國立臺灣大學電機資訊學院電信工程學研究所 碩士論文
Graduate Institute of Communication Engineering College of Electrical Engineering ＆ Computer Science
National Taiwan University Master Thesis
無線網路中以機器學習方法為基礎的新穎速率適應演算法
A Novel Rate Adaptation Algorithm via Machine Learning Approaches for Wireless Networks
陳贊羽 Zan-Yu Chen
指導教授：林宗男 博士
Advisor: Tsung-Nan Lin, Ph.D.
中華民國 98 年 7 月
July, 2009
國立臺灣大學碩士學位論文 口試委員會審定書
無線網路中以機器學習方法為基礎的新穎速率適應演算 法
A Novel Rate Adaptation Algorithm via Machine Learning Approaches for Wireless Networks
本論文係陳贊羽君（R96942099）在國立臺灣大學電信所完成之 碩士學位論文，於民國 98 年 7 月 15 日承下列考試委員審查通過及口 試及格，特此證明
口試委員：
（簽名）
（指導教授）
系主任、所長 （簽名）
誌謝
兩年的碩士生涯在口試結束的剎那，到達了尾巴，感謝指導教授林宗男老師，
在碩一開學前收留了我，並且在這二年中不時的召喚和教導，不厭其煩的憚述著 老師心中的人生大道理，令人受益匪淺、回味無窮。
在苦悶的研究生生活同時，也感謝一起打牌的 keter、昊呆、歐肥、curve 和小魚，在實驗室一起拉琴的震謙，拉著大家玩魔獸的浩儒，安排聚餐的林威，
和四個學弟的幫忙，雖然有的很吵、有的身體不好、有的愛嗆我、有的愛傻笑，
至少給日常的生活添了點不一樣的氣氛，自然還有兩個大學長柏江和士豪這兩年 來的照顧，兩個人撐起來整間 Lab 507，還有感謝光為留下的研究題目。
再來，感謝兩位口試委員陳俊良和蔡子傑教授，不遠千里前來參加我們的口 試，因為你們的建議和批評，讓論文更加臻完善。
最後，感謝我的父母從小到大給了我極大的自由，並且不辭辛勞的培育我完 成了大學和研究所的課業，更在當初決定就讀電信所時，尊重並支持我的決定，
讓我不必煩惱課業之外的事情，專心於研究的課題。
陳贊羽 謹誌 2009 年 6 月
中文摘要
IEEE 802.11 無線網路在最近幾年中成為了最受歡迎的無線網路技術，IEEE 802.11 支援了許多不同的傳輸速率，因此如何去決定一個適當的傳輸速率是一 個挑戰，在這篇論文中，我們提出了一個新穎的速率演算法來解決這個問題，我 們利用了 Maximum Likelihood Estimator 來穩健的預測出每一個速率的傳輸統 計量，再來我們利用了在 PHY 層和 MAC 層兩個跨層的關系來決定每一個速率的傳 輸代價。我們設計的目標是希望達到最大的頻譜使用效率，根據我們嚴謹的模擬 結果，我們所提出的演算法比現存知名的演算法有更佳的表現。同時，WMNs (Wireless Mesh Networks)在過去幾年中擁有可觀的成長，而 WMNs 的效能決定 在路由演算法和速率適應演算法，而各種不同的路由演算法效能已經在各種文獻 中被深度的研究，然而速率適應演算法在 WMNs 中卻只有少量的研究，因此這篇 論文也同時把我們所提出的演算法，實現在 WMNs 中，也同時比較各種演算法在 WMNs 中的表現，經由模擬結果，我們所提出的演算法在 WMNs 的環境下，同樣擁 有傑出的效能。
關鍵字：自應性調變、速率控制、無線網路、IEEE 802.11、WMN
Abstract
In recent years, IEEE 802.11 wireless networks have become the most popular wireless technology. IEEE 802.11 supports multiple transmis- sion rates. How to determine the appropriate transmission rate is chal- lenging. In this paper, we propose a novel rate adaptation algorithm to tackle this problem. We utilize the maximum likelihood estimator to robustly predict the transmission statistics for each transmission rate. Then we exploit the cross-layer correlation between PHY and MAC to determine the transmission cost for each transmission rate.
The goal of our design is to achieve the maximum spectral efficiency.
Based on extensive simulation experiments, the proposed algorithm outperforms existing well-known algorithms. Wireless mesh networks (WMNs) have experienced an enormous growth over the past few years. The performance of WMNs depends on the joint effect of both routing algorithms and rate adaptive algorithms. The performance of various routing algorithms has been studied extensively in the litera- ture. However, little work has been done to evaluate the cross-layer impact of rate adaptive algorithms in WMN environments. In this pa- per, we compare the performance of several rate adaptive algorithms to exploit the multi-hop performance in WMN environments. In ad- dition, a novel rate adaptive algorithm is proposed via the machine
learning approach to robustly reflect the channel information. The goal of our design is to maximize the spectral efficiency. Through extensive computer simulations under different channel and topology environments, experimental results demonstrate the proposed algo- rithm outperforms other existing algorithms in WMN environments.
Contents
List of Figures iii
List of Tables v
1 Introduction 1
1.1 Introduction . . . 1
2 Related Work 5 2.1 ARF and AARF . . . 5
2.2 CARA . . . 6
2.3 RRAA . . . 6
2.4 SampleRate . . . 7
2.5 SLA and SARA . . . 7
3 Proposed Algorithm 9 3.1 Cross-layer performance between MAC and PHY layers . . . 9
3.2 Algorithm Description . . . 11
3.3 Maximum Likelihood Estimator . . . 13
4 Simulation Set-up 17 5 Simulation Results 19 5.1 Single Transmission Link with the Fixed Distance . . . 19
CONTENTS
5.2 Single Transmission Link with Different Distances . . . 23
5.3 Multiple Static Stations in an Infrastructure Mode . . . 24
5.4 Multiple Mobile Stations in an Infrastructure Mode . . . 29
5.5 Topologies of Equal Distances . . . 32
5.6 Topology of Mixed Distances . . . 35
6 Conclusions 41
Bibliography 43
List of Figures
3.1 The average transmission time with respect to packet error prob- ability for different MCSs. . . 13
4.1 The simulation topology of a WMN. . . 18
5.1 The percentage of every rate chosen by algorithms without small scale fading. . . 20 5.2 The percentage of every rate chosen by algorithms with Rayleigh
fading. . . 21 5.3 Throughput in different distance in Rayleigh fading. . . 23 5.4 Multiple nodes vs. throughput in Ricean fading with distance =
200m . . . 25 5.5 The percentage of rate selection of each algorithms with 15 stations
in fig.5.4. (1) is no small scale fading and (2) is with Ricean fading. 26 5.6 Multiple nodes vs. throughput in Rayleigh fading with distance =
300m. . . 27 5.7 The percentage of chosen rates in an environment of 15 contending
stations in Fig.5.6. (1) is no small scale fading and (2) is with Rayleigh fading. . . 28
LIST OF FIGURES
5.8 The aggregate throughputs for different rate adaptation algorithms.
30
5.9 The individual throughput in every node for different rate adapta-
tion algorithms . . . 31
5.10 The aggregate throughput of each algorithm in the distance of 250 meters. . . 32
5.11 The ratio of data rate in the distance of 250 meters. . . 33
5.12 The aggregate throughput of each algorithm in the distance of 370 meters. . . 36
5.13 The ratio of data rates in the distance of 370 meters. . . 37
5.14 The aggregate throughput in the mixed type scenario. . . 38
5.15 The ratio of data rates in in the mixed type scenario. . . 39
List of Tables
3.1 Main Steps of the proposed rate adaptive algorithm . . . 14 5.1 The throughput of each algorithms in Fig.5.1, Fig.5.2, Fig.5.5, and
Fig.5.7 . . . 19
LIST OF TABLES
Chapter 1 Introduction
1.1 Introduction
In recent years, IEEE 802.11 [1] Wireless Local Area Networks (WLANs) have become the most popular technology due to the wide deployment in many ar- eas, such as home, office, airport, and so on. IEEE 802.11 [1] standards define Medium Access Control (MAC) layer and Physical layer (PHY) specifications.
IEEE 802.11 [1] PHY supports multiple transmission rates by different Modula- tions and Coding Schemes (MCS). For example, in 802.11b, 4 data rates, 1 Mbps, 2 Mbps, 5.5 Mbps, and 11 Mbps, are provided at 2.4 GHz band. IEEE 802.11g PHY offers 12 data rates from 6 Mbps to 54 Mbps at 2.4 GHz band.
802.11-based wireless mesh networks (WMNs) have experienced an enormous growth over the past few years. A mesh network consists three components:
Mesh Point (MP), Mesh Portal (MPP) and Mesh Access Point (MAP) according to IEEE 802.11s mesh standard (in draft) [2]. MP supports the mesh services of control, management, and operation of the mesh network. A mesh point which connects mesh networks to other networks is called MPP. And if MP also func- tions like an Access Point (AP), it is called MAP. The performance of WMNs
1. INTRODUCTION
depends on both the routing policy and the transmission rate. The routing policy is a critical subject in mesh networks [3–5]. In mesh networks, the route for a sta- tion to forward data may have a lot of options. In some sense, choosing a distinct route to forward data would lead to different performances. The performances of various routing policies have been studied extensively in the literature. How- ever little work has been done to exploit the cross-layer impact of rate adaptive algorithms in WMNs.
How to pick the most suitable rate is a challenging task since the highest rate may not guarantee the highest throughput. Wireless channel conditions change in time creating random fluctuations of received signal power level because of the multipath, noise, and host mobility. If the chosen rate is too high, the transmitted packet may be failed due to the corruption; on the other hand, selecting a too conservative rate may degrade the throughput significantly.
The rate adaptive algorithm is still left unspecified by IEEE 802.11 standards.
If a station incorrectly chooses the data rate, the total throughput of a system will be significantly degraded due to the phenomenon of performance anomaly [6].
Existing algorithms can be classified into two categories: Signal-to-Noise-Ratio (SNR) based approaches [7] and the statistic-based approaches [8, 9]. However, it is reported recently [10, 11] that there is little correlation between measured SNR and the transmission statistics in MAC layer. That is, it is difficult to derive the reliable transmission rate based on the information of SNR. Recently, several well-known algorithms, including Automatic RateFallback (ARF) [8], Adapta- tive ARF (AARF) [12], Collision Aware Rate Adaptation (CARA) [13], Robust Rate Adaptation Algorithm (RRAA) [9], Stochastic Automata Rate Adaptation (SARA) [14], and SampleRate [11] are proposed using the statistics of the packet
1.1 Introduction
delivery in the MAC layer. Those solutions provide a practical design guideline and report good throughput performances in some testing scenarios. However, several studies [9, 14] also point out the inadequacies of current rate adaptive al- gorithms. The inadequacies include the predefined thresholds [8], probe packets of accessing possible new rates [8, 11, 12], being unable to exploit the short-term channel gain in a dynamic fading environment [11], and the inefficiency in terms of the decision flexibility [9].
With the challenges associated with the rate adaptation in wireless 802.11 environments, we propose a novel approach to exploit the short-term characteris- tics of fading channels in order to achieve the maximum spectral efficiency. Our mechanism, first, utilizes a maximum likelihood estimator to a robust estima- tor to reflect the state of dynamic channels. Then the cross-layer correlation between Medium Access Control (MAC) layer and Physical layer (PHY) are exploited using the well-known two-dimensional Markov model [15]. Therefore, MCS achieving the maximum spectral efficiency can be chosen among all available transmission rates. Our design is designed to be adaptive and does not use any pre-defined thresholds. Additionally, our mechanism is a sender-based approach and is compatible with IEEE 802.11 standard. We evaluate the performances of various algorithms through extensive simulations to study the transmission char- acteristics. Experiments are performed using the commercially available simula- tor QualNet [16]. As shown by the experimental results, the proposed algorithm outperforms the existing algorithms.
1. INTRODUCTION
Chapter 2
Related Work
2.1 ARF and AARF
Automatic Rate Fallback (ARF) [8] is the most well-known rate adaptation al- gorithm, which is used in Lucent Technologies WaveLAN-II networking devices.
ARF uses consecutive ACKs, consecutive frame losses, or timeout to switch dif- ferent rates. Adaptive ARF (AARF) [12], an extension of ARF, doubles the threshold of the consecutively successful ACKs. AARF can achieve a better per- formance in a stable environment. AARF reduces trials for a higher rate while may degrade the system performance. ARF and AARF use consecutive ACKs to raise the rate, but consecutive ACKs is not easy to achieve. Therefore those algorithms tend to choose MCSs with lower transmission rate. Besides, consecu- tive ACKs have less and less correlation with regard to the channel conditions in a multi-station environment.
2. RELATED WORK
2.2 CARA
Collision-Aware Rate Adaptation (CARA) [13], another extention of ARF, adopts RTS/CTS to effectively avoid collisions. If the first ACK is missed, RTS is enabled at the next frame. If the next transmission is still failed, the station will decrease the current rate; otherwise, the station will continue the current rate and disable RTS. The other operations of CARA are just like ARF.
2.3 RRAA
Robust Rate Adaptation Algorithm (RRAA) [9], proposed by Wong, sets two loss ratios to determine rates and an adaptive RTS filter is used. First, RRAA calculates two loss ratio thresholds, Maximum Tolerable Loss threshold (MTL) and Opportunistic Rate Increase threshold (ORI), for each rate according to the ratio of the transmission time between adjacent rates. Then RRAA sets the estimation window size for each rate. When the number of frames transmitted attains to the estimation windows size of the current rate, the station switches the rate in terms of the loss ratio in the window. If the ratio is larger than MTL, the station decreases the rate in the next window, whose size is set by the new rate.
If the ratio is less than ORI, the station increases the rate in the next window.
The rate remains the same in other cases. An estimation window is renewed if the timer expires because RRAA wants to keep the newest statistics in the record.
In addition, RRAA uses adaptive RTS to avoid collisions. The concept is similar to CARA, but RRAA designs an RTS filter to turn on RTS filter. It can prevent from the drawback of RTS oscillation, which means RTS alternates between on and off.
2.4 SampleRate
2.4 SampleRate
SampleRate [11] is introduced by Bicket in his master thesis at MIT. SampleRate calculates the average transmission time of a packet for each rate, and switches the rate by choosing the rate with the minimum average time. SampleRate counts the transmission time in the packet level according to the retries. Per 10th packet, SampleRate will randomly choose a rate which has the transmission time less than the current rate to gather statistics of different rates. SampleRate gathers statistics in the past specified duration ( e.g. 10 seconds). If there are four successive packets failed, SampleRate will freeze the rate until the statistics is obsolete.
2.5 SLA and SARA
SARA [14] and SLA [17] are using learning-automata to adjust the probability of each rate. At the first, each rate have the same probability, and then pick one rate according to the probability vector. If the transmission of the chosen rate is successful, SARA and SLA will increase the probability of this rate and decrease the others. If the channel is static, we believe SARA and SLA will converge into the best rate. However, the channel is time-varying, the best solution is changed depending on the channel dynamic. It may take time to find the new suitable rate while using the approach of learning-automata.
2. RELATED WORK
Chapter 3
Proposed Algorithm
In this paper, we propose a machine-learning-based rate adaption algorithm. The goal of our design is to achieve the maximum throughput. Throughput is defined as the amount of the error-free information that is delivered to the upper-layer of the destination communication system. Not only PHY layer but also MAC layer has an important effect on the throughput. Therefore the cross-layer correlation between MAC and PHY layers on the throughput should be considered. To achieve the maximum throughput, our algorithm is devised to have the following characteristics: 1) to quickly respond to changing wireless channels; 2) to avoid improper adjustment of the transmission rate. Additionally, to be compatible with IEEE 802.11 standard, our algorithm is a sender-based approach.
3.1 Cross-layer performance between MAC and PHY layers
Wireless channel conditions change in time creating random fluctuations of the received power level, or fading. Packets transmitted using Modulation and Coding Scheme (MCS) of high data rate may fail due to the corruption from
3. PROPOSED ALGORITHM
bad channel conditions. If the number of failed transmissions is within the retry count, the packet will be retransmitted again. If the number of failed transmission is larger than the retry count, the packet will be dropped. And the packet drop probability Pdrop is given as:
Pdrop =Pcm+1, (3.1)
wherePc is the packet failed probability (which means a packet encounters either a corruption or a collision) and m is the retry limit. If the delay for a success- fully transmitted packet is defined as the duration from the time the packet is at the front of MAC queue ready to be transmitted, until an acknowledgement for the packet is received [18]. The result can be calculated using the well-known Bianchi’s two-dimensional Markov model [15, 18]. This model considers vari- ous aspects of CSMA/CA behaviors such as the retry count, binary exponential doubling the window length, dropping the packet etc. The main idea of the cal- culation is summarized below. For detailed calculation, please refer to [15, 18].
The average packet delay E[D], which means this packet is transmitted, is given by:
E[D] = E[Ns]∗ E[Tslot]. (3.2) E[Ns], the average number of slot times for successfully transmitting a packet, is given by:
E[Ns] =
m i=0
(Pci− Pcm+1)∗ Wi2+1
1− Pcm+1 , (3.3)
whereWi is the contention window size at the backoff stagei. The average length of a slot time E[Tslot] is equal to:
E[Tslot] = (1− Ptrσ) + Ptr∗ Ps∗ Ts+Ptr∗ (1 − Ps)∗ Tc (3.4)
3.2 Algorithm Description
Whereσ is the duration of an empty slot time, Ptr is the probability that at least one station will transmit a packet and Ps is the probability for a packet which is successfully delivered. Ts and Tc denote the average times the medium is sensed busy because of a successful transmission or collision.
Therefore, the average transmission time to deliver a packet should be cal- culated as the total delays of the successfully transmitted packets plus the time duration which the packets are dropped divided by the successfully transmitted packets. Therefore, the average transmission time E[Tt] can be computed as:
E[Tt] =E[Tdrop] +E[TD] (3.5) Where E[Tdrop], the average time to drop a packet, is equal to :
E[Tdrop] =E[Sdrop]∗ E[Tslot]. (3.6)
Since a packet reaches the retry limitm+1 times, it will be discarded right away.
And let the average number of the slot time required for a packet to be dropped is equal to:
E[Sdrop] =
m i=0
Wi+ 1
2 (3.7)
Fig.3.1shows the average transmission time (Eq.3.5) with respect to different packet failure probabilities.
3.2 Algorithm Description
Our algorithm can be denoted by Rm(C, P (C)) where Rm is the chosen trans- mission rate, C denotes an estimator of channel states, which is to predict the likelihood of success for each MCS promptly, andPC denotes the decision policy
3. PROPOSED ALGORITHM
to choose the rate. The devised mechanism is to maximize the spectral efficiency.
That is to find the minimum average system transmission time tm for a given MCS given the estimated channel information. In other words, our strategy of selecting MCSs can be described as:
Rm = arg min
m tm (3.8)
We utilize the technique of machine learning to robustly estimate the channel state. A maximum likelihood approach is devised to estimate the probability of a packet being transmitted successfully with respect to different transmission rates.
When the information is available, the average transmission time can be obtained easily from the cross-layer analysis result of Fig. 3.1. The behavior of the packet transmission is treated as a random process. The outcome of the event in such a process is either success or failure. After modeling its behavior, a robust feature of the channel can be estimated as described below.
The pseudo code of the proposed algorithm can be summarized in Table 3.1
parameters
T x rate the chosen transmission rate;
T x timei The transmission delay using rate i;
Num[i] the number of using rate i in the past 100 frames;
ACK[i] the successful transmission of using rate i in the past 100 frames;
count determine when to chooses the rate;
enableRT S record whether the RTS/CTS opened;
3.3 Maximum Likelihood Estimator
recACK record whether ACK is missed;
recCT S record whether CTS is received;
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0 20 40 60 80 100 120 140 160 180 200
Frame Error Probability
Delay(ms)
1Mbps 2Mbps 5.5Mbps 11Mbps
Figure 3.1: The average transmission time with respect to packet error probability for different MCSs.
3.3 Maximum Likelihood Estimator
There are two possible outcomes of a frame transmission, success s and failure f. The sample space contains two events, s and f. Let X be a random variable defined by X(s) = 1 and X(f) = 0. The frame success probability is defined as
3. PROPOSED ALGORITHM
Table 3.1: Main Steps of the proposed rate adaptive algorithm
Summary of the proposed algorithm
initialize: Num[i]=0,count=9,recACK=1,recCTS=0 1. repeat
2. if ACK is missed then 3. recACK = 0;
4. if CTS is received then 5. recCTS = 1;
6. if recACK = 0 and enableRTS = 0 then 7. start RTS/CTS;
8. enableRTS = 1;
9. if recCTS = 1 and enableRTS = 1 then 10. finish RTS/CTS;
11. enableRTS = 0;
12. do i = 0 to 3
13. if Num[i] = 0 then FSR[i] = 1;
14. else FSR[i] = ACK[i]/Num[i];
15. get T x time[i] from Fig.??;
16. end
17. if count = 0 then count−−;
18. else
19. do i = 0 to 2
20. ifT x time[i] < T x time[i + 1] then 21. T x rate = i;
22. else
23. T x rate = i + 1;
24. count = 9;
25. end 26.end repeat
3.3 Maximum Likelihood Estimator
the conditional probability that a frame transmission would be successful given a modulation and coding scheme MCSi. It is denoted as pi =P (X = 1|MCSi).
Suppose that there are m MCSs. When a station has a frame to transmit, we are interested in the frame success probabilities of this frame for allm MCSs.
Although this frame is not transmitted yet, an estimated frame success probability pˆi for i = 1,...,m could be obtained by a maximum likelihood estimation method.
The collected samples of X are separated according to each individual MCS, so that there are m sets, denoted as D1, ...Dm. Suppose that Di contains n samples, x1, ..., xn. In Di, α of the samples are 1’s, and the others are 0’s. Since the samples are obtained independently,
p(Di|MCSi) =
n k=1
p(X = xk|MCSi) =pαi × (1 − pi)(n−α) (3.9)
We define a log-likelihood function l(pi) as
l(pi)≡ ln p(Di|MCSi) =α ln pi+ (n − α) ln(1 − pi) (3.10)
The maximum likelihood solution could be written as the argumentpi that max- imizes the log likelihood, i.e.,
pi = arg max
pi l(pi). (3.11)
3. PROPOSED ALGORITHM
Chapter 4
Simulation Set-up
We evaluate the performances of the proposed algorithm via extensive computer simulations using Qualnet 4.0, which provide many wireless protocols and channel models. [16]. We implement several well-known algorithms, such as AARF [12], CARA [13], RRAA [9], SampleRate [11], SLA [17], and SARA [14]. We take 802.11b as MAC protocol. For reality, several time-varying wireless channel mod- els are considered. We use the two-ray ground reflection model as the large scale fading model and Rayleigh and Ricean fading model (with Ricean factor of 3) as the small scale fading model. We set the traffic CBR in a greedy mode, which means a station always has the data to send, at a duration of 100sec and the pay- load size is 1500 bytes. Four scenarios are devised to exploit the characteristics of rate adaptive algorithms: 1) single transmission link with the fixed distance 2) single transmission link with different distances; 3) several transmission links with the same distance in a infrastructure mode; 4) several transmission links with different distances and mobilities in an infrastructure mode. And then we evaluate the performance of each algorithms in 802.11-based mesh network. The topology of the wireless mesh network is shown in Fig. 4.1. Mesh Portal Point is
4. SIMULATION SET-UP
placed at the black node (in the center of the leftmost stations). Three different distances between stations are chosen in order to evaluate the decision flexibil- ity of various algorithms. Since the transmission range of 11 Mbps is about 340 meters, the experiments are performed in three scenarios of (1) 250 meters (the best fixed transmission rate is 11 Mbps), (2) 370 meters (the best fixed rate is 5.5 Mbps), and (3) the mixed distances of 250 and 370 meters. The distances are chosen because, for a single transmission link, the 11 Mbps performs the best in 250 meters and 5.5 Mbps is the best in 370 meters during a static channel condition.In addition, to exploit the responsiveness to sudden changes in channel conditions, three different channel models, including (1) large-scale fading, (2) Ricean fading with Ricean factor of 3 , and (3) Raleigh fading, are utilized. Con- stant Bit Rate traffic of a 1500-byte packet sent every 40 mili-second is adopted in the experiments. The simulation time for each scenario is 60 seconds and the routing protocol of AODV is applied.
Figure 4.1: The simulation topology of a WMN.
Chapter 5
Simulation Results
5.1 Single Transmission Link with the Fixed Dis- tance
Table 5.1: The throughput of each algorithms in Fig.5.1, Fig.5.2, Fig.5.5, and Fig.5.7
Algorithms (1) (2) (3) (4) (5) (6)
AARF 3.81 Mbps 0.73 Mbps 0.97 Mbps 0.85 Mbps 0.83 Mbps 0.39 Mbps CARA 3.69 Mbps 0.75 Mbps 0.89 Mbps 0.56 Mbps 0.35 Mbps 0.23 Mbps RRAA 2.53 Mbps 1.48 Mbps 0.29 Mbps 2.51 Mbps 0.29 Mbps 0.07 Mbps SampleRate 3.85 Mbps 1.06 Mbps 5.24 Mbps 3.23 Mbps 0.25 Mbps 0.14 Mbps Proposed 3.86 Mbps 1.39 Mbps 5.80 Mbps 4.04 Mbps 2.04 Mbps 0.78 Mbps SARA 3.07 Mbps 1.18 Mbps 3.84 Mbps 2.27 Mbps 0.30 Mbps 0.20 Mbps SLA 2.93 Mbps 1.17 Mbps 3.77 Mbps 2.26 Mbps 0.30 Mbps 0.20 Mbps a. (1) (2) is the throughput of Fig.5.1 and Fig.5.2.
b. (4) is the throughput of Fig.5.4 with 15 nodes and (3) is (4) without Ricean fading.
c. (6) is the throughput of Fig.5.6 with 15 nodes and (5) is (6) without Rayleigh fading.
In this scenario, two stations are placed with the distance of 280 meters. In this situation, the rate 5.5 Mbps is the best choice. In this experiment, we want to investigate the rates chosen by different algorithms when only the large scale
5. SIMULATION RESULTS
AARF CARA RRAA SampleRate Proposed SARA SLA
0 10 20 30 40 50 60 70 80 90 100
%
Algorithms
11Mbps 5.5Mbps 2Mbps 1Mbps
Figure 5.1: The percentage of every rate chosen by algorithms without small scale fading.
5.1 Single Transmission Link with the Fixed Distance
AARF CARA RRAA SampleRate Proposed SARA SLA
0 10 20 30 40 50 60 70 80 90 100
%
Algorithms
11Mbps 5.5Mbps 2Mbps 1Mbps
Figure 5.2: The percentage of every rate chosen by algorithms with Rayleigh fading.
5. SIMULATION RESULTS
fading, without the small scale fading, is used. Fig.5.1 shows the result of the percentage. It can be observed that the proposed algorithm and SampleRate can select 5.5 Mbps above 90%, and the proposed algorithm is even a little higher than SampleRate. SARA and SLA use stochastic automata, so both of them rarely choose 2Mbps and 1Mbps. One thing worth mentioning is that the rate chosen by RRAA would oscillate between 5.5 Mbps and 11 Mbps and that is the reason why RRAA has large percentage to choose 11 Mbps. This phenomenon is consistent with the problem of ping-pong effect for RRAA as reported in the literature [19].
Then the channel, not only the large scale fading but also Rayleigh fading, is considered. The purpose is to evaluate the decision flexibility and responsive- ness of various algorithms to explore the short term characteristics of channel dynamics. Because the channel is varying dramatically with Rayleigh fading, the rate of 5.5 Mbps is not always the best choice at some time. When the channel is getting better, 11 Mbps would become the best choice. While the channel is getting worse, transmissions may be failed with the rate 5.5Mbps. From Fig.5.2, we can see that this time the proposed algorithm doesn’t choose the rate of 5.5 Mbps with a high percentage, but SampleRate still selects 5.5 Mbps for almost 90%. When the channel is better, the proposed algorithm would choose 11 Mbps, and select 2 Mbps or 1 Mbps instead while the channel is bad. This can explain why, in Fig.5.3, it shows that the performance the proposed algorithm does ac- tually better than SampleRate does. AARF and CARA take 1Mbps for most of the time because it is difficult for their mechanism to select a high data rate in a noisy environment.
5.2 Single Transmission Link with Different Distances
5.2 Single Transmission Link with Different Dis- tances
1500 200 250 300 350 400 450 500
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Distance(m)
Throughput(Mbps)
AARF
CARA
RRAA
SampleRate
Proposed
SARA
SLA
Figure 5.3: Throughput in different distance in Rayleigh fading.
In this scenario, two static stations are placed with different distances, from 150m to 500m, and Rayleigh small-scale fading model is assumed. Fig.5.3demon- strates the throughput performance. In general, the throughput of every rate adaptive algorithm decreases as the distance increases . Among existing algo- rithms, RRAA performs the best. Similar observations has been reported [9].
We observe that the proposed algorithm performs almost the same as RRAA does in this scenario, better than SampleRate does. Because Rayleigh fading is a violent dynamic model, the timely and robust statistics is necessary in order
5. SIMULATION RESULTS
to make the decision of choosing the appropriate MCS promptly in such chan- nel conditions. The proposed algorithm exhibits such a characteristic and this is why it always produces better performance than others when the distance is below 330m. AARF and CARA are little higher than the proposed algorithm and RRAA when the distance is longer than 330m. 1 Mbps may be the best choice when the distance gets longer and longer; when the distance is longer than 330m, 1 Mbps would become the only possible choice to make a successful transmission.
AARF and CARA choose the rate of 1 Mbps because they require consecutive successful transmissions in order to choose the higher data rate. The proposed algorithm and RRAA would choose the rates other than 1 Mbps in some cases, not doing as good as AARF or CARA does.
5.3 Multiple Static Stations in an Infrastructure Mode
In the experiment, we study the performance of different rate adaptive algorithms in an extensive contention environment with error-prone fading channels. Access Point (AP) stays statically in the center and we place a number of contending hosts around AP with the same radius. Experiments are performed with two distances (200m and 300m) and different fading channels. Fig.5.4 presents the aggregate throughput performances with respect to different numbers of hosts at the distance of 200m with the small-scale Ricean fading model. It should be noticed that the best MCS should be 11Mbps at this distance. It can be observed that the proposed algorithm, RRAA, and SampleRate produce comparable results when the number of contending hosts is small (less than 8). When the number
5.3 Multiple Static Stations in an Infrastructure Mode
0 5 10 15 20 25
0 1 2 3 4 5 6
Number of Nodes
Throughput (Mbps)
AARF CARA RRAA Sample Proposed SARA SLA
Figure 5.4: Multiple nodes vs. throughput in Ricean fading with distance = 200m
5. SIMULATION RESULTS
0 10 20 30 40 50 60 70 80 90 100
Algorithms
%
11Mbps 5.5Mbps 2Mbps 1Mbps
(1) (2) (1) (2) (1) (2) (1) (2) (1) (2) (1) (2) (1) (2)
AARF CARA RRAA SampleRate Proposed SARA SLA
Figure 5.5: The percentage of rate selection of each algorithms with 15 stations in fig.5.4. (1) is no small scale fading and (2) is with Ricean fading.
5.3 Multiple Static Stations in an Infrastructure Mode
0 5 10 15 20 25
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Number of Nodes
Throughput (Mbps)
AARF CARA RRAA Sample Proposed SARA SLA
Figure 5.6: Multiple nodes vs. throughput in Rayleigh fading with distance = 300m.
5. SIMULATION RESULTS
0 10 20 30 40 50 60 70 80 90 100
Algorithms
%
11Mbps 5.5Mbps 2Mbps 1Mbps
(1) (2) (1) (2) (1) (2) (1) (2) (1) (2) (1) (2) (1) (2)
AARF CARA RRAA SampleRate Proposed SARA SLA
Figure 5.7: The percentage of chosen rates in an environment of 15 contending stations in Fig.5.6. (1) is no small scale fading and (2) is with Rayleigh fading.
5.4 Multiple Mobile Stations in an Infrastructure Mode
of contending hosts is large, RRAA breaks down quickly. The original paper of RRAA [9] also reports the weakness of the design. To further investigate the detailed behavior, we plot the percentage of the rate selection for every algorithm with 15 stations in Fig.5.4 experiment. The results are shown in Fig. 5.5. There are two bars for each algorithm. The left bar is the experimental result when only the large-scale fading is assumed, and is displayed for the purpose of comparison.
In only large scale fading channels, RRAA, SampleRate, and the proposed can easily make the correct decision (11Mbps). When the small-scale Ricean fading is applied, RRAA makes lots of wrong decisions and SampleRate can not make the flexible decision promptly. On the other hand, the proposed algorithm can well exploit the short-term characteristics of channel dynamics. Fig.5.6 displays the performance results when the distance becomes 300m and Rayleigh fading model is used. At this distance, the best MCS should be 5Mbps. The percentage results for both large-scale fading channels and Rayleigh channels are plotted in Fig. 5.7.
5.4 Multiple Mobile Stations in an Infrastruc- ture Mode
In this scenario, an 800m*800m area is created, AP is statically placed in the cen- ter, and the small scale fading model is Rayleigh. Twenty stations are randomly placed in the area, and they move randomly at speeds from 1m/s to 10m/s.
We generate the same traffic, saturated CBR traffic, from each station to AP.
Small-scale Rayleigh fading channels are used in the experiment. The aggregate throughputs of different algorithms are shown in Fig.5.8. And the Fig.5.9 shows
5. SIMULATION RESULTS
AARF RRAA CARA SAMPLE SARA SLA Proposed 0
0.5 1 1.5 2 2.5 3 3.5
Algorithms
Throughput (Mbps)
Figure 5.8: The aggregate throughputs for different rate adaptation algorithms.
5.4 Multiple Mobile Stations in an Infrastructure Mode
2 4
6 8
10 12
14 16
18 20
Proposed SLASARA
SAMPLE CARARRAA
AARF 0
0.1 0.2 0.3 0.4
Station Index Algorithms
Throughput(Mbps)
Figure 5.9: The individual throughput in every node for different rate adaptation algorithms
5. SIMULATION RESULTS
the throughput of each node. We observe that our mechanism can achieve the performance gain over the others more than 200%.
ARF AARF CARA RRAA SARA SampleRate Proposed 0
0.5 1 1.5 2
x 106
Rate Adaptation Algorithms
Aggregate Throughput (bps)
Path Loss Ricean Rayleigh
11 Mbps (Path Loss)
Figure 5.10: The aggregate throughput of each algorithm in the distance of 250 meters.
5.5 Topologies of Equal Distances
In the experiments, we compare the performances in two topologies of equal distances (250 meters and 370 meters). Fig. 5.10 and Fig. 5.12 show the aggre- gate throughputs respectively with different rate adaptation algorithms, in which
”Path loss” means two-ray ground reflection in the channel condition (large scale
5.5 Topologies of Equal Distances
ARF AARF CARA RRAA SARA SampleRate Proposed 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Rate Adaptation Algorithms
The Ratio of Data Rate
1 Mbps 2 Mbps 5.5 Mbps 11 Mbps
Path Loss
Rayleigh
Figure 5.11: The ratio of data rate in the distance of 250 meters.
5. SIMULATION RESULTS
fading) is adopted and the line acroses the graph means the performance of the best fixed-rate for the comparison purpose.
Both results show the performances of all algorithms degrade when the channel fluctuations become severe. ARF-based algorithms use consecutive transmission success or failure to determine the rate. From both results, it can be observed the performances of such algorithms do not perform well since the characteristics of such a design make it difficult to choose the correct rate in a wireless mesh environment. Those algorithms attribute all transmission failures to corruption from error-prone channels. Therefore those algorithms are likely to incorrectly choose the rate in a wireless mesh environment where the collision may occur frequently. This phenomenon can be better explained in Fig. 5.11 and Fig. 5.13.
The figures show the percentage of chosen data rates during the experiments.
For the purpose of clarity, we only plot the results of two fading models of ”Path loss” (more stable) and ”Raleigh” (more random fluctuations). From Fig. 5.11, the percentage of 11 Mbps is low even in a stable channel condition. Although RRAA [9] is reported to be the best rate adaptive algorithm in the literature and the authors ”believe the proposed algorithm to perform well in 802.11-based wireless mesh networks”, our experimental results show an important observa- tion that the performance of RRAA suffers in all channel conditions in a mesh environment. As [9] indicates that, when the channel condition becomes worse (i.e. SNR is low or contention level is high), RRAA can not deal with collision problem, the phenomenon can be easily observed from Fig.5.12. In such a harsh environment, ARF-like algorithms even perform better than RRAA. From the rate-percentage results of Fig. 5.11 and Fig. 5.13 , the likelihood of chosen the correct data rate is not high in ”Path loss” channel model. One of the reasons
5.6 Topology of Mixed Distances
can be attributed to that RRAA sets the threshold of each data rate in advance, then calculate MTL and ORI according to the estimation window [9] of each data rate. The determined threshold does not show enough flexility to sudden changes in channel conditions during the multi-hop transmission.
SARA reports better performance than RRAA in both distances. How- ever, the rate-percentage results (Fig. 5.11 and Fig. 5.13) for ”Path loss” and
”Rayleigh” is almost identical. This analysis demonstrates SARA lack the re- sponsiveness to the channel fading. Therefore, such a characteristics results in the performance degradation as observed in Fig.5.10 and Fig. 5.12.
SampleRate and the proposed algorithm demonstrate the robust performance for all channel conditions in both topologies. The capability of chosen the correct rate rate can be easily shown from Fig. 5.11 and Fig. 5.13 in ”Path loss” chan- nel model. This explains why the proposed algorithm outperforms all existing well-known algorithms. From the results of fading channel models, they further demonstrate the proposed algorithm exhibits superior responsiveness compared to other algorithms and can better exploit the short-term channel variations. That is the reason why the proposed mechanism shows the best throughput performance as shown in Fig.5.10 and Fig. 5.12.
5.6 Topology of Mixed Distances
In this experiment, the distances between stations are mixed with 250 meters and 370 meters in order to create a heterogeneous environment. Fig. 5.14 shows the aggregate throughput with different fading models. The best fixed-rate is 11 Mbps and is plotted for a reference. Fig.5.15 shows the percentage
5. SIMULATION RESULTS
ARF AARF CARA RRAA SARA SampleRate Proposed 0
0.5 1 1.5 2
x 106
Rate Adaptation Algorithms
Aggregate Throughput (bps)
Path Loss Ricean Rayleigh
5.5 Mbps (Path Loss)
Figure 5.12: The aggregate throughput of each algorithm in the distance of 370 meters.
5.6 Topology of Mixed Distances
ARF AARF CARA RRAA SARA SampleRate Proposed 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Rate Adaptation Algorithms
The Ratio of Data Rate
1 Mbps 2 Mbps 5.5 Mbps 11 Mbps
Path Loss
Rayleigh
Figure 5.13: The ratio of data rates in the distance of 370 meters.
5. SIMULATION RESULTS
ARF AARF CARA RRAA SARA SampleRate Proposed 0
0.5 1 1.5 2
x 106
Rate Adaptation Algorithms
Aggregate Throughput (bps)
Path Loss Ricean data3
11 Mbps (Path Loss)
Figure 5.14: The aggregate throughput in the mixed type scenario.
5.6 Topology of Mixed Distances
ARF AARF CARA RRAA SARA SampleRate Propsed 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Rate Adaptation Algorithms
The Ratio of Data Rate
1 Mbps2 Mbps5.5 Mbps 11 Mbps
Path Loss
Rayleigh
Figure 5.15: The ratio of data rates in in the mixed type scenario.
5. SIMULATION RESULTS
Chapter 6 Conclusions
In this paper, we study the rate adaption problem in IEEE 802.11 wireless net- works and evaluate the performance of rate adaptive algorithms in 802.11-based wireless mesh network environments. The crux of the problem is to determine the state of the communication channels correctly and make the decision promptly.
We propose a novel cross-layer approach to tackle via a machine learning ap- proach. Maximum likelihood estimator is utilized to robustly estimate the chan- nel state. Then the joint correlation between PHY and MAC is exploited in order to evaluate the performances of available MCSs. Our decision strategy is to achieve the maximum spectral efficiency. We evaluate the performances of the proposed approach as well as several existing algorithms through extensive sim- ulations. The scenarios we consider include different topologies, fading channels, mobility, and various contending nodes. Experimental results show the proposed algorithm outperforms exiting algorithms in all scenarios and various wireless mesh network topologies.
6. CONCLUSIONS
Bibliography
[1] “IEEE standard for information technology-telecommunications and infor- mation exchange between systems-local and metropolitan area networks- specific requirements - part 11: Wireless lan medium access control (mac) and physical layer (phy) specifications,” IEEE Std 802.11-2007 (Revision of IEEE Std 802.11-1999), pp. C1–1184, 12 2007.
[2] J. Camp and E. Knightly, “The IEEE 802.11s extended service set mesh networking standard,” Communications Magazine, IEEE, vol. 46, no. 8, pp.
120–126, August 2008.
[3] D. Johnson, “Routing in ad hoc networks of mobile hosts,” Mobile Computing Systems and Applications, 1994. Proceedings., Workshop on, pp. 158–163, Dec 1994.
[4] “Optimized link state routing protocol (olsr),” 2003.
[5] C. Perkins and E. Royer, “Ad-hoc on-demand distance vector routing,” Mo- bile Computing Systems and Applications, 1999. Proceedings. WMCSA ’99.
Second IEEE Workshop on, pp. 90–100, Feb 1999.
[6] G. B.-S. A. D. Martin Heusse, Franck Rousseau, “Performance anomaly of 802.11b,” IEEE INFOCOM, 2003.
[7] G. Holland, N. Vaidya, and P. Bahl, “A rate-adaptive mac protocol for multi-hop wireless networks,” in MobiCom ’01: Proceedings of the 7th annual international conference on Mobile computing and networking. New York, NY, USA: ACM, 2001, pp. 236–251.
BIBLIOGRAPHY
[8] A. Kamerman and L. Monteban, “Wavelan-ii: a high-performance wireless lan for the unlicensed band,” Bell Labs Technical Journal, vol. 2, no. 3, pp.
118–133, 1997. [Online]. Available: http://dx.doi.org/10.1002/bltj.2069 [9] S. H. Y. Wong, H. Yang, S. Lu, and V. Bharghavan, “Robust rate adapta-
tion for 802.11 wireless networks,” in MobiCom ’06: Proceedings of the 12th annual international conference on Mobile computing and networking. New York, NY, USA: ACM, 2006, pp. 146–157.
[10] D. Aguayo, J. Bicket, S. Biswas, G. Judd, and R. Morris, “Link-level mea- surements from an 802.11b mesh network,” SIGCOMM Comput. Commun.
Rev., vol. 34, no. 4, pp. 121–132, 2004.
[11] J. C. Bicket, “Bit-rate selection in wireless networks,” Master’s thesis, De- partment of EECS, MIT, February 2005, sampleRate.
[12] M. Lacage, M. H. Manshaei, and T. Turletti, “IEEE 802.11 rate adaptation:
a practical approach,” in MSWiM ’04: Proceedings of the 7th ACM inter- national symposium on Modeling, analysis and simulation of wireless and mobile systems. New York, NY, USA: ACM, 2004, pp. 126–134.
[13] J. Kim, S. Kim, S. Choi, and D. Qiao, “CARA: Collision-aware rate adapta- tion for IEEE 802.11 WLANs,” in INFOCOM 2006. 25th IEEE International Conference on Computer Communications. Proceedings, Apr 2006, pp. 1–11.
[14] D. S. Trrun Joshi, Disha Ahuja and I. Dharma P. Agrawal, Fellow,
“Sara: Stochastic automata rate adaptation for IEEE 802.11 networks,”
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYS- TEMS, vol. 19, p. 11, 2008.
[15] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed coordina- tion function,” Selected Areas in Communications, IEEE Journal on, vol. 18, no. 3, pp. 535–547, Mar 2000.
[16] http://www.scalablenetworks.com/, “Qualnet network modeler and simula- tor.”
BIBLIOGRAPHY
[17] M. Haleem and R. Chandramouli, “Adaptive downlink scheduling and rate selection: a cross-layer design,” Selected Areas in Communications, IEEE Journal on, vol. 23, no. 6, pp. 1287–1297, June 2005.
[18] P. Chatzimisios, A. Boucouvalas, and V. Vitsas, “Ieee 802.11 packet delay- a finite retry limit analysis,” Global Telecommunications Conference, 2003.
GLOBECOM ’03. IEEE, vol. 2, pp. 950–954 Vol.2, Dec. 2003.
[19] K. Ramachandran, R. Kokku, H. Zhang, and M. Gruteser, “Symphony: syn- chronous two-phase rate and power control in 802.11 wlans,” pp. 132–145, 2008.
