Allocation (SORA) whose main goal is to guarantee that most of the users achieve the minimum ... SORA is composed of two parts: Resource Allocation and.
F EDERAL U NIVERSITY OF C EAR A´ D EPARTMENT OF T ELEINFORMATICS E NGINEERING
Satisfaction Oriented Resource Allocation for Wireless OFDMA Systems
F RANCISCO R AFAEL M ARQUES L IMA
M ASTER
OF
S CIENCE T HESIS
P ROF. D R . F RANCISCO RODRIGO P ORTO C AVALCANTI A DVISOR
´ D R . WALTER C RUZ F REITAS J UNIOR C O -A DVISOR
F ORTALEZA – C EAR A´ AUGUST 2008
Universidade Federal do Cear´a Departamento de Engenharia de Teleinform´atica Programa de P´os-Graduac¸a˜ o em Engenharia de Teleinform´atica
Alocac¸a˜ o de Recursos de R´adio Orientado a` Satisfac¸a˜ o de Usu´arios em Redes OFDMA Autor Francisco Rafael Marques Lima
Orientador Prof. Dr. Francisco Rodrigo Porto Cavalcanti
Co-Orientador Dr. Walter Cruz Freitas J´unior
Dissertac¸a˜ o apresentada a` Coordenac¸a˜ o do Programa de P´os-graduac¸a˜ o em Engenharia de Teleinform´atica da Universidade Federal do Cear´a como parte dos requisitos para obtenc¸a˜ o do grau de Mestre em Engenharia ´ de Teleinform´atica - Area de concentrac¸a˜ o em Sinais e Sistemas.
F ORTALEZA – C EAR A´ AGOSTO 2008
Abstract Long Term Evolution (LTE) has adopted Orthogonal Frequency Division Multiple Access (OFDMA) as the multiple access scheme for the downlink and a pure packet-switched network by using Internet Protocol (IP) as the base protocol for transporting information. OFDMA allows the implementation of scheduling algorithms that take advantage of both time and frequency diversity and consequently utilize the system resources in a more efficient way. Moreover, an all-IP architecture brings benefits that range from reduced costs to easily integration of multimedia services. However, it also imposes challenges on Quality of Service (QoS) guarantees of applications that traditionally were provided over circuit-switched networks and now must share system resources with other services, e.g., voice. With this in mind, scheduling algorithms are of fundamental importance in order to provide a high number of users with fulfilled QoS. In this context, we propose a downlink scheduler named Satisfaction Oriented Resource Allocation (SORA) whose main goal is to guarantee that most of the users achieve the minimum QoS requirements in order to maximize the system capacity. Moreover, the SORA scheduler was developed to deal with mixed traffic scenarios where delay-sensitive and rate greedy services compete for resources in a system. SORA is composed of two parts: Resource Allocation and Resource Assignment. In the former part, SORA builds per-service prioritized lists where flows are sorted according to the current QoS conditions. Then, the most prioritized flows of each service are chosen to be scheduled. In Resource Assignment part, the flows selected in the previous part get assigned system resources in an opportunistic way so as to improve spectral efficiency. In a case study with Web and Voice over IP (VoIP) services in a modeled LTE-like system, we show that SORA outperforms reference schedulers in providing a high number of satisfied users in different mixed-traffic scenarios. The improved overall capacity is achieved by a better resource distribution among services and flow prioritization aiming at avoiding quality overprovision.
Acknowledgments My master thesis would not be finished without the special help of some people. I acknowledge the confidence of my advisor, Rodrigo Cavalcanti, in my potential since the beginning when I was only an undergraduate student. Moreover, I could not forget the guidance, support and countless discussions with all members of UFC.18 project with special thanks to Ricardo Brauner and my co-advisor, Walter Freitas. This work was partially developed at Ericsson Research Labs in Lule˚a-Sweden, where I had the opportunity to improve myself as a researcher. Among the special people that I have known there I thank Stefan W¨anstedt for the technical support and for making my adaptation in Lulea easier. I also appreciate the help of Sven-Olof Jonsson and Mats Nordberg with practical issues in my internship in Lule˚a. Being a post-graduate student and doing research in Brazil usually are not attractive activities especially because of the lack of investment. Therefore, I am very grateful to Ericsson Research and FUNCAP that financially supported this work. Finally, I would like to thank my parents, Garcia and F´atima, for all effort and sacrifice that they made in order to give me opportunity to study and finish this master thesis. They together with my sister, Jos´etima, have always believed in my capacity and understood that doing research implies in some restrictions. I also would like to express all my gratitude to my friend and wife, Cibelly, for the unlimited love and for shining my life with her smile every new day. I am also thankful for the long friendship of Daniel Barbosa and Fabio Freitas. Last but not least, I thank God for listening to me and for all the special things He has given me in my life.
Contents List of Figures
8
List of Tables
9
List of Acronyms
10
1 Introduction 1.1 Motivation . . . . . . 1.2 Objectives . . . . . . 1.3 Methodology . . . . 1.4 Related works . . . . 1.5 Scientific production 1.6 Document structure .
. . . . . .
14 14 15 16 16 17 18
. . . . . . . .
19 19 19 20 21 23 23 24 25
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
2 Radio Resource Allocation in OFDMA Systems 2.1 System description . . . . . . . . . . . . . . . . . . . . 2.1.1 Overall description . . . . . . . . . . . . . . . . 2.1.2 Orthogonal Frequency Division Multiple Access 2.1.3 Services types . . . . . . . . . . . . . . . . . . . 2.2 RRA functionalities . . . . . . . . . . . . . . . . . . . . 2.2.1 RU allocation . . . . . . . . . . . . . . . . . . . 2.2.2 Power allocation . . . . . . . . . . . . . . . . . 2.2.3 Scheduling . . . . . . . . . . . . . . . . . . . . 3 SORA Scheduler 3.1 Introduction . . . . . . . . . . . . . . . . . 3.2 General idea of SORA-NRT and SORA-RT 3.3 SORA - Non Real Time . . . . . . . . . . . 3.4 SORA - Real Time . . . . . . . . . . . . . 3.5 SORA for multi-service scenarios . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . . .
. . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
26 26 26 30 31 34
4 Application to 3GPP’s Long Term Evolution System 4.1 Simulation models and LTE description . . . . . 4.1.1 Simulator overview . . . . . . . . . . . . 4.1.2 LTE system and models . . . . . . . . . 4.2 Definitions and simulation parameters . . . . . . 4.3 Sensitivity analysis . . . . . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
37 37 37 38 44 47
5
. . . . .
. . . . .
4.4
4.3.1 Channel quality measurements 4.3.2 QoS parameters . . . . . . . . Traffic mix results . . . . . . . . . . . 4.4.1 Web only . . . . . . . . . . . 4.4.2 VoIP only . . . . . . . . . . . 4.4.3 75% VoIP 25% Web . . . . . 4.4.4 50% VoIP 50% Web . . . . . 4.4.5 25% VoIP 75% Web . . . . . 4.4.6 Capacity . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
47 49 51 51 53 56 59 63 65
5 Final Remarks 67 5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 5.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6
List of Figures 2.1 2.2 2.3 2.4
Generic multicell OFDMA system. . . . . . . . . . Frequency-time representation of an OFDM signal. Frequency-time resource grid in OFDMA. . . . . . Illustration of frequency and multiuser diversities. .
. . . .
. . . .
. . . .
. . . .
20 21 22 23
3.1 3.2 3.3 3.4 3.5 3.6 3.7
Building blocks of SORA-NRT and SORA-RT. . . . . . . . . . . . . . . . . Example of list a building in the Resource Allocation part. . . . . . . . . . . Flowchart of Resource Assignment part of SORA scheduler. . . . . . . . . . Illustration of list p building in SORA-NRT. . . . . . . . . . . . . . . . . . . Illustration of list p building in SORA-RT. . . . . . . . . . . . . . . . . . . . Illustration of SORA-RT priority within satisfied and unsatisfied flow groups. Building blocks of SORA. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
28 29 29 31 33 33 35
4.1 4.2 4.3 4.4 4.5 4.6
Main simulator network components. . . . . . . . . . . . . . . . . . . . . . . . . LTE network elements and their interfaces. . . . . . . . . . . . . . . . . . . . . . System architecture and user/control planes. . . . . . . . . . . . . . . . . . . . . . User and control planes protocol stack. . . . . . . . . . . . . . . . . . . . . . . . . Time frame and resource structure in LTE. . . . . . . . . . . . . . . . . . . . . . . User satisfaction ratio in the Web-only scenario with variable channel state reporting periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . User satisfaction ratio in the VoIP-only scenario with variable channel state reporting periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . User satisfaction ratio for different average data rate update periods in the Web-only scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . User satisfaction ratio for different FER update periods in the VoIP-only scenario. . Average number of scheduled UEs per TTI in the Web-only scenario. . . . . . . . Average cell throughput in the Web-only scenario. . . . . . . . . . . . . . . . . . . Average Web user throughput CDF at load 83 UEs per cell in the Web-only scenario. User satisfaction ratio in the Web-only scenario. . . . . . . . . . . . . . . . . . . . Average user delay of satisfied flows in the VoIP-only scenario. . . . . . . . . . . . Average number of scheduled UEs per TTI in the VoIP-only scenario. . . . . . . . VoIP FER CDF at load 289 UEs per cell in the VoIP-only scenario. . . . . . . . . . User satisfaction ratio in the VoIP-only scenario. . . . . . . . . . . . . . . . . . . . Average number of scheduled UEs per service in the mixed traffic scenario 75% VoIP and 25% Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Average number of allocated RUs per service in the mixed traffic scenario 75% VoIP and 25% Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Average cell throughput in the mixed scenario 75% VoIP and 25% Web. . . . . . .
4.7 4.8 4.9 4.10 4.11 4.12 4.13 4.14 4.15 4.16 4.17 4.18 4.19 4.20
7
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
38 39 40 41 42 47 48 50 50 51 52 53 54 55 55 56 57 58 58 59
4.21 User satisfaction ratio in the mixed traffic scenario 75% VoIP and 25% Web. 4.22 Average number of scheduled UEs per service in the mixed traffic scenario VoIP and 50% Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.23 Average number of allocated RUs per service in the mixed traffic scenario VoIP and 50% Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.24 Average cell throughput in the mixed scenario 50% VoIP and 50% Web. . . 4.25 User satisfaction ratio in the mixed traffic scenario 50% VoIP and 50% Web. 4.26 Average number of scheduled UEs per service in the mixed traffic scenario VoIP and 75% Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.27 Average number of allocated RUs per service in the mixed traffic scenario VoIP and 75% Web. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.28 Average cell throughput in the mixed scenario 25% VoIP and 75% Web. . . 4.29 User satisfaction ratio in the mixed traffic scenario 25% VoIP and 75% Web. 4.30 Capacity region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
. . . 50% . . . 50% . . . . . . . . . 25% . . . 25% . . . . . . . . . . . .
. 60 . 60 . 61 . 62 . 62 . 63 . . . .
64 64 65 66
List of Tables 4.1
Simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
9
List of Acronyms
3GPP
3rd. Generation Partnership Project
3G
Third Generation
ACK
Acknowledgement
ADSL
Asymmetric Digital Subscriber Line
AM
Acknowledged Mode
AMC
Adaptive Modulation and Coding
AMR
Adaptive Multirate
ARQ
Automatic Repeat Request
B3G
Beyond Third Generation
BER
Bit Error Rate
BS
Base Station
CAPEX
Capital Expenditure
CDF
Cumulative Distribution Function
CDMA2000 EV-DO Code-Division Multiple Access 2000 Evolution-Data only CMC
Connection Mobility Control
CN
Core Network
DS
Delay Scheduler
eNB
Enhanced Node B
E-UTRAN
Evolved UMTS Terrestrial Radio Access Network
FDD
Frequency Division Duplex
FDM
Frequency Division Multiplexing 10
FER
Frame Erasure Rate
FTP
File Transfer Protocol
GSM
Global System for Mobile communication
HARQ
Hybrid Automatic Repeat Request
HSDPA
High Speed Downlink Packet Access
HSPA
High Speed Packet Access
HTTP
Hypertext Transfer Protocol
ICIC
Inter-Cell Interference Coordination
IMT
International Mobile Telecommunications
IP
Internet Protocol
ISI
Inter Symbol Interference
L1
Physical Layer
L2
Data Link Layer
L4
Transport Layer
L7
Application Layer
LTE
Long Term Evolution
MAC
Medium Access Control
MCS
Modulation and Coding Scheme
MME
Mobility Management Entity
MR
Maximum Rate
NACK
Negative Acknowledgement
NAS
Non-Access Stratum
NRT
Non-Real Time
OFDM
Orthogonal Frequency Division Multiplexing
OFDMA
Orthogonal Frequency Division Multiple Access
OPEX
Operational Expenditure
OSI
Open Systems Interconnection
PDCP
Packet Data Convergence Protocol 11
PDCCH
Physical Downlink Control Channel
PDSCH
Physical Downlink Shared Channel
PDU
Protocol Data Unit
PF
Proportional Fair
P-GW
Packet Data Network Gateway
PHY
Physical
QoS
Quality of Service
RAC
Radio Admission Control
RAN
Radio Access Network
RBC
Radio Bearer Control
RLC
Radio Link Control
ROHC
Robust Header Compression
RR
Round Robin
RRA
Radio Resource Allocation
RRC
Radio Resource Control
RRM
Radio Resource Management
RT
Real Time
RTP
Real-Time Transport Protocol
RU
Resource Unit
SC-FDMA
Single Carrier - Frequency Division Multiple Access
SDU
Service Data Unit
S-GW
Serving Gateway
SID
Silence Insertion Description
SINR
Signal to Interference plus Noise Ratio
SORA
Satisfaction Oriented Resource Allocation
SORA-NRT
SORA - Non Real Time
SORA-RT
SORA - Real Time
TCP
Transport Control Protocol 12
TDD
Time Division Duplexing
TTI
Transmission Time Interval
UDP
User Datagram Protocol
UE
User Equipment
UM
Unacknowledged Mode
VoIP
Voice over IP
WiMAX
Worldwide Interoperability for Microwave Access
13
Chapter 1 Introduction 1.1 Motivation The cellular networks have allowed us to communicate with people who are at the most remote places in the world through mobile phone calls. Furthermore, we are used to searching for information and entertainment by utilizing fixed broadband access in our homes. With the introduction of Third Generation (3G) networks, besides phone calls the mobile phones are now capable of accessing data services such as Web browsing and e-mail. Nowadays, the mobile broadband is a reality in many countries by utilizing for example High Speed Downlink Packet Access (HSDPA) networks. However, the mobile communications are continuously evolving due to the increased demand for new multimedia services, lower costs and improved Quality of Service (QoS) provision. With this in mind, 3rd. Generation Partnership Project (3GPP) with other standardization bodies have been working in the specification of Beyond Third Generation (B3G) systems. There are several definitions for B3G systems, however, there is a unanimity when the main requirements of these networks are regarded. Among them we can mention the utilization of a completely packet-switched network based on Internet Protocol (IP), seamless connectivity and global roaming across multiple radio networks, interoperability with existing wireless standards and improved spectral efficiency in the radio networks compared with 3G ones. Long Term Evolution (LTE) has been standardized by 3GPP and it is the next major step in mobile radio communications. LTE promises to bring advantages for subscribers, with new applications such as interactive TV and user-generated videos, and for operators with backward compability with legacy networks and simpler architecture. Among the main features of LTE we can state [1]: • Downlink peak data rates over 200 Mbits/s; • Radio Access Network (RAN) round-trip times less than 10 ms; • Bandwidth flexibility ranging from less than 5 MHz to 20 MHz; • Support of both Frequency Division Duplex (FDD) and Time Division Duplexing (TDD) duplexing modes; • Reduced number of physical and logical nodes. Besides the presented LTE features, we highlight the utilization of Orthogonal Frequency
Division Multiple Access (OFDMA) as the radio access technology in the downlink and a pure packet-based All-IP architecture. OFDMA is a multiple access scheme based on Orthogonal Frequency Division Multiplexing (OFDM) digital modulation scheme where multiple User Equipments (UEs) get assigned subcarriers or subsets of them in order to receive simultaneously. This technology has also been adopted as the multiple radio access technology of other B3G systems such as Worldwide Interoperability for Microwave Access (WiMAX). One of the advantages of an OFDMA-based system is the opportunity to benefit from frequency and multiuser diversities, differently of High Speed Packet Access (HSPA) where only the last could be exploited. Frequency diversity means that it is unlikely that all frequency resources in a link have the same channel quality. Multiuser diversity occurs due to the independence of UE channels caused by distinct UE positions in a cell, therefore, frequency resources in poor channel states for some UEs possibly will be in good channel conditions for other UEs. A mechanism for taking advantage of the frequency and multiuser diversities is the employment of scheduling algorithms. Scheduling algorithms are responsible for selecting which UEs will have access to the system resources and with which configuration. In this way, scheduling algorithms have a great impact on system performance. As commented before, LTE will be based on the All-IP concept that does not only mean that the IP will be the transport protocol, but also that LTE will be included in a more general framework based on IP that provides an enhanced and integrated service set independent as far as possible to the access system used [2]. Among the advantages of an All-IP architecture we cite the efficient support to mass-market usage of any IP-based service and reduced Operational Expenditure (OPEX) and Capital Expenditure (CAPEX). However, a pure packet-switched network imposes some challenges on the provision of QoS guarantees of Real Time (RT) services that traditionally were provided over circuit-switched networks and now must share system resources with other services, e.g., voice. Once again, the design of scheduling algorithms is of utmost concern. The scheduler should deal with different flows that have different traffic patterns and QoS requirements. Therefore, we believe that scheduling algorithms play an important role in the fulfillment of the B3G requirements of higher capacity and improved QoS provision. In summary, the main motivations of this study are: • The introduction of OFDMA as the multiple access technology of B3G networks that opens opportunity to schedulers exploit some aspects that are unavailable in other technologies; • The All-IP architecture that leads to scenarios in which multiple IP-based services, such as RT ones, share the same access network; • Requirements of B3G systems such as higher system capacities.
1.2 Objectives The primary objective of this master thesis is to conceive a multicell downlink scheduler that increases the capacity of B3G networks that includes basically the following features: • OFDMA multiple access scheme; • Multicell and pure packet-switched network according to the All-IP concept; • Multiple flows of different IP services with different requirements sharing the same access network.
15
1.3 Methodology Assessing the performance of complex wireless communication systems and studying/proposing enhancements to their key functions and architectural aspects are all challenging tasks, for which the computer simulation appears as an effective and widely used analysis approach. Therefore, in this work we provide a relative performance evaluation utilizing a computational simulator. Besides that, we review some topics in the technical literature in order to support our study that include: • OFDMA aspects that impact on resource allocation and scheduling; • Services classes such as RT and Non-Real Time (NRT), their characteristics and QoS requirements; • Common and basic aspects of B3G systems; • Classical scheduling algorithms found in the related literature.
1.4 Related works Studies in scheduling algorithms for wireless networks have acquired emphasis with the introduction of packet-switched networks, such as HSDPA [3] and Code-Division Multiple Access 2000 Evolution-Data only (CDMA2000 EV-DO) [4], where the system resources are no longer dedicated to the flows but shared among them. In [5–7] we can find general requirements for wireless schedulers as well as some examples of single-carrier schedulers. With the introduction of multicarrier schemes such as OFDMA in large-scale networks, the research community started to focus in the design of new scheduling algorithms that takes advantage of a new dimension in resource allocation: The frequency. Many works have been published about this topic and a complete survey is out of scope of this master thesis. The objective is to show some guidelines and how our work innovates compared to the state of art. Some works [8, 9] have focused on natural generalization of single-carrier schedulers to the multicarrier case. In this way, classical schedulers such as Round Robin (RR), Maximum Rate (MR) and Proportional Fair (PF) were adapted to work with resource allocation in the frequency dimension. Moreover, some works have emphasized in the utilization of scheduling algorithms for QoS provision to multiple services in multicarrier networks [10–13]. In [10], the authors propose a scheduling algorithm based on PF that takes into account in its formulation the per-flow amount of data awaiting transmission at the Base Station (BS). In the scheduling process, flows belonging to a high priority service (RT service) have an explicit priority over the less prioritized ones (NRT services). The objective of the proposed scheduler is to improve the spectral efficiency while maintaining the user fairness. However, in the presented results it was not shown the robustness of the scheduler when the mix proportion is changed. In addition, the simulation scenario was restricted to single-cell system with perfect channel estimation. The authors in [11] propose a scheduling algorithm based on utility function. In this work, there are two different utility functions based on packet delay for both RT and NRT services. Besides improving the system throughput, the proposed scheduler aims at decreasing the packet drop ratio for RT. The performance of NRT service is shown in terms of fairness. As in [10], the simulation setup is too simplified not taking into account multicell interference, for example. 16
Moreover, the performance of the proposed scheduler when the service mix proportion is changed was not evaluated. Despite mixed traffic scenarios is considered in [12], the main concern is with Voice over IP (VoIP) flows. In this article, in order to guarantee the QoS of VoIP flows the scheduler controls the priority of this service based on the packet drop ratio measured at the BS. The results present a QoS improvement of VoIP flows. Nevertheless, the spectral efficiency and QoS of other services are degraded. This is also the case of the work [13] where the objective is to protect the RT services VoIP and video. In summary, the main objectives of the schedulers found in the literature are either to provide better spectral efficiency while keeping fairness among flows or protecting the QoS of high priority service classes such as RT. In this master thesis, we propose a new objective that is to maximize the number of connected flows with fulfilled QoS requirements. In this sense, the performance metrics to be improved are the user satisfaction ratio and system capacity. Moreover, we consider a scenario where the system operator intends to assure a balanced QoS provision among RT and NRT services, i.e., no explicit priority is assigned to any service type, so as to increase the overall system capacity. Finally, our work innovates compared to the majority of the studies in scheduling with mixed traffic scenarios for multicarrier systems by including in the performance evaluation study cases with different mixed traffic proportions to assess the robustness of the conceived algorithm to the traffic variability present in real networks.
1.5 Scientific production Throughout the master thesis course, we have published some articles about scheduling and resource allocation in multicarrier systems in national and international conferences. A list with these publications follows below: • Radio Resource Allocation for Maximization of User Satisfaction; Francisco Rafael M. Lima, Ricardo B. dos Santos, Francisco Rodrigo P. Cavalcanti and Walter C. Freitas; Ninth IEEE Workshop on Signal Processing Advances in Wireless Communications (SPAWC’08), Recife-Brazil, July 2008; • QoS Based Radio Resource Allocation and Scheduling with Different User Data Rate Requirements for OFDMA Systems; Ricardo B. Santos, Francisco Rafael M. Lima, Walter C. Freitas and Francisco Rodrigo P. Cavalcanti; The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07), Athens-Greece, September 2007; • Radio Resource Allocator for OFDMA Wireless Systems Based On Genetic Algorithm; Francisco Rafael M. Lima, Raimundo A. de Oliveira Neto, Ricardo B. dos Santos, Walter C. Freitas and Francisco Rodrigo P. Cavalcanti; 25th Brazilian Symposium on Telecommunications (SBrT’07), Recife-Brazil, September 2007; • Impact of Adaptive Loading in Scheduling Algorithms for OFDMA Systems; Ricardo B. dos Santos, Francisco Rafael M. Lima, Walter C. Freitas and Francisco Rodrigo P. Cavalcanti; 25th Brazilian Symposium on Telecommunications (SBrT’07), Recife-Brazil, September 2007; • Algoritmos de Alocac¸a˜ o de Recursos de R´adio em Sistemas OFDMA (Tutorial in Portuguese); Francisco Rodrigo P. Cavalcanti, Walter C. Freitas, Ricardo B. dos Santos and Francisco Rafael M. Lima; 25th Brazilian Symposium on Telecommunications (SBrT’07), 17
Recife-Brazil, September 2007; • A Resource Assignment Study on Wireless OFDMA Systems; Alex P. da Silva, Ricardo B. dos Santos, Francisco Rafael M. Lima, Francisco Rodrigo P. Cavalcanti and Walter C. Freitas; 26th Brazilian Symposium on Telecommunications (SBrT’08), Rio de Janeiro-Brazil, September 2008 (Accepted paper). A United States provisional patent of a version of the proposed scheduler was filled with the following information: • Name: A Method for Resource Allocation Driven by QoS Requirements for OFDMA Wireless Systems; • Registration date: 2007-11-05; • Registration number: PCT/BR2007/000300; • Inventors: Ricardo B. dos Santos, Francisco Rafael M. Lima, Walter C. Freitas and Francisco Rodrigo P. Cavalcanti. Finally, before studying the subject of this thesis we have done research on congestion control algorithms for HSDPA. The following work was published: • Load Control for VoIP over HSDPA in Mixed Traffic Scenarios; Emanuel B. Rodrigues, Francisco Rafael M. Lima and Francisco Rodrigo P. Cavalcanti; The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07), Athens-Greece, September 2007.
1.6 Document structure The remainder of this document is organized as follows: • Chapter 2: We present some required concepts and background information. In this chapter, we firstly describe a general OFDMA system with its characteristics and constraints. After that, we present some typical Radio Resource Allocation (RRA) functionalities in this context, including scheduling; • Chapter 3: The proposed scheduling algorithm, namely Satisfaction Oriented Resource Allocation (SORA), is presented. The first step in this chapter is to show the mathematical formulation of the problem to solve. After some analysis, we describe the proposed scheduler with its fundamental building blocks; • Chapter 4: A performance evaluation of the SORA scheduler in a case study in the LTE system is presented. Besides simulation results and its analysis, we provide a short description of the LTE system; • Chapter 5: In this chapter we summarize the main achievements of this work. Specifically, we draw some conclusions about the SORA scheduler emphasizing its practical issues and its main advantages. Moreover, we indicate some perspectives and future work based on the presented study.
18
Chapter 2 Radio Resource Allocation in OFDMA Systems 2.1 System description In this section we describe a general multicell OFDMA system and state the considered assumptions in this thesis. Firstly, a general description of the considered system is provided. Then, the OFDMA multiple access scheme is depicted with its particularities. After that, we present some key characteristics of packet-based services including a general classification.
2.1.1 Overall description In Figure 2.1 we present the considered system in which the present study is based. The system is divided into two parts: Core Network (CN) and RAN. The CN is in charge of general functions that are needed to provide services to the users. The RAN is responsible for providing access from the network subscriber to the system and comprises functions related to the optimization of the radio access. The division of the system into two parts aims at keeping the CN unaware of the radio access technology. Consequently, the CN can be utilized with any radio access technology that adopts the same functional split. The Network Controller represents physical and logical nodes that are responsible for charging, subscriber management, bearer control and mobility management. In addition, the Network Controller provides access to external networks such as internet. The CN has an entity that stores general information about user subscription and location. We consider that this CN is completely packet-switched relying on IP protocol. The RAN part comprises several BSs that are controlled by the Network Controller. Each BS can control one or more cells and is responsible for providing access from the network to the UEs. As the BS is the network nearest node to the radio interface, BS is in charge of the main RRA functionalities. We assume that a large range of mobile devices, which we call here by UE, can access the network. In our study we consider that any service session takes place between a user behind
Internet
Desktop
IP phone
BS Mobile phone
Notebook
Cell
Subscriber information Network controller
CN RAN
Figure 2.1: Generic multicell OFDMA system.
the internet utilizing, for example, an IP phone or a desktop, and another user inside the mobile network connected utilizing mobile devices such as notebooks and mobile phones. When protocol specification is concerned we consider a layered architecture based on the Open Systems Interconnection (OSI) model [14]. The Application Layer (L7) represents the traffic services provided by the network operator to the subscribers. The implementation of each protocol layer may be different according to the service type. As an example, the Transport Layer (L4) can utilize retransmission procedures for traffic services that demands reliability of data delivery while this procedure is unacceptable when services that demand short delays are provided. We also assume that Physical Layer (L1), which is implemented at BS, is capable of performing fast retransmissions jointly with soft combining procedures, e.g., Hybrid Automatic Repeat Request (HARQ).
2.1.2 Orthogonal Frequency Division Multiple Access OFDMA is a multiple access scheme based on OFDM [15]. OFDM is a transmission technology that has been utilized in wired and wireless communications. Asymmetric Digital Subscriber Line (ADSL) broadband access and power line communications are examples of applications of OFDM in wired systems. In wireless systems, the OFDM technology is utilized in IEEE 802.11 a/g and planned to be utilized in LTE and Mobile WiMAX. In OFDM, the available frequency band for transmission is divided in several subcarriers that have narrower bandwidth than the channel coherence bandwidth, as in Frequency Division Multiplexing (FDM) systems. However, the subcarriers in OFDM are designed to be orthogonal among each other, which leads to higher spectral efficiency than FDM as it is illustrated in Figure 2.2. The narrowband subcarriers also implies in simplified equalization process because 20
of the flat fading channel experienced in each subcarrier. Besides that, as the data rate transmitted in each subcarrier is low and consequently the modulated symbols are longer than the delay spreading, OFDM is robust against Inter Symbol Interference (ISI). In order to effectively mitigate the effects of ISI, a guard interval named cyclic prefix, that consists in a copy of part of the OFDM symbol, is inserted before the OFDM symbol transmission.
Figure 2.2: Frequency-time representation of an OFDM signal [16].
With OFDMA [17], the multiple access is achieved by the assignment of different subcarriers or block of them to individual UEs at different time periods. More specifically, in OFDMA the system resources can be arranged in a time-frequency grid as shown in Figure 2.3. In the frequency axis the granularity is defined by the subcarriers while in the time dimension it is defined by an OFDM symbol. We define Resource Unit (RU), which is the minimum allocable resource, as a group of one or more adjacent subcarriers in the frequency dimension and a number of consecutive OFDM symbols. The number of subcarriers and OFDM symbols in an RU depends on the system design and channel characteristics. In wireless networks, the system resources are utilized for both transmission of information (user data) and control signalling. Particularly, the higher is the number of scheduled UEs, the larger is the amount of signalling data. Among the signalling data necessary we can mention the information of which UEs were scheduled, which RUs were assigned to which UEs and transmission parameters. Therefore, we consider that there is a limitation in the number of scheduled UEs per Transmission Time Interval (TTI).
2.1.3 Services types An application or a service is defined as a task that requires communication of one or more information streams, between two or more parties that are geographically separated, being characterized by the main attributes, and also by traffic and communications characteristics [18]. The services and applications are important drivers in the evolution of wireless systems. The need to support data services with increased data rate and low latency requirements can only be satisfied with improved network architecture and optimized radio resource utilization. The evolution of wireless systems to packet-optimized CN and OFDMA-based radio access technology are consequence of this. 21
Resource unit User 1 User 2
Time (Set of OFDM symbols)
User 3
Frequency (Group of subcarriers)
Figure 2.3: Frequency-time resource grid in OFDMA.
Classifying the services and applications provided by wireless networks is not a trivial task because they are continuously evolving to integrated and complex applications. Besides that, the services can be classified in terms of time dependency (time or non-time-based), delivery requirements (RT or NRT), directionality (unidirectional or bi-directional), symmetry of the communications (symmetric or asymmetric), interactivity and number of parties [19]. Time-based services are the ones where the information should be presented at specific instants so as to have a meaning because time is an integral part of the information to be communicated, e.g., video and audio. Examples of non-time-based services are images and text. Bi-directional communication can be either symmetric or asymmetric. Web browsing is a classical example of an asymmetric application where only commands are transmitted in one link direction. Note that the classification according to the delivery requirements is different of the one regarding intrinsic time dependency. As an example, imagine an on-line game where decorrelated images are displayed to the users, and the number of points a user scores is dependent on how quickly the user reacts to the images (performs some actions). In this case, the images have to be displayed to the users with tight delay requirements so as to assure interactive response to the users (RT application). However, the images do not need any synchronization in order to make sense to the user (Non-time-based application). In this work we consider the classification according to the delivery requirements. In RT services, there is the requirement of a short time response between the communicating parts. In general, RT services impose strict requirements regarding packet delay and jitter. As examples of this kind of service, we can cite on-line games that requires quick responses from the users and VoIP. Specifically, VoIP service has been extensively studied with radio resource management so as to provide good solutions to replace the old circuit-switched voice service [20, 21]. The main challenge is to provide the same or improved QoS to VoIP compared to the 22
conventional voice service when the resources are no more dedicated but shared among many other services. In contrast, NRT services do not have tight requirements concerning packet delay although high packet delays are unacceptable. In fact, when transmitting NRT services the major constraint is the information integrity, i.e., information loss is not tolerable. Therefore, applications of this type must have error-correction or recovery mechanisms. As example of NRT services we can mention Web browsing and File Transfer Protocol (FTP) services.
2.2 RRA functionalities In order to efficiently utilize the system resources of the system described in section 2.1, RRA should be performed. In this section we will describe the RU assignment, power allocation and scheduling functionalities so as to provide general fundaments and background for the remainder of this thesis.
2.2.1 RU allocation OFDMA offers two possibilities to the RRA algorithms take advantage of. Due to the frequency selectivity of wireless channels that stems from the multipath propagation, different subcarriers of the same link experience different channel conditions. This phenomenon is denoted frequency diversity. In multicell systems this effect is even more noticeable by the irregular allocation pattern of subcarriers within different cells. Moreover, another opportunity arises from an effect called multiuser diversity [22]. Multiuser diversity consists in the fact that the perceived channel or subcarrier channel states are likely to have completely different attenuation for several UEs. The reason for this is that the fading process is statistically independent for different UEs, as long as their receive antennas are considerably separated. These diversities are illustrated in Figure 2.4 Frequency diversity
Multiuser diversity
UE 1
UE 2
Figure 2.4: Illustration of frequency and multiuser diversities.
23
The RU allocation functionality is responsible for defining which RU will be utilized in the transmission to each UE. Therefore, RU allocation has the potential to directly profit by the frequency and multiuser diversities. Due to the intrinsic combinatorial nature of the RU allocation problems, they are usually modeled in the literature as optimization problems. Another common assumption is that the RUs are not utilized at the same time by more than one UE within a cell. It has been shown that intra-cell orthogonality leads to better results than the subcarrier assignment to multiple UEs [9]. Although the power allocation, that is discussed in the section 2.2.2, could be included in the RU allocation problem formulation, we prefer to treat it separately. Hence, constant and uniform power distribution is considered in this context. Several approaches can be followed depending on the objective to be attained. In [9], the authors present a solution to the rate maximization problem that aims at maximizing the cell data rate per TTI constrained to the maximum available power in the cell. The solution to this problem is to assign the RUs to the UEs with better channel quality on them. However, the rate maximization approach leads to unfairness among users by the starvation of poor-channel-quality UEs. In this sense, the rate adaptive and margin adaptive problems were proposed so as to provide some fairness in the resource allocation. The rate adaptive problem intends to maximize the lowest data rate in the system per TTI. The objective in the margin adaptive problem is to minimize the utilized cell power constrained to the condition that all UEs achieve a required data rate per TTI. These problems are combinatorial ones whose optimum solution is not easily found. Several suboptimal solutions have been proposed to both rate adaptive [23,24] and margin adaptive [25,26] problems. Among the followed approaches we highlight the two step approach utilized in [26]. In this strategy, the RU allocation problem is split into two parts: The Resource Allocation and Resource Assignment. In the first part, the number of RUs that will be allocated to each UE is determined, while in the last one it is defined which subcarrier is associated to which UE.
2.2.2 Power allocation The power allocation is normally utilized jointly with Adaptive Modulation and Coding (AMC). As in the RU allocation problems, some problems can be formulated such as rate maximization and margin maximization [27]. In the former the objective is to provide the highest possible data rate subject to a power budget and Bit Error Rate (BER) constraints. In the latter the objective is to minimize the transmit power subject to BER and rate constraints. When the rate maximization problem is regarded, some directions to the solution can be found in the water-filling theorem from information theory [28]. Roughly speaking, this theorem states that the power should be allocated in the frequencies with better channel qualities. However, this theorem cannot be applied in a real system because it assumes that the frequency resources are infinitely divisible, i.e., the subcarrier bandwidth tends to zero. Moreover, it considers that the relationship between the allocated power and transmitted bits is continuous. However, in practice only a limited set of Modulation and Coding Scheme (MCS) are available. The solution in [29] provides an optimal solution to the problem with finite resource granularity and finite MCS without the need of a complete enumeration of all possible solutions. The performance gain of the utilization of power allocation algorithms jointly with AMC is considerable over a static scheme with fixed power and modulation scheme. However, some results points to a conclusion that the gain of power allocation schemes with AMC compared to a scheme with AMC and fixed power allocation does not pay off. The reasons are mainly the signalling 24
overhead, channel accuracy necessity and low variance of the channel state transfer function. In this way, we consider a fixed and uniformly distributed power among the RUs. A good survey on this topic can be found in [30].
2.2.3 Scheduling The previous functionalities RU and power allocation described in sections 2.2.1 and 2.2.2, respectively, do not consider the impact of the previous allocations on the system or UEs when they are performed. In other words, these functionalities are static and the objective is to solve the problem in a specific time instant. Although the scheduling functionalities performs RU allocation and possibly power allocation, it also owns the implicit feature of taking into account UE specific metrics that changes according to the previous allocations. As an example of these metrics we can cite QoS measurements such as current packet delay or average throughput. The scheduling algorithms have been utilized as an important functionality to perform QoS control among users utilizing different packet-based services. Some general requirements of a good scheduler are [7]: • Efficient link utilization: Scheduler must be opportunistic in the sense of taking advantage of multiuser diversity so as to utilize the channel efficiently; • Delay bound: The scheduler must guarantee delay bounds for individual flows in order to support delay-sensitive applications; • Fairness: A certain level of fairness should be assured in the system in order to avoid flows with quality overprovision; • Implementation complexity: A low-complexity algorithm is a necessity in high-speed networks in which scheduling decisions have to be made very rapidly; • Isolation: The algorithm should isolate a session from the ill effects of misbehaving sessions. The QoS guarantees for a session should be maintained even in the presence of sessions whose demands are in excess of their reserved values; • Delay/bandwidth decoupling: For most schedulers, the delay is tightly coupled with the reserved rate; that is, a higher reserved rate provides a lower delay. However, some high-bandwidth applications, such as Web browsing, can tolerate relatively large delays; • Scalability: The algorithm should operate efficiently as the number of sessions sharing the channel increases. Moreover, the scheduler must be flexible enough to work well in different scenarios. e.g., different traffic mix proportions. Schedulers can be classified according to the information that it utilizes to support its decision. In this way, depending on if the scheduler utilizes channel state information of the UEs it can be classified in either channel-aware or channel unaware. Another possible classification is the ability to deal with multiple services [31]. Hence, QoS-differentiated schedulers are capable of prioritize flows according to the QoS demands and service. Otherwise, the scheduler is considered non-QoS-differentiated. The SORA scheduler presented in chapter 3 belongs to the classes of channel-aware and QoS-differentiated schedulers.
25
Chapter 3 SORA Scheduler 3.1 Introduction In this chapter, we present SORA that is a multi-service downlink scheduler to be utilized in OFDMA networks. This scheduler is based on two previous schedulers developed during the master thesis course and that deal with NRT and RT services separately: SORA - Non Real Time (SORA-NRT) and SORA - Real Time (SORA-RT), respectively. Therefore, in order to provide a smooth introduction to the multi-service scheduler SORA, we firstly describe SORA-NRT and SORA-RT, and then the SORA scheduler is shown.
3.2 General idea of SORA-NRT and SORA-RT Analyzing the performance of wireless networks is not a trivial task due to their high complexity, randomness of the involved variables and different points of view that can be focused. Therefore, many performance metrics can be formulated in order to measure the efficiency of a network in achieving a given objective. One possible metric is the spectral efficiency. The maximization of spectral efficiency implies that the system resources are utilized in the best possible way by transmitting more bits per Hz/s. However, in order to attain this objective only the UEs in best channel conditions should have transmission opportunities. Therefore, another criteria has to be thought to capture fairness in some way. Roughly speaking, assuring fairness means that the system resources must be shared among the flows so as to achieve a proper balance of conflicting interests. Consequently, in some cases the best resource allocation solution in a fairness point of view can lead to resource waste by allocating too much system resources to UEs that are in poor channel conditions in order to provide the same QoS experience to all flows in the system. With this in mind, we highlight the user satisfaction ratio metric as an important one [17]. User satisfaction ratio consists in the percentage of satisfied users in the system. By maximizing the user satisfaction ratio we intend to utilize the system resources in order to provide the maximum number of flows with fulfilled QoS requirements. With this strategy we also satisfy the scheduler
requirement of isolation depicted in section 2.2.3, which states that the majority of the flows in the system should be protected from misbehaving flows whose demand for system resources are excessively high. In other words, user satisfaction ratio can quantify in which level the QoS has been guaranteed to the connected flows. In this way, the main objective of the SORA scheduler is to maximize the user satisfaction ratio. Considering that all flows belong to the same service, the problem to maximize the user satisfaction ratio is
max X[k]
X
U(j, k)
j∈Ω
subject to X xj,n [k] 6 1,
∀n ∈ N ,
(3.1)
j∈Ω
X j∈Ω
Ψ
X
n∈N
!
xj,n [k]
6 S,
where N is the RU set and Ω is the active flow set. We consider that active flows are the ones with packets awaiting transmission. One UE may bear multiple service flows. For the sake of simplicity here, only one service flow is considered per UE. However, the proposed algorithm could be easily extended to other cases. Therefore, flow and UE are interchangeable throughout the text. S is the maximum allowable number of scheduled UEs per TTI. X [k] is the assignment matrix with elements xj,n [k] defined as 1, if RU n is assigned to flow j in the TTI k (3.2) xj,n [k] = 0, otherwise, Ψ (g) is a step function that assumes 1 when g is positive and 0 otherwise. U(j, k) is the satisfaction function of flow j at TTI k and is defined as 1, if flow j is satisfied at TTI k (3.3) U(j, k) = 0, otherwise. The first constraint in problem (3.1) represents the system restriction that one RU cannot be allocated to more than one UE in a cell. In this way, no intra-cell interference is present. The second constraint we model a system restriction that the number of scheduled UEs in a TTI is limited. This is a common constraint in OFDMA system due to the limitation of the number of available control channel utilized to inform the UE transmission parameters for example. Note that problem (3.1) is a non-linear combinatorial problem whose optimal solution is not easily found. Computational heavy algorithms to search for optimum solutions may not be suitable to the small time-scale with which the scheduling takes place in B3G systems. For that reason, simple algorithms that provide sub-optimum solutions are highly recommended as stated in scheduler requirements in section 2.2.3. Both the SORA-NRT and SORA-RT schedulers have a common core that utilizes smart and simple heuristics. We have adopted the same strategy of splitting the RU allocation into two parts, as described in section 2.2.1. Thus, SORA-NRT and SORA-RT are composed of two building 27
blocks: Resource Allocation and Resource Assignment, as shown in Figure 3.1. The Resource Allocation part is responsible for defining which flows will be scheduled and determining an estimate of how many resources they will receive, while the Resource Assignment part defines which resources will be associated with which flow. Resource Assignment
Resource Allocation Define list p
List a Flow’s requirements
Assign resources to flows
Calculate flow’s requirements
Scheduled flows with assigned resources
Control the amount of assigned resources
Define list a
Flow’s information Channel state info QoS requirements of flows Flow’s current state
Figure 3.1: Building blocks of SORA-NRT and SORA-RT.
The first step of the Resource Allocation part is to define the Prioritized List p. This list is composed of all active flows ordered by a given priority. We consider that flows with pending retransmissions at physical layer have the maximum priority. Among the flows that do not have pending retransmissions, the priority is service dependent and further details will be presented in sections 3.3 and 3.4. Moreover, each flow j is also associated with an estimated required data rate ∆r j [k]. The inputs utilized in both parts of SORA-NRT and SORA-RT are QoS requirements, current flow’s state measurements and channel quality. In the end of the Resource Allocation part the Allocation List a, is defined. This list comprises the group of flows that will have assigned resources in the Resource Assignment part. Because of the limitation in the number of scheduled UEs in a TTI (S), the number of UEs in list a will also be limited. More specifically, the list a is composed of the S more prioritized flows taken from the first positions of list p. With this, we intend to preserve the QoS of the most important flows in the cell. An illustration of the process to create list a is shown in Figure 3.2. In the Resource Assignment part, it is important that each flow gets the predicted number of resources in the Resource Allocation part. Therefore, resource assignment is performed in assignment phases so that in each phase a flow can get only one resource. Also, in the beginning of a new phase the flows are prioritized according to their best RU, i.e., the flow with the best RU among all resources of all flows gets its RU first. The flows compete for resources in this part until their requirements are fulfilled. In case all flows achieve the the required rate and there are still unused resources the resource assignment starts again in the same way as described. In Figure 3.3 we show a flowchart of the Resource Assignment part.
28
Prioritized list Flow 1 Allocation list
Flow 3 Flow 4 Flow 5 Flow 6
Priority order
Flow 2
Maximum number of scheduled UEs: 4
Flow 7 Flow 8
Figure 3.2: Example of list a building in the Resource Allocation part.
Start
New phase: Prioritize the (remaining) flows according to its best resource
Higher priority flow gets its best resource
Are there available resources?
No
End
Yes
Did the selected flow achieve the desired rate?
No
Take the selected flow out of this phase
Yes
Did all flows get a resource in the current phase?
Yes
Take the selected flow out of the assignment
No
Figure 3.3: Flowchart of Resource Assignment part of SORA scheduler.
29
3.3 SORA - Non Real Time SORA-NRT intends to maximize the user satisfaction ratio of NRT services such as Transport Control Protocol (TCP)-based traffic Web browsing and FTP. NRT services do not have strict packet delay requirements. In fact, subscribers utilizing this service type only expect that their average data rate be maintained above a given target or requirement. Fluctuations of the average data rate around the target are tolerable. In problem (3.1), the U(j, k) function can be specified in this case as 1, if r j [k] ≥ rreq j (3.4) U(j, k) = 0, otherwise, where r j [k] is the average data rate of flow j at TTI k in the L4 layer computed from the session start, and rreq j is the average data rate requirement of flow j in the L4 layer. r j [k] is defined as r j [k] =
sj [k] ∀j ∈ Ω, (tj [k] · T0 )
(3.5)
where sj [k] is the number of correctly transmitted bits at the L4 layer from flow j until TTI k, tj [k] is the total active time of flow j until TTI k and T0 is the duration of one TTI. As described in section 3.2, SORA-NRT is initialized by the Resource Allocation part. ∆r j [k] in SORA-NRT consists of the rate at TTI k required to return an unsatisfied flow to a satisfied state. ∆r j [k] is defined as follows ∆r j [k] = r req j · (tj [k] + z) − r j [k − 1] · tj [k] ,
(3.6)
where z is a constant. More specifically, ∆r j [k] is the data rate that has to be allocated to flow j at TTI k in order for this flow to remain satisfied even if it does not have transmission opportunities in the next ( z − 1) TTIs. Note that if a flow is satisfied, the value of ∆r j [k] will be non-positive. In order to define the priority among flows, it is necessary to estimate the number of required resources for the flows, mj . It is calculated as ! ∆r j [k] , mj [k] = max 1, (3.7) F γ j [k] where γ j [k] is the mean channel quality among all available resources in the system of UE j at TTI k and F (·) is the link adaptation function that maps the channel quality to data rate. Note that currently satisfied flows require only one RU. The list m is composed of the elements mj [k]s. The list m is utilized to define the Prioritized List p. Among the flows that do not have pending physical-layer retransmissions, the flows with positive ∆r j [k] (unsatisfied flows) are the most prioritized ones. On the other hand, the flows with lower priority are the ones with ∆r j [k] lower than or equal to zero (satisfied flows). The unsatisfied flows are further sorted according to mj [k], i.e., the unsatisfied flows with small estimated number of required RUs have higher priorities than the other flows. By doing this, we give priority to the flows that are currently unsatisfied, but demands the lower amount of resources in order to become satisfied. The prioritization in SORA-NRT is illustrated in Figure 3.4. The Resource Assignment part works according to the explanation in section 3.2. It is performed in assignment phases so that in each phase a flow can get only one resource. The 30
Figure 3.4: Illustration of list p building in SORA-NRT.
flows are prioritized according to the channel quality of their best resources. The selected flows in the Resource Allocation part (in list a) compete for resources until ∆r j [k] is achieved. When a flow achieves the required data rate it is taken out of the assignment. In case all flows achieve the required rate ∆r j [k] and there are still unused resources, all flows that composed list a and still need more resources to transmit the remaining buffered data will compete for resources again in the same fashion as explained before.
3.4 SORA - Real Time The counterpart of SORA-NRT, SORA-RT is utilized for RT services. This service type is delay sensitive and imposes strict requirements on the packet loss rate and delay variation as described in section 2.1.3. In this work, we consider that the main QoS requirement of a RT service is to guarantee a given Frame Erasure Rate (FER) requirement. The U (j, k) function presented in problem (3.1) is defined as follows for a scenario with only an RT service 1, if F ERj [k] ≤ F ERjreq (3.8) U(j, k) = 0, otherwise, where F ERj [k] is the accumulated FER for flow j at TTI k, and F ERjreq is the FER requirement of flow j. F ERj [k] is defined as nlost j [k] ∀j ∈ Ω, F ERj [k] = lost nj [k] + nsucc [k] j
(3.9)
where nsucc [k] is the number of successfully transmitted packets from flow j until TTI k and nlost j j [k] is the number of lost packets from flow j until TTI k. From the FER metric, we can extract another important metric utilized by SORA-RT: The list w [k] whose elements are wj [k]s. This metric represents how many packets an unsatisfied flow j at TTI k has to successfully transmit in a row so as to become satisfied (with F ERj [k] ≤ F ERjreq [k]). For a satisfied flow j, wj [k] means the maximum number of packets that this flow 31
can lose successively and still be satisfied. wj [k] can be utilized to define how close or far the FER of a given flow is from F ERjreq . This variable is defined as
wj [k] =
$ % nsucc [k] + nlost [k] · F ERjreq − nlost [k] j j j , if F ERj [k] ≤ F ERjreq req 1 − F ER j
(3.10)
' & nlost [k] − nsucc [k] + nlost [k] · F ERjreq j j j , otherwise, F ERjreq
where ⌊v⌋ represents the first integer equal to or lower than v and ⌈v⌉ is the first integer equal to or greater than v. The required rate at the Data Link Layer (L2), ∆r j [k], by a flow is simply the rate necessary to transmit the oldest L2 packet. Therefore, ∆r j [k] is calculated as follows boldest [k] j ∆r j [k] = ∀j ∈ Ω, T0
(3.11)
where boldest [k] represents the number of bits of the oldest L2 packet from flow j at TTI k. j The Resource Allocation part of SORA-RT starts by defining the Prioritized List, p. Excluding the flows that have physical-layer retransmissions and consequently have the highest priority, the most prioritized flows are the ones that are currently satisfied. Note that this strategy is the opposite from SORA-NRT. We chose this procedure because the RT service does not tolerate fluctuations in the experienced QoS, which NRT services do. Therefore, the idea is to keep the highest number of flows with acceptable QoS instead of trying to recover flows from dissatisfaction as it is done in SORA-NRT. Among the most prioritized flows (satisfied flows) we further prioritize the flows with higher −1 ((D discard − Djoldest [k]) · (wj [k] + 1)) , where Djoldest [k] is the delay of the oldest L2 packet of flow j at TTI k and D discard is the maximum allowable delay at L2 for RT service before discard. Therefore, the flows with higher delays (first term) and the ones that a lost packet can be more damageable for leading this flow to an unsatisfied state (second term) are more prioritized. In the group of unsatisfied flows, we also prioritize the ones with higher ((D discard − Djoldest [k]) · (wj [k] + 1))−1 , prioritizing the flows with higher delays (first term) and that need a small quantity of successfully transmitted packets to become satisfied (second term). The prioritization among flows is summarized in Figure 3.5. In Figure 3.6, we illustrate how the priority function inside satisfied and unsatisfied groups changes with Djoldest [k] and wj [k] parameters considering D discard equal to 80 ms. As in SORA-NRT, the Resource Assignment part of SORA-RT works according to what was explained in section 3.2. It is performed in assignment phases so that in each phase a flow can get only one RU. The flows are prioritized according to the channel quality of their best resources. The selected flows in the Resource Allocation part (in list a) compete for resources until ∆r j [k] is achieved. In case all flows achieve the required rate ∆r j [k] and there are still unused resources, all flows that composed list a and still need more resources to transmit the remaining buffered data will compete for resources again in the same fashion as explained before.
32
Sorted in descending order according to
((D
Flows in retransmission
di scard
)
)
[k ] ⋅ (w j [k ] + 1) − D oldest j
Satisfied flows
−1
Unsatisfied flows
Priority order
Figure 3.5: Illustration of list p building in SORA-RT.
Priority within satisfied and unsatisfi ed fl ow groups
1000
P ri ority
800 600 400 200 0 0.08 0.07 10
0.06
8 6
0.05 D oldest j
4 0.04
2 0
w
j
Figure 3.6: Illustration of SORA-RT priority within satisfied and unsatisfied flow groups.
33
3.5 SORA for multi-service scenarios When evaluating single service scenarios, the user satisfaction ratio metric can be directly mapped to the system capacity. Consider that ρ is the number of flows in the cell with only one service type. The system capacity in this scenario can be defined as follows C = max ρ|Q (ρ) ≥ Qlimit
(3.12)
where Q (ρ) is the satisfaction rate when there are ρ flows in the system and Qlimit is the minimum acceptable satisfaction rate for the considered service. Nevertheless, when more than one service is utilized in the system another capacity metric must be utilized [32]. Consider that Ψ is the service set and that the percentage of flows from service s ∈ Ψ in the cell is given by fs =
ρs , ρtotal
(3.13)
where ρs is the number of flows from service s and ρtotal represents the total number of flows in total the cell. f is composed of elements fs . Consider that, Qs ρ , f is a function that returns the satisfaction rate of flows from service s and Qlimit consists of the minimum acceptable satisfaction s rate for service s. In this way, the individual capacity for service s is defined as Cs (f) = max ρtotal |Qs ρtotal , f ≥ Qlimit . s
(3.14)
C (f) = min (Cs (f)) .
(3.15)
In a mixed service scenario, it is important that each provided service type attains the target satisfaction rate in a balanced way, i.e., the system capacity is the maximum offered load (e.g., in number of flows) in which all provided services achieve the minimum satisfaction rate. Therefore, the system capacity, C (f), is defined as
SORA consists of a scheduling algorithm with objective to provide high system capacity in mixed service scenarios. This scheduler is composed of SORA-NRT and SORA-RT, and adds new functions regarding resource sharing among flows from different services. Figure 3.7 shows an overview of SORA. Following the general framework shown in section 3.2, SORA is also split in two main blocks: Resource Allocation and Resource Assignment. The Resource Assignment holds the main characteristics described in the previous sections, however, the Resource Allocation part was expanded to deal with multiple services. The Resource Allocation part of SORA can be divided into two main steps. The first one is the definition of the Prioritized List ps for each service type s ∈ Ψ and the calculation of required resources per flow. The second step is the definition of the Allocation list a that is utilized by the Resource Assignment part. The first step of Resource Allocation part is based either on SORA-NRT or SORA-RT depending on the flow service type. If the considered service is NRT, the definition of ps and the calculation of required resources are based on SORA-NRT scheduler presented in section 3.3. However, if the service is RT the steps defined in section 3.4 for SORA-RT are followed. In the second step of the Resource Allocation part, a resource sharing among services takes 34
NRT services
RT services RT service 1
RT service s1
NRT service 1
NRT service s2
Flows
Flows
Flows
Flows
Resource Allocation Definition of prioritized list and flow’s required resource based on SORA-NRT
Definition of prioritized list and flow’s required resources based on SORA-RT
Resource sharing among services
Resource Assignment
Figure 3.7: Building blocks of SORA.
place. In a general way, this task is the main functionality of any scheduler designed to mixed service scenarios. Roughly speaking, the function of any scheduler in a mixed scenario is to answer two questions: 1. Which flows of each service type will be scheduled? Indirectly, this question is related to define in which proportion the flows of the multiple services will be scheduled. 2. How many RUs will be assigned to each scheduled flow? In other words, how will the system resources be divided among the existing services? These questions are partially answered in the first step of the Resource Allocation part. By defining ps , we determine which flows of each service that must be prioritized. The amount of required resources for all flows is also determined. The complete answers to these questions are found in the Resource Sharing step. In order to understand which flows will compose the list a, it is necessary to define some variables. Consider that set Ωs is composed of the active flows from service s ∈ Ψ. The number of flows of each service in a is constrained by the conditions below
35
y ≈ S · P |Ωs | s p∈Ψ |Ωp | P s∈Ψ ys ≤ S,
(3.16)
where ys represents the number of selected flows from service s to be scheduled, i.e., to compose list a; and |Z| represents the cardinality of set Z. The first part in equation (3.16) states that the number of flows from each service in the list a must be almost proportional to the number of active flows from each existing service in the cell. The second part has the objective of guaranteeing that the number of scheduled flows is equal to or smaller than the maximum number of UEs that can be scheduled in a TTI. In this study, we also considered that the minimum number of scheduled flows from a service is one to avoid service starvation. The main objective of these constraints is to provide a better resource distribution among the service classes even when the number of flows of each service is unbalanced. The Resource Assignment is the same compared with the described ones in sections 3.3 and 3.4. It is performed in assignment phases so that in each phase a flow can get only one resource. The flows are prioritized according to the channel quality of their best resources. The selected flows in the Resource Allocation part (in list a) compete for resources until ∆r j [k] is achieved. In case all flows get the required rate and there are still unused RUs, all selected flows for Resource Assignment part compete for resources again in the same way as described before.
36
Chapter 4 Application to 3GPP’s Long Term Evolution System In this chapter, we evaluate the SORA scheduler in a case study considering mixed traffic scenarios in the LTE system. Before presenting the performance results, we shortly describe the LTE system. The results are divided into two parts. In the first part we provide a sensitivity analysis to assess the performance of the SORA scheduler when the availability of some metrics necessary to the algorithm are restricted. In the last part we present a performance evaluation of SORA scheduler in different traffic mixed scenarios, and compare the relative performance concerning reference schedulers.
4.1 Simulation models and LTE description In this section we provide a description of the LTE system including the system architecture and the main aspects. Moreover, the simulator and the models utilized to evaluate the SORA scheduler are depicted.
4.1.1 Simulator overview In this study, we utilize an advanced radio network platform implemented using Java and developed internally at Ericsson Research. Today, this network platform provides simulators for several radio access technologies including LTE, which is the considered system in this chapter. The simulator is event-driven and runs by sequentially executing events placed in an event queue. The events are positioned in the queue based on their defined execution. Therefore, the event queue contains those events that have been planned but not yet executed. In a simplified way, the main actions (or events) performed by the network in the simulator are: • Reception of physical channels transmitted in the preceding period; • Scheduling of transmissions to be performed in the following period; • Updating of radio propagation models for the following period.
These three operations invoke other operations, which in turn invoke other operations, and so on. In fact, a large fraction of the simulation processing is typically performed in direct call chains from these three operations. The main components of the network simulator are shown in Figure 4.1. In Deployment the simulator models mainly the number of BSs and cells, their positions in the radio network grid and wrap around techniques. The Mobility determines how the users move in the system grid and it is important for fading dynamics and handovers. The main aspects of the wireless channel such as path loss, shadowing and fast fading are modeled in Propagation and Fading. Aspects such as physical transmission and reception, channel errors and link-to-system models are present in the Physical Layer. The Transport Network contains functions needed to model the link between Enhanced Node B (eNB) and external networks. Radio Protocols such as Medium Access Control (MAC), Radio Link Control (RLC) and Packet Data Convergence Protocol (PDCP), as well as Internet Protocols are also modeled in the simulator in order to provide reliable results when evaluating traffic performance. The scheduling functionality is present in the Radio Resource Management, which besides scheduling includes handover, and link adaptation. Finally, in Application Traffic several services such as VoIP, Web surfing and streaming are modeled.
Figure 4.1: Main simulator network components.
4.1.2 LTE system and models LTE is a new system which is currently being specified by 3GPP and will be probably the basis for 3GPP’s submission for International Mobile Telecommunications (IMT)-Advanced. Several 38
MME / S-GW
MME / S-GW
S1
S1
S1
S1
E-UTRAN
X2
eNB X2
X 2
eNB
eNB
Figure 4.2: LTE network elements and their interfaces [33].
working groups have been involved in the definition of the specifications for LTE since the summer of 2006. LTE provides a new transmission access technology with OFDM for downlink and based on Single Carrier - Frequency Division Multiple Access (SC-FDMA) for uplink. More details about OFDM and SC-FDMA can be found in [15]. Also, LTE will be a pure packet-switched network and will use the IP protocol as the base protocol for transporting information. Basically, LTE is composed of the following physical network elements: • User Equipment (UE): Mobile user terminals capable of accessing the wireless services; • Enhanced Node B (eNB) or base station: Network element that serves the user terminals; • Mobility Management Entity (MME) / Serving Gateway (S-GW): Network elements responsible for mobility management and user plane functions. These physical elements may assume different roles, which characterize the logical network elements. A logical node is defined by the services it provides, and one physical element may comprise several logical nodes. The network elements and their corresponding interfaces are depicted in Figure 4.2. The eNBs are connected to each other through the X2 interface, while the connection between eNB and MME/S-GW is done through the S1 interface. Another entity, that was not shown in Figure 4.2, is the Packet Data Network Gateway (P-GW) that together with S-GW is responsible for the user plane part. The system architecture is shown in Figure 4.3 where we can see the logical nodes (yellow boxes), functional entities of the control plane (white boxes) and the protocol layers (blue boxes). The MME is the key control node for Evolved UMTS Terrestrial Radio Access Network (E-UTRAN) and mainly perform the following functions: Non-Access Stratum (NAS) signaling, UE tracking in idle mode including control and execution of paging retransmissions, bearer 39
eNB Inter Cell RRM RB Control Connection Mobility Cont . MME Radio Admission Control NAS Security eNB Measurement Configuratio n & Provision
Idle State Mobility Handling
Dynamic Resource Allo cation (Scheduler)
EPS Bearer Control RRC PDCP S-GW
P-GW
RLC Mobility Anchoring
MAC
UE IP address allo cation
S1 PHY
Packet Filtering internet
E-UTRAN
EPC
Figure 4.3: System architecture and user/control planes [33].
management and P-GW/S-GW selection. The eNB, besides selecting and routing data to/from the MME/S-GW, has the crucial tasks of Inter-cell Radio Resource Management (RRM), scheduling and transmission of information, Connection Mobility Control (CMC), AMC, HARQ, configuration and provision of measurements, Radio Bearer Control (RBC), and Radio Admission Control (RAC). Inter-cell RRM functionalities include Inter-Cell Interference Coordination (ICIC) and load balancing. Besides user data packet routing and forwarding, S-GW is the anchoring point for inter-eNB handover and inter-3GPP mobility. Finally, the P-GW provides the IP address to connect the UE to external packet data networks and is the anchoring point for mobility between 3GPP and non-3GPP technologies. In Figure 4.4, we show the protocol stack for user and control planes. In the control plane, the Radio Resource Control (RRC) protocol handles radio bearer setup, active mode mobility management and broadcast of system information. The NAS protocols deal with idle mode mobility management, service setup, authentication and security. The PDCP layer performs ciphering and integrity check in the control plane. Regarding user plane, the PDCP layer is responsible for header compression of IP packets and ciphering. The RLC, MAC and Physical (PHY) layers have the same function in the user and control planes. RLC layer focuses on lossless transmission of data, and MAC layer handles uplink and downlink scheduling and HARQ signaling. The PHY layer is responsible for protecting data against channel errors using AMC schemes based on channel conditions. In the following we describe in more details the radio interface protocols as well as transport protocols, system deployment and traffic modelling.
40
UE
eNB
MME
NAS
NAS
RRC
RRC
P DCP
P DCP
RLC
RLC
MAC
MAC
PHY
PHY
Control plane UE
e NB P DCP
P DCP
RLC
RLC
MAC
MAC
PHY
PHY
User plane Figure 4.4: User and control planes protocol stack [33].
Physical layer LTE supports both FDD and TDD duplexing schemes. In this study FDD is utilized. The time domain structure of LTE is composed of radio frames of 10 ms. Each radio frame has 10 equally-sized subframes of length 1 ms. Subframes, in turn, consist of two slots of length 0.5 ms. The scheduling takes place in a subframe basis [33]. The default subcarrier spacing is 15 kHz and all subcarriers are grouped in sets of 12 subcarriers. A resource block in LTE is defined as a two-dimensional grid with 12 subcarriers in frequency and 0.5 ms in time that corresponds to 6 or 7 OFDM symbols depending on cyclic prefix length. An RU in the system is composed of two resource blocks concatenated in the time domain, i.e., 12 subcarriers and 1 ms. The resource structure in LTE system is illustrated in Figure 4.5. The physical resources are utilized by physical channels and signals. Physical channels are utilized for transmission of data and/or control information from the MAC layer. The physical signals are used to support physical-layer functionality and do not carry any information from the MAC layer [34]. Among the physical channels, we highlight the Physical Downlink Shared Channel (PDSCH) and Physical Downlink Control Channel (PDCCH). The former is utilized for transmission of data traffic while the later is used for downlink L1/L2 control signaling. Specifically, PDCCH is used to carry uplink scheduling grants and downlink scheduling assignments, such as PDSCH resource indication, transport format, HARQ information and transport block size. Depending 41
One frame (10 ms)
One subframe (1 ms)
One resource block
12 subcarriers
One slot (0.5 ms)
Figure 4.5: Time frame and resource structure in LTE.
on the time-variant PDCCH capacity, different number of UEs can be scheduled in a given TTI. However, the maximum number of scheduled UEs in a TTI was simplified to a fixed number over a cell in this study. In this study, we consider that the allocated power per RU is fixed and is equal to the ratio between the total cell power and the number of RUs. The link adaptation utilizes the last reported channel quality. Medium access control The MAC layer in LTE is similar to the Release 6 MAC-hs/MAC-e and covers mainly similar functions such as HARQ, priority handling (scheduling) and transport format selection [35]. The HARQ is modeled as a number of processes where each process uses a simple stop-and-wait protocol. HARQ for downlink, that is the focus in this study, is asynchronous and adaptive. By asynchronous we mean that the scheduler has the freedom to choose the subframe for retransmission dynamically. In adaptive HARQ, the scheduler can use a different resource for retransmission compared to the previous (re)transmission. The version redundancy of a (re)transmission needs to be known by the receiver. In case the data is a retransmission of a previously stored data, the received data is soft combined with the data stored in the soft buffer. In case the received data is not a retransmission or a retransmission of data has not been stored, the soft buffer is cleared and only the latest received data is placed in the buffer. Due to the HARQ retransmissions, MAC Protocol Data Units (PDUs) can arrive at the receiver in a different order of the transmission, therefore, the MAC layer does not provide in-order delivery to the RLC layer. Finally, MAC performs multiplexing of RLC PDUs of different flows of a single UE. Radio link control RLC protocol supports both Unacknowledged Mode (UM) and Acknowledged Mode (AM). UM performs duplicate avoidance and reordering, and it is utilized with applications that require low delay such as VoIP. AM can perform retransmissions utilizing the Automatic Repeat Request (ARQ) protocol, which is used when reliable data transmission is required such as in Web traffic. Depending on scheduler decision, a certain amount of data is selected from the RLC Service Data Unit (SDU) buffer and segmented and/or concatenated according to the size of the SDUs. 42
This block becomes the RLC PDU that receives a sequence number and is delivered to the MAC layer. In this way, one transport block can contain only one RLC PDU per radio bearer. The exception occurs when an RLC retransmission is triggered. In this case, an RLC PDU containing new data might be multiplexed at the MAC layer with an RLC PDU. In case an RLC retransmission is configured in AM and the channel conditions have changed significantly compared to the original RLC transmission, the RLC protocol is able to perform a re-segmentation. The RLC retransmissions are initiated either by status reports sent by the RLC receiver or by local triggers from the MAC layer in case of reaching the maximum number of HARQ transmissions. Further details about the RLC layer can be found in [36]. Other protocols The PDCP layer supports Robust Header Compression (ROHC) [37] in order to reduce the overhead imposed by large IP headers. PDCP also has an SDU discard mechanism based on a discard timer that discards PDCP SDUs and PDCP PDUs when their age exceeds a certain value. Further details about PDCP can be found in [38]. TCP layer provides reliable data transfer between two peers by means of an ARQ protocol. Furthermore, it performs congestion control to avoid overloading underlying links. In this study, this layer is located below the Web application layer that consists of a Web server on the internet and a Web client at the radio network. TCP is modeled according to [39]. Real-Time Transport Protocol (RTP) layer is a standardized network protocol for audio and video transmission, and in this study, it is utilized below the VoIP application. The RTP layer can be utilized above either TCP or User Datagram Protocol (UDP), but since RTP is intended for RT applications and such applications are normally more sensitive to delay than packet loss, UDP is the natural choice. This layer is modeled according to [40]. Deployment The system consists of base stations (or eNBs) that control a certain number of hexagonal cells. All the cells have the same radius and in order to avoid border effects and obtain the same interference load in all cells a wrap around technique is utilized. The average path gain model is Okumura-Hata [41] and for channel fading and dispersion the 3GPP Typical Urban model [42] is utilized. Web traffic modelling The Web traffic model is characterized as request-response traffic: A client who utilizes a mobile station and is located in the radio network, requests one or more web pages within a session i.e., Hypertext Transfer Protocol (HTTP) requests. The server, which is behind the internet, generates and returns the web pages. Once a web page is received at the client, this user reads the Web page for some seconds and then requests another Web page. We consider that the Web pages have a fixed length and that the reading time follows an exponential distribution.
43
VoIP traffic modelling VoIP is the transmission of voice in a packet-switched way over an IP-based network. The VoIP packets are generated by a speech coder that mimics the Adaptive Multirate (AMR) codec currently used in all Global System for Mobile communication (GSM) phones [43]. This coder produces voice frames every 20 ms during speech periods and small packets, named Silence Insertion Description (SID) packets, to simulate background noise during silence periods. The IP packets that arrive at the radio network have a large overhead from IP/RTP/UDP headers. Due to this, headers are compressed using ROHC at the PDCP layer. When a VoIP user is silent, the first SID packet is sent directly after the last talk frame, the second SID packet is sent three frames later, and after that they are sent every 8 frames. In this study, we consider that there is a conversation between two clients, one behind the internet utilizing a computer (client A) and another utilizing a UE (client B) in the radio network. As downlink is focused, the performance is measured in the client located in the radio network utilizing the UE (Client B). The model for conversation has three states: Client A talking, client B talking and mutual silence. The model switches from one client talking, mutual silence, the other client talking, and so on. Finally, the time period in each state (talking and silence) is drawn from exponential distributions. With 50% activity mutual silence never occurs.
4.2 Definitions and simulation parameters In this master thesis, SORA is compared with the references RR, Delay Scheduler (DS) and MR schedulers in single and mixed scenarios. The RR scheduler gives opportunity to the UE with the longest waiting time. When a flow is scheduled, it either receives the number of RUs necessary to transmit all its RLC SDUs in the eNB buffer or it receives all the unused RUs in the cell in case all its RLC SDUs cannot be transmitted even if the flow utilizes all the unused cell resources. When the UE is selected to receive data, the scheduler chooses its best RU, which means that this scheduler is not a pure RR scheduler. DS prioritizes the flows whose headline RLC SDU has the greatest delay. One of the first references about this kind of scheduler is [44]. As in RR, the selected flow by DS receives all the necessary RUs to transmit all pending RLC SDUs. The RUs are selected in an opportunistic way, i.e., the UE gets assigned its best RUs. In order to adapt this algorithm to a mixed service scenario with Web and VoIP flows, we consider that the VoIP delay is multiplied by the constant 10.5, which is approximately the ratio between the average required rate for Web service (128 kbps) and the generated data rate by VoIP application (12.2 kbps). If the raw RLC delays were utilized, the VoIP flows would hardly ever have transmission opportunities because of their lower packet sizes compared to the Web packets. Note that the value of this constant impacts directly on the prioritization between Web and VoIP flows and the search for an optimum value is out of the scope of this work. The criterion to schedule a UE in the MR scheduler is the channel quality. More specifically, MR schedules the UE whose transport block when utilizing the available bandwidth can transmit more information bits. When a UE is selected, it gets assigned the RUs in best channel conditions until all awaiting packets in the transmit buffer fit in the transport block. Some of the metrics utilized to evaluate the performance of the schedulers are: Average Web 44
throughput, VoIP packet delay, VoIP FER, user satisfaction ratio, cell throughput and system capacity. The average Web throughput consists of the ratio between the number of correctly received bits at the TCP layer of the UE and the total session active time. By total session active time we mean the total time in which the Web flow was active. The Web flow is considered active in the period between the transmission of the HTTP request from the client (UE) to the server and the complete reception of the requested Web page at the UE. A Web flow is considered satisfied if the average Web throughput is greater than an average required throughput. The VoIP packet delay is the time length since the transmission of the VoIP frame from the user behind the internet and the reception of the frame at the UE. The VoIP FER in this study is defined as the ratio between the number of lost and the total expected packets. A packet is considered lost if it does not arrive at the receiver or its reception is performed with a delay greater than the maximum VoIP packet delay. Note that the case of a packet does not arrive at the receiver can be caused by either an RLC SDU discard at the transmitter or an HARQ failure, e.g., maximum number of HARQ retransmissions achieved. The total number of expected packets is the sum of the lost and successfully received packets. A VoIP flow is considered satisfied if its VoIP FER is lower than the required FER. Finally, it is assumed in this work that the play-out buffer in the UE is configured in such a way that it virtually eliminates the delay jitter of VoIP packets. Therefore, this work does not evaluate the delay jitter behavior. Regarding cell/system measurements, the user satisfaction ratio for a specific service is defined as the ratio between the number of satisfied flows and the total number of flows of the considered service. In case of a single service scenario, the system capacity is defined as the maximum offered load (e.g., number of flows in the cell) in which the user satisfaction ratio is greater than the user satisfaction threshold. In case of mixed service scenarios, the system capacity is the maximum offered load in which all services have a user satisfaction ratio greater than their respective satisfaction thresholds. The cell throughput is calculated at PDCP above the RLC layer for both Web and VoIP flows. The main simulation parameters are shown in Table 4.1. Table 4.1: Simulation parameters. Parameter
Value
Unit
Bandwidth
3
M Hz
Carrier frequency
2
GHz
Duplexing scheme
FDD
−
15
−
LTE Network
Number of RUs Total cell power
20
W
Noise factor
2.3
dB
Internet packet delay
5
ms
Transport network packet delay (including Internet and CN)
9
ms
Number of scheduled user per TTI (PDCCH limit)
5
−
Path gain at 1 meter distance
-29.03
dB
Path gain per dB distance
-3.52
dB
Propagation
45
Table 4.1: Simulation parameters (continued). Parameter
Value
Unit
8
dB
SCM 3GPP
−
Number of eNBs
3
−
Number of cells per eNB
3
−
Number of UE antennas
2
−
Number of cell antennas
1
−
Cell radius
500
m
Frequency reuse
Shadowing standard deviation Antenna type Deployment
1/3
−
User speed
0
m/s
Fast fading speed
3
km/h
Number of HARQ processes
16
−
Maximum number of HARQ retransmissions
10
−
RLC ARQ
Enabled
−
Web page size (fixed)
Web Flows
10,000
bytes
Mean reading time
1.5
s
Average required throughput
128
kbps
Number of HARQ processes
16
−
Maximum number of HARQ retransmissions
10
−
RLC SDU discard period
80
ms
Disabled
−
VoIP Flows
RLC ARQ
5
s
Voice activity
0.5
−
Frame size
264
bits
Frame period
20
ms
Maximum end-to-end VoIP frame delay
140
ms
SID frame size
39
bits
Required FER
1
%
Mean talk period time
Other Parameters Simulation time
20
s
Number of samples per load point
3
−
VoIP satisfaction threshold
95
%
Web satisfaction threshold
90
%
Finally, in each simulation we have a fixed number of UEs that arrive in the system within the first 3 s, to avoid traffic synchronization and system overload in the simulation beginning. The flows generate their packets according to the service model until the end of the simulation.
46
4.3 Sensitivity analysis The SORA algorithm depends on some measurements in order to work correctly. In the following we comment about the availability of these measurements and provide some results to assess the SORA robustness when these measurements are regarded.
4.3.1 Channel quality measurements SORA is a channel aware scheduler, i.e., it relies on channel quality measurements. In Figure 4.6, we show the user satisfaction ratio in the Web-only scenario when the channel state reporting period is increased. The SORA scheduler utilizes channel quality measurements for Web flows in two parts: In the priority calculation of Web flows and in the resource assignment. In the priority calculation, an average channel quality measurement is considered. Therefore, it is expected that the dependence of this part on channel measurements is not critical. In the Resource Assignment, a per-RU channel quality measurement is utilized in order to assign the best resources to the UEs. Consequently, this part must be more affected by higher channel reporting periods. However, as it can be seen in Figure 4.6, the degradation in capacity of SORA considering a satisfaction threshold of 90% is only 2 UEs, which represents a capacity loss of approximately 2% when the channel reporting period is changed from 10 ms to 25 ms. RR, DS and MR also suffer a capacity loss of approximately 2 UEs considering the same satisfaction threshold. This similar performance among the schedulers points to a degradation in the link adaptation as the main reason for the capacity loss. Link adaptation, which is common for any scheduler, also utilizes channel quality measurements. 100
Percentage of satisfied users (%)
90 80 70 60 50 40 30 20 10 55
Period 10 ms SORA Period 10 ms DS Period 10 ms RR Period 10 ms MR Period 25 ms SORA Period 25 ms DS Period 25 ms RR Period 25 ms MR 60
65
70
75
80
85
90
Load (# of UEs per cell)
Figure 4.6: User satisfaction ratio in the Web-only scenario with variable channel state reporting periods.
In Figure 4.7, we show the user satisfaction ratio with different channel state reporting periods in a VoIP only scenario. The channel quality measurements are only utilized in the Resource 47
Assignment part of SORA when VoIP flows are concerned, that is a common part for any service type. The degradation observed in the VoIP service is similar to the one observed in the Web-only scenario in Figure 4.6. Changing the channel reporting period from 10 ms to 25 ms caused similar capacity decreases of approximately 2% for SORA and RR and 3% for DS considering the satisfaction threshold of 95%. The capacity loss for MR scheduler cannot be measured in the satisfaction threshold of 95% because of its poor performance in the simulated load range. However, the degradation in satisfaction ratio is similar to the visualized in the other schedulers. The performance loss was mainly caused by a degradation of link adaptation, as in the previous scenario. 100 90
Percentage of satisfied users (%)
80 70 60 50 40 30 20
Period 10 ms SORA Period 10 ms DS Period 10 ms RR
10
Period 10 ms MR
0
Period 25 ms SORA Period 25 ms DS Period 25 ms RR Period 25 ms MR
200
220
240
260
280
300
320
Load (# of UEs per cell)
Figure 4.7: User satisfaction ratio in the VoIP-only scenario with variable channel state reporting periods.
48
4.3.2 QoS parameters When VoIP flows are concerned, SORA utilizes the RLC SDUs delays to estimate VoIP packets’s delay. This measurement is feasible since it is only necessary that the RLC layer marks incoming packets with the arrival time. Another important metric is the estimate of the receiver FER. In the results presented in section 4.4, we consider that this measurement/estimate is present in the cell scheduler by some mean. In a practical implementation, at least two approaches can be followed. One is the report of this metric by the UE and the other is to utilize HARQ Acknowledgements (ACKs) and Negative Acknowledgements (NACKs) to estimate the receiver FER. The latter is subject to feedback errors, however, the planned error rates for HARQ feedback is about 1% and we do not expect this to be a problem [15]. Regarding NRT flows, the main considered metric is the estimate of receiver average throughput. As the TCP layer is not located at eNB and performs retransmission of segments, the estimate from HARQ ACKs and NACKs is only possible if the eNB has a way to identify when the incoming data to transmit at RLC is new data or a retransmission, and then tracks the average throughput by monitoring HARQ ACKs and NACKs. Another possibility is the feedback of this estimate from the receiver. As the uplink traffic when TCP-based services are regarded is very low, this solution is not critical. Both required average throughput and FER can be known at the eNB when the bearer setup takes place. In order to verify how the SORA scheduler behaves when the Web average data rate is not known every TTI, we performed some simulations varying the period in which the scheduler has access to an updated average data rate. In the time between two update periods, the scheduler considers that the average data rate has not changed. The user satisfaction ratio for different update periods is shown in Figure 4.8 for the Web-only scenario. We can see that the knowledge of an updated average throughput is not as important as the channel quality is. Even with a large difference in the average throughput update period, the capacity was not severely degraded. The capacity loss at the satisfaction threshold of 90% was about 1% and 5% when the update period is 300 ms and 500 ms, respectively. In Figure 4.9, we show the user satisfaction ratio for VoIP flows when the current FER is known at the scheduler according to different update periods. In the time between two update periods it is assumed that the current FER has not changed. As in the case with the Web service, we can see that to observe some performance degradation in capacity updates periods longer than channel quality update periods are necessary. No capacity loss was experienced when the update period was 100 ms. When the update periods were 200 ms and 300 ms the capacity loss was about 3% and 7%, respectively. Comparing the last two figures, we can see that the VoIP flows seem to be more affected by the update period. In fact, it was observed in the simulations that the VoIP packets are usually smaller than the TCP segments of Web service and consequently, the VoIP FER changes more often than the TCP throughput.
49
Percentage of satisfied users (%)
100
95
90
85
80 Update period 1 ms Update period 300 ms Update period 500 ms 75 55
60
65
70
75
80
85
90
Load (# of UEs per cell)
Figure 4.8: User satisfaction ratio for different average data rate update periods in the Web-only scenario.
Percentage of satisfied users (%)
100
95
90
85 Update period 1 ms Update period 100 ms Update period 200 ms Update period 300 ms 80 200
220
240
260
280
300
320
Load (# of UEs per cell)
Figure 4.9: User satisfaction ratio for different FER update periods in the VoIP-only scenario.
50
4.4 Traffic mix results In the following, we present the results per mix of service where the proportion of the number of flows from VoIP and Web services is changed. The considered mixes of service are: Web only; VoIP only; 75% of VoIP and 25% of Web; 50% of VoIP and 50% of Web; and 25% of VoIP and 75% of Web. Finally, we present the capacity region.
4.4.1 Web only In this scenario with only Web flows, we start showing the average number of scheduled UEs per TTI in Figure 4.10. As it can be seen, the average number of scheduled UEs by the reference schedulers is very similar. In these algorithms, almost only one UE is scheduled per TTI. This happens because when a flow is chosen to have transmission opportunities, it has a large amount of bits to be transmitted at the RLC layer. Therefore, all RUs are usually utilized and no other flow can be scheduled. Another reason that contributes to this fact is the scheduling criteria of RR and DS: Both algorithms choose flows that are starving for transmission opportunities and probably have a large amount of data awaiting transmission. 5
# of scheduled UEs
4
SORA DS RR MR
3
2
1
0 55
60
65
70
75
80
85
90
Load (# of UEs per cell)
Figure 4.10: Average number of scheduled UEs per TTI in the Web-only scenario.
Regarding SORA scheduler in Figure 4.10, we can observe that the average number of scheduled flows per TTI increases with load and is higher than for the three other schedulers. Of course, if higher load points would have been provided in this plot, the number of scheduled flows would saturate in the maximum allowed number of scheduled UEs per TTI that is 5 in this study. The reason for the higher number of scheduled UEs is that the SORA scheduler always chooses the 5 (the number of PDCCH channels) most prioritized UEs to be scheduled, in case there are more than 4 active UEs in the cell. This allows taking advantage of multiuser and frequency diversities in the Resource Assignment part of SORA. 51
The analysis of Figure 4.10 gives support to the understanding of the gains observed in Figure 4.11. In this figure we can see the average cell throughput. As a consequence of the Resource Assignment part that exploits multiuser diversity and channel conditions, the SORA scheduler provides the second best cell throughput above RLC/MAC compared with the reference schedulers. Another important fact that contributes to the high cell data rate is the criterion utilized in the flow prioritization. The most prioritized flows are the ones that can transmit the required rate with the lower number of resources, i.e., the flows with better channel conditions. The best cell throughput is achieved with the MR scheduler that schedules the UEs in best channel states improving the spectral efficiency. In this way, MR can be seen as an upper bound in the cell data rate for the Web service. DS presents the worst cell throughput of the four schedulers. The Web traffic is characterized by generating large packets and even with the TCP segmentation the RLC SDUs are relatively larger than the ones provided in the VoIP traffic. As flows with poor channels conditions tend to have more queued data because of small transport blocks, DS wastes system resources leading to low cell throughput. 3 2.9
Throughput (Mbps)
2.8
SORA DS RR MR
2.7 2.6 2.5 2.4 2.3 2.2 2.1 2 55
60
65
70
75
80
85
90
Load (# of UEs per cell)
Figure 4.11: Average cell throughput in the Web-only scenario.
Figure 4.12 shows the Cumulative Distribution Function (CDF) of throughput calculated at the application layer for Web flows in the offered load of 83 UEs per cell. The poor performance of DS, as explained before, is caused by the waste of system resources with flows in bad channel conditions. Regarding the RR scheduler, all flows have the same opportunity to transmit their queued packets. However, due to the link adaptation, flows in good channel conditions are more likely to transmit utilizing high order modulations and code rates. Consequently, the user average throughput distribution in the RR scheduler conforms to the Signal to Interference plus Noise Ratio (SINR) distribution in the system. With the MR scheduler, we can observe that high throughputs can be achieved as a consequence of the channel opportunistic scheduling criterion. When the SORA scheduler is concerned, some observations can be drawn. In the higher throughput region the performance is similar to the RR scheduler and worse than the one provided by the MR scheduler. However, at the medium and lower throughputs a higher concentration above 52
the average data rate requirement (128 kbps) can be seen, i.e., with the SORA scheduler the number of flows with throughput greater than the average data rate requirement is higher. This is a consequence of SORA Resource Allocation part explained in the section 3.3 that prioritizes the unsatisfied flows that needs fewer resources to achieve the minimum QoS requirement. In fact, this is a smart strategy when the objective is to increase the number of satisfied flows: To assure that a higher number of flows achieve the minimum QoS requirement avoiding quality overprovision. This feature will lead to a good user satisfaction ratio result. 1 0.9 0.8 0.7
CDF
0.6 0.5 0.4 0.3 SORA DS RR MR
0.2 0.1 0 0
100
200
300
400
500
600
700
UE average throughput (kbps)
Figure 4.12: Average Web user throughput CDF at load 83 UEs per cell in the Web-only scenario.
In Figure 4.13, we present the user satisfaction ratio in the Web-only scenario. Firstly, we highlight the good performance of MR scheduler. The main reason is the burst nature of Web traffic. This feature avoids that the UE in best channel condition always gets the system resources, because during the inactive periods other UEs can be scheduled (time multiplexing). Regarding SORA scheduler, it provides a higher number of satisfied Web flows than the reference schedulers. Considering the satisfaction threshold of 90% the gain in capacity is about 12% compared with the second best that is the MR scheduler. By prioritizing the current unsatisfied flows that need the lower amount of system resources to fulfill the minimum required rate, SORA performs a fair resource distribution avoiding that some flows have an excessively good QoS.
4.4.2 VoIP only In Figure 4.14 we can observe the average VoIP packet delays of the satisfied flows. Note that only the delays of the successfully received VoIP packets were computed. The RR and MR schedulers provided lower delays than SORA and DS. This is explained by the fact that both SORA and DS prioritizes flows with high packet delays. Therefore, packets with low delays have to wait in the transmit buffer until becoming an urgent packet and then be transmitted when SORA and DS are utilized. This is a smart strategy when VoIP service is regarded since VoIP packets 53
100
Percentage of satisfied users (%)
90 80 70 60 50 40 30 20
SORA DS RR MR
10 55
60
65
70
75
80
85
90
Load (# of UEs per cell)
Figure 4.13: User satisfaction ratio in the Web-only scenario.
that arrive at the receiver with delays lower than a maximum tolerable delay are still useful. In this way, by transmitting packets later but within the maximum delay threshold, more successful packets can arrive at the receiver and packet discard at the transmitter occurs less often. Another reason to prioritize flows with high headline packet delays is that these flows usually have more than one buffered packet to transmit. The transport block size in LTE utilizing one RU can, depending on the modulation order and code rate, be greater than one RLC SDU that is mapped one to one with VoIP frames. As a result, scheduling flows with high headline packet delays increases the efficiency by reducing the protocol layer overheads and padding rate per sent VoIP packet [45]. With MR scheduler, the VoIP packets of the flows in best channel conditions are quickly transmitted due to large transport blocks leading to small average delays. Comparing SORA and DS in Figure 4.14, we can observe that at low and medium loads SORA provides lower delays. However, at high loads the average delays of the satisfied flows with SORA are higher. Although SORA takes into account the packet delays in its prioritization, the metric wj [k] defined in section 3.4 also plays an important role. At low and medium loads wj [k] usually assumes higher values due to the high number of satisfied flows with low FERs, therefore, the role of delay in the prioritization is not as strong as in DS. However, when the load increases and the FERs get closer to the required FER, wj [k] starts to have lower values and a smaller standard deviation. In this way, the effect of the packet delays in the SORA priority is more perceptible leading to higher packet delays. In Figure 4.15 we present the average number of scheduled UEs per TTI. The SORA scheduler, as explained in section 3.2 always selects the five (number of PDCCH in the cell) most prioritized UEs. These UEs get assigned resources in the Resource Assignment part. The number of scheduled UEs with RR and DS is also five at low and medium loads. However, at high loads it can be seen that the number of scheduled UEs decreases. This is due the fact that in these schedulers, when a UE is selected it gets assigned the number of RUs necessary to transmit all buffered data. 54
80
Average user delay (ms)
70 60 50 40 30 20 SORA DS RR MR
10 0 200
220
240
260
280
300
320
Load (# of UEs per cell)
Figure 4.14: Average user delay of satisfied flows in the VoIP-only scenario.
As in the higher loads the number of buffered bits is higher, the number of scheduled UEs per TTI starts to become limited by the number of RUs instead of number of PDCCH channels. This is not observed in MR scheduler where the scheduled UEs utilize high order modulations and code rates and the scheduling process finishes due to the limitation in the maximum number of scheduled UEs. 5
# of scheduled UEs
4
3
2
1
0 200
SORA DS RR MR 220
240
260
280
300
320
Load (# of UEs per cell)
Figure 4.15: Average number of scheduled UEs per TTI in the VoIP-only scenario.
55
In Figure 4.16 we show the CDF of VoIP FER at load of 289 flows per cell in order to illustrate how SORA performs a QoS control in the cell. As it can be seen, most of the VoIP FERs provided by SORA scheduler are concentrated below the FER threshold of 1%. This is one of the key factors to provide a high number of satisfied users. It occurs because SORA besides prioritizing VoIP flows with larger delays, it also preempts the flows that have an excessively good QoS to give transmission opportunities to the flows that have critical packets to transmit.
1 0.9 0.8 0.7
CDF
0.6 0.5 0.4 0.3 SORA DS RR MR
0.2 0.1 0 0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15
UE FER (%)
Figure 4.16: VoIP FER CDF at load 289 UEs per cell in the VoIP-only scenario.
The user satisfaction ratio in the VoIP-only scenario is shown in Figure 4.17. The MR scheduler has a very poor performance in this scenario differently of the one observed in the Web-only scenario shown in Figure 4.13. One of the reasons is the low traffic generation rate of VoIP service. Since the transport block of the scheduled flows with MR scheduler is in general much larger than the size of VoIP packets with protocol overheads at layer 2, the scheduling process finishes due to the limitation in the number of scheduled UEs leading to low resource utilization, i.e., many unused RUs. Another important reason is the fact that MR is QoS unaware and do not take into account packet delays, for example. User satisfaction ratio is the main focus of the SORA scheduler and as it can be seen, the performance compared with reference schedulers is quite good. With a satisfaction threshold of 95%, the number of flows supported in a cell with SORA scheduler is about 297, which is 30 flows more than the number of supported flows when DS is used. This gain comes from the smart strategy of SORA to perform a QoS balance in order to guarantee that a high number of flows achieve the minimum acceptable QoS.
4.4.3 75% VoIP 25% Web From this point, we start to show performance results when there are both Web and VoIP flows in the system. In this section we start with a likely scenario in the future where most of the users 56
100 95
Percentage of satisfied users (%)
90 85 80 75 70 65 60 55 50 45 40
SORA
35
DS
30
RR
25
MR 20 200 210 220 230 240 250 260 270 280 290 300 310 320
Load (# of UEs per cell)
Figure 4.17: User satisfaction ratio in the VoIP-only scenario.
will be utilizing the VoIP service that will have replaced completely the old circuit-switched voice. The first result of the mixed service scenario in which the majority of the connected flows are VoIP users is the average number of scheduled UEs per TTI presented in Figure 4.18. All schedulers granted more VoIP UEs than Web ones because of the higher number of active VoIP flows compared with Web ones in this mix. The number of scheduled flows of each service with SORA is higher compared to the reference schedulers. As explained in section 3.5, the number of scheduled flows from each service with SORA is proportional to the percentage of active flows from each service. In the Resource Assignment, the flows get assigned RUs until the data rate (∆r j [k]) defined in the Resource Assignment part is achieved. This is a different strategy than the one adopted by the reference schedulers, i.e., once a UE is selected it receives the number of RUs necessary to empty its transmission buffer. Therefore, in some cases the assignment process for these algorithms finishes due to the limitation in the number of RUs and not because of the number of available PDCCH channels. In Figure 4.19 we show the resource usage according to the service. As expected, the resource usage increases with the load and saturates in the number of available RUs in the cell. Despite the higher number of VoIP flows in the system most of the system resources are allocated to Web flows. The reason is that Web service tends to generate a larger amount of traffic than VoIP. Note that the reference schedulers utilize less cell resources than SORA. The explanation is that in some cases the limitation in the number of PDCCH channels happens before the constraint in the number of available RUs in the cell occurs. The low number of utilized resources by VoIP flows when MR scheduler is utilized is explained by the resource waste caused by small buffered data size and large transport blocks of the scheduled UEs, i.e., the ones in best channel conditions. The average cell throughput in this scenario is presented in Figure 4.20. The cell data rate, in a general way, increases with system load mainly due to the higher resource utilization (RUs and power). SORA presents higher cell data rate than RR and DS because of the higher and 57
4
VoIP SORA
# of scheduled UEs
3
VoIP DS VoIP RR VoIP MR 2
Web SORA Web DS Web RR Web MR
1
0 120
130
140
150
160
170
180
190
200
210
220
Load (# of UEs per cell)
Figure 4.18: Average number of scheduled UEs per service in the mixed traffic scenario 75% VoIP and 25% Web.
15 14 13
# of assigned RUs
12 11 10
VoIP SORA VoIP DS VoIP RR VoIP MR Web SORA Web DS Web RR Web MR
9 8 7 6 5 4 3 120
130
140
150
160
170
180
190
200
210
220
Load (# of UEs per cell)
Figure 4.19: Average number of allocated RUs per service in the mixed traffic scenario 75% VoIP and 25% Web.
better resource utilization. By “better” we mean the opportunistic way in which the resources are allocated in the Resource Assignment part. Although MR scheduler gives transmission opportunities to the flows in best channel conditions and is supposed to maximize the spectral efficiency, its performance regarding cell throughput is not the best. The reason for that is, as 58
explained before, the resource waste when VoIP flows are scheduled. 3.8
Throughput (Mbps)
3.6 3.4 3.2 3 2.8 SORA DS RR MR
2.6
120
130
140
150
160
170
180
190
200
210
220
Load (# of UEs per cell)
Figure 4.20: Average cell throughput in the mixed scenario 75% VoIP and 25% Web.
In Figure 4.21 we present the user satisfaction ratio for Web and VoIP flows. Despite the good performance of RR and DS for VoIP flows, the QoS of the Web flows was compromised. Since Web flows require a higher data rate compared with VoIP flows, the RR strategy of granting equal transmission opportunities to the flows does not work well in mixed traffic scenarios. Note that, scheduling algorithms that give absolute priority to RT services such as VoIP normally have a poor performance in this scenario due to the higher number of VoIP flows than the number of Web flows. The inefficiency of DS with Web flows has penalized the overall system performance. Regarding MR scheduler, the situation is the opposite compared to the other reference schedulers: Web service achieves good satisfaction ratio with a degraded performance of VoIP one. With the SORA scheduler, the performance of Web traffic is improved while the VoIP performance is maintained. The resource distribution in SORA scheduler is accomplished in a more intelligent way, by considering the number of active flows from each service (in the resource sharing among services, section 3.5) and the flow priority based on current QoS conditions of each flow (in the definition of the prioritized list, sections 3.3 and 3.4). In this way, a capacity gain can be experienced with a better service balance.
4.4.4 50% VoIP 50% Web In this mixed traffic scenario, the number of connected flows from VoIP and Web services is the same. The average number of scheduled flows per service is shown in Figure 4.22. If compared with the same result shown in the mixed scenario with a majority of VoIP flows, we can see that the number of scheduled UEs from each service is more balanced. Regarding reference schedulers, the number of scheduled VoIP UEs has a decreasing pattern especially at high loads. This can be explained by the fact that scheduled Web flows have more buffered data at higher loads and also 59
100
Percentage of satisfied users (%)
90 80 70 60 50 40 30 20 10 0 120
VoIP SORA VoIP DS VoIP RR VoIP MR Web SORA Web DS Web RR Web MR 130
140
150
160
170
180
190
200
210
220
Load (# of UEs per cell)
Figure 4.21: User satisfaction ratio in the mixed traffic scenario 75% VoIP and 25% Web.
5 VoIP SORA VoIP DS 4
VoIP RR
# of scheduled UEs
VoIP MR Web SORA 3
Web DS Web RR Web MR
2
1
0 80
90
100
110
120
130
140
150
160
170
180
Load (# of UEs per cell)
Figure 4.22: Average number of scheduled UEs per service in the mixed traffic scenario 50% VoIP and 50% Web.
need more system resources to transmit. In this way the number of scheduled UEs per TTI starts to be limited by the number of RUs and no more by the number of PDCCH channels. When SORA is concerned, the limit in the number of scheduled UEs per TTI continues to be exploited. In Figure 4.23 we present the average number of assigned RUs per service. The resource utilization is similar for all scheduling algorithms, with higher resource utilization of Web flows. 60
Note that, differently of the results in the mix 75% VoIP and 25% Web, in this scenario the cell resources are almost completely utilized. This is explained by the presence of more data rate demanding Web flows. Still comparing with the mixed service scenario 75% VoIP and 25% Web, we can observe that the percentage of utilized RUs by Web flows is higher. 12
VoIP SORA
10
# of assigned RUs
VoIP DS VoIP RR
8
VoIP MR Web SORA
6
Web DS Web RR Web MR
4
2
0 80
90
100
110
120
130
140
150
160
170
180
Load (# of UEs per cell)
Figure 4.23: Average number of allocated RUs per service in the mixed traffic scenario 50% VoIP and 50% Web.
The average cell throughput is shown in Figure 4.24. SORA presents higher cell data rate than the DS and RR schedulers and the explanation for this is the same as stated before. However, differently of the scenario 75% VoIP and 25% Web, the best cell throughput is achieved by MR scheduler. This better performance of MR scheduler is due to the decrease in the number of VoIP flows and increase in the number of Web flows in the system. Even with most of the cell resources being utilized, SORA, RR and MR still present an increasing cell data rate. However, the prioritization of Web flows with high packet delays, large amount of packets awaiting transmission and poor channel conditions have lead the cell data rate of DS to saturation. In Figure 4.25 we show the user satisfaction ratio for the mixed scenario 50% VoIP and 50% Web. As in the mixed traffic scenario 75% VoIP and 25% Web, SORA achieves a good balance in user satisfaction ratio by providing a better performance of Web flows compared with the reference schedulers. As in the previous mixed scenario, the Web service limits the capacity of all schedulers except MR. Considering a satisfaction threshold of 95% for VoIP service, we can see that SORA accepts the offered load of 173 connected flows while RR and DS allows 136 and 117 flows, respectively. The capacity of MR scheduler for VoIP service is very poor. Regarding Web service with satisfaction threshold of 90%, the maximum cell load supported is 120, 101 and 97 for MR, RR and DS, respectively, with advantage for SORA with 128 connected flows.
61
3.8 SORA DS RR MR
3.6
Throughput (Mbps)
3.4 3.2 3 2.8 2.6 2.4 2.2 80
90
100
110
120
130
140
150
160
170
180
Load (# of UEs per cell)
Figure 4.24: Average cell throughput in the mixed scenario 50% VoIP and 50% Web.
100
Percentage of satisfied users (%)
90 80 70 60 50 40 30 20 10 0 80
VoIP SORA VoIP DS VoIP RR VoIP MR Web SORA Web DS Web RR Web MR 90
100
110
120
130
140
150
160
170
180
Load (# of UEs per cell)
Figure 4.25: User satisfaction ratio in the mixed traffic scenario 50% VoIP and 50% Web.
62
4.4.5 25% VoIP 75% Web In this last traffic mixed scenario the number of connected Web flows in the system is higher than the number of VoIP flows. In Figure 4.26 we present the number of scheduled flows per TTI according to the service type. The reference schedulers maintain their properties of a low number of scheduled flows per TTI for both service types. On the other hand, with SORA scheduler the number of scheduled flows per TTI is higher with advantage for Web flows. The higher number of scheduled Web flows with SORA is explained by the resource sharing method that selects the flows from each service type according to the number of active UEs of each service. The objective of this strategy is to prioritize the flows of the service that have more active flows in order to balance the user satisfaction ratio. 4 VoIP SORA VoIP DS
# of scheduled UEs
3
VoIP RR VoIP MR Web SORA Web DS
2
Web RR Web MR
1
0 60
80
100
120
140
160
Load (# of UEs per cell)
Figure 4.26: Average number of scheduled UEs per service in the mixed traffic scenario 25% VoIP and 75% Web.
The resource utilization per service is presented in Figure 4.27. Following the tendency depicted in the previous mixed service scenarios, the increasing number of Web flows in the system provides a higher utilization of the system resources. In fact, the average number of assigned RUs for Web flows at high loads is almost 14 while VoIP flows get approximately one RU. All simulated scheduling algorithms have a similar performance in this figure. The average cell throughput is shown in Figure 4.28. As in the previous results, the SORA scheduler presents better spectral efficiency than RR and DS schedulers. The reasons are the same stated before: Exploitation of multiuser diversity by scheduling more UEs and opportunistic resource assignment. MR scheduler provided the best cell throughput due to the pure channel opportunistic scheduling criterion and the low percentage of VoIP flows in the system. DS provides a poor cell throughput and proves not to be suitable for scenarios with Web service. The user satisfaction ratio in this scenario is plotted in Figure 4.29. The SORA scheduler, as in the previous mixed service scenarios, outperforms the reference schedulers for both VoIP and Web services. Also, the Web service limits the capacity for all simulated schedulers except MR. 63
15 14 13 12
# of assigned RUs
11
VoIP SORA
10
VoIP DS
9
VoIP RR
8
VoIP MR
7
Web SORA
6
Web DS
5
Web RR
4
Web MR
3 2 1 0 60
80
100
120
140
160
Load (# of UEs per cell)
Figure 4.27: Average number of allocated RUs per service in the mixed traffic scenario 25% VoIP and 75% Web.
3.8 3.6
Throughput (Mbps)
3.4
SORA DS RR MR
3.2 3 2.8 2.6 2.4 2.2 2 60
80
100
120
140
160
Load (# of UEs per cell)
Figure 4.28: Average cell throughput in the mixed scenario 25% VoIP and 75% Web.
Considering only VoIP satisfaction, the gains of SORA at the satisfaction threshold of 95% are of approximately 29% and 47% compared with RR and DS. When Web satisfaction is regarded, the gains are of 8%, 23% and 24% compared with MR, RR and DS, respectively. In the next section, we summarize the capacity achieved in all simulated scenarios by presenting the capacity region.
64
100
Percentage of satisfied users (%)
90 80 70 60 50 40 30
VoIP SORA VoIP DS VoIP RR VoIP MR Web SORA
20
Web DS
10
Web RR Web MR
0 60
70
80
90
100
110
120
130
140
150
160
Load (# of UEs per cell)
Figure 4.29: User satisfaction ratio in the mixed traffic scenario 25% VoIP and 75% Web.
4.4.6 Capacity In order to summarize all the previously presented results we show in Figure 4.30 the system capacity. This is an important result since it allows us to assess the performance of the schedulers when the network is submitted to different traffic mixes. The MR scheduler is not included in Figure 4.30 due to the poor performance shown in the simulated load ranges for the different mixes. As the main objective of SORA scheduler is to maximize the number of satisfied flows in the system and the user satisfaction ratio defines the system capacity, SORA scheduler also provides a greater capacity region. DS and RR schedulers have almost the same performance with a small advantage for DS when there are more VoIP flows than Web flows. When the Web flows are the majority, RR scheduler performs slightly better than DS. The reason for this behavior is the degradation in the performance of DS scheduler when the Web flows are the majority. The overall capacity of MR scheduler was compromised due to the poor performance of this scheduler when VoIP flows are present in the system. The higher system capacity with SORA is obtained by its intelligent strategy of prioritizing the most important flows from each service type. The priority of a flow in SORA is service dependent and aims at keeping a high number of satisfied flows in the system by guaranteeing the minimum QoS requirement of the flows.
65
Web offered load (# of Web users per cell)
90 SORA DS RR
80 70 60 50 40 30 20 10 0 0
50
100
150
200
250
VoIP offered load (# of VoIP users per cell)
Figure 4.30: Capacity region.
66
300
Chapter 5 Final Remarks 5.1 Conclusions In this master thesis, we proposed SORA that is a downlink scheduler to be utilized in OFDMA and packet-switched networks. The main objective of SORA scheduler is to increase the number of satisfied flows in the network. The SORA scheduler is designed to benefit from the main characteristics of OFDMA scheme, such as the multiuser and frequency diversities. In addition, the SORA scheduler can deal with multi-service scenarios where many packet-based service classes share the same system resources. The SORA scheduler is split into two parts: Resource Allocation and Resource Assignment. In the first part, the flows that will be scheduled are defined based on their current satisfaction status, channel quality state and QoS parameters. In the second part, the selected flows get assigned system resources in such a way to take advantage of multiuser and frequency diversities. In order to evaluate the performance of the proposed scheduler, we have shown a case study where the SORA scheduler is applied in the LTE system. In order to be able to assess relative performance, reference schedulers were also implemented: DS, RR and MR. From the performance of DS with VoIP service, we have concluded that it is of utmost importance the utilization of packet delay when scheduling RT services. However, the scheduling criteria of DS is definitely not suitable to NRT services. This has lead to a poor performance of DS in mixed traffic scenarios. RR scheduler has a similar performance to DS. In fact, the starvation period, that is the scheduling criterion of RR, and packet delay have some similarities. The MR scheduler presented a good performance with rate-greedy services such as Web. In this way, we highlight the need to take channel state conditions into account so as to profit by the time multiplexing originated from the burst nature of NRT services. Nevertheless, with VoIP service that is characterized by small and regularly time-spaced packets the performance of MR scheduler was degraded mainly due to the limitation in the number of scheduled UEs per TTI. When there is this limitation, it is important that the selected UEs to be scheduled have data awaiting transmission enough so as to avoid padding and waste system resources. Therefore, the scheduling criterion of MR, that is the channel quality, failed in this point. When SORA is concerned, we have achieved the following conclusions:
• The SORA scheduler is robust regarding its main input variables such as channel quality state and service specific measurements; • The number of satisfied flows with SORA scheduler is higher than the one provided by reference schedulers in different traffic mixed scenarios; • An improved QoS balancing among different services, including RT and NRT services, is reached with SORA compared to reference schedulers; • The proposed schedulers is composed by simple and smart heuristics leading to a low complexity; • The overall system capacity in mixed traffic scenarios is improved by the SORA scheduler To sum up, we highlight here the new approach of the SORA scheduler that differs of most of the works in the literature. With SORA, the objective is to maximize the number of satisfied flows in scenarios where several packet-based services demands system resources. Even if compared with schedulers that intends to maximize the user satisfaction, SORA is distinguished by the property of achieving good results in different and unbalanced mixed traffic scenarios as shown in the presented case study.
5.2 Perspectives The natural extension of this work is the generalization of SORA scheduler for other services such as streaming. In order to do that, the main aspects that interfere in the user QoS perception of the considered services must be taken into account. The application of the SORA concept to the uplink is another interesting perspective. This would enable the analysis of the combined system capacity in downlink and uplink directions. However, there are some remarkable differences between the two links such as the multiple access scheme, that is the SC-FDMA in the uplink of LTE system. Since the SORA capacity in all simulated service mixes was limited by Web service, higher system capacity would be achieved if the VoIP service quality was degraded in a controlled manner. Therefore, we believe that the system capacity of SORA scheduler could be further improved. This QoS control could be performed, for example, by a power allocation algorithm that distributes the power differently among services. Another possible study is to analyze the performance of SORA scheduler when interference coordination schemes are employed. Such schemes can impose constraints in the availability of cell resources such as number of RUs and total cell power. Finally, we believe that an interesting continuation of this work would be the adaptation of some steps of the SORA scheduler that are necessary to a real implementation such as the utilization of moving average and filtering before utilizing the estimates in the algorithm.
68
Bibliography [1] 3GPP, “Requirements for Evolved UTRA (E-UTRA) and Evolved UTRAN (E-UTRAN),” 3rd Generation Partnership Project, Tech. Rep. TR 25.913 V7.3.0 - Release 7, March 2006. [2] ——, “All-IP Network (AIPN) Feasibility Study,” 3rd Generation Partnership Project, Tech. Rep. TR 22.978 V7.1.0 - Release 7, June 2005. [3] H. Holma and A. Toskala, WCDMA for UMTS: Radio Access for Third Generation Mobile Communications, 3rd ed. John Wiley & Sons, Ltd, 2004. [4] 3GPP2, “CDMA2000 High Rate Packet Data Air Interface Specification,” Third Generation Partnership Project 2, Tech. Rep. C.S20024-A V1.0, March 2004. [5] Y. Cao and V. O. K. Li, “Scheduling Algorithms in Broadband Wireless Networks,” Proceedings of the IEEE, vol. 89, no. 1, pp. 76–87, January 2001. [6] P. J. A. Guti´errez, “Packet Scheduling and Quality of Service in HSDPA,” Ph.D. dissertation, Aalborg University, Department of Communication Technology, Institute of Electronic Systems, October 2003. [7] H. Fattah and C. Leung, “An Overview of Scheduling Algorithms in Wireless Multimedia Networks,” IEEE Wireless Communications, vol. 9, no. 5, pp. 76–83, October 2002. [8] H. Kim, K. Kim, Y. Han, and S. Yun, “A Proportional Fair Scheduling for Multicarrier Transmission Systems,” in Vehicular Technology Conference, 2004. VTC2004-Fall. 2004 IEEE 60th, vol. 1, September 2004, pp. 409–413. [9] J. Jang and K. B. Lee, “Transmit Power Adaptation for Multiuser OFDM Systems,” IEEE Journal on Selected Areas in Communications, vol. 21, no. 2, pp. 171–178, February 2003. [10] B. Chen, H. Hu, B. Wang, and H. Wang, “A Novel Multi-Service Scheduling Scheme for E-UTRA,” in Internet, 2007. ICI 2007. 3rd IEEE/IFIP International Conference in Central Asia on, September 2007, pp. 1–5. [11] H. Lei, L. Zhang, X. Zhang, and D. Yang, “A Packet Scheduling Algorithm Using Utility Function for Mixed Services in the Downlink of OFDMA Systems,” in Vehicular Technology Conference, 2007. VTC-2007 Fall. 2007 IEEE 66th, October 2007, pp. 1664–1668. [12] S. Choi, K. Jun, Y. Shin, S. Kang, and B. Choi, “MAC Scheduling Scheme for VoIP Traffic Service in 3G LTE,” in Vehicular Technology Conference, 2007. VTC-2007 Fall. 2007 IEEE 66th, October 2007, pp. 1441–1445. 69
[13] W. Shuang, G. Youjun, G. Xuelin, T. Hui, and Z. Ping, “Packet Scheduling for Multimedia Traffics in Downlink Multi-User OFDM Systems,” in Wireless Communications, Networking and Mobile Computing, 2006. WiCOM 2006.International Conference on, September 2006, pp. 1–4. [14] J. D. Day and H. Zimmermann, “The OSI Reference Model,” Proceedings of the IEEE, vol. 71, no. 12, pp. 1334–1340, December 1983. [15] J. S. E. Dahlman, S. Parkvall and P. Beming, 3G Evolution: HSPA and LTE for Mobile Broadband, 1st ed. Academic Press, July 2007. [16] 3GPP, “Feasibility Study for Evolved Universal Terrestrial Radio Access (UTRA) and Evolved Universal Terrestrial Radio Access Network (E-UTRAN),” 3rd Generation Partnership Project, Tech. Rep. TR 25.912 V7.2.0 - Release 7, September 2006. [17] M. Sternad, T. Svensson, T. Ottosson, A. Ahlen, A. Svensson, and A. Brunstrom, “Towards Systems Beyond 3G Based on Adaptive OFDMA Transmission,” Proceedings of the IEEE, vol. 95, no. 12, pp. 2432–2455, December 2007. [18] T. Kwok, “A Vision for Residential Broadband Services: ATM-to-the-Home,” in Community Networking, 1995. ’Integrated Multimedia Services to the Home’., Proceedings of the Second International Workshop on, June 1995, pp. 119–129. [19] F. J. Velez and L. M. Correia, “Classification and Characterisation of Mobile Broadband Services,” in Vehicular Technology Conference, 2000. IEEE VTS-Fall VTC 2000. 52nd, vol. 3, September 2000, pp. 1417–1423. [20] M. Ericson and S. Wanstedt, “Mixed Traffic HSDPA Scheduling - Impact on VoIP Capacity,” in Vehicular Technology Conference, 2007. VTC2007-Spring. IEEE 65th, Dublin, April 2007, pp. 1282–1286. [21] Y.-J. Choi and S. Bahk, “Scheduling for VoIP Service in CDMA2000 1x EV-DO,” in Communications, 2004 IEEE International Conference on, vol. 3, June 2004, pp. 1495–1499. [22] R. Knopp and P. A. Humblet, “Information Capacity and Power Control in Single-Cell Multiuser Communications,” in Communications, 1995. ICC ’95 Seattle, ’Gateway to Globalization’, 1995 IEEE International Conference on, vol. 1, June 1995, pp. 331–335. [23] H. Yin and H. Liu, “An Efficient Multiuser Loading Algorithm for OFDM-based Broadband Wireless Systems,” in Global Telecommunications Conference, 2000. GLOBECOM ’00. IEEE, vol. 1, December 2000, pp. 103–107. [24] W. Rhee and J. M. Cioffi, “Increase in Capacity of Multiuser OFDM System Using Dynamic Subchannel Allocation,” in Vehicular Technology Conference Proceedings, 2000. VTC 2000-Spring Tokyo. 2000 IEEE 51st, vol. 2, May 2000, pp. 1085–1089. [25] C. Y. Wong, R. S. Cheng, K. B. Lataief, and R. D. Murch, “Multiuser OFDM with Adaptive Subcarrier, Bit, and Power Allocation,” IEEE Journal on Selected Areas in Communications, vol. 17, no. 10, pp. 1747–1758, October 1999.
70
[26] D. Kivanc and H. Liu, “Subcarrier Allocation and Power Control for OFDMA,” in Signals, Systems and Computers, 2000. Conference Record of the Thirty-Fourth Asilomar Conference on, vol. 1, November 2000, pp. 147–151. [27] J. A. C. Bingham, “Multicarrier Modulation for Data Transmission: An Idea whose Time Has Come,” IEEE Communications Magazine, vol. 28, no. 5, pp. 5–14, May 1990. [28] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. Sons, 1991.
John Wiley &
[29] D. Hughes-Hartogs, “Ensemble Modem Structure for Imperfect Transmission Media,” United States Patent, no. 4.679.227, July 1987. [30] J. Gross and M. Bohge, “Dynamic Mechanisms in OFDM Wireless Systems: A Survey on Mathematical and System Engineering Contributions,” Technical University Berlin - Telecommunication Networks Group, Tech. Rep. TKN-06-001, May 2006. [Online]. Available: http://www.tkn.tu-berlin.de/publications/papers/TKN Report 06 001.pdf [31] A. R. Braga, E. B. Rodrigues, and F. R. P. Cavalcanti, “Packet Scheduling for VOIP Over HSDPA in Mixed Traffic Scenarios,” in Personal, Indoor and Mobile Radio Communications, 2006 IEEE 17th International Symposium on, Helsinki, September 2006, pp. 1–5. [32] A. Furusk¨ar, “Radio Resource Sharing and Bearer Service Allocation for Multi-Bearer Service, Multi-Access Wireless Networks,” Ph.D. dissertation, Royal Institute of Technology (KTH), Radio Communication Systems, April 2003. [33] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved Universal Terrestrial Radio Access Network (E-UTRAN); Overall Description,” 3rd Generation Partnership Project, Tech. Rep. TS 36.300 V8.4.0 - Release 8, March 2008. [34] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); Long Term Evolution (LTE) Physical Layer; General Description,” 3rd Generation Partnership Project, Tech. Rep. TS 36.201 V8.1.0 - Release 8, December 2007. [35] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); Medium Access Control (MAC) Protocol Specification,” 3rd Generation Partnership Project, Tech. Rep. TS 36.321 V8.1.0 - Release 8, March 2008. [36] ——, “Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Link Control (RLC) Protocol Specification,” 3rd Generation Partnership Project, Tech. Rep. TS 36.322 V8.1.0 Release 8, March 2008. [37] C. Bormann et al., “Robust Header Compression,” Network Working Group, Tech. Rep. IETF RFC 3095, July 2001. [Online]. Available: http://www.ietf.org/rfc/rfc3095.txt [38] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); Packet Data Convergence Protocol (PDCP) Specification,” 3rd Generation Partnership Project, Tech. Rep. TS 36.323 V8.1.0 - Release 8, March 2008.
71
[39] J. Postel et al., “Transmission Control Protocol,” DARPA Internet Program Protocol Specification, Tech. Rep. IETF RFC 793, September 1981. [Online]. Available: http://tools.ietf.org/html/rfc793 [40] H. Schulzrinne et al., “RTP: A Transport Protocol for Real-Time Applications,” Network Working Group, Tech. Rep. IETF RFC 3550, July 2003. [Online]. Available: http://tools.ietf.org/rfc/rfc3550.txt [41] Y. Okumura, E. Ohmori, T. Kaetano, and K. Fukuda, “Field Strength and its Variability in VHF and UHF Land Mobile Radio Service,” Review of the Electrical Communication Laboratory, vol. 16, 1968. [42] 3GPP, “Deployment Aspects,” 3rd Generation Partnership Project, Tech. Rep. TS 25.943 V6.0.0 - Release 6, December 2004. [43] ITU-T, “Wideband Coding of Speech at Around 16 kbit/s Using Adaptive Multi-Rate Wideband (AMR-WB),” International Telecommunication Union, Tech. Rep. G.722.2, July 2003. [44] P. Hosein, “Scheduling of VoIP Traffic over a Time-Shared Wireless Packet Data Channel,” in Personal Wireless Communications, 2005. ICPWC 2005. 2005 IEEE International Conference on, January 2005, pp. 38–41. [45] F. Persson, “Voice over IP Realized for the 3GPP Long Term Evolution,” in Vehicular Technology Conference, 2007. VTC-2007 Fall. 2007 IEEE 66th, September 2007, pp. 1436–1440.
72
