UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Improved channel state dependent scheduling for WLANs Zhang, Limei 2005

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2005-0724.pdf [ 3.28MB ]
Metadata
JSON: 831-1.0064772.json
JSON-LD: 831-1.0064772-ld.json
RDF/XML (Pretty): 831-1.0064772-rdf.xml
RDF/JSON: 831-1.0064772-rdf.json
Turtle: 831-1.0064772-turtle.txt
N-Triples: 831-1.0064772-rdf-ntriples.txt
Original Record: 831-1.0064772-source.json
Full Text
831-1.0064772-fulltext.txt
Citation
831-1.0064772.ris

Full Text

Improved Channel State Dependent Scheduling for WLANs By Limei Zhang B.Eng., Northern Jiaotong University, 1994 A THESIS SUBMITTED IN PARTIAL F U L F I L L M E N T OF THE REQUIREMENTS FOR THE D E G R E E OF Master of Applied Science in THE F A C U L T Y OF G R A D U A T E STUDIES (Electrical and Computer Engineering) T h e U n i v e r s i t y o f B r i t i s h C o l u m b i a September 2005 © Limei Zhang, 2005 11 Abstract This thesis focuses on the study of Channel State Dependent Scheduling (CSDS) based scheduling schemes for Wireless Local Area Networks ( W L A N ) . C S D S based scheduling schemes can improve the usage efficiency of scarce radio resource since packets transmitted on wireless channels are often subject to burst errors which cause back-to-back packet losses. B y taking into account the channel state in scheduling decisions, C S D S based schemes can achieve a large improvement in total throughput compared to FIFO scheme, which is currently used in W L A N systems. However, i f not properly designed, a C S D S scheme can cause unfair resource allocation among users. Various mechanisms have been proposed for addressing the fairness problem in the literature, but unfortunately most of them are too complicated to be employed in W L A N . The main contribution of this thesis is to propose a simple compensation mechanism which is easy to implement. Based on the compensation mechanism, three C S D S based scheduling schemes are designed for W L A N . Among them, Max S* selects the best channel state S*, which is calculated according to the compensation mechanism; I W R R (Improved Weighted Round Robin) adopts dynamically changing weights for users. The weight is derived based on the compensation mechanism. A hybrid scheme, IWRR+Max S*, is proposed which possesses the advantages of both of I W R R and M a x S*. A W L A N system simulation is used to show that the proposed schemes can yield substantial improvements in short-term fairness compared to C S D S + R R (Round Robin) in most test scenarios, while still maintaining high channel efficiency. i i i T a b l e o f C o n t e n t s Abstract ii Table of Contents iii List of tables v List of figures vi Glossary ...viii Acknowledgments xi Chapter 1 Introduction 1 1.1 Background 1 1.2 C S D S Based Scheduling Schemes 2 1.3 Motivation 5 1.4 Thesis Contributions 6 1.5 Overview of Thesis 8 Chapter 2 Related works 9 2.1 Threshold-based C S D S scheduling algorithms 10 2.2 Weight-based C S D S Scheduling algorithms 13 Chapter 3 New CSDS scheduling algorithm developments 17 3.1 Overview of 802.11b W L A N [19] : 17 3.1.1 Network architecture 17 3.1.2 M A C sublayer functions 18 3.2 Wireless channel model 20 3.3 Fairness definition and Compensation mechanisms 22 iv 3.3.1 Fairness definition 22 3.3.2 Compensation mechanisms 23 3.4 New scheduling schemes 26 3.4.1 M a x ' S * : :. .' • 26 3.4.2 Improved W R R (IWRR) 27 3.4.3 Combined scheduling scheme (IWRR+Max S*) 30 Chapter 4 Simulation environment 33 4.1 Fading channel module implementation and verification 33 4.2 I E E E 802.1 l b W L A N simulation system implementation and verification 36 4.2.1 System implementation 36 4.2.2 Verification of C++ simulation model with simple O P N E T test scenarios 42 Chapter 5 Simulation results and analysis 48 5.1 Simulation assumptions and test scenarios 48 5.2 Results for different Doppler frequencies and fairness window size 50 5.3 Effect of number of users 59 5.4 Selection of y for IWRR+Max S* 64 Chapter 6 Conclusions and future work 71 References 75 List of tables Table 4.1: Throughput comparison between O P N E T and C++ simulation for test scenario 1 46 Table 4.2: P E R comparison between O P N E T and C++ simulation 46 Table 4.3: Throughput comparison between O P N E T and C++ simulation for test scenario 2 47 Table 5.1: Simulation model parameter values.... 48 Table 5.2: Other simulation parameters 49 Table 5.3: The fairness index improvement of IWRR+Max S* to C S D S + R R 63 Table 5.4: y o p t based on the different value of Pdw 68 vi List of figures Figure 1.1: The architecture o f a C S D S based scheduler 4 Figure 3.1: F low chart of I W R R . : 29 Figure 3.2: Functions of y in (3.10) and (3.11) 31 Figure 3.3: Flow chart of IWRR+Max S* 32 Figure 4.1: C D F o f Rayleigh distributed fading channel 34 Figure 4.2: L C R functions of Rayleigh fading channel 35 Figure 4.3: Topology of an infrastructure 802.11b W L A N 36 Figure 4.4: The flow chart of the operation at a Node / 37 Figure 4.5: Downlink function block of modules 38 Figure 4.6: Uplink function block of modules 38 Figure 4.7: Test topology in O P N E T environment with one A P and two user nodes 45 Figure 5.1: Deployment of an equi-distant 8-user W L A N 51 Figure 5.2: The comparison results of (10 H z , 1 s) in the 8-user test scenario 54 Figure 5.3: The comparison results for (1 Hz , 10 s) in the 8-user test scenario 55 Figure 5.4: The comparison results for (2.5 H z , 4 s) in the 8-user test scenario 55 Figure 5.5: The comparison results of (1 Hz , 1 s) in the 8-user test scenario 56 Figure 5.6: The comparison results for (1 Hz , 5 s) in the 8-user test scenario 57 Figure 5.7: The comparison results for (5 Hz , 5 s) in the 8-user test scenario 57 Figure 5.8: The comparison results for (10 Hz , 5 s) in the 8-user test scenario 58 Figure 5.9: The comparison results of (10 Hz , 10 s) in the 8-user test scenario 58 Figure.5.10: The results of (10Hz , 1 s) in 2-user test scenario 61 Figure 5.11: The results of (10 Hz , 1 s) in 4-user test scenario 61 Vll Figure 5.12: The results o f (10 H z , 1 s) in 12-user test scenario 62 Figure 5.13: The results of (10 Hz , 1 s) in 16-user test scenario 62 Figure 5.14: The results of (10 Hz , 1 s) in 20-user test scenario 63 Figure 5.15: The performance curves of IWRR+Max S* parameterized by y for Pdw = 1, N u = 8 65 Figure 5.16: The performance curves o f IWRR+Max S* parameterized by y for Pdw = 5, N u = 8 :.. 66 Figure 5.17: The performance curves of IWRR+Max S* parameterized by y for Pdw =10, N u = 8 66 Figure 5.18: The performance curves of IWRR+Max S* parameterized by y for Pdw = 25, N u = 8 67 Figure 5.19: The performance curves of IWRR+Max S* parameterized by y for Pdw = 1, N u = 2....: 69 Figure 5.20: The performance curves of IWRR+Max S* parameterized by y for Pdw = 10, N u = 2 69 G l o s s a r y Acronyms A C K . Acknowledgement A P Access Point A W G N Additive White Gaussian Noise B P S K Binary Phase Shift Keying BSS Basic Service Area C D F Cumulative Distribution Function C S M A / C A Carrier Sense Multiple Access with Coll ision Avoidance C S D S Channel State Dependent Scheduling C W Contention Window D C F Distributed Coordination Function DIFS Distributed Coordination Function Interframe Space DSSS Direct Sequence Spread Spectrum EIFS Extended Interframe Space F H S S Frequency Hopping Spread Spectrum FIFO First Come First Out GPS Generalized Processor Sharing H O L Head of Line L C R Level Cross Rate L L C Logical Link Control L Q F Longest Queue First M A C Medium Access Control M P D U M A C Protocol Data Unit O P N E T Optimized Network Engineering Tools P C F Point Coordination Function P E R Packet Error Rate P H Y Physical Layer PIFS Point Coordination Function Interframe Space P L C P Physical Layer Convergence Protocol P S D U P L C P Service Data Unit QoS Quality of Service R R Round Robin R T S / C T S Request To Send/Clear To Send SIFS Short Interframe Space S N R Signal to Noise Ratio S T A (wireless) Station T C P Transmission Control Protocol W F Q Weighted Fair Queuing W L A N Wireless Local Area Network W R R Weighted Round Robin X Symbols fd Doppler Frequency fwin Fairness window Pdw The product offd and fwin N u The number of users Sav Average S N R S Measured S N R TT Transmission Threshold TTopt The T T value corresponding to an optimal performance T T n i 0 p t The TT value corresponding to an optimal total throughput xi Acknowledgments I would like to express my sincerest gratitude to my supervisor, Dr. Cyr i l Leung. His in-depth knowledge and serious working attitude inspired me and affected me a lot during my graduate studies. His inexhaustible encouragement, constructive suggestions and great patience helped me to overcome all the difficulties I faced on my thesis working. I would also like to thank my parents and my husband. Their constant encouragement and dedicating support are precious treasure to me. Finally, I want to thank my friends and fellow students, who discussed study topics with me arid gave me so many valuable comments. Chapter 1 Introduction 1 Chapter 1 Introduction The primary objective of this thesis is to study the employment of C S D S (Channel State Dependent Scheduling) schemes in improving the performance of wireless communication networks such as wireless L A N s . In this chapter, a brief description of relevant previous works, motivation and contributions of the research project is given. 1.1 Background Wireless L A N has gained in popularity since it was introduced in the late 1980s. It is an attractive alternative to replace wired L A N at hotspots such as airports, office buildings, and residential areas especially when wires and cables are difficult to deploy and/or the network structure is frequently changing. The I E E E 802.11 standards were first published in 1997 by I E E E 802.11 committee and some of the sub work groups are still working on new drafts. Now most of the Wireless L A N products support the 802.11b and g standards. The key difference between them is that 802.1 l g supports higher data rates. In this thesis, we focus on the 802.11b standard. The 802.11 W L A N network architectures can be ad-hoc or infrastructure, the latter being usually deployed at hot spots, and is the only type of architecture considered in this thesis. The default scheduling scheme in 802.11 W L A N is FIFO (First Come First Out); it is simple and uses only one queue for all incoming packets, and yields good performance in a wired network. However, in a wireless environment, due to the time-varying channel capacity and the location dependent bursty errors, a F IFO scheme w i l l cause a H O L (Head of Line) blocking effect under the control of retransmission mechanism applied in 802.11 M A C protocols, and lead to a large number of Chapter 1 Introduction 2 retransmissions and low channel efficiency. This is typically not a problem in a wire-line network because error frames tend to occur at random. In order to improve the performance of the system, scheduling mechanisms based on channel state information - C S D S (Channel State Dependent Scheduling) have been proposed in the literature. In the next section, we w i l l briefly introduce fair packet-scheduling algorithms proposed for wired networks and the problems encountered when employed in wireless networks, C S D S schemes are introduced followed by a description of how they operate. 1.2 CSDS Based Scheduling Schemes In a W L A N , the radio channel is shared by multiple mobile terminals. Therefore, efficient utilization ofthe wireless channel is important. Two common objectives of a wel l -designed scheduling algorithm focus on two subjects: i) fairness, or controlled sharing of the channel among multiple packet streams, and ii) high system throughput. In the literature, fairness can be evaluated by throughput discrepancy or channel capacity (i.e. allocated service time in W L A N s ) discrepancy [38], [6]. In this thesis, we w i l l adopt throughput based fairness constraint. And a popular fairness index which is used to measure fairness performance w i l l be introduced in Chapter 3. Besides, fairness can refer to both long-term and short-term. Generally, short-term fairness implies long-term fairness, but not vice-versa. In this thesis, both are considered. Many efficient and fair packet-scheduling algorithms have been proposed for wire-line networks. Among them are Generalized Processor Sharing (GPS), Packet-by-packet GPS Chapter 1 Introduction 3 also known as Weighted Fair Queuing (WFQ), Start-Time Fair Queuing (STFQ), Worst-case Fair Weighted Fair Queuing ( W F 2 Q , or called W G P S ) , Self-Clocked Fair Queuing (SCFQ), etc. Most of the proposed packet scheduling algorithms imitate fair scheduling scheme for stream traffic - Fair F low Queuing (FFQ), and try to obtain a similar fairness performance. Another popular scheduling scheme family, which is widely used in a packet switching environment, is Round Robin (RR)-based. Its advantages include simple implementation and relatively good performance. However, these scheduling algorithms proposed for wire-line networks cannot be directly applied to wireless networks due to the time-varying, location dependent channel capacity. Therefore channel-state-dependent packet scheduling (CSDS) mechanism was proposed [1] to improve channel utilization. The basic idea behind C S D S is to consider channel conditions when selecting the Head-Of-Line (HOL) packet for next transmission. Since links between different pairs of communication terminals (typically between base station/access point and mobile stations) undergo independent fading and exhibit bursty error characteristics, the probability of success transmission wi l l be enhanced by avoiding the transmissions to those terminals suffering a deep fading channel condition and furthermore channel utilization can be enhanced. Architecture of a CSDS-based scheduler is shown in Figure 1.1. Besides a high throughput, service fairness is often an important system requirement in a wireless network. For example, users located near the Access Point (AP) or Base Station (BS) w i l l generally earn better channel states than distant ones. Hence, in the long run users Chapter 1 Introduction 4 near the A P may obtain much more services. To remedy this problem, fairness control is needed in the scheduling mechanism. C S D S can be combined with many different fair packet-scheduling algorithms used in wire-line networks such as WFQ-based, RR-based schemes etc. to achieve somewhat fairness performance. Per-destination queues Channel State information Figure 1.1: The architecture of a C S D S based scheduler There are two common methods for combining C S D S and proposed fairness policies: "threshold-based" and "weight-based". In "threshold -based" schemes, the channel state is used as a transmission threshold and the combined scheduling algorithm selects H O L according to its rule among queues eligible for transmission. Examples include I W F Q [2], and W D R R [5]. Usually they can guarantee long-term fairness among all sessions through service compensation, and short-term fairness among error-free sessions. In order to further provide graceful service degradation for leading sessions, in other words, control and obtain a reasonable compensation period, some of such schemes (CIF-Q [3], W D R R [5], L T F S [4]) also involve special-designed compensation mechanisms. The disadvantages of threshold-based schemes include computation overhead and complexity. Chapter 1 Introduction 5 In a "weight-based" scheme, the channel state is considered as a part of transmission weight; the other part of weight is decided by the employed scheduling algorithm. The next packet to be transmitted is the.HOL packet from the queue with the highest weight. Examples include OSP [6], W C F Q [7], which can achieved a tradeoff between fairness and total throughput. N o special compensation mechanism is involved. Generally, the fairness measures used in these schemes are long-term fairness. At one extreme is a scheme in which the channel state completely determines the weight, results in maximum total throughput. A t the other extreme is absolute fairness when the weight is completely determined by the fair scheduling algorithm. Since threshold-based combination method is simple and can achieve good total throughput, our proposal w i l l be developed based on that method. 1.3 Motivation The basic idea behind C S D S scheduling is that users with better channel states can obtain higher bandwidth, so that the total throughput is enhanced. However fairness may be sacrificed. The threshold-based C S D S schemes mentioned above improve fairness compared to the original scheme in [1] by means of various service compensations. However, they are too complicated to be deployed in 802.11 W L A N . A suitable scheduling algorithm should be designed based on the features of 802.11 W L A N M A C layer and should not modify it or add much scheduling computation overhead since an important feature of the I E E E 8 0 2 . i l standards is a very simple M A C layer. Moreover, the main traffic type in a W L A N system is packet data traffic, which requires high reliability. Therefore in the 802.11 standards, retransmission at the M A C layer instead of higher layer is provided to support fast recovery from errors. If the scheduling scheme applied in the system is too complicated, the efficiency Chapter I Introduction 6 of the whole system w i l l be greatly reduced. Therefore, a simple and efficient scheduling scheme with good total throughput and fairness performance is especially suitable for such a system like W L A N . T i l l now, most of the proposed schemes involve complicated compensation schemes to guarantee long-term fairness among sessions and incur a high computation overhead cost. Returning to the fairness problem mentioned in the previous section, namely users near the A P receive more service; a simple and direct solution about compensation mechanism is to consider "relative channel state" or "average channel state" (referred to as Sav) of each user. "Average channel state" here means average S N R level of the user in a relatively long period. Under the assumption of infrastructure W L A N with fixed or very slow moving users, Sav is mainly determined by the distance between the A P and the user. Because every user has independent position and channel characteristics, the Sav should be a reasonable parameter to decide how much compensation it should obtain under good channel states. For instance, a distant user with a certain channel state (expressed by S N R level at the receiver) should have a higher priority for transmission than a nearby user with the same S N R level. This compensation mechanism can be applied together with other scheduling algorithms like Round Robin. Hence, our goal is to design an algorithm which provides good fairness while maintaining a high overall throughput by taking the Sav into account. 1.4 Thesis Contributions In this thesis, first part of the contribution is the proposed compensation mechanism, which is based on the average S N R (Sav) of a fading channel and is simple to implement. Chapter 1 Introduction 7 The principle of the compensation mechanism is to provide higher priority to users with low Sav i f they have similar current S N R level with users with high Sav. To achieve this target, a modified S N R level S' o f the current S N R S is calculated based on Sav of that channel and a certain value of current S N R with a lower Sav wi l l results in a higher S'. For S > Sav, S' > S, otherwise S' < S. Moreover, a parameter cp is employed to control the amount of compensation, which increases with (p. The compensation mechanism is proposed to improve the long-term fairness. However, short-term fairness is more important in a real network. Therefore, short-term fairness is used as a criterion to evaluate the performance of scheduling schemes in this thesis. The second part of the contribution is that three CSDS-based scheduling schemes are developed for 802.11 W L A N network based on the proposed compensation mechanism. A l l three schemes are simple for implementation and manage to yield an improved short-term fairness performance and maintain a high total throughput simultaneously. Among the schemes, M a x S* attempts to maximize the total throughput by selecting the H O L with the best modified S N R value S'; I W R R scheme also employs the compensation mechanism, but in this scheme the modified S N R level S' acts as transmission weight. IWRR+Max S* represents a combination of the above two schemes to benefit from the advantages of both. In this thesis we mainly studied the effect of Doppler frequency (fd), short-term fairness window size (fwin) and the number of users N u to the performance ofthe preposed schemes. The simulation results show that IWRR+Max S* yield no worse performance than the traditional scheme C S D S + R R for any value of the product Pdw = fd x fwin (e [1, 100]) and N u (e [2, 20]) and can achieve substantial improvement for Pdw > 5, and any Nu . For N u = Chapter 1 Introduction 8 8 and Pdw > 25, the three proposed schemes can achieve similar performance, which shows a fairness improvement to C S D S + R R over 15%. A n d the fairness improvement increases with the decreasing of Nu . We did not compare their performance with that of the default F IFO scheme in W L A N since CSDS+RR shows much better performance than FIFO did in [1]. In conclusion, our proposed scheduling schemes can supply good performance not only for long-term fairness but also for short-term one; the benefit of high throughput from C S D S is maintained as well . " v" 1.5 Overview of Thesis The rest of this thesis is organized as follows. In Chapter 2, related works of the C S D S based scheduling schemes in the literature are presented. Then the proposed compensation mechanism and scheduling schemes based on the compensation mechanism are developed in Chapter 3. Additionally, channel model and fairness index definition are also included. In Chapter 4, simulation system is build up to imitate a real 802.11b W L A N environment. Verification by simple tests in O P N E T is also involved. Chapter 5 presents simulation results to investigate the effect of Doppler frequency, fairness window and the number of users to the performance of the proposed schemes and the performance comparison is also studied. Finally, a conclusion of the presented work and suggestions to future work are given in Chapter 6. Chapter 2 Related works 9 Chapter 2 Related works The original C S D S scheme was proposed in [1] for wireless L A N network. The key feature of C S D S is to avoid bursty errors at the link layer instead of relying on the transport or application layer for error recovery to improve overall throughput performance. The basic idea behind C S D S scheme is to defer the transmission effort when the channel is in a bad state and the destination is currently unreachable, to proceed with transmission to receivers enjoying good channel states. A C S D S scheduler consists of two modules. One is channel-state measuring or predicting module. The other is scheduling (dequeuing) module. The former performs the measurement of the channel state. Based on the information, the prediction of channel state is made for the next packet selection. This step is not necessary when measurement value is obtained frequently or the channel is changing slowly relative to transmission rate. The scheduling module implements dequeuing decision. C S D S schemes always avoid or limit allocating resource to queues in poor states. In [1], C S D S is combined with R R (Round Robin), E T F (Earliest Timestamp First) and L Q F (Longest Queue First) scheduling algorithms. A threshold is used to indicate i f a channel state is "Good" or "Bad". If "Good", transmission of the session is allowed based on the rule of the scheduling algorithm, otherwise transmission is delayed. No compensation mechanisms is designed in [1]. Typically in proposed schemes, C S D S works with fair packet scheduling algorithms proposed for wired networks or with QoS requirements to achieve acceptable resource allocation. Some of them also include compensation mechanisms to improve fairness. Chapter 2 Related works 10 Proposed C S D S based scheduling schemes in the literature can be classified as "threshold-based" or "weight-based". 2.1 Threshold-based CSDS scheduling algorithms In "threshold-based" schemes, the policy is similar to the C S D S scheme in [1]: channel state is used to set a transmission threshold (TT). If the state is good enough, the session transmits as determined by the scheduling algorithm. Otherwise transmission is denied, except when none of users is above TT. A n y of the scheduling algorithms can be employed: R R based, GPS based, and QoS requirements based schemes, such as W R R (weighted round robin), D R R (Deficit Round Robin), W F Q (Weighted Fair Queue), W F 2 Q (Worst-case Weighted Fair Queue), C B Q (Class-Based or Credit-Based queue), L Q F (Longest Queue First), etc. Some proposed schemes [2], [3], [4], [5] use compensation mechanisms, which implement service compensation to lagging sessions. Generally, scheduling algorithms such as W F Q and L Q F that include tag or stamp factor to indicate historical service allocation can naturally implement service compensation. Therefore, the compensation mechanism designed in these schemes is to achieve a controlled compensation period. For example, in I W F Q [2], the H O L (head of line) of each queue is tagged with a time stamp, which is a factor in queue selection. If a queue is denied transmission due to poor channel state, its H O L time stamp w i l l be unchanged and the queue wi l l be given higher transmission probability when its channel state rises above TT. However, i f not controlled, the compensation period may cause starvation of in-sync or severe penalty to leading sessions, which is called Chapter 2 Related works 11 "blackout" problem. To ease the problem that can lead to poor short-term fairness, a compensation mechanism was proposed in [2]. Typically, this kind of compensation mechanism is complicated but can achieve good fairness performance. However, scheduling algorithms such as R R do not include any compensation policy. Therefore, some of the compensation mechanisms designed in RR-based schemes attempts not only to improve short-term fairness performance of pure R R but also to achieve a reasonable compensation period as in IWFQ. Others are simple and suboptimal but efficient and working well with the R R policy. Some of the threshold-based schemes are introduced respectively below. IWFQ (idealized wireless fair queuing) uses W F Q or W F 2 Q and includes a compensation policy to control the "blackout" problem in an error-prone scenario. I W F Q and W F Q are identical in error-free scenario. To avoid unlimited channel occupation by lagging sessions and provide a reasonable compensation period, I W F Q sets bounds on the lead and lag of each session. However, this mechanism does not guarantee graceful service degradation for leading sessions. CIF-Q (Channel-condition Independent packet Fair Queuing) and L T F S (Long-Term Fairness Server) were proposed in [3] and [4] to achieve more acceptable service degradation for leading sessions. CIF-Q uses Start Time Fair Queuing (STFQ) as its error-free service model while L T F S can use any wired fairness scheduler as its error-free model. CIF-Q uses a parameter to control the rate at which a leading session gives up its lead to a lagging session, while L T F S reserves bandwidth for compensation in advance. Chapter 2 Related works 12 WDRR+CSDS [5] scheme uses W D R R (Wireless Deficit Round Robin) as its error-free fairness algorithm, which is derived from D R R (Deficit Round Robin). Combined with C S D S , the compensation mechanism enables the scheduler to provide short-term fairness among sessions that perceive a clean channel, long-term fairness among all sessions, and graceful service degradation among sessions that received excessive service. W D R R uses a parameter to keep historical records of throughput obtained by each session so that long-term fairness can be achieved among all sessions. Furthermore, to achieve graceful service degradation of leading sessions and further to obtain better short-term fairness in error-prone scenarios, an additional parameter is used to control the maximum amount of service a leading session has to give up in each round. In [9] a simple compensation mechanism was proposed for T C P traffic short-term fairness improvement in the 802.11b W L A N system. The proposed scheme is an L L C (Logical Link Control)-layer algorithm that aims at enhancing fair access to the medium for every user, by awarding longer transmission opportunities to sessions that experiences short channel failures. The award mechanism works by monitoring the successful medium accesses over a measurement window, and allowing short or long transmission bursts depending on the session's recent history (which means, how much service it received during the previous measurement window). Same as [1], the applied scheduling algorithm is Round-Robin. Upon entering a good channel state, a queue that was in a bad state is rewarded with the possibility to send to the M A C layer a maximum number of back-to-back frames (hereinafter refer to as "TXburst"), Chapter 2 Related works 13 proportional to its forced silence period. Based on the fractional throughput a queue enjoyed in the previous measurement window, the TXburst length w i l l be set longer (as a reward) or shorter (as a penalty) than the default -value to improve short-term fairness performance among sessions. Although this simple compensation mechanism cannot guarantee tight short-term and long-term fairness among users, simulation results show a much better fairness performance than with the default scheduling scheme (FIFO) of W L A N . Moreover, in a simple network like 802.11 W L A N , a scheme which is uncomplicated and efficient is especially expected. Among all proposed schemes, [9] is most closely related to our research. 2.2 Weight-based CSDS Scheduling algorithms In weight-based C S D S , the channel state is transformed to a transmission cost (or weight), and the scheduling algorithm contributes the other part of whole weight. The H O L selection decision is based on the total weight, and the next packet to be transmitted is the H O L with the highest weight. These schemes can attempt to achieve balance between fairness and total throughput performance. One extreme example is C S D S with optimal total throughput performance. Here, the H O L selection is based solely on the channel state of each queue. A t each packet transmission moment, the queue with the best channel state is selected (referred to as M a x S, where S is the S N R level of the channel). Max S yields the highest total throughput but provides no fairness feature. At the other extreme is fairness scheduling only, where the H O L Chapter 2 Related works 14 is selected only by the applied fairness scheduling algorithm. In this case, fairness as defined by the adopted scheduling scheme can be guaranteed at the cost of low total throughput. Some proposed schemes that attempt to strike a balance between fairness and total throughput are described below. The objective of O S P (Opportunistic Scheduling Policy) [6] is to maximize the average system performance while satisfying pre-determined fairness constraints. A n opportunistic policy is defined to achieve this goal: e , (C/) = argmax(L/ l .+v;) (2.1) i where £/,• is the channel-state-dependent throughput value achieved by user i, v* is the offset used to satisfy the time-fraction assignment constraint, i.e. for any e>0: P{Ui+v]-e>m^M(Uj+v]-e)}<ri<P{(UiWi>m^j^Uj+v)^ (2.2) where r,is the time-fraction assignment of user i. B y doing so, P(Q* (U) = /) = rt is satisfied so that fairness constraint is guaranteed. Under this constraint, the "relative-best" user is scheduled for transmission. Therefore, the opportunistic transmission scheduling policy maximizes the average network performance under the fairness requirement (e.g. fair sharing). W C F Q (Wireless Credit-based Fair Queuing) [7] applies C B F Q (Credit Based Fair Queuing) as its fairness scheduling algorithm. The packet selection criterion is to select the H O L packet that satisfies: / = a r g m i n I ; ~ ^ ' + C / ' (2.3) Chapter 2 Related works 15 where Z, is the H O L packet length for session i; Kj is the credit counter, which calculates the accumulated credit according to C B F Q to provide a balanced bandwidth sharing among sessions; £/, is the transmission cost; and (£», denotes the assigned weight for session i. A t each decision moment, scheduler w i l l select the H O L with the lowest channel cost (i.e. best channel state), unless one session's cumulated credits outweigh a poor channel state. Therefore, transmission cost function (which decides transmission cost under each channel state) can control the tradeoff between fairness performance and total throughput. [7] shows that a transmission cost function can be derived to satisfy a given fairness requirement so that a loose or tight short-term fairness is guaranteed. However, due to the model's complexity it is generally difficult to derive such a function in practice. SPS (SNR-based Packet Scheduling) scheme [8] can be viewed as a C S D S + W F Q policy. Each S N R level (dB) has a corresponding weight according to a piece-wise linear function. A good channel state (i.e. high S N R level) w i l l be given a heavy weight. A s in W F Q , the H O L with the earliest virtual finish time w i l l be selected . for transmission. However, the virtual finish time is calculated by: F" (0 = max(v(0, (0) + A /r* (2-4) where F" (t) is the virtual finish time of nth packet in session i, v(t) is the arrival time of packet n, r" is the bandwidth currently allocated to session rbased oh its weight, L , i s packet length of session i. Note that r, is the deserved bandwidth of session i. Therefore, an active session's allocated bandwidth is dynamic according to its channel state instead of constant as in W F Q . The difference between SPS and the threshold-based I W F Q scheme is that SPS does not avoid transmission when user's channel state is low, but allocate low bandwidth Chapter 2 Related works 16 (priority) instead. Therefore, a tradeoff between total throughput and fairness is achieved. The difference between SPS and W C F Q is the transmission cost (weight) function, which in SPS is identical for all sessions. In all the weight-based C S D S schemes discussed above, the sessions with relatively poor channel, states can still get a transmission opportunity, when the fairness control factor overweighs the channel condition factor. The main difference among the schemes is the function used to select the H O L for transmission. N o special compensation mechanism is involved. Chapter 3 New CSDS scheduling algorithm developments 17 Chapter 3 New CSDS scheduling algorithm developments Most of the proposed C S D S scheduling algorithms described in Chapter 2 are designed for cellular networks, and are not very applicable to W L A N s . The common characteristics of several proposals for W L A N are simplicity and low overhead of scheduling computation, e -g- [9], [1]. Due to its simplicity, the network can be rapidly developed and widely deployed. A low protocol overhead also results in high system efficiency. The limitation is that the network does not provide tight QoS guarantees. GPS based scheduling algorithms seem too complex for W L A N , due to the need for time stamping of each packet, updating the virtual time function, and so on. R R based scheduling policies are simpler and adopted in our proposed scheduling algorithm for W L A N . In this chapter we describe the new C S D S based scheduling algorithm in detail. But before this we briefly discuss the 802.11 standards and the wireless channel models used in W L A N simulation. Then a new compensation mechanism and the corresponding scheduling schemes are introduced. 3.1 Overview of 802.11b WLAN [19] 3.1.1 Network architecture A basic service set (BSS) is the basic unit ofthe network. It consists of wireless stations (STA) that compete for access to the shared wireless medium. Two types of W L A N can be deployed: Ad-hoc or infrastructure. A n Independent B S S (IBSS) is an ad-hoc network, in which all stations are able to communicate with each other directly. In an infrastructure Chapter 3 New CSDS scheduling algorithm developments 18 W L A N the BSS also includes an access point (AP) that provides access to backbone distribution systems. A P is also a S T A , but it simultaneously owns access ports to the fixed network. In reality, its function is similar to that of a base station (BS) in a cellular network. Infrastructure networks are currently more popular than ad-hoc networks, because they meet the needs of coverage networks of large size and complexity. In this thesis we w i l l focus on infrastructure networks with one A P and several stationary or low mobility STAs . Transmission from the A P is referred to as downlink; transmission from any other STAs is termed uplink. In infrastructure network, communications are only allowed between the A P and the non-AP STAs . 3.1.2 M A C sublayer functions The Distributed Access Function (DCF) is the basic protocol of the 802.11 M A C layer and uses a carrier-sense multiple access/collision avoidance ( C S M A / C A ) [19] protocol as the fundamental access method. A n optional access method is the Point Coordination Function (PCF), which is deployed at the A P and resolves contention by polling stations in turn. In this thesis, only D C F is studied. It operates as follows: when a S T A is ready to transmit a M A C protocol data unit ( M P D U ) , which is a complete data unit passing from M A C layer to P H Y layer, it must sense the medium to determine i f it is busy. If the medium is idle for at least the D C F interframe space (DIFS) duration, transmission is allowed. If the medium is sensed busy, after detecting the channel to be idle without interruption for a period equals to DIFS the S T A wi l l backoff for an additional period of n times time_slot, where n is an integer randomly selected over the interval [0, C W (Contention Window)] and iime_slot is decided by the correspondingly named P H Y characteristic (In Direct Sequence Spread Spectrum Chapter 3 New CSDS scheduling algorithm developments 19 (DSSS), time_slot = 20 us). The size of C W , which is bounded by a maximum value CWmax, is doubled after each unsuccessful transmission to reduce the collision probability. Following each successful transmission, C W is reset to the minimum value C W m i n . During backoff period, a S T A decreases its backoff counter by one i f the medium is idle for a time_slot and keeps it frozen when the medium is busy. When the backoff counter reaches zero, the S T A w i l l transmit its M P D U immediately. If the M P D U is received successfully by the destination S T A , an Acknowledgement ( A C K ) frame w i l l be sent back to the source after a Short Interframe Space (SIFS). In the case of unsuccessful transmissions requiring A C K , backoff w i l l begin at the end of A C K timeout interval, i.e. after an Extended IFS (ELFS) period following detection of a frame that was not received correctly. Therefore, after a successful or unsuccessful transmission, another contention procedure w i l l begin for all STAs ready for transmission. Retransmission may be triggered at the source S T A i f it does not receive an A C K within A C K t i m e o u t period. A frame is dropped when its retry limit is reached. The parameters dotUShortRetryLimit and dotlILongRetryLimit are specified in the I E E E 802.11b W L A N standard as retry limits corresponding to different frame lengths. When a M A C frame of length is greater than dotl IRTSThreshold, which is the threshold of frame length to apply R T S / C T S method, dotl ILongRetryLimit is used; otherwise dotl IShortRetryLimit is applied. The values of the three parameters are defined as P H Y management information base (MLB) attributes and are given in Annex C of [19], which equal to 7, 4, and 3000 bytes respectively. Chapter 3 New CSDS scheduling algorithm developments 20 D C F also includes an optional transmission method of R T S / C T S , which requires transmission of Request To Send (RTS) and Clear To Send (CTS) frames prior to the transmission of actual M P D U . R T S / C T S can alleviate the hidden termimal problem and can reduce the waste in transmission time of long frames caused by collisions. However the efficiency is reduced for short frames. R T S / C T S mechanism is not used in this project since it is optional. -3.2 Wireless channel model A s mentioned above, a wireless channel has properties which are quite different than wire-line channel. Fading wi l l cause a time-varying channel capacity. There are various mathematical models to simulate the properties of a wireless channel. The Gilbert model [17] is a wireless channel model that is frequently used in the study of scheduling algorithms. It is a discrete-time Markov chain with two states: error-free ("good") or error-prone ("bad"). The channel moves between the two states according to a certain transition probability matrix. The bad state (good state) is often assumed to have a packet error rate (PER) = 1(0). In a variant of Gilbert model, packets are successfully received with high probability when channel is in Good state (PER w 0 ) and with low probability when channel is in Bad state (PER w 1). Another popular variant is Finite-state Markov Chain with n states. It has been shown in [18] that this model can simulate Rayleigh fading channel well . Each of n states corresponds to a small range of signal levels. A nxn matrix of state transition probabilities characterizes the wireless channel behavior. One simplified version is a birth-death chain in which only transitions between neighbor Chapter 3 New CSDS scheduling algorithm developments 21 states are possible. Usually the B E R for each state depends on the corresponding signal level range. Since the channel states in Markov chain based channel models are discrete, it seems they are too simple to be accurate. Therefore, in our simulation we use a practical multipath fading channel model as in [25], in which the signal power level at the receiver is given by: where Po is the power received at a close-in reference point at a small distance do from the transmitting station, dr is the distance from the receiver to the transmitter, and B is the attenuation factor. Note that where D is the largest physical linear dimension of the antenna and X is the wave length. Additionally, do is chosen to be smaller than any practical distance used in the mobile communication system [36]. The first two terms model the channel path loss by propagation. Fr is the envelope of the time-correlated multipath fading signal and is simulated by Jake's model [26]. In this model, according to mathematical analysis, a complex signal T(t), which is expressed by: Pr=P0-\0/3\og(dr/d0) + Fr (3.1) 2D2 (3.2) + e ] + e ,N0 =1(1^-1) (3.3) is approximately a complex Gaussian process i f N is big enough based on the Central Limit Theorem, so that \T\ is Rayleigh distributed. Typically TV > 6 is sufficient to obtain a good Chapter 3 New CSDS scheduling algorithm developments 22 Rayleigh approximation. No is the number of frequency components involved and (om is the maximum Doppler frequency. Figure 1.7-2 in [26] shows the block diagram of such a simulator. No low-frequency oscillators with frequencies equal to the Doppler shifts comcos(2itn/N), n=l,2,...,No , plus one with frequency com are used to generate signals frequency-shifted from a carrier frequency coc. The phases are uniformly distributed. In our simulation model, No=38 is used as the model in [31]. The advantage of this Rayleigh fading channel model is that it can provide continuously changing channel states so as to achieve better precision than those model mentioned above. Similar to them, S N R level is used to represent the dynamic channel states. 3.3 Fairness definition and Compensation mechanisms 3.3.1 Fairness definition Fairness can be measured by service allocation difference. There are at least two ways to describe service difference, one is by throughput, and the other is by service time. Throughput discrepancy is a more popular way for fairness measurement. Therefore, we use traditional throughput difference as the fairness measure. Fairness can be short-term or long-term. [27] gives definitions of both: a system is said to be long-term fair i f all users gain equitable access to its resources in the long run, although there may be transient periods of unbalanced access; short-term fairness, instead, refers to equitable sharing of resources in the short run. Usually, short-term fairness implies long-term fairness, but not vice-versa. Actually, the difference between them is not precisely defined. In most cases, it depends on the type of applications. For example, in [9] 10 s is specified as the Chapter 3 New CSDS scheduling algorithm developments 23 fairness window for short-term fairness for T C P application. In [1] it is also noted that T C P applications can be tolerant to several seconds delay. However, this may be different for other applications. In our work, we assume T C P traffic, and short-term fairness is measured in seconds. The fairness index used in this thesis to indicate fairness performance is the definition proposed by Jain et.al in [28], which is widely used by researchers. The definition is as follows: in a system allocating resources to n contending users, such that user i receives an allocation x„ the fairness index is defined as /(*) 12 x ; > 0 ,x = (xi, X2, .... x„) (3.4) n This index measures the "equality" of user allocation x. If all users get the same service amount, i.e. x,-'s are all equal, then the fairness index is 1, which means the system is 100% fair. A s the disparity increases, fairness decreases and a scheme which favors only a selected few users has fairness index near 0. 3.3.2 Compensation mechanisms A s described in Chapter 2, simple fairness scheduling algorithms like Round Robin are preferable to complicated scheduling algorithms for W L A N system. To support good fairness performance, a compensation mechanism should be included in such a scheduler. Chapter 3 New CSDS scheduling algorithm developments 24 The goal of a compensation mechanism is to improve fairness performance between users, not only long-term fairness but also to support good enough short-term fairness. Generally a compensation mechanism can only guarantee long-term fairness in an error-prone environment because short-term fairness is related to fluctuating channel states. In I W F Q [2], long-term fairness in error-prone conditions is guaranteed whereas short-term fairness is supported only among error-free users. Short-term fairness in error-prone cases cannot be guaranteed due to the time-varying nature of the channel. In [2] for error-prone scenarios, a lower bound is obtained on the short-term throughput, which depends on the accumulative period a user stays in good state during the measurement interval. Subsequently proposed WFQ-based compensation mechanisms support graceful service degradation of leading sessions at the cost of more complexity and worse short-term fairness. In such a scheduling policy, not only is timestamp employed, but also leading and lagging sessions need to be marked to calculate service gap. RR-based scheduling algorithms need specially designed compensation mechanism to improve fairness performance, such as W D R R [5] etc. Performance comparable to W F Q -based compensation mechanisms can be achieved. The drawbacks are complexity and computation overhead. In T C P short-term fair scheme [9], R R is applied as the basic scheduling rule and a simple compensation mechanism is proposed to improve short-term performance of T C P applications. This compensation mechanism is simple to adopt in a W L A N environment. Compensation is implemented based on service allocation during the previous period of Chapter 3 New CSDS scheduling algorithm developments 25 fairness measurement window. Users who obtained lower throughput than average during the previous measurement window are considered as lagging and are rewarded by additional packet transmissions during the current window. On the other hand the leading sessions w i l l be penalized by reducing their transmission times. The number of additional packets transmissions are chosen empirically from simulation results. This mechanism is simple and easy to deploy in a W L A N system, and can improve the short-term fairness. However it w i l l not provide a big improvement in fairness, especially long-term fairness because compensation depends only on the service allocation in the previous measurement window of 10 seconds, which means that service is considered fair before that. This is in contrast to more complex compensation mechanisms in which the service gap is determined by service obtained in the whole past. In this thesis we propose a simple compensation mechanism for 802.11 W L A N . The basic idea is to give high priority to users with low average S N R level (Sav) when they have similar channel state to those users with high Sav. This target can be achieved by calculating "Modified SNR Level" S' for each user. In other words, according to the relative channel condition (difference between measured S N R level and the average S N R level), the measured S N R level is assigned a weight - "modified SNR level" according to: s ; = s , + ( - ^ m - s n (3.5) where St and S"v are the measured and average S N R level of user i respectively and <p is the compensation control parameter. The difference between measured and modified S N R increases w i t h ^ . Thus a user with a low average S N R level, which currently enjoys a good channel state, w i l l be assigned a high priority to transmit provided that the long-term fairness Chapter 3 New CSDS scheduling algorithm developments 26 performance can be improved. From (3.5), i f S"v < SjV and 5,. = Sj , S] > S] , also i f (si-srr<o,s'i><si. 3.4 New scheduling schemes Based on the proposed compensation mechanism, new scheduling schemes can be designed for 802.11 W L A N networks which are expected to have better fairness performance while maintaining a high total throughput. In this thesis three new scheduling schemes are proposed. The first one is based on "Maximum S N R first" and the second on Weighted Round Robin (WRR) . The third scheme is a combination of the first two schemes. 3.4.1 Max S* In this scheme, the scheduling algorithm is the simple "Max S" policy. The original algorithm selects the H O L packet with the best channel state each time. This yields an optimal total throughput performance at the cost of poor fairness performance. The M a x S* scheme is the Max S algorithm combined with the proposed compensation mechanism to improve fairness performance. Specifically, each time Max S* chooses the H O L packet with the highest Modified SNR level S' for next transmission. A s discussed in the previous section, the compensation mechanism gives precedence to a user with a low average S N R when it has a similar channel state to a user with a high average SNR. Hence it is possible that the compensation results in a decrease in total throughput. To avoid large total throughput degradation, a threshold is used to control the compensation procedure. If Chapter 3 New CSDS scheduling algorithm developments 27 the measured S N R level (referred to as S) is higher than the threshold, the modified S N R level S' w i l l be used; otherwise, the S N R level is not modified. The Max S* scheme thus operates as follows: A t each scheduling instant, the queue with the highest S* is selected for transmission. If there is more than one such queue, one is selected in a R R fashion. The threshold is chosen based on the packet error rate (PER), which is related to packet length. In our simulations the threshold is set to 10.5 dB, at which the P E R is close to 10"2 for 8800 bits packet length. This P E R value is considered as a bound of good channel state in [39]. If the channel state of a user with low average S N R is good enough (i.e. S is higher than the threshold), it can receive service compensation from those with a high average S N R so that better long-term fairness performance can be achieved and simultaneously a high total throughput performance can also be expected. 3.4.2 Improved W R R ( I W R R ) ' "As in the original GSDS scheme [1], backlogged users are selected in a Round Robin fashion. However, in I W R R , a Weighted Round Robin including the proposed compensation mechanism is applied to obtain "a better fairness performance than the original C S D S + R R scheme [1]. In contrast to traditional W R R schemes [41], a dynamic weight is used instead of a predetermined one: the weight of each user is calculated per packet transmission slot Set S > threshold S < threshold (3.6) Chapter 3 New CSDS scheduling algorithm developments 28 according to its current channel state. For user i whose S N R level is higher than TT, the weight is calculated by weight j = (3.7) S,ZTT For a user with S N R level lower than TT, its weight is set to 0. If the weights of all users are zero, a user with the best channel state is selected. If there is more than one such user, the tie is broken by R R method. The implementation of this scheme is as follows: A round is adopted and its length, Lround, is fixed to a certain number of packet transmission slots. For each backlogged user i, a counter C, is employed to indicate the number of packet transmission slots used so far by user i in the current round. A t the beginning of each round, C, = 0. In each round, users are served in cycles; during each cycle, user i can be allocated at most one packet transmission slot, i f user i has a weight greater than 0, its allocated packet transmission slots w i l l be calculated as At = Lround *weighth then i f C, < Ai, the user w i l l be allowed for transmission and C,- is increased by one. If none of users satisfies above conditions of C, < Ai, a user with none zero weight is selected in order. Otherwise, the user with the highest S is selected since in this case all weights are zeros. The flow chart of the operation of I W R R is shown in Figure 3.1, where "picked" is the queue which obtained previous transmission slot, N is the number of users, R is the counter of a round. Chapter 3 New CSDS scheduling algorithm developments 29 Init R=Lround; A l l Q = 0 / = picked i = i+ 1; i f i > N , i = mod(N) 1 C,= C ,+1; Ui transmits i R = 0? N Y Pick up the first user j with weight j > 0 Pick up user k with the best S R = R - 1 Figure 3.1: Flow chart of I W R R scheme Chapter 3 New CSDS scheduling algorithm developments 30 3.4.3 Combined scheduling scheme ( I W R R + M a x S*) In order to inherit the advantages of the two above schemes, a combined scheme is proposed in which the weight of user i is calculated in the case that some users have channel states above T T as follows: weight , = W - ( 3 . 8 ) i where i o ( r 0 , - r ) ^ + (i + / ) s ; r e [ 0 > 1 ] S i >_ T T B:={ ' L ' J (3.9) S,ZTT l o In the case that no user has an S greater than TT, the weight of user i is calculated by (3.8), where Bi=\0(r0A-r)^-+(i + r4)Si ye[o,i] (3.10) Note that i f none of the users has an S above the TT, the weight of each user is not set to zero but is decided by S. The maximum S wi l l corresponds to the heaviest weight to approximate the rule (selecting the best S when all S < TT) applied in Max S* and IWRR. The first term in the R H S of (3.9) reflects the I W R R , and the second part reflects the M a x S*. W h e n ^ = 0 , the first term is 0, and the second part is a constant 1; the combined scheme approximates R R since all users have the same weight. A s y increases from 0 to 0.1, the first term rapidly overweighs and the scheme approaches I W R R . When / « 0.1, the first term dominates the whole weight and the scheme approximates I W R R . A s y increases to 1, Chapter 3 New CSDS scheduling algorithm developments 31 the scheme approaches Max S*. When y = 1 the first term is again 0 and the second term becomes dominant and the scheme approximates Max S*, in which the weight is determined by the modified S N R 5,; the exponential function amplifies the gap between different £• values. The functions fl = lO(yOA-y) (3.11) f2 = l + r 4 (3-12) are shown as a function of y in Figure 3.2. * : 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 y Figure 3.2: Functions of y in (3.10) and (3.11) : • The implementation of IWRR+Max S* is similar to IWRR. The only difference is that the case that all weights of users are zero wi l l never happen. In the case that all users' channel states are lower than TT, transmission slot w i l l be given to a user whose current At is not zero in a R R fashion. The flow chart of IWRR+Max S* is shown in Figure 3.3. Chapter 3 New CSDS scheduling algorithm developments 32 Init R=Lround; A l l Q = 0 i = picked i = i + 1; i f i > N , i = mod(N) 1 Y C,= C,+1; Uj transmits r R = 0? r N N r Y Pick up the first user j with weight] > 0 R = R - 1 Figure 3.3: Flow chart of IWRR+Max S* Chapter 4 Simulation environment 33 Chapter 4 Simulation environment The most popular W L A N products are based on the I E E E 802.1 l b W L A N standard and are deployed in the 2.4 G H z . The newly proposed 802.1 l g standard in 2003 uses the same frequency band; the only difference is the P H Y layer modulation scheme which allows 802.1 l g to support higher transmission rates. We base the simulations on 802.11b. In the following sections, we describe the function of different modules in the simulation system, following by the verification tests, which were done with the help of Optimized Network Engineering Tools (OPNET) . 4.1 Fading channel module implementation arid verification The channel module is an important component whose output is input to both the scheduling algorithm and traffic sink modules. In our simulation scenarios, the Jake's model is employed to simulate Rayleigh fading channel as described in Chapter 3. Inputs of the module are average received signal level Pr and Doppler frequency fd. In [29], the range of fd is set to 1.34 H z to 80 Hz , corresponding to user speeds of 0.4 mph (0.18 m/s) and 22.4 mph (10 m/s). Pedestrian speeds of 1~2 m/s [30] correspond to Doppler frequencies of 8~16 H z approximately. In our simulation, the maximum Doppler frequency is set to 10 Hz . The Rayleigh fading channel simulator was obtained from [31]. A W G N can also be added to the received signal. To verify the Rayleigh distribution of the channel model, the empirical Cumulative Distribution Function (CDF) of the simulation results in our C++ programming system, which are the samples of dynamically changing received signal level, can be obtained and Chapter 4 Simulation environment 34 compared with the standard function in M A T L A B . Verification parameters are as follows: = 10 Hz , average received power Pr = 20 dBm, simulation time ts = 100 s, sample rate sr = 1250 points/s. The C D F comparison results are shown in Figure 4.1, it can be seen that our channel model simulates Rayleigh distributed signal levels well . 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 CDF function comparison 10 15 signal level Ir 20 1 >."•••'"' ' I 1 1 1 ^ ' _ 1 _ _ . 1 s' 1 / ' ! 1 1 . . . . / . . ! —- ~ — simulation standard — / : j ! / / f | f / / < 1 / / _ n / 1 r 25 30 Figure 4.1: C D F of Rayleigh distributed fading channel Another key feature to evaluate the channel module is the level crossing rate (LCR) function, which shows how frequently the channel state crosses a certain threshold level in the downward direction. Assuming a fixed packet length of 8800 bits, and a transmission rate of 11 Mb/s , the duration of one packet transmission is 800 us. Figure 4.2 shows sample L C R results as a Chapter 4 Simulation environment 35 function of the received power Pr (dBm) for sampling rate sr = 1250, 12500, and 125000 pps (fd = 10 H z and average received power Pr = 20 dBm). These rates correspond to 1 points/pkt, 10 points/pkt and 100 points/pkt respectively. L C R function shows difference for the first two scenarios, which means the channel is changing during a packet transmission period and sampling rate of 1250 pps is not accurate enough. However the function shows no difference when sampling rate is changed from 12500 points/s to 125000 points/s. To ensure accurate simulation P E R results, the sampling rate sr is chosen to be 100 points/pkt, i.e. 125000 pps through all simulation results in this thesis. fd=10 Hz st=100 s avg Pr=20 dBm w CL cr o 10 9 8 7 6 5 4 3 2 1 sr = 1250pps sr= 12500pps sr = 1250000pps A -30 -20 -10 • 0 , '• 10 received power Pr (dBm) 20 30 Figure 4.2: L C R functions of Rayleigh fading channel Chapter 4 Simulation environment 4.2 IEEE 802.11b W L A N simulation system implementation and verification 36 4.2.1 System implementation The simulation is programmed in C++. The network is one BSS with one A P and N user nodes as shown in Figure 4.3. We assume that the distances between A P and user nodes are not identical. Communication is only allowed between A P and user nodes. No direct communication among user nodes is possible. In this thesis the focus is on downlink scheduling policy at the AP, but to simulate a more realistic 802.11 environment with contention, a light uplink traffic load is included. STAs AP Figure 4.3: Topology of an infrastructure 802.1 l b W L A N The simulation system includes two-way communication. A l l the N+l nodes including the A P contend for the shared wireless channel according to the C S M A / C A protocol specified in 802.11 standards. If no collision is detected and the A P gets the opportunity of transmission, downlink transmission is triggered and the scheduling algorithm is employed to select the H O L of queues; i f a user node gets the opportunity of transmission, uplink transmission is triggered. In the uplink direction a FIFO scheduling scheme is employed. A Chapter 4 Simulation environment flow chart of the operation at a node / is shown in Figure 4.4. 37 Figure 4.4: The flow chart of the operation at a Node i The simulation system consists of several modules: node module, which includes traffic-arrival sub module, enqueue/dequeue sub module, and scheduling-algorithm sub module, channel module, and traffic sink module. On the downlink transmission, all modules are involved. On the uplink direction, the node module only includes the first two sub modules. The function blocks of both directions are shown in Figure 4.5 and Figure 4.6. Chapter 4 Simulation environment 38 Traffic arrival Traffic 1 Traffic N Enq/Deq Queue 1 Queue N I I I I A P node o 60 00 a "3 -a CJ JS o 00 Fading Channels Traffic sink Sink 1 SinkN Figure 4.5: Downlink function block of modules Traffic arrival Enq/Deq Queue 1 1 1 Traffic > Node i Figure 4.6: Uplink function block of modules Traffic Sink (AP) Fading channel On the downlink, at each scheduling moment, the algorithm sub module selects the H O L of one queue for transmission based on channel state information. The traffic sink module is used to determine whether the incoming packet is successfully received. If not, a retransmission procedure would be triggered according to .the rule of the scheduling Chapter 4 Simulation environment 39 mechanism. When the retransmission limit is reached, the packet is dropped. Similar events occur on the uplink, except that only one queue exists at each node and the scheduling algorithm is FIFO. The functions of each module in the system are introduced next. • Node module The main function of a node module is to set the backoff counter for channel contention. Each node involves several sub modules. On the downlink, the A P node includes the traffic arrival sub module, enqueue/dequeue sub module and scheduling algorithm sub module; on the uplink, only the traffic arrival and enqueue/dequeue sub modules are needed. • Traffic arrival sub module The traffic arrival module generates - packets. On the downlink, N traffic arrival processes are built up at the A P corresponding to N traffic sinks. In our simulation this function can be ignored since saturated downlink traffic is assumed so that queues are never empty. On the uplink, independent, identically distributed (iid) Bernoulli arrival process is assumed in each node. The sub module also records the retransmission time of H O L in the case of an unsuccessful transmission. If the retransmission limit reaches, the H O L wi l l be dropped, and the information is recorded in this module as well . • Enqueue/dequeue sub module The main functions of the enqueue/dequeue sub module are to set up queues and deal with enqueue and dequeue cases. For the downlink, N queues need to be set up in the A P node. A n enqueue process is triggered by the traffic arrival process and a dequeue process is Chapter 4 Simulation environment 40 based on the output of the scheduling algorithm sub module. In our test scenario, enqueue cases are ignored for downlink traffic since a saturated traffic load is assumed. On the uplink, only one queue is used per node and the enqueue process is triggered by a traffic arrival. Dequeue follows the FIFO rule. • Scheduling algori thm sub module This module is only involved in downlink transmission. For algorithm comparison, Scheduling algorithms including CSDS+RR, MAX S, Max S/Sav, Max S*. Improved WRR (IWRR), and combined scheme IWRR+Max S* are simulated. According to the channel states of the TV sessions, the scheduling algorithm selects an appropriate H O L for next transmission and the results are sent back to the enqueue/dequeue sub module, implementations of the scheduling algorithms are introduced next. • Traffic sink module The traffic sink module is used to determine whether the transmission of the selected H O L is successful or not, based on the current channel state. In the case that N active sessions are in the simulation scenario, N+I traffic sinks are built up for bidirectional communication. The traffic sink module of A P is invoked in an uplink transmission; other N sink modules are involved in downlink transmission. The P E R of a channel state is calculated by the channel module and transferred to the sink for reception decision. If the reception is not successful, information is sent to the corresponding node module: A P node i f downlink, other nodes i f uplink. A n d then the retransmission process is implemented based on the scheduling rule applied at the node. Chapter 4 Simulation environment 41 • Channel module The functions of the channel module are to generate time-correlated Rayleigh distributed signal level and calculate P E R based on packet length, modulation mode and current channel S N R level. The P E R is transferred to the traffic sink module and the current S N R level is used by the scheduling algorithm module at the A P for H O L selection. In our simulation, perfect channel knowledge is assumed. C S D S + R R : In this algorithm, a S N R threshold is chosen. Channels with SNRs higher than the threshold are in "Good" state; otherwise the channel is marked as in "Bad" state. The scheduling algorithm selects the queues with Good channel state in a Round Robin fashion. If all queues are in bad state, a queue is selected for transmission according to R R as well . We simulated the CSDS+RR scheme according to the original policy proposed in [1]. M A X S: This algorithm was proposed as an extreme case in weight-based C S D S schemes, such as [7], [38]. At each scheduling moment, the queue with the highest S N R is selected for transmission. If there is more than one such queue, one is selected at a R R fashion. A s mentioned in Section 2.2, optimal total throughput can be achieved by selecting the user with the best channel state each time. Therefore, this algorithm can be employed as a performance criterion of total throughput of other C S D S scheduling schemes. M A X S/Sav: At each scheduling moment, the queue with the best relative channel state defined as the ratio of current S N R state (S) to the queue's average S N R (Sav) is selected for transmission. As in the M A X S scheme, ties are broken by a R R fashion. This scheme was Chapter 4 Simulation environment 42 mentioned in [35] and is naturally related to our new schemes since all o f them include the factor of average channel states. M a x S/Sav is simple and was introduced to be fair in [35] for STAs with different average received powers, so here it is involved for performance comparison. 4.2.2 Verification of C++ simulation model with simple OPNET test scenarios In order to verify that our C++ I E E E 802.1 l b W L A N M A C layer simulation program (hereafter referred to C++) is correct, simple test scenarios were implemented and compared with reference scenarios built up in O P N E T . The objective is to verify that the C++ simulated M A C layer protocols perform in a comparable way to the O P N E T results. • Introduction to OPNET WLAN simulation environment O P N E T supports an 802.11b W L A N simulation module. Unfortunately, no wireless multipath fading channel model is included and only a FIFO policy is implemented as the scheduling algorithm on both directions. However, the M A C layer protocol performance of our C++ simulator can be verified by comparison with test scenarios supported in O P N E T . The M A C layer functions of W L A N module in O P N E T are based on 802.11 standards. The main features are as follows [32]: Access Mechanism: Carrier sense multiple access and collision avoidance ( C S M A / C A ) distributed coordination function (DCF) scheme as defined in the standard. The point coordination function (PCF) is also supported. Chapter 4 Simulation environment 43 Frame Exchange Sequence: Data and Acknowledgment frame exchange to ensure the reliability of data transfer. Optional R T S / C T S frame exchange for media reservation. Deference And Backoff: Interframe spacing: DIFS, SIFS, EIFS for D C F and PIFS for P C F implementation. The values of the intervals are selected based on the P H Y layer characteristics. Data Rate: 1 Mbps, 2 Mbps, 5.5 Mbps, 11 Mbps Recovery mechanisms: Retransmission mechanism for data frames used when A C K frame is not received. Long and Short retry counters as defined in the standard. Fragmentation and Reassembly: optional, based on data packet size Buffer Size: length defined by the user For P H Y layer, O P N E T can model FHSS , DSSS , and Infrared technologies from the I E E E 8 0 2 . i l standard specification. In order to model the 802.11b physical layer, we set the channel data rate at 11 Mbps with DSSS physical characteristics. However all control packets are transmitted at a data rate of 1 Mbps as specified by the standard. The radio transceiver pipeline, which consists of 14 stages, is used to model wireless transmission of packets [33]. These stages include Received power, Background noise, Bi t error rate, Signal to noise ratio etc. Based on the outputs of these stages, the pipeline can provide error information of received packets to the M A C layer. Background noise and Interference stages are used to model wireless channel characteristics. Both of them are constant values, which mean no multipath fading effect is included in the module. In the W L A N model of O P N E T , path-loss propagation model is included in the received power stage, and the received signal power is computed by: Chapter 4 Simulation environment 44 Reception power (mW) = Transmission power (mW) xpathjoss (4.1) where path Joss is calculated by: path Joss = /iV/oV/2 (4.2) where / is the distance between the transmission node and the receiving node. • Comparable test scenarios in OPNET and C++ for verification We can see from (4.1) and (4.2) that the received power is only a function of the distance between the transmitter and the receiver.. No fading is included in the O P N E T simulation system. N o w the comparable test scenarios can be built up in C++ and O P N E T simulation systems by translating test parameters in O P N E T to the inputs of C++ simulation. Given the distance between the A P and the S T A , according to the pipeline introductions in O P N E T [33], the received S N R is calculated as: SNR (dB) = reception power (dBm) - noise (dBm) (4.3) where the reception power can be computed by (4.1) and noise can be obtained by background and interference noise stages in pipeline [33]. The transmission power is set to 0.1 m W in our verification tests. Two verification scenarios are considered, one is simple with only one A P and one S T A , and only downlink traffic is considered. The S T A ' s channel state is set to be "Bad", which means P E R is higher than 10"2. For packet length of 8800 bits, the S N R level is lower than 10.5 dB. The other is with one A P and two STAs , one with a "good" channel state, and the other with a "bad" state. In this test bidirectional traffic is considered. The topology of the second test scenario is shown in Figure 4.7. The environment is assumed a free-space office. Chapter 4 Simulation environment 45 The distance between the AP and the wlanjstal is about 50 m (received S N R = 20 dB) and the distance between the AP and wlan_sta2 is about 160 m (received S N R = 8 dB). Note that ifwlan_stal is removed, the topology is exact of the test scenario 1. Figure 4.7: Test topology in O P N E T environment with one A P and two user nodes For each scenario, tests are repeated five times to obtain average system downlink throughput for comparison. In the first scenario, P E R comparison is also included. The details of simulation parameters are introduced respectively as follows. Test scenario 1:1 A P and 1 S T A . The parameters of both C++ and O P N E T are as follows: Channel condition: Sav = 8 dB (no fading effect is considered); traffic source: 200 pkt/s, constant interval, downlink traffic only; packet length: 8800 bits; modulation mode: B P S K ; data transmission rate: 11 Mb/s; control message transmission rate: 1 Mb/s; simulation time: Chapter 4 Simulation environment 46 25 s. The results of downlink total throughput and overall P E R are shown in Table 4.1 and Table 4.2 respectively. Note that the theoretical result of P E R is also provided as a reference by the calculation in Matlab as follows: Fo r S N R = 8 d B , pktlength = 8800 bits, modulation mode: B P S K , P E R = 0.7728. Throughput (pkt/s) Testl Test2 Test3 Test4 Test5 A v g O P N E T 56.96 56.8 59.08 58.32 56.4 57.512 C++ 58.6 57.24 56.04 56.92 59.4 57.64 Table 4.1: Throughput comparison between O P N E T and C++ simulation for test scenario 1 P E R Testl Test2 Test3 Test4 Test5 A v g O P N E T 0.7746 0.764 0.770 0.7733 0.7718 0.7707 C++ 0.7704 0.7722 0.7749 0.7761 0.7684 0.7724 Table 4.2: P E R comparison between O P N E T and C++ simulation Test scenario 2: 1 A P and 2 S T A s . The parameter values are as follows: Channel condition: Savl = 20 dB, Sav2 = 8 dB, both downlink and uplink traffic are involved. Downlink traffic is as in scenario 1, uplink traffic: 5 pkts/s, constant interval, Chapter 4 Simulation environment Al scheduling scheme in A P for downlink: FIFO (destination is randomly selected between STA1 and STA2) . Other parameters are set as in scenario 1. The results of downlink throughput obtained by STA1 and S T A 2 are shown in Table 4.3 as below. Throughput (pkt/s) Testl Test2 Test3 Test4 Test5 A v g O P N E T STA1 58.96 60.32 58.12 57.72 56.08 58.24 S T A 2 49.08 50.6 48.14 48.36 47.36 48.708 C++ STA1 59.36 59.88 60.12 57.12 56.64 58.624 S T A 2 50.08 49.76 50.16 48.28 47.64 49.184 Table 4.3: Throughput comparison between O P N E T and C++ simulation for test scenario 2 From the results in Table 4.1, Table 4.2, and Table 4.3, it can be seen that the C++ simulation results are quite similar to those obtained by O P N E T . Chapter 5 Simulation results and analysis 48 Chapter 5 Simulation results and analysis 5.1 Simulation assumptions and test scenarios In the simulation, only results related to downlink scheduling is generated, even though bidirectional traffic is included. The downlink traffic queues are assumed to be always backlogged (i.e. queues are never empty); while on the uplink a light traffic load is assumed. Scheduling decisions are made on per packet basis and perfect knowledge of channel states is available. Our 802.11b M A C layer simulation system includes C S M A / C A access mechanism, backoff and retransmission procedures, and DIFS, SIFS, EIFS, and M A C / P H Y overhead of M P D U . The assumptions and parameters used in every test scenario are specified in Table 5.1. Network structure Infrastructure, one A P with several user STAs Channel model Rayleigh fading + A W G N noise Downlink traffic backlogged Uplink traffic i.i .d Bernoulli distribution with generation probability 0.05% (packets per time_sloi) Packet length: 8800 bits Modulation mode B P S K R T S / C T S N O Data transmission rate 11 Mbps Fragmentation N O Control frame transmission rate 1 Mbps Buffer size Unlimited Retransmission limit 7 Table 5.1: Simulation model parameter values Chapter 5 Simulation results and analysis 49 Other parameter values as speeified in [34] are shown in Table 5.2. T i m e s l o t 20 us PLCP_overhead_control 192 us SIFS 10 us Default_PLCP_overhead 57 bits DIPS 50 us MSDU_header_size 224 bits EIFS 112 us A C K timeout 112 us C W m i n 31 CW_max 1023 Table 5.2: Other simulation parameters In this chapter, simulation results are provided to illustrate the effect of Doppler frequency, fairness window size and the number of users on our proposed scheduling schemes. Results for CSDS+RR, Max S and Max S/Sav are also presented for comparison. Although our compensation mechanism is targeted to improve long-term fairness, for real applications, short-term fairness is more important and should be considered. So for all simulation scenarios, the size of the short-term fairness window is specified in seconds according to T C P traffic requirements. Since good short-term fairness implies good long-term fairness, long-term fairness results are not explicitly included. The simulation time is set to 300 s for all scenarios to obtain meaningful statistic results. To evaluate the average short-term performance over the whole simulation period, simulation time is divided into disjoint intervals, each equal in duration to the short-term fairness window. For each interval, the fairness index and the normalized total throughput on the downlink are calculated. Recall that the fairness index is defined in Section 3.2 as Chapter 5 Simulation results and analysis 50 / ( * ) = xi>Q,x = (x,, x2, .... x„) where xt. is the throughput that user i obtains during the interval of fairness window. The normalized total throughput is equal to the ratio of total throughput in bps to channel capacity (i.e. U M b / s in our simulation). Then the average values of them over the whole period are obtained and are used as criterions for performance comparison. Tests scenarios including 1 A P and the number of users N u = 2, 4, 8, 12, 16, 20 users are simulated separately; in each scenario, results for a given Doppler frequency and fairness window size pair are collected. Given the fading characteristics in 802.11b indoor W L A N [30], the Doppler frequency (fd) is chosen from the range 1 H z to 10 Hz . A n d based on the T C P traffic delay requirements, the short-term fairness window (/win) is set to between 1 s and 10 s. The results with different (fd, fwin) pairs for the proposed and reference schemes are first presented. Then the effect of different N u is studied. Finally, the selection of the parameter y in the IWRR+Max S* scheme is described based on the simulation results. 5.2 Results for different Doppler frequencies and fairness window size We assume a scenario with one A P and eight equi-distant user nodes as shown in Figure 5.1. Assuming that the received power at 1 m from the A P is 1 dBm, the W L A N coverage range is approximately 200 m according to the receiver minimum input level Chapter 5 Simulation results and analysis 51 sensitivity of -76 dBm specified in the standard [19]. Then according to the path-loss model mentioned in Chapter 3, the average S N R level at the receiver of a user can be calculated as [37]: S-v (dB) = RxPowen (dBm) - KTB (dBm) - NF (dB) (5.1) where K = 1.379xIO" 2 3 is Boltzmann's constant, T is the temperature (290°AT), B is the bandwidth (20 M H z ) , NF is the receiver noise figure (20 dB) and RxPowert (dBm) = P0 (dBm) - 10j3log(dt I d0) (5.2) where Po = 1 dBm is the received power at the reference point, which is at distance do = 1 m from the A P , di is the distance from the node to the A P and = 3.3 is the attenuation factor [25]. Therefore the average S N R values at the eight equi-distant nodes are: 35.4 dB, 25.5 dB, 19.7 dB, 15.6 dB, 12.4 dB, 9.7 dB, 7.5 dB, 5.6 dB Figure 5.1: Deployment of an equi-distant 8-user W L A N Chapter 5 Simulation results and analysis 52 Simulation results were obtained for the following (fd,fwin) pairs: (10 Hz , 1 s), (1 Hz , 10 s), (2.5 Hz , 4 s), (1 Hz , 1 s), (1 Hz , 5 s), (5 Hz , 5 s), (10 Hz , 5 s), (10 Hz , 10 s). Test results are shown in Figure 5.2 - Figure 5.9, in which the x-axis shows the average total throughput and the y-axis shows the average fairness index. In each figure, in addition to the three proposed scheduling schemes, three other C S D S based schemes are also involved. They are (see Section 4.2.1): Max S, Max S/Sav, and CSDS+RR. . The performance of the Max S and M a x S/Sav schemes are represented by two individual points since no tuning parameters is involved in these schemes. For CSDS+RR, I W R R , and IWRR+Max S*, test results with different transmission threshold (TT) values of 12 dB, 10.5 dB, 9.5 dB, 8.5 dB, 7.5 dB are shown as three curves. For (fd,fwin) = (10 Hz , 1 s), additional T T values of 16 dB and 22 dB are also included to study the differences among these algorithms with high T T values. The compensation control parameter cp in I W R R and IWRR+Max S* is set to 10.5 for all test cases. For Max S*, the threshold as defined in (3.6) is set to 10.5 dB for all test cases and the test results for <p values of 2.5, 5.5, 7.5, 10.5, and 20.5 are shown. In Figure 5.2, which shows the results for (fd,fwin) = (10 Hz , 1 s), it can be seen that the proposed schemes (i.e. Max S*, I W R R , and IWRR+Max S*) can provide better performance than CSDS+RR. The Max S scheme achieves the highest total throughput as expected, but has the poorest fairness performance among the schemes. Max S/Sav scheme has a good fairness performance but poor total throughput. For Max S*, when <p is small, the performance is close to that of Max S; as p increases, the fairness is greatly improved while Chapter 5 Simulation results and analysis 53 the total throughput decreases only by a small amount. The reason is that the service compensation to sessions with poor average S N R is small for small cp, but increases with tp. A n d the compensation procedure is only performed when channel S N R exceeds the threshold. If all of them are below the threshold, the session with the best channel S N R is selected. In the case that the channel S N R exceeds the threshold, the P E R is low (In this test, it is lower than 10" since the threshold is set to 10.5 dB) so that the waste of channel resource is limited. Therefore, the compensation mechanism can improve fairness performance whilst maintaining high channel efficiency. CSDS+RR, I W R R and IWRR+Max S* were tested for different TT values. Figure 5.2 also shows that the performance of CSDS+RR decreases when the threshold exceeds a certain value, T T n > o p t , defined as the S N R level beyond which the total throughput decreases as the threshold increases. This is because when T T is less than TT^ > o p t , an increase in T T causes the average P E R of channel states above T T to decrease, thus resulting in a lower probability of failed transmission. However, the fairness index decreases since the users have different average SNR ' s . When TT is above TT n > o p t , it is quite likely that all user S N R ' s wi l l be below TT, resulting in a R R allocation. Therefore, the total throughput decreases with T T when T T is greater than T T n j 0 p t . In Figure 5.2, it can be seen that T T n , o p t = 1 6 dB. From Figure 5.2 we observe that I W R R also degrades in a similar manner as C S D S + R R but IWRR+Max S* achieves the best performance among all compared schemes when y is chosen optimally. In Figure 5.2, y is set to 0.1 when T T < T T n ; 0 p t and y is set to 1.0 otherwise. In the following tests an empirically optimal value of y is used. These values are tabulated in Section 5.4. The combined scheme performs similar to M a x S* in all tests when T T is higher Chapter 5 Simulation results and analysis ' 54 than the TT^opt, therefore, in the following figures only the test results with the T T values lower than the T T n i 0 p t are included. fd=10Hzfwin=l s 1 0.9 0.8 0.7 x o I 0 , 6 g 0.5 > ca 0.3 0.2 0.1 0 0 .35 0.36 0.37 0.38 0.39 a v g t o t a l t h r o u g h p u t Figure 5.2: The comparison results of (10 Hz, 1 s) in the 8-user test scenario Figure 5.3 and Figure 5.4 show the results for (fd,fwin) = (1 Hz , 10 s) and (2.5 Hz , 4 s) respectively. From the figures we observe that all the schemes have similar performances as in Figure 5.2. The fairness index improvement of I W R R and IWRR+Max S* over CSDS+RR is about 0.1. Max S/Sav achieves a performance similar to that of C S D S + R R with a T T between 7.5 dB and 8.5 dB. Figure 5.2 - Figure 5.4 suggest that different test scenarios with the same value of Pdw = fd x fwin = 10 yield similar performance results. It is thus necessary to only study the effect of the product Pdw rather than of each of these two parameters. In our simulation,y# is in the range of 1 H z to 10 Hz, fwin is between' 1 s and 10 s and thus Pdw is between 1 and 100. The effect of Pdw is studied in the following tests. TT=7.5 dB Max S/Sav 0 Max S - %r — Max S* - _l _ CSDS+RR r\ IWRR+Max S* - * IWRR TT=9.5 dB _! I I !_ Chapter 5 Simulation results and analysis fd= 1 H z fwin = 10 s 0.33 0.34 0.35 0.36 avg total throughput 0.37 0.38 Figure 5.3: The comparison results for (1 Hz, 10 s) in the 8-user test scenario fd = 2.5 Hz fwin = 4 s 0.34 0.35 0.36 0.37 avg total throughput 0.38 0.39 Figure 5.4: The comparison results for (2.5 Hz, 4 s) in the 8-user test scenario Chapter 5 Simulation results and analysis 56 fd = 1 H z fwin = 1 s 0.33 0.34 0.35 0.36 0.37 avg total throughput 0.38 Figure 5.5: The comparison results of (1 Hz, 1 s) in the 8-user test scenario Figure 5.5 shows the results for (fd,fwin) = (1 Hz , 1 s), i.e. Pdw = 1. IWRR+Max S* exhibits similar performances as CSDS+RR, whereas the other four schemes have worse performance. Max S/Sav has a poorer fairness index than CSDS+RR because it always selects the user with the best S/Sav. Thus, when the channel is changing slowly (i.e. fd = 1 Hz) service allocation fluctuates widely during short intervals (i.e. fwin = 1 s) yielding a poor fairness index value. Max S* is also based on the "best channel state selection" method. Therefore, although its overall performance is better than those of Max S and Max S/Sav, it is still worse than that of CSDS+RR. I W R R shows better performance than the other three although it is worse than CSDS+RR. Compensation procedure applied in I W R R also causes a bigger fluctuation of service allocation than CSDS+RR when Pdw is small, however the effect to RR-based schemes is less than that to "best channel state based" schemes. In Figure 5.6 - Figure 5.9, the results of Pdw = 5, 25, 50, 100 are shown respectively. Chapter 5 Simulation results and analysis fd = 1 H z fwin = 5 s 0.35 0.36 0.37 avg overall throughput 0.38 0.39 Figure 5.6: The comparison results for (1 Hz, 5 s) in the 8-user test scenario fd = 5 H z fwin = 5 s 0.33 0.34 0.35 0.36 0.37 avg overall throughput 0.38 0.39 Figure 5.7: The comparison results for (5 Hz, 5 s) in the 8-user test scenario Chapter 5 Simulation results and analysis 58 fd = 10Hzfwin = 5 s 0.34 0.35 0.36 0.37 0.38 avg overall throughput 0.39 Figure 5.8: The comparison results for (10 Hz, 5 s) in the 8-user test scenario fd= 10Hzfwin= 10 s 0.34 .-. 0.35 0.36 0.37 ; avg total throughput 0.38 0.39 Figure 5.9: The comparison results of (10 Hz, 10 s) in the 8-user test scenario Chapter 5 Simulation results and analysis 59 Figure 5.6 to Figure 5.9 show that the performance of C S D S + R R does not change much with Pdw; the fairness index difference between I W R R and IWRR+Max S* on the one hand and C S D S + R R on the other increases, as Pdw increases from 5 to 25. Beyond Pdw = 25, the difference remains fairly constant. The fairness index of M a x S/Sav and M a x S* increases rapidly with Pdw for Pdw < 25; for Pdw >25, the improvement is smaller. From Figure 5.2 to Figure 5.9, it can be concluded that when Pdw is greater than 10, the new schemes yield improved fairness index and total throughput simultaneously. For 10 > Pdw > 5, I W R R and IWRR+Max S* have the best performance. Overall, IWRR+Max S* provides a reasonably good performance among the three new schemes in any test scenario. Besides, it can be seen that in each test case at the optimal TT value TTopt « 10 dB, I W R R and IWRR+Max S* yield the best overall performance which means both total throughput and fairness index are close to the highest values. This means that Pdw w i l l not affect TTopt much. Note that when S N R = TTopt, P E R « 10"2, which is considered as a bound of good channel state in [39], can be obtained for 8800 bits packet length. 5.3 Effect of number of users In this section, the Doppler frequency is fixed at 10 Hz , and the fairness window size is set to 1 s for all test cases. The number of users, N u , is set to 2, 4, 8, 12, 16, 20. For N u = 4 to 20, the user nodes are equally spaced in the coverage area of the A P . For N u = 2, we assume that one user has a good average channel state (25 dB) and the other has a poorer one (8 dB). The parameter values used for the different schemes are the same as those in the previous Chapter 5 Simulation results and analysis " 60 section except that the TT value used in Max S* is 8.5 dB instead of 10.5 dB when N u = 2 and 4. The results with 8 users are shown as Figure 5.2. The results for 2, 4, 12, 16, 20 users are shown in Figure 5.10 - Figure 5.14. The results for the 2-user and 4-user test cases are shown in Figure 5.10 and Figure 5.11 respectively. Comparing with Figure 5.2, we see that the value of TTopt for CSDS+RR, I W R R , and IWRR+Max S* is lower than that in the 8-user case. In the two-user case, T T ^ ) 0 p t is around 9.5 dB and in the four-user case, it is around 10.5 dB. The reason for this is that when N u is small, it is more likely that the SNRs of all users fall below the TT simultaneously i f the T T is kept fixed. In this case, the performance of C S D S + R R tends to that of pure R R : channel resource is wasted and both total throughput and fairness are impaired. In Figure 5.10 M a x S/Sav exhibits the highest fairness index and the lowest total throughput among the compared schemes. Among the other four schemes, Max S* and IWRR+Max S* have the best performances for all T T values. The fairness index improvement of M a x S* and IWRR+Max S* over C S D S + R R is approximately 0.15. In Figure 5.11, I W R R , M a x S* and IWRR+Max S* have similar performances and are generally better than CSDS+RR. The fairness index improvement of IWRR+Max S* to C S D S + R R is about 0.12. Chapter 5 Simulation results and analysis Nu = 2 Sav = 8 dB, 25 dB 0.34 0.36 0.38 0.4 avg total throughput 0.42 0.44 Figure 5.10: The results of (10 Hz, 1 s) in 2-user test scenario Nu = 4 Sav = 25.5,15.6, 9.7, 5.6 dB 0.32 0.34 0.36 0.38 avg total throughput 0.4 0.42 Figure 5.11: The results of (10 Hz, 1 s) in 4-user test scenario Chapter 5 Simulation results and analysis Nu=12, Sav=41.2, 31.2, 25.5, 21.3, 18.1, 15.6,13.3, 11.4, 9.7, 8.2, 6.8, 5.6 dB 1 0.32 0.33 0.34 avg total throughput 0.35 0.36 F i g u r e 5.12: T h e results o f (10 H z , 1 s) i n 12-user test scenario Nu=16, Sav = 45.4,35.4,29.6,25.5,22.3,19.7,17.5,15.6,13.9,12.4,11, 9.7,8.6,7.5,6.5,5.6 dB 1 0.28 0.29 0.3 0.31 A avg total throughput . 0.32 0.33 F i g u r e 5.13: T h e results o f (10 H z , 1 s) i n 16-user test scenario Chapter 5 Simulation results and analysis 63 Nu=20, Sav=48.6, 38.6, 32.8, 28.7, 25.6, 22.9, 20.7, 18.8, 17, 15.6, 14.2, 13, 11.8, 10.7, 8.8, 8, 7.1, 6.4, 5.6 dB 1 0.9 0.8 0.7 o> ^3 a 0.6 « a 0.5 0.4 M >• cs 0.3 0.2 0.1 0 - © -® k ® « M a x S/Sav 0 M a x S -+— C S D S + R R -+ - M a x S* • I W R R © _ IWRR+Max S* 0.25 0.26 0.27 0.28 0.29 avg total throughput 0.3 0.31 Figure 5.14: The results of (10 Hz, 1 s) in 20-user test scenario The results for 12, 16 and 20 users are shown in Figure 5.12 to Figure 5.14 respectively. It can be seen that I W R R and IWRR+Max S* have the best performances in each test scenario; the fairness index improvement of I W R R and IWRR+Max S* compared to C S D S + R R decreases somewhat with N u as shown in Table 5.3. N u 2 4 8 12 16 20 Fairness index Improvement 0.15 0.12 0,1 0.08 0.06 0.05 Table 5.3: The fairness index improvement of IWRR+Max S* to C S D S + R R Figure 5.2, Figure 5.10 to Figure 5.14 also show that M a x S provides the best total throughput as expected; however it decreases with Nu . The reasons for this are that more uplink traffic is carried due to a larger number of users involved in the network and the Chapter 5 Simulation results and analysis 64 channel resource is wasted by more collisions as well . Furthermore, we observe that the fairness index for M a x S/Sav and Max S* decreases with N u while that for C S D S + R R changes slightly except that when N u increases from 4 to 8. Among all the compared scheduling schemes, IWRR+Max S* and I W R R yield the best overall performance at TTopt. From Figure 5.2, Figure 5.10 to Figure 5.14 it can be seen that when N u > 8, TTopt « 10 dB, which is fairly constant, and decreases when N u is small. For N u = 2 and 4, TTopt « 9 dB. 5.4 Selection of y for IWRR+Max S* For rVVRR+Max S*, the parameter y e [0,l] should be chosen to provide close to optimal performance. In this section, we focus on the effect of Doppler frequency, fairness window, and number of users to the selection of y. From the discussion in Section 5.2, the effect of Doppler frequency (fd) and fairness window size (fwin) can be simplified to considering the product Pdw =fd* fwin. In this section, we use the test scenario of one A P and eight user nodes in Section 5.2 to study the effect of Pdw on the selection of y. The. results with Pdw = 1, 5, 10, 25 are shown in Figure 5.15 to Figure 5.18 respectively. In each test case, for a given value of TT, a performance curve of IWRR+Max S* parameterized by y is plotted. A s in Section 5.2, the values of TT are equal to 7.5, 8.5, 9.5, 10.5, 12 dB (The curve for 10.5 dB is not included since the other four are enough to illuminate the results.) and two additional values of 16 and 22 dB are included in Figure 5.17, where Pdw =10 . Besides, the curves for Max S/Sav, Max S, CSDS+RR schemes are also included for comparison proposes. Chapter 5 Simulation results and analysis- - ;- . .. - 65 Figure 5.15 shows the performance curves of IWRR+Max S* parameterized by y under different T T values for Pdw =1. For each TT value, the performance points for y = {0, 10"", 10"6, 0.001, 0.1, 0.3, 1.0} are plotted. From the results we observe that for each T T value, IWRR+Max S* with y = 0 shows the best performance, which is similar with that of CSDS+RR. This means that when Pdw = 1, the best choice is y = 0 at which IWRR+Max S* has a similar performance to CSDS+RR. From Figure 5.16 which shows the test results for Pdw =5, it can be seen that when 10"11 < y < 0.1 IWRR+Max S* exhibits better fairness index than CSDS+RR. The average total throughput of IWRR+Max S* increases with y for T T = 7.5 dB and 8.5 dB, and remains fairly constant for TT = 9.5 dB and 12 dB. When y > 0.3, the performance of IWRR+Max S* is somewhat worse than that of CSDS+RR. 8 U fd=1 H z (win=1 s y se lec t ion for I W R R + M a x S * s c h e m e 0 . 3 5 0 .36 0 .37 a v g t o t a l t h r o u g h p u t Figure 5.15: The performance curves of rWRR+Max S* parameterized by y for Pdw = 1, Nu = 8 Chapter 5 Simulation results and analysis 66 8U fd=l Hz fwin=5 s y selection for IWRR+Max S* scheme li 1 1 , , , 0.5 It Max S/Sav 0.4 Max S —t— CSDS+RR 0.3 X IWRR+Max S*, TT= 7.5 dB 0.2 —*— TT= 8.5 dB - - • T T = 9.5 dB 0.1 - —•— T T = 12.0 dB 0> 1 1 ' ' ' 0.33 0.34 0.35 0.36 0.37 0.38 avg total throughput Figure 5.16: The performance curves of IWRR+Max S* parameterized by y for Pdw = 5, Nu = 8 8U fd=10 Hz fwin=ls y selection for IWRR+Max S* 1 0.9 0.8 0.7 41 g 0.5 0.4 0.3 0.2 0.1 0 x 0 0.34 Max S/Sav Max S IWRR+Max S* TT=7.5 dB TT=8.5 dB TT=9.5 dB T I M 2.0 dB TT=16.0dB TT=22.0dB CSDS+RR 0.35 0.36 0.37 0.38 0.39-avg total throughput Figure 5.17: The performance curves of IWRR+Max S* parameterized by y for Pdw = 10, Nu = 8 Chapter 5 Simulation results and analysis 61 The results for Pdw = 10 are shown in Figure 5.17. A s in Figure 5.15 and Figure 5.16, when 7 = 0, IWRR+Max S* has a similar performance to CSDS+RR, but for y > 0, IWRR+Max S* has a better performance. For 0 < y < 0.1, the fairness index increases rapidly then remains fairly unchanged with y; for y > 0.1, it decreases with y. In Figure 5.17, two additional curves with T T = 16 dB and 22 dB are included to show the effect of T T ^ o p t t o the y selection. It can be seen that for TT = 22 dB, y = 1 achieves the best performance and for T T =16 dB, the curve is similar to that for TT = 12 dB. The best average total throughput is achieved for IWRR+Max S* with y = 1 and when T T > T T n i 0 p t , the average total throughput is decreases. To keep the good performance of IWRR+Max S*, y = 1 is chosen for TT > T T n ) 0 p t . 8 U fd=5 H z fwin=5 s y se lec t i on to I W R R + M a x S * s c h e m e 1, , , > , , » Max S/Sav • Max S -•+ CSDS+RR . IWRR+Max S*, TT=9.5 dB - - x - - TT=8.5 dB 0 * II =7.5 dB * TT=12.0 dB fjl 1 1 1 1 1 0.33 0.34 0.35 0.36 0.37 0.38 0.39 avg total throughput Figure 5.18: The performance curves of IWRR+Max S* parameterized by y for Pdw = 25, Nu = 8 Chapter 5 Simulation results and analysis 68 Figure 5.18 shows the results for Pdw = 25. Since the test results for Pdw > 25 are similar to Pdw = 25, they are not presented in this thesis. In Figure 5.18 it can be seen that the fairness index of IWRR+Max S* increases with y. When y > 0.3,-IWRR+Max S* achieve the best performance. Based on the results above we can choose y to achieve close to optimal performance as in Table 5.4. Pdw 1 5 10 25-100 Yopt 0 [10"", 0.1] [10"b, 0.1] [0.3, 1.0] Table 5.4: y o pt based on the different value of Pdw It can be seen that y o p t increases with Pdw. Note that when Pdw = 5, 10 and > 25, a range of y is given because the fairness index of IWRR+Max S* changes slightly in this range and the average total throughput is almost constant for T T > 9.5 dB and increases slightly for TT = 7.5 dB and 8.5 dB. Also note that Table 5.4 does not include the y selection for TT > T T n , o p t , in which case y is set to 1 for all Pdw values. For simplicity, in Section 5.2 for T T < 16 dB, y = 0.1 for all Pdw > 10 since there is slight performance difference between y=0.1 and 1.0 at TTopt; y = 10"11 for Pdw = 5 and y = 0 for Pdw = 1. For T T > 16 dB, y = 1.0. To study the effect of the number of users on y o p t , test scenarios with Nu=2, 4, 8, 12, 16, 20 as in Section 5.3 are considered. It was found that for N u = 4, 8, 12, 16, and 20, y o p t is as given in Table 5.4. The value of y o p t for N u = 2 can be deduced from the results in Figure 5.19 and Figure 5.20 which are for Pdw = 1 and 10 respectively. The figures show that in both cases, y o p t = 1 for any TT values shown. Therefore, it can be extended to other values of Pdw e [1, 100], y o p t = 1 for any TT values of N u = 2 test case. Chapter 5 Simulation results and analysis 2 U fd=1 H z twin=1 s y se lec t ion for I W R R + M a x S * s c h e m e 0.45 0.32 0.34 0.36 0.38 0.4 avg total throughput 0.42 0.44 F i g u r e 5.19: T h e performance curves o f I W R R + M a x S* parameter ized b y y for P d w = 1, N u = 2 2 U fd=10 H z fwin=1 s y se lec t ion for I W R R + M a x S * s c h e m e •O c 0.9 0.85 0.8 0.75 0.7 3 0.65 Ml « 0.6 0.55 0.5 0.45 0 X Max S/Sav 0 Max S CSDS+RR IWRR+Max S*, TT=9.5 dB -x - TT=8.5 dB a - TT=12 dB TT=7.5 dB 0.32 0.34 0.36 0.38 0.4 avg total thrughput 0.42 0.44 F i g u r e 5.20: T h e performance curves o f I W R R + M a x S* parameter ized b y y for P d w = 10, N u = 2 Chapter 5 Simulation results and analysis 70 Besides, the number of users may also affect the value of T T n ! 0 p t . Usually for N u > 8, TT^optis at least 10.5 dB, which achieves 10"2 P E R for 8800 bits.long packet. When N u < 8, T T n , o p t ^ 9 d B . Chapter 6 Conclusions and future work 71 Chapter 6 Conclusions and future work In this thesis, based on the proposed compensation mechanism, which is simple to implement, three scheduling schemes are proposed for 802.11 W L A N network. The key idea behind the compensation mechanism is to provide somewhat service compensation to those users with low average S N R (Sav) by considering the Sav values of the corresponding fading channels. B y doing this, a better long-term fairness among users can be expected. Certainly, the short-term fairness is a more important criterion than long-term fairness in a real network. Therefore, to achieve a good short-term fairness and maintain a good total throughput simultaneously, three scheduling schemes are developed by employing the principle of the compensation mechanism. Among the schemes, Max S* attempts to maximize the total throughput by selecting the H O L with the best modified channel state value; the value depends on the compensation mechanism. I W R R scheme also employs the compensation mechanism, but in this scheme the modified channel state value acts as a transmission weight. IWRR+Max S* represents a combination of the above two schemes to benefit from the advantages of both. W L A N simulation models in C++ are built up to study the performance of the proposed schemes in terms of a short-term fairness index and a total throughput, which are the two fundamental issues when evaluating the performance of a W L A N . The effect of Doppler frequency, fairness window and number of users to the performance of the proposed schemes are obtained and analyzed by simulation. Fairness window is used to indicate the short-term fairness requirement among users. The fairness window size is set to between 1 s and 10 s for T C P traffic delay requirements. Doppler frequency is set to between 1 H z and 10 H z Chapter 6 Conclusions and future work 72 according to the channel characteristics of an Indoor W L A N . The number of users is set to typically between 2 and 20 with the equi-distance distribution over the A P ' s coverage area. We compared the performance of the three proposed schemes with three other CSDS-based schemes: Max S, Max S/Sav and CSDS+RR. FIFO which is the default scheduling scheme employed in W L A N is not included because in [1] CSDS+RR achieves a much better performance than that of it. From the simulation results, we observe that the effect of Doppler frequency and fairness window size can be simplified to considering the product of these two parameters (referred to Pdw). Several conclusions can be drawn based on the simulation results. Here the overall performance refers to the total throughput and fairness index pair. • For any value of Pdw and N u , It can be see that Max S* can substantially improve the fairness index of Max S and maintain a close to optimal total throughput. A n d its fairness index can increase close to that of Max S/Sav, which has a worse total throughput performance. Therefore, it can be concluded that Max S* can achieve a better overall performance than both Max S and Max S/Sav. • For any value of Pdw and N u , IWRR+Max S* can yield the best or close to best fairness index and maintain a close to optimal total throughput simultaneously with a T T = TTopt. A t the point of TTopt, the overall performance improvement of IWRR+Max S* to CSDS+RR reflects mainly on fairness index, because the difference between their total throughput is not much. Pdw and N u can affect the improvement of fairness index. When N u is fixed, the improvement Chapter 6 Conclusions and future work 73 increases with Pdw; when Pdw is fixed, the improvement decreases with Nu . According to the simulation results, the range of the fairness index improvement of IWRR+Max S* to CSDS+RR is from 0% to over 20% depending on the values of Pdw and Nu . Typically, for Pdw = 1 and N u > 4, the improvement is 0%; and for Pdw > 25 and N u < 8, the improvement is over 20%. • For N u = 2, IWRR+Max S*, shows a similar performance as Max S*, can achieve the best overall performance among the all compared schemes for any value of Pdw. The TTopt is approximately 9 dB for IWRR+Max S*, I W R R and CSDS+RR, where the fairness index improvement of IWRR+Max S* and I W R R to CSDS+RR are about 12% and 10% respectively for Pdw = 1 and increase with Pdw. • For N u > 4, when Pdw = 1, IWRR+Max S* has a similar performance to C S D S + R R and is better than the other compared schemes. For 25 > Pdw > 5, IWRR+Max S* and I W R R show the best overall performance among the schemes, and for 25 > Pdw > 10, M a x S* also has a better overall performance than CSDS+RR. Then for Pdw > 25, the three proposed schemes can achieve similar performance which is the best among the compared schemes and the fairness index improvement to CSDS+RR is about 0.2 (25%) for Pdw = 100 and N u = 8. The TTopt is about 9.5 dB for N u = 4, and 10.0 dB for N u > 8. Note that 10.0 dB is the S N R value corresponding to approximate 10"2 P E R for the 8800 bit packet length. • The results shows that when N u is fixed, Pdw wi l l not give much effect to the performance of C S D S + R R and Max S, but w i l l improve the fairness index Chapter 6 Conclusions and future work 74 performance of the other schemes with the increasing of it. When Pdw is fixed, the total throughput of M a x S decreases with Nu . Besides, the fairness index of Max S*, Max S/Sav, I W R R and IWRR+Max S* also decreases with N u ; but the fairness index of CSDS+RR slightly increases with it. Therefore, it can be concluded that the effect of Pdw and N u to C S D S + R R and Max S is small, to the other schemes that are related to Sav is large. Among them, Max S/Sav and M a x S* show the biggest changes on the fairness index performance. In general, the proposed IWRR+Max S* scheme combined the advantages of the other two scheduling schemes and is simple and suitable for slow fading W L A N . It can achieve a better total throughput and short-term fairness performance than all the other compared scheduling schemes. In our test scenarios where Rayleigh fading is assumed, we found that the combined scheme is especially suitable for Pdw > 5 and any value of N u although its performance is no worse than that of CSDS+RR for small Pdw values and large Nu . For future work, one direction is to investigate the effect of imperfect knowledge of channel information. In this thesis we assume that perfect knowledge of channel state can be obtained per packet, but in a real system, the channel state information may be erroneous or outdated. These features can affect the performance of the new proposed schemes. Another direction is to consider the effect of different packet lengths. In this study a fixed packet length is assumed. In practice, packet lengths may change in W L A N networks. The performance of the proposed schemes with dynamic packet lengths should be assessed. References 75 References [1] P. Bhagwat, P. Bhattacharya, A . Krishna, and S. Tripathi, "Enhancing throughput over wireless L A N s using channel state dependent packet scheduling", IEEE INFOCOM, vol. 3, pp. 1133-1140, Apr. 1996. [2] S. L u , V . Bharghavan, and R. Sirkant, "Fair scheduling in wireless packet networks", IEEE/ACM Trans. Networks, vol. 7, no. 4, pp. 473-489, Aug . 1999. [3] T.S.E. Ng , I. Stoica, and H . Zhang, "Packet fair queueing algorithms for wireless networks with location-dependent errors", IEEE INFOCOM, vol . 3, pp. 1103-1111, Mar. 1998. [4] P. Ramanathan and P. Agrawal, "Adapting packet fair queuing algorithms to wireless networks", ACMMOBICOM, pp. 1-9, Dallas, U S , Oct. 1998. [5] H . Fattah and C. Leung, " A n efficient scheduling algorithm for packet cellular networks", IEEE Veh. Technol. Conf, vol . 4, pp. 24-28, Sept. 2002. [6] X . L i u , E . K . P . Chong, and N . B . Shroff, "Transmission scheduling for efficient wireless utilization", IEEE INFOCOM, vol . 2, pp. 776-785, Apr. 2001.. [7] Y . L i u , S. Gruhl, and E . W . Knightly, " W C F Q : ah opportunistic wireless scheduler with statistical fairness bounds", IEEE Trans. On Wireless Commun., vol . 2, Issue: 5, pp. 1017-1028, Sept 2003. [8] H . Aida , Y . Tamura, Y . Tobe, and H . Tokuda, "Wireless packet scheduling with signal-to-noise ratio monitoring", IEEE 25th Annu. Local Computer Networks Conf, pp. 32-41, Nov. 2000. [9] M . Bottigleliengo, C. Casetti, C . - F . Chiasserini, and M . Meo, "Short-term fairness for T C P flows in 802.11b W L A N s " , IEEE INFOCOM, vol . 2, pp. 1383 - 1392, Mar. 2004. References 76 [10] H . Aida , Y . Tobe, M . Saito, and H . Tokuda, " A software approach to channel-state dependent scheduling for wireless L A N s " , ACM Workshop On Wireless Mobile Multimedia, 2001, pp. 33-42, Rome, Italy,2001. [11] C. Fragouli, V . Sivaraman and M . B . Srivastava, "Controlled multimedia wireless link •sharing via enhanced class-based queuing with channel-state-dependent packet scheduling", IEEEINFOCOM, vol . 2, pp. 572-580, Apr. 1998. [12] S. W u , Q. Ding, and C C . K o , "Improving the performance of wireless L A N using a new scheduling algorithm", IEEE Int. Conf. on Distributed Computing Systems Workshop, pp. 455-460, Apr. 2001. [13] A . Gyasi-Agyei, " B L 2 X F - channel state-dependent scheduling algorithms for wireless IP networks", IEEE 11th Int. Conf. on Networks, pp. 623-628, Sept. 2003. [14] P. L i n , B . Benssou, Q .L . Ding, and K . C . Chua, " C S - W F Q : a wireless fair scheduling algorithms for. error-prone wireless channels", IEEE Proc. 9th Int. Conf. on Computer Commun. and Networks, pp. 276-281, Oct. 2000. [15] D . A . Eckhardt and P. Steenkiste, "Effort-limited fair (ELF) scheduling for wireless networks", IEEE INFOCOM, vol . 3, pp. 1097-1106, Mar. 2000. [16] H . Fattah and C. Leung, " A n overview of Scheduling algorithms in wireless multimedia networks", IEEE Wireless Commun., vol . 9, Issue 5, pp. 76 - 83, Oct. 2002. [17] E . N . Gilbert, "Capacity of a burst-noise channel", The Bell Sys. Tech. Journal, pp. 1253-1265, Sept. 1960. [18] H.S. Wang and N . Moayeri "Finite-state Markov channel - a useful model for radio communication channels", IEEE Trans. Veh. TechnoL, vol. 44, No. 1, pp. 163-171, Feb. 1995. References 11 [19] I E E E Std. 802.1 lb-1999 - Wireless L A N Medium Access Control ( M A C ) and Physical Layer ( P H Y ) specifications, Aug. 20, 1999. [20] I E E E Std. 802.11 a-l999 - Wireless L A N Medium Access Control ( M A C ) and Physical Layer ( P H Y ) specifications, 1999. [21] I E E E Std. 802.1 lg-2003 - Wireless L A N Medium Access Control ( M A C ) and Physical Layer ( P H Y ) specifications, Jun. 27, 2003. [22] S. X u , "Advances in W L A N QoS for 802.11: an overview", IEEE 14th Int. Symposium on Personal, Indoor and Mobile Radio Commun. Proc, vol . 3, pp. 2297-2301, 2003. [23] P. Garg, R. Doshi, R. Greene, M . Baker, M . Malek, and X . Cheng, "Using I E E E 802.1 l e M A C for QoS over wireless", IEEE Int. Proc. Conf. on Performance, Computing, and Commun., pp. 537-542, Apr. 2003. [24] S. Mangold, S. Choi , P. May, O. Kle in , G . Hiertz, and L . Stibor, " I E E E 802.1 le wireless L A N for Quality of Service", Proc. European Wireless 2002, Florence, Italy, Feb. 2002. [25] H . - H . L i u , J . -L .C . Wu , and W . - Y . Chen, "New frame-based network allocation vector for 802.1 l b multirate wireless L A N s " , IEE Proc. Commun., vol . 149, Issue 3, pp. 147 - 151, Jun. 2002. [26] W . C . Jakes, Microwave Mobile Communications. New York: Wiley, 1974. [27] C. Koksal , H . Kassab, and H . Balakrishnan, " A n analysis of short-term fairness in wireless media access protocols", ACMSIGMETRICS'00, Santa Clara, C A , U S A , Jun., 2000. References 78 [28] R. Jain, D - M . Chiu, and W . Hawe, " A quantitative measure of fairness and discrimination for resource allocation i n shared computer systems", DEC (Digital Equipment Corporation) Research Report TR-301, Sept. 1984. [29] K . Lee, M . Cheng, and L . E . Chang, "Wireless QoS analysis for a^Rayleigh fading channel", IEEE Int. Conf. Commun., vol . 2, pp. 1089 - 1093, Jun. 1998. [30] J. Tellado-Mourelo, E . K . Wesel, and J. M . Cioffi , "Adaptive D F E for G M S K in indoor radio channels", IEEE Journal on selected areas in Commun. vol. 14, issue 3, pp. 492-501, Apr. 1996. [31 ] http://www.ee.columbia.edU/~pwhiting/rayleigh.c [32] O P N E T Wireless L A N s Model User Guide, O P N E T Technologies Inc. 2003. [33] O P N E T W M User Guide-Radio Transceiver Pipeline, O P N E T Technologies Inc. 2003. [34] I E E E Std. 802.11b 1999 - Amendment 1: High-speed Physical Layer ( P H Y ) extension in the 2.4 G H z band - Corrigendum 1, 2001. [35] Z . Chen and A . Khokhar, "Improved Mac protocols for D C F and P C F modes over fading channels in wireless L A N s " , IEEE Wireless Commun. and Networking Conf, vol. 2, pp. 1297-1302, Mar. 2003. [36] T.S. Rappaport, Wireless Communications: Principles And Practice. Upper Saddle River, N J : Prentice Hal l , 1996. [37] A . Doufexi, E . Tameh, A . N i x , and S. Armour, "Hotspot wireless L A N s to enhance the performance of 3G and beyond cellular networks", IEEE Commun. Mag., vol. 41, issue 7, pp. 58 - 6 5 , July 2003. [38] S.H. Shah, and K . Nahrstedt, "Channel-aware throughput fairness in multi-cell wireless L A N s " , IEEE Veh. Technol. Conf, vol . 5, pp. 3511 - 3515, Sept. 2004. References 79 [39] B . Bing, Wireless Local Area Networks: The New Wireless Revolution. New York: Wiley, 2002. [40] L . B . Jiang and S. C. Liew, "Proportional fairness in wireless L A N s and ad-hoc networks", IEEE Wireless Commun. and Networking Conf, vol . 3, pp. 1551-1556, 2005. [41] Z . Zhao, L . Zhang, L . Hao and Y . Shu, " A n efficient real-time traffic scheduling algorithm in wireless networks", IEEE Canadian Conf. on Electrical and Compu. Eng., vol. 3, pp. 1543 - 1546, M a y 2003. 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0064772/manifest

Comment

Related Items