Efficient Wireless Link Bandwidth Detection for IEEE Networks

Similar documents
A New Space-Repetition Code Based on One Bit Feedback Compared to Alamouti Space-Time Code

A SELECTIVE POINTER FORWARDING STRATEGY FOR LOCATION TRACKING IN PERSONAL COMMUNICATION SYSTEMS

Application of Improved Genetic Algorithm to Two-side Assembly Line Balancing

CHAPTER 5 A NEAR-LOSSLESS RUN-LENGTH CODER

APPLICATION NOTE UNDERSTANDING EFFECTIVE BITS

4. INTERSYMBOL INTERFERENCE

Analysis of SDR GNSS Using MATLAB

x 1 + x x n n = x 1 x 2 + x x n n = x 2 x 3 + x x n n = x 3 x 5 + x x n = x n

WSN Node Localization Regularization Algorithm Based on Quasi Optimal Criterion Parameter Selection

Cross-Layer Performance of a Distributed Real-Time MAC Protocol Supporting Variable Bit Rate Multiclass Services in WPANs

High Speed Area Efficient Modulo 2 1

SHORT-TERM TRAVEL TIME PREDICTION USING A NEURAL NETWORK

Radar emitter recognition method based on AdaBoost and decision tree Tang Xiaojing1, a, Chen Weigao1 and Zhu Weigang1 1

Efficient Feedback-Based Scheduling Policies for Chunked Network Codes over Networks with Loss and Delay

Broadcasting in Multichannel Cognitive Radio Ad Hoc Networks

Importance Analysis of Urban Rail Transit Network Station Based on Passenger

Sapana P. Dubey. (Department of applied mathematics,piet, Nagpur,India) I. INTRODUCTION

ECE 333: Introduction to Communication Networks Fall Lecture 4: Physical layer II

Compound Controller for DC Motor Servo System Based on Inner-Loop Extended State Observer

Logarithms APPENDIX IV. 265 Appendix

Capacity of Large-scale CSMA Wireless Networks

EMU-Synchronization Enhanced Mobile Underwater Networks for Assisting Time Synchronization Scheme in Sensors

A New Energy Efficient Data Gathering Approach in Wireless Sensor Networks

PHY-MAC dialogue with Multi-Packet Reception

Lecture 4: Frequency Reuse Concepts

Novel pseudo random number generation using variant logic framework

A study on the efficient compression algorithm of the voice/data integrated multiplexer

The Detection of Abrupt Changes in Fatigue Data by Using Cumulative Sum (CUSUM) Method

LETTER A Novel Adaptive Channel Estimation Scheme for DS-CDMA

Spread Spectrum Signal for Digital Communications

X-Bar and S-Squared Charts

MEASUREMENT AND CONTORL OF TOTAL HARMONIC DISTORTION IN FREQUENCY RANGE 0,02-10KHZ.

Data Mining of Bayesian Networks to Select Fusion Nodes from Wireless Sensor Networks

SELEX Elsag. 5/18/2012 R. Pucci SDR 12 WinnComm 1

Adaptive Resource Allocation in Multiuser OFDM Systems

x y z HD(x, y) + HD(y, z) HD(x, z)

Wi-Fi or Femtocell: User Choice and Pricing Strategy of Wireless Service Provider

Unit 5: Estimating with Confidence

Subcarriers and Bits Allocation in Multiuser Orthogonal Frequency Division Multiplexing System

AkinwaJe, A.T., IbharaJu, F.T. and Arogundade, 0.1'. Department of Computer Sciences University of Agriculture, Abeokuta, Nigeria

COMPRESSION OF TRANSMULTIPLEXED ACOUSTIC SIGNALS

Data Acquisition System for Electric Vehicle s Driving Motor Test Bench Based on VC++ *

Sensors & Transducers 2015 by IFSA Publishing, S. L.

Is Diversity Gain Worth the Pain: Performance Comparison Between Opportunistic Multi-Channel MAC and Single-Channel MAC

Distributed Resource Management in Multi-hop Cognitive Radio Networks for Delay Sensitive Transmission

SCALABLE MODEL FOR THE SIMULATION OF OLSR AND FAST-OLSR PROTOCOLS

Enhancement of the IEEE MAC Protocol for Scalable Data Collection in Dense Sensor Networks

lecture notes September 2, Sequential Choice

Methods to Reduce Arc-Flash Hazards

Design of FPGA Based SPWM Single Phase Inverter

General Model :Algorithms in the Real World. Applications. Block Codes

A New Design of Log-Periodic Dipole Array (LPDA) Antenna

Comparison of Frequency Offset Estimation Methods for OFDM Burst Transmission in the Selective Fading Channels

A novel adaptive modulation and coding strategy based on partial feedback for enhanced MBMS network

Performance Analysis of Channel Switching with Various Bandwidths in Cognitive Radio

Time-Space Opportunistic Routing in Wireless Ad Hoc Networks, Algorithms and Performance

Optimal Arrangement of Buoys Observable by Means of Radar

Measurements of the Communications Environment in Medium Voltage Power Distribution Lines for Wide-Band Power Line Communications

An Adaptive Image Denoising Method based on Thresholding

CHAPTER 8 JOINT PAPR REDUCTION AND ICI CANCELLATION IN OFDM SYSTEMS

INCREASE OF STRAIN GAGE OUTPUT VOLTAGE SIGNALS ACCURACY USING VIRTUAL INSTRUMENT WITH HARMONIC EXCITATION

Measurement of Equivalent Input Distortion AN 20

Introduction to Wireless Communication Systems ECE 476/ECE 501C/CS 513 Winter 2003

ELEC 350 Electronics I Fall 2014

Politecnico di Milano Facoltà di Ingegneria dell Informazione. Wireless Networks. Prof. Antonio Capone

A Radio Resource Allocation Algorithm for QoS Provision in PMP-based Systems

Distributed Resource Management in Multi-hop Cognitive Radio Networks for Delay Sensitive Transmission

Sea Depth Measurement with Restricted Floating Sensors

Spectrum Access Games for Cognitive Radio Networks

Design of FPGA- Based SPWM Single Phase Full-Bridge Inverter

Outline. Motivation. Analog Functional Testing in Mixed-Signal Systems. Motivation and Background. Built-In Self-Test Architecture

CCD Image Processing: Issues & Solutions

Peer-to-Peer Protocols and Data Link Layer

Density Slicing Reference Manual

Summary of Random Variable Concepts April 19, 2000

Fingerprint Classification Based on Directional Image Constructed Using Wavelet Transform Domains

Ch 9 Sequences, Series, and Probability

BOTTLENECK BRANCH MARKING FOR NOISE CONSOLIDATION

ON THE FUNDAMENTAL RELATIONSHIP BETWEEN THE ACHIEVABLE CAPACITY AND DELAY IN MOBILE WIRELESS NETWORKS

Optimization of Base Station and Maximizing the Lifetime of Wireless Sensor Network

Optimal Geolocation Updating for Location Aware Service Provisioning in Wireless Networks

An Optimal Test Pattern Selection Method to Improve the Defect Coverage

The Potential of Dynamic Power and Sub-carrier Assignments in Multi-User OFDM-FDMA Cells

Adaptive MMSE Rake-Equalizer Receiver Design with Channel Estimation for DS-UWB System

Neighbor Discovery for Cognitive Radio Ad Hoc Networks

Efficient Energy Consumption Scheduling: Towards Effective Load Leveling

Secret Searching in Wireless Sensor Networks with RFIDs

CAEN Tools for Discovery

Fast Sensor Deployment for Fusion-based Target Detection

Solution 2 Discussion:

Low Latency Random Access with TTI Bundling in LTE/LTE-A

The Fundamental Capacity-Delay Tradeoff in Large Mobile Ad Hoc Networks

Broadcast Throughput Capacity of Wireless Ad Hoc Networks with Multipacket Reception

Capacity of Large-scale CSMA Wireless Networks

Intermediate Information Structures

Estimation of non Distortion Audio Signal Compression

Information-Theoretic Analysis of an Energy Harvesting Communication System

AME50461 SERIES EMI FILTER HYBRID-HIGH RELIABILITY

Faulty Clock Detection for Crypto Circuits Against Differential Faulty Analysis Attack

Tehrani N Journal of Scientific and Engineering Research, 2018, 5(7):1-7

Transcription:

Efficiet Wireless Lik Badwidth Detectio for IEEE 80. Networks Haohua Fu, Lidog Li ad Weijia Jia Dept. of Computer Eg & IT, City Uiverty of Hog Kog 83 Tat Chee Ave, Kowloo, Hog Kog, SAR Chia Email: fu.haohua@studet.cityu.edu.hk; Li.lidog@studet.cityu.edu.hk; itjia@cityu.edu.hk Abstract-I order to provide accurate ad real-time badwidth iformatio ad ehace the QoS service for badwidth-setive applicatios i the dyamic chagig wireless etwork, this paper proposes a efficiet method for Wireless Badwidth Detectio (WBD) ug packet probig approach o mobile odes. The ovelty ad cotributios of WBD are three-fold: () Efficiecy: by meas of sedig probe packets with various zes, WBD ca determie the wireless lik badwidth with light load ad short time duratio; () Accuracy: ug differet mechaisms to filter out the radom time variatio, a high detectio accuracy ca be achieved; (3) Stability: our algorithm attais stable results uder differet cross traffic coditios. Experimetal data observed through exteve mulatios shows the effectiveess ad efficiecy of WBD. Keywords-IEEE 80.; Probe packet; Wireless badwidth detectio.. INTRODUCTION Badwidth has bee a critical resource i various kids of etworks. Havig a good kowledge about the real-time badwidth value ca greatly improve the performace of data-iteve applicatios such as file trasfers or multimedia streamig []. Bedes, may QoS maagemet operatios, such as admiso cotrol, resource reservatio, etc, make their decios accordig to the characteristics ad chages of the lik badwidth. Thus, badwidth is a key cocept i cotet distributio, itelliget routig systems, ed-to-ed admiso cotrol, video/audio streamig ad may other etwork fields []. I most wired etworks, the lik badwidth is assumed to be a fixed value over period of time. However, this assumptio may ot hold i the wireless etworks where lik badwidth ca be temporarily degraded due to evirometal reasos such as gal fadig, lik sharig or mobility of the mobile odes, etc [3]. Accordig to IEEE 80.b stadards, the lik badwidth may chage from Mbps dow to Mbps with the variatios of gal ad oise stregths. This idicates that the commoly kow values of the phycal badwidth caot reflect the actual badwidth used i the etwork at a istat time. Thus, i order to provide more accurate wireless badwidth iformatio for mobile odes (MN), this paper presets a ovel algorithm for the MNs to detect the real-time badwidth value. The ovelty ad cotributios of our algorithm are: () Efficiecy: by meas of sedig probe packets with various zes from a source (MN) to a destiatio (i.e. base statio, BS), efficiet computatio ug liear regreso is applied to fast determie the wireless lik badwidth withi a short period of time. () Accuracy: ug two differet kids of mechaisms to filter out the radom time variatio, our algorithm ca achieve a high accuracy; (3) Stability: our algorithm ca achieve a stable result uder differet cross traffic coditios, i.e., the detectio results do ot fluctuate with the chages of cross traffic. The rest of the paper is orgaized as follows. Related work is discussed i Sectio. Sectio 3 itroduces a efficiet algorithm for lik badwidth detectio ug packet probig for IEEE 80. wireless etwork. Sectio 4 presets the experimetal evaluatios ad the fial sectio cocludes the paper.. RELATED WORK I geeral, the techiques of lik badwidth detectio fall ito two categories: itruve [6][7][8][] ad o-itruve detectios [4]. Itruve detectios refer to the cotrolled ijectio of packets ito the etwork ad the subsequet collectio of packets (the same packet or a remotely geerated ACK packet). Based o the iformatio collected from these packets, such as roud trip time delays, oe may estimate the lik badwidth [5]. There are a umber of itruve approaches for estimatig lik badwidth i wired etwork. Typically, packet pair techology [6][7] estimates the bottleeck lik badwidth with a relative error about 7% by measurig the roud-trip-time (RTT) of the back-to-back packets. However, these wired methods geerally require a log time to coverge to a accurate result ad are ot suitable for badwidth detectio of wireless etwork which may chage frequetly with time. A umber of itruve approaches have also bee proposed to detect the available badwidth, such as the scheme i [], which estimate the curret available badwidth of wireless etwork based o the recorded throughput. No-itruve detectio methods measure the etwork behavior by observig the packet-arrival rate or other data collected at a ed-system, makig some deductio o the state of the etwork, thereby deducig the lik badwidth o the bas of these observatios. For wireless etwork, a o-itruve detectio scheme is proposed i [4], which uses Sigal-to-Noise-Ratio (SNR) to estimate wireless lik badwidth. The disadvatage of this approach is that it eeds to apply BP-eural-etwork to trai a set of data first, which is quite time-cosumig. I additio, the approach geerates

relative errors ragig from 0% to 5%, which is relatively high compared with the itruve detectio. No-itruve schemes have relatively slight ifluece o the etwork. However, they ormally require moitorig at the receivig ode, which is ot suitable for the o-demad applicatios implemeted at the sedig ode. Bedes, the relative error of the result is also high due to the way of pasve detectio. I this paper we focus o the phycal badwidth detectio ug itruve probig scheme. The probig techiques may itroduce some overhead to etwork durig the active detectio. Therefore, the itruve traffic should be kept small so as ot to add too much overhead to the etwork. 3. BANDWIDTH DETECTION FOR 80. NETWORKS This sectio presets our efficiet method for wireless lik badwidth detectio o IEEE 80. wireless etwork. Probig packets are set to acquire the RTT delay samples, which will later be used for the calculatio of the lik badwidth. We will first review the MAC mechaisms of IEEE 80. ad the give the details of our scheme. 3. MAC Mechaisms of IEEE 80. I 80. WLAN, the fudametal MAC mechaism is a Distributed Coordiatio Fuctio (DCF) kow as carrier sese multiple access with collio avoidace (CSMA/CA). There is also a optioal Poit Coordiatio Fuctio (PCF), which is a cetralized MAC protocol to support collio-free ad time-bouded services. I this paper, we limit our discuso to DCF mode. DCF defies two differet mechaisms for packet trasmiso as pictured i Fig.. The bac access mechaism is a two-way hadshakig scheme, which is characterized by the immediate trasmiso of a potive ackowledgemet (ACK) from the destiatio statio upo successful receptio of a packet trasmitted by the seder statio. Aother mechaism is a four-way hadshakig scheme, kow as request-to-sed/clear-to-sed (RTS/CTS) mechaism. I this mechaism, the source statio will first sed a special RTS short frame to reserve the wireless chael. The the destiatio statio ackowledges the receipt of the RTS frame by sedig back a CTS frame, after which the data packet ad ACK respose will be trasmitted. I the RTS/CTS mechaism (Fig. (b)), the roud trip time of a packet costs of three differet parts: () t b : the waitig time before a successful packet trasmiso, which icludes the time that the chael is detected busy, the time for usuccessful trasmisos (collided trasmisos) ad the radom cotetio time; () t d : the trasmiso time of the packet, defied as t d = s/c L, where s is the ze of the packet, C L is the wireless lik badwidth to be detected; ad (3) t c : a costat icludig the time spacigs such as SIFS ad DIFS, the time duratios for trasmittig fixed-legth short frames such as RTS, CTS, ad ACK, ad the propagatio delays. For the bac access mechaism (Fig. (a)), the total roud trip time ca be milarly established as RTC/CTS mechaism. The oly differece is that the t c i bac access mechaism does ot cotai the trasmiso delay for sychroous cotrol packets RTS/CTS. Fig.. Two differet Access Mechaisms of DCF. 3. Our approach A mobile ode (MN) that iteds to detect the badwidth should sed the probe packets to the BS ad collect the RTT data upo receptio of ACK messages from the BS. The MN uses the RTT to estimate the badwidth o the wireless lik betwee the MN ad BS. To compute the RTT, deote t s as the time whe the packet is ready to be set at the MAC layer of the MN ad t r as the time whe the correspodig ACK is received as show i Fig.. The the RTT of a probe packet with ze s ca be expressed as: s RTT = tr ts = tb + td = tb + (3.) CL Note that expreso (3.) will be used for the calculatio of the badwidth C L o the wireless lik L. As we have discussed before, t b is the waitig time due to cotetio of the lik, ad it may chage radomly with the evirometal coditios, thus it is a ucotrollable factor whe we estimate C L. Sice our purpose is to detect the value of C L, we are ot iterested i fidig the value of t b. Therefore, we first discuss some approaches to reduce the ifluece of t b by covertig t b ito a costat ad the give our algorithm to calculate the badwidth with the RTT values detected. 3.. Reducig t b s Ifluece Geerally, t b ca be regarded as radom values for differet packet s RTTs. However, ce it is the waitig time before the trasmiso of the data packet, its variatios are idepedet of the packet zes. I fact, the differet values of t b are mostly caused by the cross competig traffic. The heavy cross traffic ca make the chael busy for a loger time, resultig i collios ad larger cotetio widow. However, i this paper, we focus o the detectio of phycal badwidth istead

of available badwidth ad wish to mitigate the ifluece of the cross traffic. To achieve this goal, we assume that, i the detectio period, which is supposed to be relatively short, the etwork coditio does ot chage. Ad we ited to fid out the chage patter of t b. Our projectio is that t b may take the values with some fixed patter so that we ca use a costat to replace t b. We test our guess by measurig values of t b uder some predefied cross traffic i our mulatio test bed described i Sec. 4. More tha 8,000 packets of differet zes (ragig from 50 to 50 bytes) are set from a MN to the BS to acquire the samples of t b withi 0s. From the measured values show i Fig., we ca see that values of t b for differet packet zes mostly fall withi the same rage betwee 0.35ms ad ms evely. Based o this observatio, we thik that the patter of t b is relatively regular for differet packet ze. Thus, we ca take two mple approaches to elimiate the variatios of t b from (3.) by choog the appropriate values of RTTs. Fig. 3. Two approaches to elimiate the variatios of t b. 3.. Badwidth Detectio Algorithm After takig the Mi/Mea approaches, t b is regarded as a costat for differet probe packet zes ad thus we merge it ito the costat t c for mplicity. Let k=/c L, the (3.) ca be represeted as: Fig. Distributio of t b values uder predefied cross traffic coditio. () Selectig the miimum RTT value for each differet packet ze (Fig. 3(a)): Because for the same packet ze, t b value is the oly dyamic part of the RTT value, thus the miimum RTT values must also cotai the miimum value of t b. As values of t b for differet packet zes geerally fall ito the same rage, ad their miimum value may have the same value as each other, thus selectig the miimum RTT values will miimize the ifluece of t b ad make it a costat for differet probig packet zes. () Calculate the mea of RTT values for each differet packet ze (Fig. 3 (b)): As t b is the oly dyamic part of the RTT value for the same packet ze, calculatig the mea of RTT is equivalet to takig the mea of t b. Sice the t b values of differet packet zes are distributed evely i the same rage, the meas of RTT will cacel the variatios of t b. I this way, radom values of t b are tured ito a costat for differet probig packet zes. For mplicity, we ame the two approaches as the Mi/Mea i our algorithm. RTT = k s + t c (3.) I order to calculate the badwidth (C L =/k) accurately, RTT samples will be gathered for a umber of probig packet zes. Ad to quickly calculate the accurate value for C L, we apply liear regreso for the estimatio of badwidth based o the RTT ad their correspodig packet zes. Assume that RTTs for differet probe packet zes are obtaied, thus we ca set up the liear equatio system with probig packet zes as follows: RTT = k s RTT = k s (3.3) M RTT = k s We use matrix variables to deote the set of packet zes, RTTs, t c ad k show below: RTT s = RTT RTT s tc s = c = M M M (3.4) k RTT s 3

Thus (3.4) ca be shorteed as RTT = s c, ad the ormal equatio below is used to solve the solutios for the liear equatios: = s i tc (3.5) k Therefore, the coefficiet of (3.5) is resolved as: t c c = = k s i = (3.6) s + i ( ) I this way, the lik badwidth ca be calculated efficietly based o pairs of differet packet zes ad its correspodig RTTs. Based o the derivatios of (3.6), the WBD algorithm is give as follows: Algorithm: WBD at Mobile Node. REPEAT:. Geerate a probe packet radomly with a ze i the rage of 50 ~ 50 bytes, ad sed it to BS;. Collect the ACK from the BS ad calculate RTT;.3 Maipulate the RTT values with Mi/Mea; UNTIL the measuremet duratio is up;. Use (3.6) to calculate k; 3. Set C=/k ad output C as wireless lik badwidth; Ed. Fig. 4. Bac WBD Algorithm. The workig procedure of the algorithm ca be divided ito two phases. The first oe is the samplig phase. Probe packets of differet zes are set to the BS to attai the RTTs; meawhile, Mi/Mea approach is take to hadle the RTTs for the purpose of elimiatig the effect of t b. The secod phase calculates the badwidth of the wireless lik betwee the MN ad BS through liear regreso computatio. We would like to poit out that the probig packets are relatively small as they are set i the rage of [50, 50] bytes. This will ot itroduce too much overhead to the etwork. I this algorithm, oly oe badwidth value is calculated at the ed of the algorithm ad values detected may ot accurately reflect the actual badwidth o the wireless lik. To achieve a smooth value of the detectio, we exted our algorithm to calculate a badwidth every time a ew RTT sample is collected. Expoetial smoothig method is applied to combie the curretly calculated value of C L with the historical oe to attai a more accurate result. The exteded algorithm with expoetial smoothig approach is give i Fig. 5. Algorithm: WBD with Expoetial Smoothig (WBD-ES) Iitializatio: C = 0; = 0.; REPEAT:. Data Samplig ad Maipulatio. Geerate a probe packet radomly with a ze i 50 ~ 50 bytes, ad sed it to BS;. Collect ACK from BS ad calculate RTT;.3 Maipulate RTT with Mi/Mea;. Expoetial Smooth of Badwidth Value If the RTT value is modified i step.3 the Use (3.6) to calculate a ew lik badwidth value C t=/k; C = (-)*C +*C t; UNTIL the measuremet duratio is up. Retur C as the fial result of lik badwidth; Ed. Fig. 5. WBD Algorithm with expoetial smoothig for detected badwidth. From the algorithm descriptios i Fig. 4 ad Fig. 5, WBD algorithm ca be further clasfied ito four variatios: () V Mi without expoetial smoothig; () V Mi with expoetial smoothig; (3) V3 Mea without expoetial smoothig; (4) V4 Mea with expoetial smoothig. I Sectio 4, performace comparisos will be give for the four variatios of WBD algorithm ad we will see that V4 is the best amog the four variatios i terms of efficiecy, accuracy ad stability. 4. EXPERIMENTAL PERFORMANCE EVALUATION We have extevely evaluated the performace of WBD ug NS mulatio tool [9]. I this sectio, we first describe the experimet model ad performace metrics, ad the preset the evaluatio results. 4. Experimetal Model The mulatio is ru o a etwork model show i Fig. 6: Fig. 6. The Experimetal Network. 4

The etwork costs of two domais: wireless ad wired domais. I the wireless domai, there are 0 wireless mobile odes, MN -0, ad oe BS. I the wired domai, there are 0 wired fixed hosts, FH -0, ad oe gateway ode, G. The wired ad wireless domais are coected to each other through a wired lik betwee the BS ad G. Packets exchaged betwee two domais are trasmitted through this lik. Oe of the 0 mobile odes, MN, is chose as the probig ode to detect the lik badwidth. The cross traffic will be geerated amog other wireless odes ad wired odes. The cross traffic is geerated by a Poisso process with rate, which is a parameter uder our cotrol. The pair of source ad destiatio is chose radomly amog the wireless odes ad fixed hosts. Duratio of the cross traffic trasmiso is also a radom umber which obeys a expoetial distributio with average of 0 secods. 4. Performace Metrics For the performace of our detectio methods, there are three aspects to coder: efficiecy, accuracy ad stability. For efficiecy ad accuracy, the detectio method should achieve accurate detectio result withi a short period of time, facig the quick chage of wireless etwork status. Meawhile, the method is also required to have a stable performace uder various traffic coditios. I order to evaluate the performace of WBD algorithm, the followig performace metrics are used. () Relative Error (RE). Let the phycal badwidth be B ad the measured badwidth be B*, the the RE is defied as * B B RE = (4.) B I order to judge the accuracy ad efficiecy of the algorithm, the REs of detectio results attaied with differet time duratios will be evaluated. () Measuremet Stability (MS). It expresses the stability of the detectio method uder various measuremet coditios. Deote the measured values as B,..., B respectively ad B as the average value of the measured values. MS is defied as: MS = [ ( Bi B / B)]*00% (4.) i= 4.3 Performace of WBD Fig. 7 shows the measuremet process of the four variatios with the actual badwidth set to be 6 Mbps. From the figure we ca observe that, at the begiig of the process, the detected values of the four variatios vary greatly from the actual badwidth (with a RE more tha 50%). However, with more samples acquired, the detected values will quickly coverge to the actual badwidth. At the time of 5s, REs of all the four variatios have already coverged to the rage of 0%. More specific experimet results are illustrated below. Fig. 7. Measuremet Process of the Four Differet Variatios. A. Efficiecy: I order to evaluate the efficiecy, we have recorded the shortest time duratios that the four algorithms eed to take to meet the differet relative error requiremets. TABLE I TIME NEEDED TO MEET DIFFERENT RELATIVE ERROR REQUIREMENTS WITH DIFFERENT BANDWIDTHS 0% 0% V V V3 V4 V V V3 V4 6Mbps 3.5s 3.5s.89s 3.5s 0.7s 3.37s 6.s 6.7s 9Mbps 9.6s.0s 6.75s 6.98s 3.s 3.8s 8.83s 7.33s Mbps 0.8s 8.68s 7.75s 8.09s 4.8s 5.s 5.8s 6.7s 8Mbps 6.s 6.6s 6.8s 7.0s 34.8s 35.6s 7.s 3.7s Table I shows that with the badwidth set to be differet values (6Mbps ~ 8Mbps), the best of the four algorithms ca achieve a value with RE withi 0% i aroud 0s. Thus WBD performs fast eough to adapt with the dyamic wireless etwork. It is observed that the algorithms ug Mea (V3 ad V4) geerally take shorter time to approach to the actual badwidth as compared with Mi (V ad V). O the other had, algorithms with expoetial smoothig (V ad V4) take loger time tha without (V ad V3). B. Accuracy: To evaluate the accuracy, we also tested the four algorithm variatios with badwidth set to be differet values betwee 6Mbps ad 36Mbps. The REs of detectio results with duratio of 50s are show i Table II. TABLE II RELATIVE ERRORS OF DIFFERENT ALGORITHMS UNDER DIFFERENT NETWORK CONDITIONS 6Mbps 9Mbps Mbps 8Mbps 4Mbps 36Mbps V 8.63% 0.76%.5%.% 8.0% 36.4% V 6.96% 0.3% 0.56%.3% 6.% 34.% V3 3.98% 0.4% 6.48% 5.06% 5.46% 5.3% V4 3.99%.9% 7.45% 5.44%.85% 8.77% We have the followig observatios: whe the actual badwidth is relatively small (from 6Mbps to Mbps), all four algorithms obtaied results with REs aroud 5%. However, whe the actual badwidth is large (from 8Mbps to 5

36Mbps), V ad V may geerate REs aroud 0%, sometimes eve more tha 30%, but V3 ad V4 ca keep their REs aroud 0%. Coderig all the differet tuatios, it ca be see that V4 (Mea with expoetial smoothig) performs the best of all, with the RE less tha 0%. C. Stability: I order to evaluate the measuremet stability, we test our algorithms uder differet cross traffic. As we have oted that cross traffic is geerated amog the odes i a Poisso process with rate, we set the rate varyig from 0 to 50 to mulate differet cross traffic ad calculate the MS of the detectio algorithms. Table III shows the MS of the four algorithms for differet badwidth settigs. TABLE III MEASUREMENT STABILITY (MS) OF DIFFERENT ALGORITHMS UNDER DIFFERENT NETWORK CONDITIONS 6Mbps 9Mbps Mbps 8Mbps 4Mbps 36Mbps V 95.% 95.0% 96.7% 87.7% 87.8% 78.9% V 94.8% 95.0% 97.0% 88.4% 87.% 77.9% V3 96.6% 94.7% 93.3% 90.% 90.7% 74.6% V4 96.5% 95.0% 94.0% 90.3% 90.% 73.7% From Table III, we have the followig observatios. All these four algorithms preseted a stable performace (MS aroud 90%) i most cases. Oly whe the actual badwidth settig is large (e.g. 36Mbps), the stability is comparatively low (but still above 70%). Compared with other three algorithms, V4 is still the best. 5. CONCLUSIONS I this paper, we have preseted a efficiet active WBD algorithm for wireless lik badwidth detectio. By sedig probe packets with various zes from MN to BS, liear regreso is used to determie the lik badwidth based o the measured RTT delays. Two differet schemes Mi/Mea, have bee employed to attai stable RTT detectio results, so that our algorithm ca achieve a high accurate detectio result. Expoetial smoothig techique is also used to improve the accuracy ad stability of the origial algorithm. Ad we foud that WBD with Mea ad expoetial smoothig (V4) outperforms other variatios i terms of efficiecy, accuracy ad stability. Our work ca be exteded to the available badwidth detectio i wireless etwork. We will try to fid a efficiet method to detect available badwidth for differet wireless odes, which is useful for admiso cotrol ad QoS maagemet for wireless multimedia commuicatios. 6. ACKNOWLEDGEMENTS The work is partially supported by Research grat coucil (PGC) Hog Kog, SAR Chia uder grat Nos CityU 055/00E (9040687) ad CityU 039/0E (9040596). REFERENCES [] S. Verdú, Wireless Badwidth i the Makig, IEEE Commuicatios Magazie, Volume 38, Issue 7, July 000, pp 53-58. [] R. Prasad ad C. Dovrolis, M. Murray ad Kc claffy, Badwidth Estimatio: Metrics, Measuremet Techiques, ad Tools, IEEE Network, Volume 7, Issue 6, November/December 003, pp 7-35. [3] S. Wag, R. Nathuji, R. Bettati ad W. Zhao, Providig Statistical Delay Guaratees i Wireless Networks, Proc. of 4 th Iteratioal Coferece o Distributed Computig Systems, March 004, pp48-55. [4] L. Cheg, J. Zhag, ad I. Marc, Models for o-itruve estimatio of wireless lik badwidth, Spriger-Verlag Lecture Notes i Computer Sciece (LNCS) for Persoal Wireless Commuicatio 003, Vol. 775, pp. 334-348, 003. [5] P. Ferguso, G. Husto, Quality of Service Deliverig QoS o the Iteret ad i Corporate Networks, Joh Wiley ad Sos, 998, pp5-53. [6] V. Jacobso, pathchar-a tool to Ifer Characteristics of Iteret paths, 997. Preseted at the Mathematical Scieces Research Istitute. Available from http://ftp.ee.lbl.gov/pathchar. [7] K. Lai ad M. Baker, Nettimer: A tool for measurig bottleeck lik badwidth, Proceedig of the USENIX Sympoum o Iteret Techologies ad Systems, March 00, pp-33. [8] K. Lai ad M. Baker, Measurig Lik Badwidth ug a Determiistic Model of Packet Dealy, Proc. ACMSIGCOM, Sept 000, pp 83-94. [9] NSv. Network Simulator. http://www.i.edu/sam/s/ [0] G.A.F, Seber, Liear regreso aalys, New York: Wiley, 977, pp4-5. [] A. Bakre & B.R. Badriath, I-TCP: Idirect TCP for Mobile Hosts, Proc. ICDCS, May, 995, pp.36 ~ 43. [] S. H. Shah, K. Che, K. Nahrstedt, Available Badwidth Estimatio i IEEE 80.-based Wireless Networks, i Proc. of the st ISMA/CAIDA Workshop o Badwidth Estimatio (BEst 003), Sa Diego, CA, December 003. 6