bandwidth. In LTE/LTE Advanced, resource block is the minimum resource unit and subframe is the minimum time slot that can be allocated by scheduler. The LTE/LTE Advanced systems with 20MHz channel bandwidth have 100 resource blocks. Assume 10 terminals are scheduled in one subframe in the case of downlink scheduling. It needs at least = 1000 bits per subframe for scheduling map in individual transmission mode to inform terminals about the allocation of resource blocks. As a result, about 1 Mbit throughput per second is used for control signaling overhead. It is necessary to find a good solution to reduce overhead caused by scheduling maps while keeping flexibility of resource blocks allocation. Although resource allocation types 0, 1, 2 are specified by 3GPP release 8 [1] to mitigate control signaling overhead, each type has its drawbacks. In other words, they do not perform as well as 3GPP expected. This thesis will analyze three types of solutions proposed recently and come up with new algorithms to alleviate the influence of control signaling overhead on the overall system performance. 1.3 Thesis Scope The research on control signaling overhead and scheduling in this thesis is only for the downlink of LTE/LTE Advanced system. Since this project is at its primary stage, all the scenarios are using traditional single antenna transmitter and receiver. 1.4 Thesis Layout Chapter 1 gives a short introduction to LOLA of which this thesis is a part. After analyzing why LTE/LTE Advanced probably will become the most popular communication system in the world, this chapter points out the control signaling overhead problem that LTE/LTE Advanced is suffering from. Chapter 2 introduces the technical background of LTE and LTE Advanced. Furthermore, the radio interface architecture of LTE, specifically the physical layer is presented. Chapter 3 states several classic scheduling algorithms for Real-Time (RT) and Non-Real-Time (NRT) flows. Chapter 4 analyzes recent research on lightening the influence of signaling overhead on the overall system performance for wideband communication systems. It also demonstrates two proposed scheduling algorithms under a control signaling cost constraint for LTE/LTE Advanced.

18 6 Introduction Chapter 5 presents a novel adaptive subframe scheduling algorithm on the basis of Chapter 4, which brings up a novel hypothesis beyond 3GPP LTE standard. Chapter 6 evaluates the fairness of the proposed algorithms by comparing it with some reference algorithms. Chapter 7 summarizes the research findings of Chapter 4, 5 and 6. Chapter 8 describes some issues left by this thesis. Hopefully, future work will solve these problems.

19 Chapter 2 LTE and LTE Advanced Background 2.1 LTE and LTE Advanced As the minor update of Release 9 of current LTE specification, LTE Advanced will be coincident with LTE Release 10. The work on LTE Advanced standard like IMT Advanced is ongoing. The LTE Advanced finalization in 3GPP is expected to be accomplished in 2011, aligning with finalization of IMT Advanced in ITU. The LTE Advanced can achieve much better system performance than LTE. It supports for peak data up to 1Gbps in the downlink and 500 Mbps in the uplink, in contrast with 100Mbps in the downlink and 50 Mbps in the uplink of LTE. In addition to substantial improvement for cells and terminals, the LTE Advanced provides not only more efficient spectrum utilization, but also the higher power efficiency for both infrastructure and terminals. To carry out above key performance, the following potential technology components of LTE Advanced are introduced. Wider bandwidth and carrier aggregation In order to fulfil the higher peak data rate, Carrier Aggregation (CA) is introduced to obtain wider bandwidth. Instead of adjacent component carriers, non adjacent component carriers are supported by CA for the LTE Advanced. Extended multi antenna transmission Support for spatial multiplexing on the uplink and extension of spatial multiplexing on the downlink are expected in the LTE Advanced. Advanced repeaters and relay networks [13] Allowing for cost of network infrastructure, relays are proposed in the LTE Advanced to extend cell coverage and increase data rate. The relay is intermediate post between enodeb and terminals for cell edge. 7

20 8 LTE and LTE Advanced Background Figure 2.1. LTE protocol architecture for the downlink [11] 2.2 LTE Protocol Architecture The protocol architecture is a little different between the downlink and the uplink. This thesis is focused on the downlink scheduling, so only LTE radio interface architecture for the downlink is introduced. The overview of LTE protocol architecture is shown in Figure LTE Protocol Architecture for the Downlink RRC layer Radio Resource Control (RRC) handles control plane signaling. In addition to security and Quality of Service (QoS) management, RRC is in charge of establishment, maintenance, releasing of RRC connection and mobility management. PDCP layer Core network sends data to enodeb in the form of IP packets. In order to reduce the number of bits transmitted in radio interface, enodeb uses Package Data Convergence Protocol (PDCP) to compress IP header and cipher the transmitted data at the transmitters side. At the receivers side, PDCP performs decompression and deciphering corresponding to transmitters.

21 2.2 LTE Protocol Architecture 9 Figure 2.2. Physical channel, transport channel and logical channel RLC layer Data is transmitted to Radio Link Control (RLC) Service Data Unit (SDU) buffer from PDCP layer. RLC is responsible for segmentation or concatenating data selected from SDU buffer and creating the RLC Protocol Data Unit (PDU), according to scheduler s command. In addition, retransmission mechanism is supported by RLC to assure delivering error free data to upper level. MAC layer Medium Access Control (MAC) offers service to RLC in the form of logical channels. It is shown in Figure 2.2. Several types logical channels are defined in MAC for different types of information carried by RLC. In general, logical channels are classified into control channels and traffic channels. Control and configuration information is transmitted via control channels and terminals data is transmitted via traffic channels. Furthermore, MAC deals with logical channel multiplexing for the purpose of making full use of resources and handles hybrid Automatic Repeat-reQuest (ARQ) retransmission. Uplink and downlink scheduling is supported by MAC as well. It will be particularly introduced in Section Physical layer Physical layer (PHY) offers service to MAC in the form of transport channels. There are certain rules on how to map the logical channels to transport channels regarding of types of information. Mapping of the logical channels to transport channels is implemented in MAC, by following above rules. The PHY is used to handle coding/decoding, modulation/demodulation, multi antenna mapping and other typical physical layer functions. Finally, the signals are mapped to physical channels and transmitted over the radio interface.

22 10 LTE and LTE Advanced Background Figure 2.3. LTE frame structure and physical resource [11] Scheduler The scheduler is in the MAC layer, but it controls MAC layer and Physical layer at the same time. Scheduler can be divided into downlink scheduler for downlink scheduling and uplink scheduler for uplink scheduling. Downlink scheduler The enodeb periodically receives Channel State Information (CSI) reports, which are the feedback from terminal to report the downlink channel conditions. For channel-dependent scheduling, downlink scheduler takes channel state, buffer status and priorities into account. Then it decides resource blocks allocation, the modulation scheme, and antenna mapping for terminals. As a result, downlink scheduler decision controls RLC segmentation, MAC multiplexing and Hybrid ARQ, PHY channel coding, PHY modulation and antenna mapping. Uplink scheduler Similar to downlink, enodeb uplink scheduler decides resource blocks allocation. However, logical channel multiplexing is controlled by terminals and channel state estimation is done for channel-dependent scheduling by enodeb with reference signals transmitted from each terminal covered by this enodeb. With the knowledge of channel conditions, uplink scheduler makes decisions to control channel coding and modulation scheme of terminals. 2.3 LTE Physical Layer Frame Structure The LTE frame structure is illustrated in Figure 2.3. Each LTE frame is 10 milliseconds in duration. It consists of ten subframes of length 1 millisecond with

23 2.3 LTE Physical Layer 11 Figure 2.4. Resource block [11] equal size. Two slots of equal size constitute a subframe. Each slot consists of several Orthogonal Frequency Division Multiplexing (OFDM) symbols including cyclic prefix. In LTE, Physical Resource Block (PRB) is the minimum resource unit that can be allocated by scheduler. In contrast with 48 subcarriers per slot in WiMAX, PRB consists of 12 consecutive subcarriers for one slot in duration [17] and one slot is constituted by seven or six symbols depending on the cyclic prefix length (normal or extended), seen in Figure Resource Allocation for Signals The subframe is divided into control region and data region. Control signaling is allocated to the control region. Terminals data and uplink control signaling are allocated to data region. Reference signals are allocated to control region and data region. The size of control region can be adjusted according to the amount of control signaling which is used to maintain communication between enodeb and terminals, more specifically one, two, or three symbols Downlink Reference Signals The enodeb inserts reference symbols that are known by the receivers and the transmitters into resource blocks with a constant power so that terminals can estimate instantaneous downlink channel conditions. Information about instantaneous downlink channel conditions that can be used for coherent demodulation is reported to enodeb for scheduling. Cell specific downlink reference signals

24 12 LTE and LTE Advanced Background In case of the absence of multi antenna, each resource block consists of four reference symbols. In order to estimate downlink channel precisely, four reference symbols are distributed uniformly in the frequency and time domain. In time domain, two reference symbols are inserted in the first OFDM symbol and two reference symbols are located in the last third OFDM symbol. In frequency domain, 12 carriers of each resource block are divided into two successive regions of six carriers. Each region includes a reference symbol for the first and the last third OFDM symbol. Which carrier conveys a reference symbol is dependent on cell identity. In case of multi antenna, the location of reference symbol is determined by antenna port identity. To eliminate from other antenna ports in the same cell, the resource element is vacant for other antenna ports, if this resource element is used by a certain antenna port. For a four antenna ports cell, two antenna ports only have two reference symbols per resource block [11]. The terminal can receive the cell identity in the procedure of cell search and estimate the frame timing of the cell. Thus, terminals obtain the reference signal sequence and the start of the reference signal sequence. A UE specific downlink reference signals In contrast with other reference signals, it is intended to be sent to a specific terminal based on beam forming transmission for the purpose of channel estimation. Furthermore, UE specific reference signals are inserted into data region of assigned resource block, instead of signal region. MBSFN reference signals It is used to estimate channel conditions for Multicast/Broadcast over Single Frequency Network. MBSFN reference signals are inserted in the MBSFN part of MBSFN subframe [11] Downlink Control signaling Control signaling consists of downlink scheduling, uplink scheduling, hybrid ARQ acknowledge and power control command. The following three physical channels are designed to carry downlink control signaling. Physical Control Format Indicator Channel By decoding the information of Physical Control Format Indicator Channel (PCFICH), terminals can know the size of control region, therefore the start of data region can be concluded. Two bits for information about size of control region are coded in 32 bits codeword. With scrambling and QPSK modulation, it becomes 16 bits. 16 resource elements are divided into 4 groups of four resource elements. The four groups are distributed in the control region. Similar to reference signals, the location is decided by cell identity and antenna ports in case of multi antenna so as to achieve frequency diversity and avoid interference of inter cell PCFICH.

25 2.3 LTE Physical Layer 13 Type 0 Type Bitmap Type 1 Type Subset L/R Bitmap Type 2 Start Length Figure 2.5. Resource allocation type 0, 1 and 2 Physical Hybrid ARQ Indicator Channel It is used to response uplink UL-SCH transmission with Hybrid ARQ acknowledgement. The Physical Hybrid ARQ Indicator Channel (PHICH) are mapped to resource elements in the first symbol of control region, after finishing PCFICH allocation. Physical Downlink Control Channel The resource block assignment, modulation scheme, channel coding information as well as power control commands are packaged into Downlink Control Information (DCI) and sent to terminals with Physical Downlink Control Channel (PDCCH). DCI supports several formats with different size in order to trade off between control signaling cost and scheduling flexibility. DCI format 1C is used for special applications that carry relatively small information to multiple terminals, such as transmission system information. Compared with 0 and 1C which only support allocation of frequency contiguous resource blocks with resource allocation type 2, DCI format 1 has larger control signaling cost, since it supports non contiguous resource blocks resource allocation type 1 and 0. DCI format 2 can support spatial multiplexing based on DCI format 1. In the case that system bandwidth is less or equal than ten PRBs, the resource allocation in each PDCCH only contains information of actual resource allocation. For resource allocation type 0, each bit of bitmap is pointed to a group of contiguous resource blocks, instead of an individual resource block. The size of bitmap is decreased, with the increasing of size of group of contiguous resource blocks. However, as a result of it, single resource block cannot be scheduled even for small pay load. Resource allocation type 1 mitigates this issue by dividing the total resource blocks into several subsets with a flag indicating that bitmap relate to either left or right part of resource blocks. Type 2 only encodes the start position and length of the allocated

26 14 LTE and LTE Advanced Background Figure 2.6. Overlapping spectrum of an OFDM signal resource block. The illustration of resource block allocation types is shown in Figure 2.5. Although these resource allocation types mitigate the cost of control signaling, they are still not optimal. One of the main objectives of this thesis is to further reduce control signaling cost yielded by Downlink Control Channel, thus to improve system performance. More detail will be discussed in Section 3.2, 3.3 and 3.4. In this thesis, control signaling overhead only considers delivery scheduling map cost of Downlink Control Channel due to its large throughput, compared with other control signaling cost generated by PCFIC and PHICH. 2.4 Multiple Access Technique for LTE/LTE Adavanced OFDM OFDM is a popular scheme for wideband digital communication, applied in LTE/ LTE Advanced network. Its advantage lies in dealing with frequency selective fading and intersymbol interference with high spectral efficiency, whereas its disadvantage is high peak to average power ratio as well as sensitivity to Doppler shift and to frequency synchronization OFDM Principle The basic idea of OFDM is breaking available bandwidth into several sub-carriers. Based on the fundamental idea that several slow speed streams in parallel are equal to one high speed stream, the data stream is divided into several small data streams. Each small data stream is mapped to one sub-carrier and modulated by digital modulation scheme, such as QAM, QPSK. From the perspective of frequency, OFDM has high spectral efficiency. Each sub-carrier is closer to each other without guard band, as sub-carriers are orthogonal. In other words, the peak of one sub-carrier is intersecting with the null of

27 2.4 Multiple Access Technique for LTE/LTE Adavanced 15 Figure 2.7. OFDM symbol structure [19] Figure 2.8. OFDMA VS OFDM [31] neighbour sub-carriers, as illustrated in Figure 2.6. Therefore, OFDM is one of the most efficient spectrum modulation in wideband wireless communication Cyclic Prefix Under the principle of OFDM, high speed data stream is divided into several lower speed narrow band streams. The advantage of longer duration of lower speed streams is utilized for eliminating Inter Symbol Interference (ISI), since symbol duration of lower speed streams is much larger than the maximum delay spread in general. In wireless telecommunication, transmission suffers multipath. Multipath causes lost of orthogonality between sub-carriers. Meanwhile lost of orthogonality between sub-carriers can result in Inter Carrier Interference (ICI). In order to overcome ICI, the cyclic code is introduced to relieve issue in Figure 2.7. The content of OFDM information consists of cyclic prefix and OFDM symbol. The cyclic prefix is a copy of the end part of OFDM symbol. It plays two roles here. One of them is to convert linear convolution into circular convolution in order to keep sub-carriers orthogonal in multipath conditions. The other is to prevent ISI as a guard space between successive symbols. On one hand, the period of cyclic prefix should be selected as long as anticipated degree of delay spread at least for overcoming ICI/ISI. On the other hand, the period of cyclic prefix should be limited, because it reduces the data rate up to R((T T cp )/T ).

28 16 LTE and LTE Advanced Background Figure 2.9. OFDMA model [16] OFDMA Orthogonal Frequency Division Multiple Access (OFDMA) is a combination of OFDM and FDMA. OFDM assigns all sub-carriers to one terminal in a symbol in the time domain (see Figure 2.8). In OFDMA, sub-carriers are divided into groups named sub-channels. Next, OFDMA assigns sub-channels with proper power to different terminals based on channel knowledge from CSI intending to maximize the system throughput, seen in Figure 2.9.

29 Chapter 3 Scheduling Algorithms The key to achieve optimal performance of base station is dynamically scheduling limited resources like power and bandwidth to offer the best service for terminals with the lowest cost. For scheduling based on OFDMA, it balances maximum throughput and fairness by scheduling time slots, sub-channels, modulation scheme and power with frequency diversity and multiuser diversity. Frequency diversity is done by utilizing the fact that each sub-channel suffers different attenuation in different time and frequency, due to shadowing, fast fading, multipath and so on. In a similar way, multiuser diversity is obtained by opportunistic user scheduling, since different users locate different places leading to different channel gains of an identical sub-channel for different users. By analyzing CSI, base station recognizes variation of time, frequency, space and adjusts scheduling to keep optimal performance. In order to follow variation of channel conditions, scheduling should be done within the coherent time, so it requires that allocation algorithms must be fast, especially time varying channel. As mentioned in Section 2.3.1, PRB is formed from 12 consecutive sub-carriers in the frequency domain and seven or six consecutive symbols in the time domain. The smallest resource unit which is allowed to be assigned to one terminal is Schedule Block (SB) constituted by two successive PRBs. The fastest scheduling is required to be done within 1ms according to the symbol length of SB. After scheduling, scheduling map will be sent to all terminals. Individual terminal only decodes received data in certain particular time and frequency based on scheduling map. 3.1 Scheduling Algorithms Classification Generally, scheduling can be divided into two classes: channel-independent scheduling and channel-dependent scheduling. Channel-independent scheduling does not take channel conditions into account. Therefore the performance of this kind of scheduling can never be optimal. On the contrary, channel-dependent scheduling can achieve better performance by allocating resources based on channel conditions with optimal algorithms. The following discussion about scheduling will focus on 17

30 18 Scheduling Algorithms channel-dependent scheduling. Furthermore scheduling can be classified in the light of application scenarios, since the performance of scheduling algorithms highly relies on the type of incoming flows. In order to implement maximum performance of system, it is important to select suitable algorithms according to flows, in term of applications. The flows can be divided into Real-Time flows, Non-Real-Time flows and mixture of Real- Time as well as Non-Real-Time flows due to the consideration of main services of LTE, including voice service, data service, and live video service [22]. 3.2 Scheduling Algorithms for Non-Real-Time Flows The problem of scheduling for Non-Real-Time flows can be considered as a concave optimization problem with utility functions on the basis of Proportional Fair algorithm, which will be presented later. The optimization problem can be stated as objective function and constraints in mathematics [7]. Objective function is maximizing utility function. Constraints are limitations of bandwidth and power. Scheduler pursues the optimal or suboptimal solution of above functions, while takes computation overhead into account. The most popular scheduling algorithms for Non-Real-Time flows are Max- Rate, Round Robin, and Proportional Fair. The performance of Non-Real-Time algorithms is measured with throughput and fairness Max Rate As a channel dependant scheduling, Max Rate takes advantage of multiuser diversity to carry out maximum system throughput. First, scheduler analyzes CSI from terminals to obtain data rate of an identical sub-channel for different terminals. Then scheduler assigns this sub-channel to the terminal which can achieve the highest data rate in this sub-channel based on SNR. The Max Rate can be described as i = arg max R k,n (t) k R k,n (t) is the data rate of terminal k for one sub-channel n in time slot t. Max- Rate algorithm causes starving terminals while it obtains maximum throughout. Terminals in bad channel conditions are never considered by scheduler, so it is not a fair algorithm. However, it can be considered as a good reference to measure total throughput. In Section 4.4, Max Rate is referred in order to compare with our algorithms Round Robin Round Robin is one of the simplest resources scheduling algorithms. It is commonly applied in operating systems and computer networks. In the beginning, terminals are ordered randomly in a queue. The new terminals are inserted at the end of the queue. The first terminal of this queue is assigned all the available resources by scheduler, and then put it at the rear of the queue. The rest of steps

31 3.2 Scheduling Algorithms for Non-Real-Time Flows 19 Timer is out. Next time slot for scheduling Calculate priority based on priority function Yes All resources are assigned All users requirements are satisfied No Assign resource block according to _ priority and update all users R k Figure 3.1. Proportional Fair follow the same way, until no terminal applies for resources. On one hand, it is a fair scheduling, since every terminal is given the same amount of resources. On the other hand, it neglects the fact that terminals in bad channel conditions need more resources to carry out the same rate, so it is not absolutely fair. However, it is much fairer than Max Rate. Although Round Robin gives every terminal equal chance to obtain resources, the overall throughput is much lower than Max Rate, because scheduler does not consider channel conditions. In LTE, different terminals have different service with different QoS requirements. It is impossible to allow every terminal to take up the same resources in the same possibility, because it will decrease efficiency of resources. Nevertheless, Round Robin is a good reference to measure the fairness of scheduling for LTE, as we did in Chapter Proportional Fair Proportional Fair is a compromise between Maximum Rate and Round Robin [27]. It pursues the maximum rate, and meanwhile assure that none of terminals is starving. The terminals are ranked according to the priority function. Then scheduler assigns resources to terminal with highest priority. Repeat the last two steps until all the resources are used up or all the resources requirements of terminals are satisfied. It is shown in Figure 3.1. The priority function is following i = arg max k R k,n (m) R k (m) R k,n (m) is the estimation of supported data rate of terminal k for the resource block n. R k (m) is the average data rate of terminal k over a windows in the past.

32 20 Scheduling Algorithms T P F is the windows size of average throughput and can be adjusted to maintain fairness. Normally T P F should be limited in a reasonable range so that terminals cannot notice the quality variation of the channels. { (1 1 T R k (m + 1) = P F )R k (m) + 1 T P F R k,n (m), if user k is selected (1 1 T P F )R k (m), if user k is not selected In LTE, two strategies are considered for R k update. The one is that R k is updated after all resource blocks are allocated. The other is that R k is updated after each resource block allocation. Because of its better performance [26], this thesis only introduces the second one, which also will be used in Chapter 6 as a reference algorithm. 3.3 Scheduling Algorithms for Real-Time Flows Real-Time flows have more strict delay restraint than Non-Real-Time flows resulting in the reduction of influence of error correction. In wireless communication, Real-Time flows can be modeled as arrival process of independent packets to respective queues, under a delay target constraint. For different wireless network models, there are several stabilizing policies to satisfy different delay target constraints [22]. For example, the stabilizing policy of system whose delay target is the maximum delay deadline is to make sure that the length of queues is in a certain bound. The most popular scheduling algorithms for Real-Time flows are Largest Delay First, Modified Largest Weighted Delay First. Delay experience, packet loss rate and fairness are main performance metrics for Real-Time flows Largest Delay First The first package received by base station will be sent out in the first place. i = arg max W k (t) k W k (t) is the time that package of terminal k has spent at the base station waiting for scheduling [24]. Similar to Round Robin, neglecting channel conditions leads to poor throughput. Even so, it is still a good reference for delay experience evaluation Modified Largest Weighted Delay First MLWDF attempts to balance the weighted delays of packets, while tries to make use of channel resource efficiently. In the mathematic expression, the delay possibility should satisfy P r{w k > τ k } ρ k and the requirement of throughput is R k > r k. The definition of W k (t) is the same as Largest Delay First and τ k is the threshold of delay for terminal k. ρ k is the maximum allowable possibility of exceeding τ k. R k is average throughput of terminal k. r k is a predefined minimum

33 3.4 Scheduling Algorithms for Mixture of Real-Time and Non-Real- Time Flows 21 throughput for terminal k. In each time slot t, terminal i is selected according to following priority function. i = arg max γ k R k (t)w k (t) k Where delay factor γ k = a k is an arbitrary constant. a k is characterizing the R k (t) desired QOS for terminal k [8]. Although this thesis has not had enough time to extend our algorithms for Real-Time applications, MLWDF are studied here to prepare for the future work in Chapter Scheduling Algorithms for Mixture of Real- Time and Non-Real-Time Flows In realistic network, it supports both RT and NRT flows at the same time. Much work [8][9][14][24] has been done so as to seek better algorithms to maximize throughput by making use of resource efficiently and keep delay as low as possible by identifying urgency flows at same time, taking fairness into account. The most popular scheduling algorithms for mixture of Real-Time and Non- Real-Time flows are Exponential Rule, Utility function. The key performance is measured by throughput, delay experience and fairness Exponential Rule The Exponential Rule is a modified version of Proportional Fair [24]. where aw = 1 N i = arg max γ k R k (t)exp( a kw k (t) aw k 1 + aw ) k a kw k (t) and γ k = a k R k (t). R k(t) is the estimation of supported data rate of terminal k in time slot t. R k (t) is the average data rate of terminal k over period T P F. a k is priority coefficient, which reflects the desired QOS of terminal k. W k (t) is the longest time that terminal k s packets are spent on base station. If a k W k (t) of terminal k with high priority a k or large delay W k (t) is greater than average value of all terminals up to aw, the variation of exponential term dominates. Thus this terminal has more opportunity to be scheduled, prior to other terminals. In other words, scheduler is to adjust resources to take care terminals with larger delay or higher priority. On the contrary, if a k W k (t) of terminal k is much smaller than aw, the exponential term is closer to 1. Thus above formula becomes Proportional Fair Utility Function Utility theory is introduced to wireless communication to evaluate the degree of network satisfaction for different applications associated with users in possession

34 22 Scheduling Algorithms of different QoS and balance throughput of NRT flows and delay of RT flows. The resource consumptions such as power, bandwidth and key performance such as throughput, delay are mapped into utility function. The optimal or suboptimal resource allocation can be obtained by maximizing utility function with amount of physical resources and QoS requirement as constraints [14]. Literature[25] presents a utility based algorithm following below utility function to determine resource allocation so that maximally overall system performance are achieved and the performance of both Real-Time and Non-Real-Time individual applications are closer to QoS requirement. U(t) = R(t) i A(t) i F (t) i where R(t) i is radio physical resource function and F (t) i is fairness function. More detail can be referred from [14]. A(t) i is QoS requirement deviation function to evaluate how much correction i deviates from its QoS requirements. It guarantees the time delay for Real-Time users and the transmission data rate for Non-Real- Time users according to QoS requirements.

35 Chapter 4 Scheduling Under a Control Signaling Cost Constraint 4.1 Background Chapter 3 brings in several classic scheduling algorithms that make fully use of time, frequency and space diversity to achieve the best performance. These algorithms are widely used in narrowband systems, but they cannot be applied to wideband systems directly. One of the reasons is that these algorithms ignore the signaling overhead caused by periodic scheduling map transmissions. In real applications, terminals must exactly know which carrier conveys their data in certain particular time slot for decoding. This information is packaged into scheduling map and transmitted to terminals via Physical Downlink Control Channel of LTE. As discussed in Section 1.2, the control signaling overhead associated with the scheduling maps can be significant, if the number of terminals or carriers is large. Much prior work has been done in order to mitigate the influence of signaling overhead on the overall system performance. It can be classified into three approaches. The first approach is to reduce the amount of control signaling by merging contiguous resources into large blocks. As Section said, 3GPP Long Term Evolution is implementing this technique to reduce the size of scheduling map [1], thus smaller signaling overhead is achieved. However, it causes that single resource block cannot be scheduled even for small payload. To avoid the influence of reduction of control signaling overhead on scheduling, the second approach is brought in. The reduction of control signaling overhead is carried out by making use of correlation and source coding. Reference [28] uses data compression technique to encode the control signaling. The compression scheme composes of a run length coding, followed by a universal variable length code. Reference [18] utilizes correlation of scheduling assignments for different users. The main idea is allowing a user with high SNR to overhear the scheduling information sent to other users with weaker SNR. Thus this user can deduce which 23

36 24 Scheduling Under a Control Signaling Cost Constraint resource this user will not be scheduled based on the fact that no two users can share the same resource at the same time. As a result, the cost of control signaling transmitted to that user with high SNR can be avoided. The correlation between scheduling decision and CSI report is utilized in reference [21]. In the proposal of [21], mobiles make tentative scheduling decision and send to base station, and then base station sends "agreement maps" to mobiles. As a result of correlation between tentative scheduling decision and agreement map, the signaling overhead can be decreased with the help of source coding scheme. For the last two approaches, the overall performance is increased after signal overhead local optimization. To pursue global optimization, the third approach is exploited. In addition to encoding control signaling with compression scheme, signaling overhead as well as throughput is taken into account by scheduler. It is modeled as nonlinear integer programming problem in reference [15]. Objective function is maximizing net throughput instead of gross throughput. There are two constraints. The former one is the maximum number of subcarriers each terminal is allowed to have. The latter one is an objective fact that each subcarrier can be assigned to a terminal at most. The optimal solution for above model results in huge computation. In order to simplify the computation, an approximation model is also proposed in [15], but the objective function value is not net throughput anymore. Even so, it is still hard to compute. Reference [12] formulates it as a combination problem, which is resolved by dynamic programming approach Viterbi algorithm. The algorithm proposed by [12] is much flexible and closer to practical applications, compared with [15]. In [12], the number of subcarriers each terminal can have is not fixed and the set of terminals selected for transmission are depending on CSI and control signaling overhead. The most important advantage is that the optimal scheduling can be computed efficiently. 4.2 Proposed Scheduling Algorithm Under a Control Signaling Cost Constraint for LTE/LTE Advanced Notation A subframe is a set of consecutively transmitted OFDM symbols, i.e. 7 2 = 14 symbols. Referring to the introduction part of Chapter 3, the smallest resource unit allowed to be assigned to one terminal is the schedule block, which is constituted by two successive PRBs. For convenience, this thesis still use resource block instead of schedule block as a unit for scheduling so that each resource block consists of 12 carriers for one subframe in duration. N is the amount of resource blocks in one subframe. K is the largest number of users that are allowed to be scheduled in a subframe. U represents the set of users that will be scheduled in the subframe. number of users in set U. u is the

37 4.2 Proposed Scheduling Algorithm Under a Control Signaling Cost Constraint for LTE/LTE Advanced 25 S stands for resource block assignment in one subframe based on U. S n denotes the index of user assigned to resource block n. C n (k) is the gross capacity of resource block n, if this resource block is assigned to terminal k. ρ represents the set of control signaling coefficient ρ i. ρ i is control signaling coefficient for user i (i [1, K]). The value of ρ i depends on service type. ρ max is the maximum value of control signaling coefficient of scheduled users set U. A is control signaling cost penalty coefficient. The intention of introducing A is to evaluate the influence of changing overall control signaling cost on throughput. External water-filling represents the algorithm presented in Section 4.3.2, used to find the optimal power assignment P for a given optimal resouce block assignment S. Internal water-filling represents the algorithm introduced in Section 4.3.3, used for obtaining the optimal resource block assignment S and optimal power allocation P for a given user set U Individual Transmission of Scheduling Maps{M k } After the scheduling, transmitter must send the resource block assignment S to terminals. Resource block assignment S can be described as a scheduling map. This thesis will present two ways to implement delivery of scheduling map on the basis of [12]. The first way is individual transmission of scheduling maps, which transmits scheduling map M k to each user k. If scheduler assigns resource block n to user k (S n = k), the n th entry of the binary map M k is "1". Otherwise the n th entry of the binary map M k is "0". It is also the way carried out by LTE for delivery of scheduling map. In the following example, 23 resource blocks are allocated to 1, 2, 3, 4, 5 users with individual transmission of scheduling maps. The scheduler decided resource blocks allocation like The first 2 bits are 1. It indicates the first two resource blocks are assigned to user 1. User 1 will receive scheduling information 110.., so he knows the first and second resource blocks are assigned to him, but third resource block is not assigned to him User User User User User5

38 26 Scheduling Under a Control Signaling Cost Constraint a) log 2 (N) bits are assigned to each user and used for expressing the first interval length. Since it is a constant value, add it to R. The initial value of n is equal to 0. b) n=n+1. c) Assign bits for the next interval length to each user, until n is equal to N. f k (S n, S n 1, n) = { 0, if S n = S n 1 P n, if S n S n 1 (S n = k S n 1 = k) When S n =S n 1, n th resource block like (n 1) th resource block is still assigned to the user S n 1. There is no switching in M k (k U). When S n S n 1, it means (n 1) th resource block is assigned to user S n 1 and n th resource block is assigned to user S n. Switch only happens in M Sn and M Sn 1. So P n bits are needful for user S n 1 to describe the possible maximum length of the next active interval. It is the same as user S n. ρ k f k (S n, S n 1, n) = f indiv (S n, S n 1, n)) (4.1) k u f indiv (S n, S n 1, n) = User User User User User5 { 0, if S n = S n 1 (ρ sn + ρ sn 1 )P n, if S n S n 1 (S n = k S n 1 = k) So the control signaling cost of individual user k can be described as SIZE Mk = R k + N f k (S n, S n 1, n) n=2 If consider a control signaling cost coefficient ρ k for user k, then above expression is SIZE Mk = ρ k (R k + N f k (S n, S n 1, n)) n=2 Where P n = log 2 (N n) and R k = N fec log 2 (N n) The maximum net capacity function of individual scheduling map transmission mode is

39 4.2 Proposed Scheduling Algorithm Under a Control Signaling Cost Constraint for LTE/LTE Advanced 27 max S C indiv (S) (4.2) Where N N C indiv (S) = ρ k (R k + f k (S n, S n 1, n)) (4.3) C n (S n ) n=1 k u N = n=2 n=1 C n (S n ) ( k uρ k R k + k u N ρ k f k (S n, S n 1, n)) (4.4) n=2 Since all R k (k u) are equal, it can be replaced by R. C indiv (S) is simplified with function (4.1). The second term of N C indiv (S) = C n (S n ) (R k + n=1 k uρ N ρ k f k (S n, S n 1, n)) (4.5) k u n=2 N N = ρ k + f indiv (S n, S n 1, n)) (4.6) n=1 C n (S n ) (R k u n=2 Insert a control signaling cost penalty A for all the users. C indiv (S) = N n=1 C n (S n ) (AR k u ρ k + A N f indiv (S n, S n 1, n)) (4.7) n= Broadcasting of Joint Scheduling Map M The entire resource block assignment S can also be represented with one joint scheduling map. This joint scheduling map is broadcasted to all the users. Using the same example as 4.2.2, 23 resource blocks are allocated to 1, 2, 3, 4, 5 users with broadcasting of joint scheduling map. The scheduler decided resource block allocation like After run length encoding, runlengths of above example is [2, 1, 1, 1, 1, 1, 4, 4, 1, 2, 2, 1, 2] and indices of the users that are assigned to the corresponding resource blocks is [1, 3, 2, 1, 5, 3, 2, 3, 4, 2, 1, 2, 1]. a) Assign log 2 (N) bits to each user for expressing the first interval length. Since it is a constant value, add it to R. The initial value of n is equal to 0. R = u L id + N fec + log 2 (K) + log 2 (N) b) If n is not equal to N, increase n by 1. If yes, bits allocation is end.

40 28 Scheduling Under a Control Signaling Cost Constraint c) If S n = S n 1, go back b). Otherwise, implement d) and e). d) Assign P n bits to represent the possible maximum length of the next active interval when switching between two users in the map occurs. P n = log 2 (N n) e) Assign Q bits to identify which user is being assigned to the resource blocks. Q = log 2 u f) Repeat from b). The control signaling cost of all the users can be concluded as N SIZE M = R + f bro (S n, S n 1, n) n=2 where f bro (S n, S n 1, n) = { 0, if S n = S n 1 P n + Q, if S n S n 1 Consider that the same network offers different service for different users, different service has different QoS requirement. It is more realistic to assign a specific ρ k to a particular user k based on his service type. In broadcast mode, one scheduling map may include more than one user. The maximum ρ among users U in scheduling map is selected as a control signaling coefficient to measure how much it costs to transmit one bit of control information in this map. The maximum net capacity function of broadcast scheduling map transmission mode is max S C bro (S) (4.8) Where N N C bro (S) = C n (S n ) ρ max (R + f bro (S n, S n 1, n)) (4.9) n=1 n=2 Insert control signaling cost penalty A N N C bro (S) = C n (S n ) Aρ max (R + f bro (S n, S n 1, n)) (4.10) n=1 n=2

41 4.2 Proposed Scheduling Algorithm Under a Control Signaling Cost Constraint for LTE/LTE Advanced 29 Figure 4.1. Illustration of the trellis for the algorithm in Section 4.2.4, for U = 3. The figure only shows n th instance, using the notation for broadcast scheduling map transmission mode. [12] Optimal Assignment S Given the User Set U Dynamic programming approach Viterbi algorithm [12] can be used to solve (4.2) and (4.8) for fixed U. a) Construct a trellis like Figure 4.1 with u 2 states and N time slots. Each time slot n(n [1, N]) represents one resource block n. Each state stands for the pair {S n, S n 1 } at time slot n. b) u candidate assignments exist for each state at resource block n. Only the best of these candidates for each state will be recorded into back tracing table. The metric is measured with the function C n (S n ) Af indiv (S n, S n 1, n) for individual scheduling map transmission and C n (S n ) Aρ max f bro (S n, S n 1, n) for broadcast scheduling map transmission respectively. c) Move current state to the next state at resource block n and repeat step a to b until it goes through all u 2 states. d) Move resource block n to n + 1 and repeat step a to c until n = N. e) The optimal assignment S is found from back tracing table Selecting the Set of Active Users U a) Find the first user who can achieve the maximum net capability, if all the resource blocks are assigned to this user. ˆk = arg max k {1,2,...K} C ({k})

42 30 Scheduling Under a Control Signaling Cost Constraint } Set m = 1 and U (1) = {ˆk b) Calculate possible overall capacity of u + 1 users, if one particular user in the rest user list is added into U. Find the next user who would contribute the most to the overall capacity. ˆk = arg max C (U (m) k) k {1,2,...K},k / U (m) c) Check if C (U (m) ˆk) > C (U (m) ) If it is true, then user ˆk will be added into U and m is increased by 1. Repeat from step b). Otherwise, return(u (m), S (U (m) ), C (U (m) )) [12]. 4.3 Power Optimization Let us assume an extreme situation. There is a resource block that causes extremely low data rate for any user. According to the strategy of Section 4.2, that resource block must be allocated to one user in U. Is it possible to ignore that resource block? The answer is yes. We can save the power of certain resource blocks, which yield low data rate due to some reason like distortion, and allocate these power to other resource block which can bring higher data rate. Thus the total throughput can be increased, especially at low average power Water-filling The philosophy of water-filling algorithm is to allocate more power to sub-channel with higher Signal to Noise Ratio (SNR) in order to maximize total throughput of all sub-channels, in the case of constant overall power. Water-filling is the optimal algorithm for power allocation. It can be verified with following mathematic derivative process. For OFDM transmission, a channel is divided into N c independent sub-channels. Each sub-channel is corrupted by independent white noise N 0. P n is the power allocated to n th sub-channel. h n is the channel impulse response between n th sub-channel and the terminal. The below throughput expression of OFDM can be obtained with Shannon capacity bound formula. Nc 1 n=0 log 2(1 + Pn h n N 0 2 ) bits/ofdm symbol Considering that the total power is fixed and h n is known from CSI, the maximum throughput can be achieved by choosing proper P n. This is a optimization problem. Objective function is (4.11), which subjects to (4.12). N c 1 C Nc = max log 2 (1 + P n p 0...p Nc 1 n=0 hn N 0 2 ) (4.11)

43 4.3 Power Optimization 31 N c 1 n=0 P n = N c P, P n 0, n = 0,...N c 1 (4.12) Since objective function is a jointly concave function. It can be solved with Lagrangian method. L(λ, P 0,..., P Nc 1) = Where λ is Lagrange multiplier. N c 1 n=0 log 2 (1 + P n Pn = ( 1 λ N + 0 hn 2 ) hn N 0 2 N c 1 ) λ P n (4.13) n=0 (4.14) P n is the optimal power allocation of (4.11). If P n < 0, P n is assigned with 0, since power is impossible to be negative External Water-filling The basic idea of external water-filling is to implement water-filling after the optimal assignment S of the first proposal in Section 4.2. Thus external water-filling has nothing to do with resource block allocation. However, it still can improve the gross capacity with very low computation overhead, especially for large control signaling cost penalty A or low average power [10] [29]. a) Base station uses CSI to calculate sub-channel gains h k,n (k U, n [1, N]) which is used for resource block allocation and power allocation. b) Select the optimum set of active users U with Section c) Find the optimal assignment S in view of Section d) Optimize power allocation for S. Because resource block assignment S is known, sub-channel gain of each resource block can be deduced from h k,n (k U, n [1, N]). Shannon s capacity formulate tells us that the throughput is proportional to the power. In case of a fixed overall power, one way for achieving maximum throughput is to assign more power to the sub-channel with high h. C = Blog 2 (1 + SNR) = Blog 2 (1 + P h 2 )(Noise = 1) There are several classic algorithms of capability to carry out this function. The well known water-filling algorithm is chosen, because of its low computation cost. Water-filling algorithm makes use of Lagrange multiplier to calculate optimal power allocation P. e) Return optimal assignment S and optimal power assignment P.

44 32 Scheduling Under a Control Signaling Cost Constraint Internal Water-filling The key factors of the scheduling in the first proposal are the data rate of each user achieved in a particular resource block and control signaling cost caused by assigning this particular resource block to a certain user. This is a two dimensions mathematic model. Internal water-filling is the upgraded version of the first proposal. It extends mathematic model from two dimensions to three dimensions. Power control is the third dimension, besides channel condition and control signaling cost. Complexity mathematic model leads to huge computation cost. Even so, it is still suboptimal solution due to mix integer programming problem. a) Construct a trellis with u 2 states and N time slots. Each time slot n (n [1, N]) is corresponded to one resource block n. Each state represents the pair {S n, S n 1 } at time slot n. b) At the resource block n(n (1, N]), each state has u candidate assignments. The best one is picked with the below metric functions and recorded into back tracing table. C indiv (S) = N N ρ k f k (S n, S n 1, n)) n=2 C n (S n ) (AR k + n=1 k uρ A k u N N = ρ k + A f indiv (S n, S n 1, n)) C n (S n ) AR n=1 k u N N = C n (S n ) A n=2 C n (S n ) (AR n=1 k u N N = ρ k A f indiv (S n, S n 1, n) n=2 n=2 n=2 f indiv (S n, S n 1, n) + C n (S 1 ) AR k u ρ k C bro (S) = = = N N C n (S n ) Aρ max (R + f bro (S n, S n 1, n)) n=1 n=2 N C n (S n ) Aρ max R Aρ max n=1 N C n (S n ) Aρ max N n=2 n=2 The metric associated with each branch is N n=2 f bro (S n, S n 1, n)) f bro (S n, S n 1, n)) Aρ max R + C n (S 1 ) (C n (S n ) Af indiv (S n, S n 1, n)) + C n (S 1 ) n n=2

45 4.4 Simulation Results 33 for individual scheduling map transmission mode and (C n (S n ) Aρ max f bro (S n, S n 1, n)) + C n (S 1 ) n n=2 for broadcast scheduling map transmission mode. c) Find resource block assignment S i (i [1, n ]).S n is the user occupying the current resource block. S i (i [1, n )). S n is obtained with back tracing table. d) Water-filling is implemented to optimize power allocation for S i (i [1, n ]).S n. After power optimization, the gross capacity is obtained. The control signaling cost is calculated with f(s n, S n 1, n) and S i (i [1, n ]). Gross capacity minus control signaling cost is net capacity. Fill net capacity into Viterbi table. If n is the last resource block (n = N), power allocation scheme is recorded in the power table. e) Move current state to the next state at resource block n and repeat step a) to d) until current state is equal to u 2. f) Move resource block n to n + 1 and repeat step a) to e) until n = N. g) Deduct extra bits for map initialization from net capacity. Extra bits are AR k uρ k for individual scheduling map transmission mode and Aρ max R for broadcast scheduling map transmission mode. 4.4 Simulation Results Individual and Broadcast Scheduling Map Transmission The algorithms of individual scheduling map transmission and broadcast scheduling map transmission are designed for LTE OFDMA system. Figure 4.2 illustrates the operation of resource block assignment for a LTE OFDMA system with 18 MHz bandwidth using the proposed algorithm of individual scheduling map transmission. The bandwidth per resource block is 180kHz and each frame has ten OFDM symbols. The independent and identically distributed (i.i.d.) Rayleigh fading channel is chosen for this simulation. When A is near zero, scheduler selects the users with slightly better channel conditions due to negligible control signaling cost. It leads to high gross capacity and high control signaling cost at the same time, as shown in Figure 4.2. As A increasing, control signaling cost becomes the primary factor for scheduler. It indicates that the scheduler prefers to assign resource block to the owner of the last resource block in order to avoid the cost of control signaling. That is why the control signaling cost is decreasing instead

46 34 Scheduling Under a Control Signaling Cost Constraint Figure 4.2. Illustration of average throughput in individual scheduling map transmission mode. Figure 4.3. Illustration of average throughput in individual scheduling map transmission mode and broadcast scheduling map transmission mode respectively.

47 4.4 Simulation Results 35 of increasing during the rising of A. Even so, the net capacity is still reducing, because of the sharply decreasing of gross capacity. The comparison of individual scheduling map transmission and broadcast scheduling map transmission is plotted in Figure 4.3. Both modes are simulated in the i.i.d. Rayleigh fading channel condition with the same A and ρ. The bandwidth per resource block is 180kHz and every frame has ten OFDM symbols. Solid line represents individual scheduling map transmission mode. Dashed line represents broadcast scheduling map transmission mode. In the region where A is less than 1, there is not doubt that the control signaling cost of individual scheduling map transmission is larger than that of broadcast scheduling map transmission in the light of the control signaling cost function of Section and Once A is larger than 1.5, Figure 4.3 indicates that the control signaling cost of broadcast scheduling map transmission is beyond that of individual scheduling map transmission. Since scheduling map becomes less and less fragmented and thereby control signaling cost of individual scheduling map transmission is smaller than signaling overhead of broadcast scheduling map transmission. Eventually, only one single user is scheduled per frame and the gap of control signaling cost between individual scheduling map transmission and broadcast scheduling map transmission mode is constant. N fec =24, N=100, k=10, u=1 SC indiv = A N n=2 f indiv (S n, S n 1, n) + AR k u = AR k uρ k ρ k = (N fec log 2 (N) )A k uρ k = 32A k uρ k N SC bro = Aρ max (R + f bro (S n, S n 1, n)) n=2 = Aρ max R = Aρ max ( u L i d + N fec + log 2 (K) ) + log 2 (N) ) = 51Aρ max ρ i (i [1, k])is a random value between 0 and 1. The mean value of ρ i (i [1, k]) K is 0.5 and the mean value of ρ max is K+1, which is approximate to 1. Introduce above mean value to SC indiv and SC bro SC indiv =16

48 36 Scheduling Under a Control Signaling Cost Constraint Figure 4.4. Illustration of average throughput in individual scheduling map transmission mode and broadcast scheduling map transmission mode based on individual subcarriers. Solid line represents individual scheduling map transmission mode. Dashed line represents broadcast scheduling map transmission mode. SC bro =51 When only one signal user is scheduled per frame, the control signaling cost of individual scheduling map transmission and broadcast scheduling map transmission are 16 and 51 respectively. These derivations can be proven by Figure 4.3 The similar configuration as Figure 4.3 is implemented in Figure 4.4. The only difference is that scheduling is based on individual subcarriers, instead of resource block, which consists of 12 subcarriers. It is natural that the control signaling cost is much larger. Comparing with 4.3, the net capacity of both modes fall down sharply, when A is small. That is because of large control signaling overhead. The simulation environment of Figure 4.5 likes Figure 4.3, except the parameters of Rayleigh fading channel. The 3GPP Extended Vehicular A power delay profile is set for the Rayleigh fading channels of the simulation. Figure 4.5 points out that its gross capacity and net capacity outperform those of Figure 4.3. It profits from the correlation between channel and frequency. The correlation makes the users have an opportunity to have high channel gains at numerous frequencies External Water-filling The reason why we present plots of gross capacity instead of net capacity is that external water-filling does not bring any additional control signaling cost so that increased gross capacity can be consider as increased net capacity. From Figure

49 4.4 Simulation Results 37 Figure 4.5. Illustration of average throughput of individual scheduling map transmission mode and broadcast scheduling map mode in the Rayleigh fading channel following the 3GPP Extended Vehicular A power delay profile. Figure 4.6. Illustration of increased gross capacity ratio of external water-filling in individual scheduling map transmission mode, referring to gross capacity of equal power assignment. The bandwidth per resource block is 180kHz and every frame has ten OFDM symbols. The channel condition is Rayleigh fading channel.

50 38 Scheduling Under a Control Signaling Cost Constraint 4.6, we can see that the gross capacity of individual scheduling map transmission is improved apparently by using external water-filling algorithm. The plots of Figure 4.6 is obtained by the below function. G Increased [%] = G2 G1 G1 where G1 is the gross capacity of the first proposal without power optimization, G2 is the gross capacity of external water-filling. It is obvious that as the average power is increasing, the increased gross capacity is decreasing. Finally, the increased gross capacity is close to 0 due to the fact that the individual power is increasing corresponding to the rising of the overall average power. Thus power optimization has less and less impact on gross capacity. On the contrary, the external water-filling generates large gain for the case of low average power. Furthermore, increased gross capacity relies on the control signaling cost, since control signaling cost affects scheduling. The control signaling cost is relating to control signaling cost coefficient ρ and control signaling cost penalty coefficient A. With the growing of ρ and A, the control signaling cost due to switching between users may be larger than increased data rate caused by switching users. In this situation, scheduler prefers to select the same user as the last resource block rather than switching to another user with the higher data rate. It changes resource block assignment S, which results in variation in sub-channel gain distribution of S. The intention of external water-filling is to make use of this variation to achieve higher gross capacity. When the most sub-channels of S are located in lower gain in distribution, external water-filling can have excellent performance. The sub-channel gain distribution will be discussed in the Figure 4.7. Figure 4.6 concludes that external water-filling can contribute more extra gain by optimizing power distribution. It can be explained with Figure 4.7. For A=0, control signaling cost is for free. Scheduler chooses the user who can achieve maximum data rate in that resource block. Data rate is determined by C = Blog 2 (1 + SNR) = Blog 2 (1 + P h 2 ) where B, P is fixed for each user, so the user with the higher sub-channel gain h 2 in that resource block will be chosen. It is verified by Figure 4.7. When A=0, the most sub-channel gains of S are distributed in the high level. It leads to a very slight improvement of gross capacity for A=0 in Figure 4.6. For A =12, control signaling penalty is significant. Scheduler selects users to obtain maximum net capacity, which is gross capacity minus control signaling cost. The most sub-channel gains of S are distributed in the low level in order to avoid switching in view of control signaling cost. Hence, there is a large space to improve gross capacity by optimizing power allocation. However, Figure 4.6 shows that it is merely a slight raise for increased data rate with the increasing of A from 3 to 12. It can be explained easily with Figure 4.7. Sub-channel gain distribution is almost not changing, when A reaches the threshold value that the first factor considered by scheduler is switched from data rate of individual resource block to control signaling cost. The increased gross capacity of external water-filling and sub-channel gain

51 4.4 Simulation Results 39 Figure 4.7. Illustration of sub-channel gain distribution of optimal assignment S in individual scheduling map transmission mode. Figure 4.7 is simulated in the same configuration as Figure 4.6, but sub-channel gains are measured before optimizing power. Figure 4.8. Illustration of increased gross capacity of external water-filling in broadcast scheduling map transmission mode.

52 40 Scheduling Under a Control Signaling Cost Constraint Figure 4.9. Illustration of sub-channel gain distribution of optimal assignment S in broadcast scheduling map transmission mode. distribution of optimal assignment S in broadcast scheduling map transmission mode are illustrated in 4.8 and 4.9. The both plots tell the similar story as Figure 4.6 and Figure 4.7, so the analysis will not be repeated again Internal Water-filling In Figure 4.12, the solid line stands for the net capacity of external water-filling and the dashed line represents the net capacity of internal water-filling. It shows that net capacity of internal water-filling is higher than that of external waterfilling. It is because power control becomes the third dimension of internal waterfilling, besides channel condition and control signaling cost, to decide resource block assignment S. In other words, power control is involved in the process to find the resource block assignment S for a given user set U. More specifically, when A=1.2, the sub-channel gain distribution of S i (i [1, n ]) makes power control have enough gains to vote the selection of S i (i [n, N]). It causes that sometimes internal water-filling picks S i with low gross capacity to avoids some expensive control signaling cost penalties in order to attain larger net capacity. On the contrary, external water-filling has nothing to do with assignment S. It is verified by Figure 4.10 and Figure Figure 4.10 points out that the gross capacity of external water-filling is higher than that of internal water-filling, as A=1.2. For some specific average power values between 0 and 5, the gross capacity of internal water-filling is even worse than that without power optimization. Figure 4.11 depicts that the control signaling cost of internal water-filling is much lower than that of external water-filling.

53 4.4 Simulation Results 41 Figure Illustration of increased gross capacity of internal water-filling vs external water-filling in individual scheduling map transmission mode with A =1.2. Figure Illustration of increased control signaling cost of internal water-filling vs external water-filling in individual scheduling map transmission mode with A =1.2.

54 42 Scheduling Under a Control Signaling Cost Constraint Figure Illustration of increased net capacity of internal water-filling vs external water-filling in individual scheduling map transmission mode with A =1.2. Figure Illustration of increased net capacity of internal water-filling vs external water-filling in individual scheduling map transmission mode.

55 4.4 Simulation Results 43 The net capacities of individual scheduling map transmission using internal water-filling and external water-filling are plotted in Figure Both methods are simulated with A =0, 0.5, 1.2, 5. The net capacities of internal water-filling are represented with dashed lines and the net capacity of external water-filling are expressed with solid lines. Figure 4.13 presents that internal water-filling and external water-filling can provide higher gains in the region of lower average power. As average power is increasing, the net capacities of both methods are closing to the net capacity without power optimization. It implies that the influence of power control of internal water-filling and external water-filling becomes less and less, because individual user already can receive strong enough signals. For A near 0, the control signaling cost penalty is free. In each resource block n, first scheduler picks users who can achieve maximum data rate in that resource block. Then run power optimization of internal water-filling. Due to fact that the most sub-channel gains of S i (i [1, n ]) are distributed in the high level, the power control cannot produce enough gains to change the selection of S i (i [n, N]). The resource block assignment S obtained by internal water-filling and S attained by power optimization are almost the same, so the performance of internal waterfilling is as bad as external water-filling. For A=1.2, the control signaling cost penalty is not significant. During the increasing of control signaling cost penalty A, some sub-channel gains of S are distributed in the low level in order to avoid switching. It implies that power control has ability to change the selection of S i (i [n, N]) by producing extra gains, so as to achieve higher net capacity. As A increases, the control signaling cost is so large that scheduler only selects a user set of small size, since switching user even cannot bring high data rate to counteract control signaling overhead. By observation, we find that the alteration of the size of user set is very slow, when A is larger than 3. It means that, no matter how large A is, sub-channel gain distribution of S i (i [1, n ]) is almost not change any longer. The power optimization of A=5 only can provide similar gains as A=3, but it suffers larger control signaling penalty than A=3. Consequently, power control of internal water-filling is incapable of affecting the selection of S i (i [n, N]). As a result, internal water-filling has the same effect as external water-filling at high computation expense.

56 44 Scheduling Under a Control Signaling Cost Constraint

57 Chapter 5 Adaptive Length Subframe 5.1 Rayleigh Fading Channel Rayleigh fading channel model is a suitable model for urban area where there is no line of sight propagation between transmitters and receivers. Transmitted signals suffer attenuation, reflection, refraction and diffraction, before reaching the receivers. If the number of reflections is large, the channel impulse response can be considered as Gaussian process according to the central limit theorem. Since there is no direct signal component in Rayleigh fading channel model, the process has zero mean and evenly distributed phase between 0 and 2π. Therefore, the envelope of channel response is distributed corresponding to Rayleigh distribution. There are many approaches to model and simulate Rayleigh fading channel. In this thesis, an improved sum-of-sinusoids simulation model [30] is used to simulate Rayleigh fading channel. The improved sum-of-sinusoids simulation model proposes a new way to simplify the Clark s reference model. The reference nonselective fading of Clark s reference model is expressed by g(t) = 2 N N n=1 e [j(w dtcosα n+φ n)] (5.1) where α n is the angle of incoming wave, φ n is the initial phase for n th propagation path, and w d is the maximum angular Doppler frequency. (5.1) can be transformed to (5.2) by reasonable assumptions and algebraic manipulation [30]. g(t) = { 2 M 2cos(w d tcosα n + φ n ) + j N n=1 } M 2cos(w d tsinα n + φ n ) n=1 (5.2) Except N = 4M, other parameters of (5.2) have the same meaning as those of (5.1). For simplification, the real part and imaginary part of (5.2) are expressed with Z c and Z s in (5.3)respectively. 45

58 46 Adaptive Length Subframe Figure 5.1. Possibility density function of Rayleigh fading channel with a maximum Doppler shift of 222 Hz and sampling frequency of 14 khz. g(t) = Z c (t) + jz s (t) (5.3) Z c (t) = 2 M M cos(w d tcosα n + φ n ) (5.4) n=1 Z s (t) = 2 M M cos(w d tsinα n + φ n ) (5.5) n=1 Simulations are run to prove that Rayleigh channel model selected by this thesis is correct. In Figure 5.1, the red line is Rayleigh distribution function x σ 2 e x2 f(x; σ) = 2σ 2 (σ 2 =0.5) and the black x represents approximate possibility density function of selected Rayleigh channel model. It shows that the envelop distribution of Rayleigh channel model applied to this thesis is corresponding to Rayleigh distribution. Figure 5.2 illustrates that our Rayleigh channel model is able to fulfil another important property of Rayleigh channel that the normalized autocorrelation function of a Rayleigh faded channel with motion at a constant velocity is a zeroth order Bessel function [20]. The power spectral density of terminals moving at the speed of 120 km/hour is plotted in Figure 5.3. Maximum Doppler shift obtained by calculation is 222 Hz. It shows that the power spectrum density is zero beyond the range from -222 Hz to 222 Hz, as expected.

59 5.1 Rayleigh Fading Channel 47 Figure 5.2. Autocorrelations of simulated quadrature components of Rayleigh fading channel with a maximum Doppler shift of 222 Hz and sampling frequency of 14 khz. The red reference line is a zeroth order Bessel function. Figure 5.3. Power spectral density of Rayleigh fading with a maximum Doppler shift of 222 Hz and sampling frequency of 14 khz.

60 48 Adaptive Length Subframe Figure 5.4. Subframe data structure 5.2 Proposed Scheduling Algorithm of Adaptive Length Subframe for LTE Advanced Motivation As discussed in Section 4.1, the influence of control signaling cost can be mitigated substantially by taking control signaling cost into account during scheduling. However, control signaling cost still can be squeezed further, if we look into the structure of subframe in Section Each subframe consists of control signaling region and data region. Subframe efficiency is defined to be the ratio of the length of data region to the length of subframe. For control region, its length can be dynamically adjusted according to traffic condition, but the length is fixed, after resource block assignment is done. Furthermore, the maximum three OFDM symbols per subframe are allowed to be assigned to control signaling region for control signaling transmission. As illustrated in Figure 5-4, if physical channel can support flexible subframe length, the higher subframe efficiency is achieved by extending the length of data region. It is more obvious, when efficiency is written in the mathematic expression. η subframe = L d L t = 1 L s L t (5.6) where L s is the length of control signaling region, L d is the length of data region and L t is the length of total subframe. In practical applications, we pay more attention on the performance like throughput instead of the subframe length of control signaling region or data region. The throughput carried by both regions is gross capacity and the throughput carried by data region is net capacity. So we redefine subframe efficiency as net capacity of one subframe divided by the gross capacity of this subframe, expressed with the following formula.