Networks

Quality-Differentiated Video Multicastin Multirate Wireless Networks

Kate Ching-Ju Lin, Member, IEEE, Wei-Liang Shen,

Chih-Cheng Hsu, and Cheng-Fu Chou, Member, IEEE

Abstract—Adaptation of modulation and transmission bit-rates for video multicast in a multirate wireless network is a challenging

problem because of network dynamics, variable video bit-rates, and heterogeneous clients who may expect differentiated video

qualities. Prior work on the leader-based schemes selects the transmission bit-rate that provides reliable transmission for the node that

experiences the worst channel condition. However, this may penalize other nodes that can achieve a higher throughput by receiving at

a higher rate. In this work, we investigate a rate-adaptive video multicast scheme that can provide heterogeneous clients differentiated

visual qualities matching their channel conditions. We first propose a rate scheduling model that selects the optimal transmission bit-

rate for each video frame to maximize the total visual quality for a multicast group subject to the minimum-visual-quality-guaranteed

constraint. We then present a practical and easy-to-implement protocol, called QDM, which constructs a cluster-based structure to

characterize node heterogeneity and adapts the transmission bit-rate to network dynamics based on video quality perceived by the

representative cluster heads. Since QDM selects the rate by a sample-based technique, it is suitable for real-time streaming even

without any preprocess. We show that QDM can adapt to network dynamics and variable video-bit rates efficiently, and produce a gain

of 2-5 dB in terms of the average video quality as compared to the leader-based approach.

Index Terms—Wireless video multicast, rate adaptation, QoS

Ç

1 INTRODUCTION

VIDEO streaming is arguably one of the most popularmultimedia applications over wireless networks today.

With the ubiquitous of such applications, constant demandfor better wireless access technology has resulted in severalgenerations of new access point (AP) products, e.g., 802.11 a/b/g/n, in a relatively short time. Future-generation APs areexpected to have much greater computation capability andstorage capacity. They offer new opportunities to incorpo-rate even more advanced features to support a variety ofapplications (e.g., [1]). In this paper, we consider a new APfeature, namely quality-differentiated video multicast, toallow better utilization of limited wireless resources forvideo streaming.

Taking advantage of the wireless broadcast nature, avideo source can multicast a video object to a group ofmulticast members in order to reduce the bandwidthrequirement, as compared with unicasting the data to eachindividual member. However, current commercial networkdevices typically transmit multicast packets at the base ratein the MAC layer, even though 802.11 standard supportsmultiple bit-rates up to 11 Mb/s for 802.11b or 54 Mb/sfor 802.11 a/g, each of which has a different modulation

scheme. This is a waste of wireless bandwidth if certainmembers in the multicast group are capable of receivingpackets at a higher bit rate, and desire a better visualquality. To address this problem, we propose a dynamicrate adaptation scheme with quality-differentiated featuresto better support heterogeneity in the clients.

The auto rate selection algorithms for unicast have beenexamined in [2], [3], [4], [5], [6], [7], [8]. Other work [9],[10], [11] studies the rate adaptation schemes for unicastvideo streaming. They measure the loss probability accord-ing to feedback acknowledged from the receiver, andpredict a rate that can achieve the highest throughput. Suchfeedback-based schemes cannot be extended to multicastscenarios because concurrent feedback from several multi-cast members can lead to severe collision. To avoid thiseffect, some multicast rate adaptation schemes [12], [13],[14], [15], [16] select the member who experiences the worstchannel condition as the leader of the multicast group, andpredict a bit rate that can reach this leader (the worstnode). Such a leader-based approach is particularlysuitable for applications that need to deliver data to allmembers reliably, e.g., data dissemination. However, thisapproach may not be efficient for video multicast because itmerely selects the rate that maximizes the throughput ofthe worst node. In doing so, it penalizes those nodes whocan receive data at a higher bit rate.

The rate selection problem1 in video multicast scenariosis more challenging due to heterogeneity of receivers. Sincevarious members may observe different channel conditions,they can receive data sent at different bit-rates as illustrated

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013 21

. K.C.-J. Lin and W.-L. Shen are with the Research Center for IT Innovation,Academia Sinica, No. 128, Academia Road, Sec. 2, Nankang, Taipei 115,Taiwan. E-mail: {katelin, wlshen}@citi.sinica.edu.tw.

. C.-C. Hsu and C.-F. Chou are with the Department of Computer Scienceand Information Engineering, National Taiwan University, No. 1Roosevelt Rd., Sec. 4, Taipei 106, Taiwan.E-mail: [email protected], [email protected].

Manuscript received 6 Nov. 2010; revised 4 Oct. 2011; accepted 12 Oct. 2011;published online 11 Nov. 2011.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-2010-11-0508.Digital Object Identifier no. 10.1109/TMC.2011.242.

1. We let “video bit-rate” denote the streaming rate encoded in theapplication layer, and represent the transmission bit-rate, i.e., determinedby the modulation and channel coding rate, used to send MAC-layerpackets as “MAC-layer bit-rate” or “bit-rate.”

1536-1233/13/$31.00 � 2013 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

in Fig. 1. To leverage this performance factor, an ideal wayis to select a different transmission bit-rate for each videoframe according to its importance (defined as the ratescheduling problem), so that members can receive differ-entiated video quality that best takes advantage of theirchannel conditions. An intuitive solution of the ratescheduling problem is to select the rates that can achievethe maximal peak signal-to-noise ratio (PSNR) value (a metricstandardized by ITU [17] and used to evaluate videoquality [18]). However, under some scenarios (e.g., Fig. 1),selecting a higher rate to maximize the PSNR may result inthe dilemma that the nodes who have worse channelconditions (e.g., node D) cannot receive any packetcorrectly due to a high bit error ratio (BER) of the MAC-layer bit-rate and, thus, cannot even obtain the minimalvisual quality.

In this paper, we propose a visual-quality-based rateadaptation protocol, which makes tradeoff between twogoals: providing differentiated video quality for variousclients to match their heterogeneous channel conditions andguaranteeing minimum visual quality for each client.Specifically, our goal is to develop a software-based ratescheduling protocol in order to produce the maximal totalvisual quality for quality-differentiated video multicastunder the constraint of ensuring at least minimum visualquality for each member. The contributions of this paperare as follows:

. We model the rate scheduling problem as a variationof the Knapsack problem, and propose a dynamicprogramming solution to solve it optimally.

. We propose a practical protocol, called Quality-Differentiated Multicast (QDM), which exploits asample-based technique to adapt the transmissionbit-rate of each video frame to variable video bit-rates and client mobility without the need of anypreprocess. Thus, it can be applied to real-time videostreaming.

Our simulation results show that not only can QDMprovide users differentiated video quality matching theirchannel conditions, but also produce a better average visualquality as compared with the leader-based schemes.

The remainder of this paper is organized as follows: Wediscuss related work on rate selection for wireless multicastin Section 2. We formulate the rate scheduling problem forquality-differentiated video multicast in Section 3, andintroduce the QDM framework in Section 4. In Sections 5and 6, we present simulation and experimental results,respectively, to compare the performance of QDM with thatof the leader-based approach. Finally, we give our conclu-sions in Section 7.

2 RELATED WORK

Due to the heterogeneity among multicast members,different multicast recipients may observe dissimilar linkqualities. Recently, several works [19], [20] have focused onhow to allocate bandwidth efficiently for broadcastingvideo streams to clients either with heterogeneous re-sources, e.g., screen resolution or decoding capability, orwith multiple access technologies, e.g., 3G and WLAN.Compared to those works that do no take channelcondition, packet losses and transmission bit-rates intoaccount, this work allocates bandwidth for video multi-casting with consideration of multiple available transmis-sion bit-rates and the corresponding loss probability. Ourgoal is to provide clients heterogeneous visual qualitymatching their channel conditions.

Most work on rate selection for wireless multicastfocuses on achieving multicast reliability by selecting therate that can deliver data reliably to the member with theworst channel condition. In [12], the Leader-Based Protocol(LBP) is the first leader-based approach proposed toovercome the problem of feedback collision. It selects theworst node as the leader to acknowledge multicast packets.Other members can issue negative acknowledgements tocollide the acknowledgement sent by the leader and, thus,trigger the sender to retransmit the lost packets. The goal ofLBP is to support reliability by a single feedback. However,it does not adapt the transmission bit rate to dynamicchannel conditions, but only sends data at the base rate.Thus, the rate adaptation algorithms, such as RAM [13] andARSM [15], are proposed for the leader-based multicastprotocol. They estimate link quality of the leader anddetermine a proper rate that can better reach the leader.Both of these techniques let each receiver embed theinformation about it receiving SNR value in the CTS frame.The sender can infer the leader’s SNR upon receivingthe CTS frames, and predict a suitable rate accordingly.

Some other work [21], [22] extends the leader-basedmulticast protocol to an environment with multiple APsthat operate on nonoverlapping channels, and investigatesthe optimal client-AP association control problem in such amulti-AP scenario. A central coordinator is deployed toconnect all APs, and assigns each client to associate with anAP such that the channel airtime occupied by APs can beminimized. The association control scheme still requireseach AP to transmit data at the transmission rate that canreach the worst client associated with it, but attempts toassociate each client with an appropriate AP to minimizethe client heterogeneity. Our work differs from the associa-tion control scheme along two axis. First, it focuses on thevideo streaming application, and provides different clientsheterogeneous visual quality by assigning each video framea different transmission bit-rates based on its importance.Second, our work considers a single-AP scenario where APcannot cooperate with each other; however, it can beintegrated with the above association control schemes toenable each AP to provide its clients differential visualquality after the optimal client-AP association.

SARM [14] and ARSM [15] further adapt the leader-based approach to rate selection for video multicast. Basedon experiments, SARM constructs a SRN-PSNR table that

22 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013

Fig. 1. Heterogeneity of Multicast Members. Users A and B can receivepackets sent at 11 mb/s in the MAC layer, while user D can only receivepackets at the base-rate.

maps a SNR to a visual quality for each video content.Hence, it can predict the rate that provides everyone aminimum visual quality. ARSM proposes a variant, calledH-ARSM, to select another best node who has the bestchannel condition. It then selects a bit-rate for the base layerof a stream according to the channel quality of the worstnode, while transmits the enhancement layer at another bit-rate based on the channel quality of the best node.However, since such a two-level scheme does not considerall multicast clients, it cannot maximize the overall videoquality for the entire group by differentiating the visualqualities among the individual users based on their channelconditions.

The work closest to ours is presented in [23], whichformulates the resource allocation problem as an optimiza-tion model for video multicast in WiMax. Unlike [23] whichutilizes multiple subcarriers in a WiMax to broadcast avideo stream, we consider a single-channel wireless net-work such as the infrastructure mode in 802.11, wheredifferent frames in a group of pictures (GOP) need to contendfor the channel time before their playback time out andmust occupy various transmission times based on theirimportance and the designated transmission bit-rate.Furthermore, the work [23] formulates the theoretical modelthat assumes all information, such as the channel conditionand the best bit-rate of each user, is given and, in addition,considers a constant scenario, e.g., constant bit-rate videoand stationary topologies. Our goal, however, is to developan easy-to-implement protocol that considers the practicalissues, including how to measure the channel conditions,how to detect channel variation, how to adapt to clientmobility, how to adjust the transmission schedule accordingto variable video bit-rate, and how to tradeoff between theonline measurement overhead and system performance.

Another originality of our work is as follows: All theabove schemes predict the bit rate based on SNR of a link.However, earlier work [24], [25] has shown that not onlymay bit error and link quality depend on SNR, but are alsorelated to environmental factors such as multipath andinterference. To avoid this uncertainty, we exploit a sample-based approach to measure the bit-rate that produces thebest visual quality for the heterogeneous clients, instead ofpredicting a rate based on the theoretical SNR-BERmapping function.

Finally, some protocols, such as APEX [26] and Softcast[27], exploit physical-layer information to enable differen-tial-quality video multicasting. Such a cross-layer designrequires the modification of hardware, and cannot bedeployed in existing WiFi APs. Our work, however, focuseson developing a software-based solution that can berealized in current WiFi devices.

3 RATE SCHEDULING MODEL

In this section, we formulate the rate scheduling problem asa theoretical optimization model, and utilize a dynamicprogramming algorithm to compute the solution usingoracle information. This model needs a high computationalcomplexity and complete information about the packet lossprobability of each wireless link. Our motivation is to usethe model as a reference to assess the performance of our

new technique, called QDM, introduced in the next section.We will discuss their performance comparison in Section 5.

3.1 Problem Formulation

We consider an environment where the video serversforward the stream to AP that can help broadcast data to all

multicast members. We assume that AP and the videoserver are interconnected by wired networks, which are not

the bottleneck; hence, we only focus on transmissionbetween AP and multicast members. The rate scheduler

can be installed either in the video server that collocateswith the AP, as shown in Fig. 2, or in multiple proxy servers

that share the workload of scheduling. It can alternativelybe implemented as a driver run in the AP. The rate

scheduler, which is run in user space, can then notify thenetwork driver of AP, which is usually run in kernel, to

send each packet at the selected transmission bit-rate. Suchinformation exchange between the user space and the

kernel can be realized by socket programming.We consider a group of multicast members M over a

wireless network supporting a set of transmission bit-rates

R ¼ fr1; r2; . . .g, where ri is a lower rate than rj if the indexi<j. We let prðmÞ denote the packet loss ratio of the link

from the AP to member m 2M, as the AP sends data at rater 2 R. We will describe how to estimate prðmÞ in Section 4.1.

Suppose the frames of a video stream are partitionedtemporally into multiple groups of pictures, each of which

contains a set of frames. Let G be this set of GOPs. GOPn 2G denotes the nth GOP that includes K frames, i.e., GOPn ¼ffn1 ; fn2 ; . . . ; fnKg. Based on the loss probability, the schedulercan determine the rate schedule ~frn¼<frn1 ; frn2 ; . . . ; frnK>,

where frnk is the bit rate assigned to transmit frame fnk , foreach GOPn such that the frames can be transmitted to the

receivers before the playback deadline TGOPn .The goal of this paper is to find a rate schedule ~fr for a

given video stream such that the maximal video quality can

be achieved for the multicast group, subject to theconstraint that each member must get at least the minimum

video quality PSNRmin. Even though we formulate the ratescheduling problem for a single video stream, the rate

scheduler can still solve the rate schedule for each ofmultiple video streams separately by considering the other

streams as background traffic. We model the rate schedul-ing problem as follows:

maxE½PSNRtotal� ¼Xm2M

E½PSNRðmÞ�; ð1aÞ

subject to

LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 23

Fig. 2. Framework architecture.

E½PSNRðmÞ� � PSNRmin; 8m 2M; ð1bÞ

Xfnk2GOPn

len�fnk Þ!nkfrnk

� TGOPnð1� �bÞ; 8GOPn 2 G: ð1cÞ

The constraint in (1b) ensures that each member m canobtain a higher expected visual quality E½PSNRðmÞ� thanthe minimum video quality PSNRmin. The PSNR of a m�nvideo frame [17], [18] is computed as follows:

MSE ¼ 1

mn

Xmi¼1

Xnj¼1

½Iði; jÞ �Kði; jÞ�2;

PSNR ¼ 10 log10

P 2max

MSE

� �;

where IðÞ is the original frame, KðÞ is the distorted frame,and P 2

max is the maximum pixel value of the image. For eachGOPn 2 G, frame fnk 2 GOPn is only useful if it is deliveredto the receivers successfully before the playback deadlineTGOPn . Besides, we let �b denote the proportion of time thatthe channel is occupied by the background traffic; therefore,the residual time available for transmitting GOPn isTGOPnð1� �bÞ. Since sending a frame that arrives after thedeadline does not help improve the video quality, theconstraint in (1c) determines whether each frame fnk 2GOPn with a packet size lenðfnk Þ is tardy and should bedropped; otherwise, this constraint selects a suitabletransmission rate for this frame. In (1c), !nk is a binaryvariable that indicates whether fnk is dropped or not. Thatis, !nk ¼ 1 if the AP can transmit fnk at rate frnk ; otherwise,!nk ¼ 0 and fnk must be dropped. This constraint ensuresthat the AP does not transmit frames that cannot bedelivered to the group members on time. Given a feasiblesolution of the rate schedule ~fr and the loss probabilityprðmÞ for all m 2M, the scheduler can compute theexpected PSNR E½PSNRðmÞ� and estimate the overallvideo quality E½PSNRtotal� by

Pm2M E½PSNRðmÞ�. Finally,

the goal is to find the optimal rate schedule ~fr� that canmaximize E½PSNRtotal� as shown in (1a).

3.2 Dynamic-Programming-Based Solution

Note that the model in (1) is a variation of the 0-1 knapsack

problem. Given a set of items (frames) in each GOP, the

scheduler must determine the number of frames to transmit

so that the total weight (transmission time) does not exceed

Tdiff¼ TGOPnð1� �bÞ � Tmin, where Tmin is the time used to

send a subset of frames in each GOP at the rate of the worst

node in order to satisfy (1b); and, hence, Tdiff is the time

allowed to send the remaining frames in the GOP at a

higher rate to provide heterogeneous clients differentiated

video quality. Nevertheless, the difference between the rate

scheduling problem and the knapsack problem is that each

item (frame fnk ) has jRj different weights (transmission

timelenðfn

kÞ

ri; 8ri 2 R) and values (total incremental quality

�Qðfnk ; riÞ) because the number of clients who can receive

fnk is determined by the rate ri used for the packet

transmission. Specifically, since each client m 2M can

receive a frame sent at rate ri with the probability priðmÞ,the expected value of �Qðfnk ; riÞ can be approximated by

Pm2M �qðfnk ÞpriðmÞ, where �qðfnk Þ denotes the incremental

quality if a client successfully receives fnk . �qðfnk Þ can be

estimated by the rate-distortion function studied in [28],

[29]. In other words, sending a frame at a higher rate

spends less transmission time. However, this results in a

lower incremental quality �Q for a multicast group

because only a few clients who have good enough channel

conditions can receive that frame.Thus, the rate scheduling model must not only deter-

mine whether a frame should be transmitted, but also selecta suitable rate for that frame. Consequently, the ratescheduling problem is at least as hard as the knapsackproblem, which is NP-complete, because the knapsackproblem is a special case of the rate scheduling problem asjRj ¼ 1. In this work, we propose a novel dynamicprogramming solution that extends the conventionalknapsack algorithms to solve the rate scheduling problemin pseudopolynomial time. Our solution considers thefollowing two types of coding techniques:

1. Predictive coding: Predictive coding (e.g., MPEG4) isa coding scheme that predicts a frame by referencinga previously coded frame. Thus, if a reference frameis not received correctly, those frames that referenceit cannot be decoded either. In general, the playbackorder and the decoding order of the video framesare not the same. To facilitate decoding, thetransmission order of the frames is the same astheir decoding order. We can sort the frames inGOPn in the decoding order such that frame fnk canhelp improve the video quality by �Qðfnk ; rÞ only ifboth of fnk and fnl are reconstructed correctly for alll < k. Hence, we let the video server send the firstk� frames of GOPn when Tdiff is only available fortransmitting k� frames. Because the number offrames k� that can be sent before TGOPn dependson the transmission time required for each frame,which is determined by the transmission bit-rateused to send that frame, we must select the optimalk� and rates for each frame so as to maximize thetotal video quality.

2. Multiple description coding (MDC): MDC encodes amedia stream into multiple independent substreams(i.e., descriptions), any part of which can be decodedindependently. Receivers can get minimum qualityby decoding one arbitrary description, and achieveincremental improvement by receiving additionaldescriptions. For simplicity, we deem each descrip-tion as a frame. That is, description fnk 2 GOPn canhelp improve the video quality by �Qðfnk ; rÞ if it isdelivered to the client correctly. Hence, we candetermine whether each frame should be trans-mitted independently.

Algorithm 1 shows the proposed algorithm for two

coding schemes. To apply the dynamic programming

solution, we represent the transmission time of each frame

fnk 2 GOPn as a discrete variable, which is approximated

by d lenðfnkÞ

r e if it is sent at rate r 2 R. Define Q½k; t� as the

maximum visual quality that can be obtained with weight

(time slots) less than or equal to t�bTdiffc by sending up to


k frames. When using MDC, the solution is similar to the

dynamic programming method used to solve the knapsack

problem except that, as computing the value of Q½k; t�, it

must consider each frame fnk by comparing jRj þ 1 alter-

natives, i.e., dropping it or sending it at jRj various rates,

instead of only 0-1 choices in the knapsack problem. It can

finally get that the maximal PSNRtotal of the rate scheduling

problem equals Q½K; bTdiffc�.

Algorithm 1. Dynamic programming solution to the ratescheduling problem

1: Initialize Q½k; t� ¼ 0 for 0 � k � K; 0 � t � bTdiffc2: for t ¼ 1 to bTdiffc (Test each size of knapsack) do

3: for k ¼ 1 to K (Go through each frame fnk ) do

4: if Use MDC then

5: Q½k; t� ¼ maxfQ½k� 1; t�;�Q�fnk ; ri

�þ Q½k� 1; t� d lenðf

nkÞ

rie� : ri 2 R; t � d

lenðfnkÞ

rieg

6: else if Use predictive coding then

7: r ¼ argri maxf�Q�fnk ; ri

�þQ½k� 1;

t� d len�fnk

�rie�:ri 2 R; t �

d lenðfnkÞ

rie; Q½k� 1; t� d lenðf

nkÞ

rie�6¼ � 1g

8: if r 6¼ � then Q½k; t� ¼ �Q�fnk ; r

�þ

Q½k� 1; t� d lenðfnkÞ

r e�; else Q½k; t� ¼ �1; endif

9: end if

10: end for

11: end for

12: if Use MDC thenPSNRtotal ¼ Q½K; bTdiffc�;else PSNRtotal ¼ max0�k�KQ½k; bTdiffc�; endif

By contrast, in predictive coding, since frames in a GOP

are dependent on their reference frames, we must include

fnk only if its reference frames fnl (l < k) are also transmitted.

To enforce Algorithm 1 to select the first k� frames, as

computing Q½k; t�, we do not consider the alternative of

dropping the current frame fnk (i.e., Q½k� 1; t�, which

indicates that it only includes the first k� 1 frames, while

discards fnk ). This is because we can always benefit from

sending it if there is enough time to transmit it. In other

words, if there is still enough time to transmit some frames,

dropping the current frame (fnk ) and sending the following

ones (fnk0 ; k0 > k) does not help improve the video quality

due to the lack of the reference frame fnk . Therefore, we set

Q½k; t� to �1 if the set of finding the maximum shown in

line 7 of Algorithm 1 is empty, which means that there is not

enough time to send fnk no matter which rate r 2 R is used.

Otherwise, as computing Q½k; t�, we only consider the

rates ri 2 R whose Q½k� 1; t� b lenðfnkÞ

ric� is not equal to �1,

and select the one that can produce the maximal incre-

mental quality. Because Q½k� 1; t� b lenðfnkÞ

ric� ¼ �1 means

that fnk�1 is already excluded by Q½k� 1; t� b lenðfnkÞ

ric�, in

this case, we do not need to include fnk as well. Finally,

unlike in the MDC scheme, since there is no guarantee that

Q½k; t� is always larger than or equal to Q½k� 1; t� in

predictive coding, PSNRtotal is set to the maximum of

fQ½k; bTdiffc� : 0� k� Kg. That is, after solving each element

Q½k; t�, Algorithm 1 must finally find the number of frames

required to be transmitted so that PSNRtotal can be

maximized.The running time of Algorithm 1 equals OðTdiffKjRjÞ.

The algorithm is polynomial only if Tdiff is polynomial.Therefore, the proposed dynamic programming basedscheme can be solved in pseudopolynomial time. Since theavailable transmission time, Tdiff , is limited, the algorithmcan be run with a low complexity.

4 QDM FRAMEWORK

The model in the last section is proposed to find the mostefficient solution, which needs complete information aboutthe loss probability prðmÞ, and incremental quality �qðÞ ofeach frame. However, to realize a practical protocol, wemust consider the following problems.

. How can the scheduler collect the visual qualityperceived by each receiver efficiently?

. How can the scheduler determine a suitable rateschedule for real-time streaming if the informationabout incremental quality �qðÞ is not computed bypreprocessing?

. How can the scheduler adapt the rate schedule tonetwork dynamics, such as channel variation, nodemobility, and variable video bit rates?

To cope with the above practical issues, we proposeQDM, a practical video multicast framework includingthree components: 1) cluster construction: it clusters clientsaccording to their channel conditions in order to character-ize the heterogeneity of clients; 2) sample-based ratescheduling: it predicts the rate schedule by real-timesampling, and, thus, can estimate visual quality even ifinformation about �qðÞ is not given; 3) two-stage rateadaptation: using the finite-state machine, it adapts the rateschedule to variable video bit rates and channel conditions,and, at the same time, avoids the unnecessary samplingoverhead.

4.1 Cluster Construction

A key challenge to characterizing heterogeneous channelconditions of multicast members is that simultaneousfeedback from all members causes collision. Even thoughtraditional delayed-based methods, which schedule time-division-based feedback, can avoid collision, it is still timeconsuming for collecting responses from all members.Thus, the design of QDM is to separate members tomultiple clusters C such that the receivers with similarchannel conditions can be classified into the same cluster.Specifically, we let rðmÞ denote the best rate of memberm 2M if it can receive the maximal throughput from AP atrðmÞ. Those who have the same best rate ri 2 R can begrouped into cluster Ci ¼ fm : m 2M; rðmÞ ¼ rig. Hence, amulticast group can be divided to jRj clusters, each ofwhich can select the one in Ci who experiences the worstchannel condition as the cluster head (CH) of Ci (denotedby CHi) to represent all members in Ci. The scheduler cancollect heterogeneous channel conditions from the CHs,and reduce the overhead of information exchange sig-nificantly. Based on the performance reported by the


representative nodes (i.e., CHs), the scheduler can approx-imate the overall video quality of a multicast group by

~PSNRtotal ¼XCi2C

PSNRðCHiÞ � jCij;

where jCij is the number of members in cluster Ci.In order to collect information about jCij and rðmÞ for

each m 2M, AP can send a probe packet at each rater 2 R before broadcasting the video stream. Based onthose probe packets, each member m can estimate the biterror rate of each rate r 2 R, denoted by BERrðmÞ. An100-byte probe packet for each rate is long enough tocapture the average bit error rate accurately. Each clientcan then estimate the throughput of each rate based on [5],which is computed by r � ð1�BERrðmÞÞN , where ð1�BERrðmÞÞN is the probability that the N-bits probe packetis received correctly without any bit errors. Each client canthen select its best rate rðmÞ that can produce the highestthroughput, i.e., rðmÞ ¼ arg maxr2Rr � ð1�BERrðmÞÞN ,and responds a control message <m; rðmÞ; SNRm> tothe rate scheduler, where SNRm is the average SNRobserved by m. The scheduler can then use the feedbackof control messages to construct the clusters, select theCHs, and broadcast the information <ri; CHi; SNRCHi

> toeach cluster Ci 2 C. Probe packets are only sent beforecluster construction, so the overhead is relatively smallerthan periodical channel estimation.

However, we note that the channel condition of eachclient might change dynamically due to channel variationor client mobility. To adapt the rate schedule to suchchannel dynamics, the scheduler needs to update theclusters as the channel condition of certain nodes changessubstantially. For a multicast application, it is howeverdifficult for the AP to monitor the channel condition ofevery member all the time. Therefore, we let each memberm 2 Ci measure the SNR value of its received data packets(i.e., video frames), and issue a reclustering request to thescheduler if it detects that its average SNR is smaller than� � SNRCHi

, where 0 � � � 1 is a sensitivity coefficient. Asa result, detecting channel variation can be done by eachclient without any extra overhead. Once the schedulerdecides to perform reclustering, it executes the sameprobing procedure as what it did in the initialization stagesuch that the clients who experience channel variation canupdate its channel condition to the scheduler. During theprobing period, the scheduler pauses the transmission ofvideo stream, and resumes the streaming after the newtransmission schedule is recomputed. We note that, sincewe only need to perform clustering in the start-up stage orwhen the network changes significantly, the overhead ofclustering should be negligible for the static environmentsand reasonable for the dynamic environments.2

4.2 PSNR-Guaranteed Quality-Differentiated RateSelection

We note that even though the solution proposed in Section 3can provide an efficient rate schedule, it may only be suitablewhen the information about �qðÞ is computed by preproces-sing. Specifically, if the information about �qðÞ is given, wecan vary Algorithm 1 to solve the rate schedule andmaximize ~PSNRtotal for the cluster-based framework. How-ever, if the information about �qðÞ is not given, it ischallenging to online analyze the incremental quality �qðÞand compute E½PSNRtotal� for various rate schedules ~fr assolving the dynamic programming algorithm. Therefore,instead of online computing incremental quality �qðÞ costly,we propose QDM, a heuristic two-level rate schedulingscheme that can adapt the rates dynamically based on asampling technique. The goal of QDM is to find a suitablerate for each frame efficiently for real-time video streaming.

To reduce the complexity of rate scheduling, we sort theframes of each GOP according to their decoding order, andpartition each GOP into three sets: essential frames (Fe),supplementary frames (Fs), and discarded frames (Fd), asshown in Fig. 3, where variables ne, ns, and nd are thenumber of frames in the sets Fe, Fs, and Fd, respectively. Themain idea is that, first, we tend to transmit the frames in Feat rate Re to all m 2M, so that each member can achieve atleast the minimum visual quality PSNGmin. We note that Re

must be set to a fixed rate minm2MrðmÞ because the essentialframes should be received by every member m 2M.Second, the frames in Fs can be sent at a higher rate (sayrs) in order to provide differentiated quality for hetero-geneous clients. Finally, the rest of frames are put in Fd andmust be discarded if the available bandwidth is not enoughto send every frame in a GOP. Therefore, we can relax theoriginal problem shown in (1) as the following one:

max ~PSNRtotal ¼ maxXCi2C

PSNRðCHiÞ � jCij; ð2aÞ

subject to

PSNRðmÞ � PSNRmin; 8m 2M; ð2bÞ

Xf2Fe

lenðfÞRe

þXf2Fs

lenðfÞrs

� TGOPnð1� �bÞ; 8GOPn 2 G: ð2cÞ

In the relaxed problem, the goal is to find the requirednumber of essential frames (ne) and the transmission rate ofsupplementary frames (rs) such that the approximate overallvisual quality ~PSNRtotal can be maximized. Equation (2c)bounds the time used to transmit the frames in each GOP upto the available transmission time TGOPnð1� �bÞ. Once thescheduler determines ne to satisfy PSNRmin, as shown in(2b), it computes the time used to send the essential frames,which equals


2. For 802.11b, the overhead of sending 100-byte probe packets equalsð1=1þ 1=2þ 1=5:5þ 1=11Þ � 800 � 10�6 ¼ 1:42 ms. Even if probing is per-formed periodically per GOP, the overhead is about 2.8 percent if each GOPis 0.5 second long. The response from multicast clients might need a higheroverhead. However, we note that only the first clustering in the start-upstage requires every client to report its channel condition; for reclustering,only the clients that experience channel variation need to report theirupdated channel conditions to the scheduler. Thus, the response overheadis relatively small for reclustering.

Fig. 3. Three blocks in each GOP.

Tmin ¼P

f2Fe lenðfÞRe

;

where lenðfÞ is the number of bits in frame f . The AP canthen send as many supplementary frames as possible at raters within the residual available time Tdiff ¼ TGOPnð1 ��bÞ � Tmin. Specifically, since each supplementary frame f 2Fs takes lenðfÞ=rs (seconds) for transmission, the AP cansend at most ns (i.e., jFsj) frames such that

Pf2Fs lenðfÞ

rs� Tdiff :

Finally, we discard the rest of frames (i.e., Fd) becausethere is no time to transmit any data before the playbackdeadline TGOPn .

Instead of solving the optimal solution of ne and rsdirectly, we propose a sample-based scheme that enablesthe scheduler to explore the suitable value of ne and rsbased on feedback from the cluster heads. As Fig. 4 shows,we divide a video stream into multiple sampling intervals,each of which contains more than three GOPs. For eachsampling interval, the scheduler sends Fs of some GOPs at abit-rate other than the current rs to gather information aboutother rates. Specifically, we let the first three GOPs of asampling interval be the samples that help the schedulerfigure out a suitable rate rs. The scheduler sets thetransmission bit-rate of Fs in GOP1, GOP2, and GOP3 atthe current rate rs, the next higher rate r0s ¼ rindexðrsÞþ1, andthe next lower rate r00s ¼ rindexðrsÞ�1, respectively, whereindexðrsÞ is the index of the current rate rs, e.g., indexðrsÞ ¼1 if rs ¼ r1. We only send three sample GOPs, i.e., samplingthe next higher rate and next lower rate, such that thesystem can conservatively adapt the transmission bit-rateand gradually converge to the optimal solution.

Instead of requesting each cluster head to compute theaverage visual quality of all clients in the same cluster, we letit respond its own reception quality to the scheduler. Therational behind this design is that each client in the samecluster has a similar channel condition and thereby couldreceive a comparable visual quality. Therefore, we can usethe reception quality of the cluster head to estimate that ofother clients in the same cluster. To be more conservativeabout the estimation reported by the cluster heads, we selectthe client that has the worst channel condition in a cluster toact as the representative cluster head who would likely tohave the worst visual quality in the corresponding cluster.

To this end, upon receiving the sample GOPs, each CHresponds the scheduler a mask, which is an array of binaryvariables where the ith bit equals 1 if the CH receives the ithpacket of a GOP correctly, and 0 otherwise. We use each bitin the mask to acknowledge the reception of each packet,instead of each frame, because a frame may be segmentedinto several packets. The scheduler can compute ~PSNRtotal

of each sample GOP by decoding the stream with the packet

losses pattern logged in the masks reported by CHs andcomparing the decoded GOP with the original stream. Webelieve that the scheduler has strong enough computationcapability for real-time decoding few sample GOPs re-ported by jRj representative CHs. By comparing ~PSNRtotal

of three sample GOPs, the scheduler can set rs to the rateused by the sample GOP that achieves the highest ~PSNRtotal

(called GOPbest-sample for short). For example, if the channelrate of GOPbest-sample is r0s, the scheduler updates rs ¼ r0s. Inaddition, if any of CHs cannot achieve PSNRmin inGOPbest-sample, it means that the current number of essentialframes (ne) is not enough to provide everyone PSNRmin.Therefore, the scheduler will increase ne by 1 (i.e.,ne ¼ ne þ 1) in order to improve the basic video qualityfor each multicast member. After the sampling stage (i.e.,the first three sample GOPs), the scheduler can use theupdated rs and ne for the rest of GOPs in the currentsampling interval. The whole procedure repeats for eachsampling interval.

4.3 Two-Stage Rate Adaptation

Since the sampled rate (i.e., a higher and a lower rate of thecurrent rs) may not be suitable for the multicast group,QDM should avoid unnecessary samples if both of thenetwork topology and the video rate are static. Specifically,the scheduler must adjust the sampling interval so as toadapt the rate to network dynamics and variable video bitrates; however, it must keep the selected rate unchanged ifthe environment is stable. Here, we propose a two-state rateadaptation scheme, which is a finite-state machine designedto determine the size of the sampling interval that canreflect variation of channel conditions or video bit rates.

As Fig. 5 shows, there are two states and two transitionfunctions in the proposed two-state machine. We expectthat the system stays in the active state if it still needs tosearch a suitable rate due to unstable environments (i.e.,network dynamics or variable video rates); otherwise, thesystem can stay in the static state, and use the currentlyselected rs and ne. The functions of two states and twotransitions are specified as follows:

. Active state: The system stay in the active state if theclients are mobile or the video bit-rate varies withtime. In this state, we repeat the sampling procedurefor each sampling interval, which is set to a fixedsize (e.g., set to six GOPs in our simulations). Hence,the system can sample a better rate periodically.

. Active state ! static state: If the system selects thesame rate rs for k (set to 2 in our simulations)continuous sampling intervals, it then switches fromthe active state to the static state because the selected


Fig. 4. Rate selection by sampling.

Fig. 5. Two-state rate adaptation.

rate works well for the current environment. As thetransition occurs, the system records the PSNR ofGOPbest-sample of the last sampling interval in theactive state as PSNRactiveðCHiÞ for each CHi.

. static state: The system stops sampling, and uses theselected rs and ne for the following GOPs.

. static state ! active state: When the system is in thestatic state, it notifies the cluster heads to report themask of each GOP such that it can track the visualquality of each CH. Once the system detects that anyCH gets a PSNRðCHiÞ less than PSNRactiveðCHiÞ ��, where � is a constant set to 10 dB in oursimulations, the system switches from the staticstate back to the active state.

Even though the cluster heads must respond the maskmore frequently, however, the size of a mask (i.e., x bits forx packets) is much smaller than the size of supplementaryframes Fs in a sample GOP. If the scheduler samples aninappropriate rs during the sampling procedure, receiversmay not achieve an adequate video quality. In this case,transmission of Fs in sample GOPs at an inappropriate ratecan be deemed as a sampling overhead. That is, thescheduler may waste a significant sampling overhead if itexecutes the sampling procedure when the environment isstable and suitable to use the current rate rs. Thus, weprefer to let the system enter the static state, which mayneed a few overhead of responding the masks, butsignificantly save the sampling overhead. By controllingthe size of the sampling interval, the two-state machine canadapt rs to network dynamics with a reasonable samplingoverhead.

5 SIMULATION-BASED EVALUATION

We first conduct simulations to evaluate the performance ofQDM in various scenarios using simulations, and thencheck the feasibility of QDM using trace-based implementa-tion in the following section. The simulations are imple-mented in NS2 [30], which is modified to support multiplerates. In each simulation, AP is deployed in the center of thetopology as the video source; 30 clients are uniformlydistributed at random over the square field topology. Eachclient is equipped with an 802.11b interface, which supportsrates 1, 2, 5.5, and 11 Mb/s. Since the default NS2 does notconsider bit error of each packet transmission, we use the802.11b PHY Simulink Model [31] as the channel errormodel, and set the transmission power of each node to10 dB. We use JM15.1 to encode and decode the H.264 videostreams, each of which has 15 frames in each GOP and aduration of 30 seconds. Each video is encoded as a variablebit-rate stream with the average video bit rate 300 Kbps. Wetest several video sequences, including forman qcif:yuv,akiyo qcif:yuv, and grandma qcif:yuv, each of which isencoded as the QCIF format at a video frame rate of30 (frames/s). The background traffic is set to 750 Kb/s, sothe available bandwidth is not enough to multicast everyframe at the base rate, i.e., 1 Mb/s. PSNRmin is set to 30 dB.The sensitivity coefficient � used to recluster clients is set to70 percent. In the two-state adaptation scheme, the thresh-old of duplicated samples k is set to 2, and � is set to 10 dB.

The default sampling interval (in the active state) is sixGOPs. All figures represent the average results over10 random topologies.

We evaluate the performance of QDM in terms ofPSNRtotal and the CDF of PSNR perceived by each client.We compare the following schemes:

1. Oracle: the dynamic-programming-based solutionwith oracle information, as shown in Algorithm 1;

2. QDM: the proposed quality-differentiated multicast;3. ARSM: the leader-based solution, which selects the

transmission bit-rate based on the worst node;4. H-ARSM: the two-level solution, which selects the

rate for the base layer according to the channelcondition of the worst node, while selects the rate forthe enhancement layer based on the best node;

5. BASE-RATE: the baseline scheme that always sendspackets at the base rate.

5.1 Video Quality Comparison

We uniformly distribute 30 clients at random in a 100 m by100 m field. Fig. 6 shows the CDF of PSNR perceived byeach client. Since the best rate of the worst node in thisscenario is the base rate (1 Mb/s), ARSM performs the samewith BASE-RATE. Thus, we exclude the results of BASE-RATE from Fig. 6. In ARSM, because the stream is sent atthe fixed base rate, all clients can receive the same subset offrames and get the same quality (about 32 dB). On the otherhand, since H-ARSM sends the base-layer and the enhance-ment-layer at two different transmission bit-rates, it canprovide a better video quality for the clients locating nearbyAP. However, the side effect is that it instead penalizesmost of clients who cannot receive the enhancement layer,which is sent at a higher rate (i.e., the transmission bit-rateof the best node). Therefore, in H-ARSM, 47 percent ofclients cannot achieve PSNRmin (i.e., 30 dB), and, evenworse, get lower PSNR than those in ARSM. Unlike H-ARSM that produces two extreme qualities (extremely highand low PSNR), QDM can exploit the cluster-basedstructure to characterize the heterogeneity of clients, and,thereby, provide each user a visual quality matching itschannel condition. Besides, almost all clients in QDM canachieve PSNRmin.

In Figs. 7a, 7b, and 7c, we compare the performance of allschemes under three different node distributions, i.e.,uniform, normal, and zipf, respectively. The zipf distribu-tion is defined as P ½20 � ði� 1Þ � dist � 20 � i� ¼ c

i , fori ¼ 1; 2, . . . , K, where dist is the distance between the clientand AP, c ¼ ð

PKi¼1

1iÞ�1, and K is the network size (defined


Fig. 6. CDF of PSNR in uniform distribution.

as the radius of a topology) divided by 20 m. Thecharacteristic of the zipf distribution is that the probabilityof placing a client nearby AP is higher than that of placing itfar from AP; hence, it can represent an environment wheremost of clients are close to AP, while only a few clients aredistant from AP and experience a worse channel condition.On the other hand, the normal distribution means that thedistance between each client and AP follows the normaldistribution, which means that only few clients areextremely close to or far from AP.

In the uniform and zipf distributions, we vary thenetwork size from 20 m to 160 m; in the normal distribution,we fix the mean of the network size to 60 m and vary thestandard deviation from five to 40 since the maximal radiuscannot be known in the normal distribution. We note that,in all distributions, BASE-RATE provides clients similarvideo qualities because AP always sends the stream at thebase rate without considering various channel conditions ofclients. On the other hand, ARSM can outperform BASE-RATE when the network size is small because it could selectthe rate according to the channel condition of the worstnode. Hence, when the network size is small, the rate usedby the leader could be higher than the base rate. Thus,ARSM allows clients to receive a higher throughput andvideo quality as compared to BASE-RATE. However, whenthe network size grows, the sender will transmit at thelowest base-rate even if there is only a single client, theworst client, that can only receive at the base-rate. In thiscase, ARSM performs exactly the same with BASE-RATEwhen the network size is larger than 80 m.

Even though H-ARSM notices the heterogeneity ofclients, it enables clients using the same rate with the bestnode to receive a better visual quality by penalizing otherclients. Such a two-level strategy could be suitable for thezipf distribution, where most nodes gather nearby thesender. However, it may perform worse than BASE-RATEwhen most of nodes locate between the best node and theworst node, such as in the normal or uniform distribution.By contrast, the proposed QDM can provide a higheraverage PSNR than other schemes no matter in which nodedistribution. This is because the cluster-based structure cancapture the heterogeneity of clients, i.e., the number ofclients in each cluster; hence, it can select an appropriate rateto transmit each video frame, and maximize video qualityfor each cluster. Besides, unlike H-ARSM, which does notconsider the minimum video quality requirement, QDMfirst provides each node a minimum video quality, and then

attempts to improve the total PSNR. Hence, the worstnodes in QDM do not suffer from the starvation problem.

Finally, the proposed protocol, QDM, can providealmost the same visual quality as the oracle solution whenthe network size is smaller than 100 m. It indicates that thescheduler can use the efficient sampling mechanism tofigure out the appropriate transmission bit-rates for eachindividual frame. The reason of the small gap betweenQDM and the oracle solution is because of the estimationerror of the incremental quality �qðÞ used in the oraclesolution. Specifically, the dynamic programming solutionconsiders the influence of missing each individual framebased on its incremental quality �qðÞ. However, such anincremental quality �qðÞ can only be estimated by existingdistortion functions. In addition, the incremental qualityof each frame can only be used as the “first-moment”estimate; that is, the actual video quality is not exactlyequal to the summation of the incremental quality of allreceived frames. As a result, the performance of thedynamic programming-based solution is determined byaccuracy of the estimate, �qðÞ. By contrast, our cluster-based algorithm applies the sampling mechanism to real-time compute the decoded visual quality of the represen-tative cluster heads, and adjusts the transmission schedulebased on online feedback. Therefore, it can avoid theestimation error of �qðÞ and, in a few cases, can slightlyoutperform the dynamic-programming-based solution.

The performance gap increases slightly when the net-work size grows and, in turn, the heterogeneity of clientsbecomes more obvious. This is because the cluster-basedmethod used in QDM can only select representative clusterheads to estimate channel quality of all users. By contrast,the dynamic-programming-based algorithm proposed inAlgorithm 1 can use the oracle information to explicitlycalculate visual quality of each multicast member and solvethe rate schedule explicitly. However, we note that theperformance gap between the oracle solution and QDM isat most 1.5 dB. Another observation is that the oraclesolution estimates incremental visual quality �qðÞ based onthe existing rate-distortion functions, which still have somelimitation to reflect real visual quality accurately. There-fore, it sometimes performs worse than QDM becauseQDM inspects the actual visual quality by computingPSNR based on the acknowledgement responded by CHs.

5.2 Impact of Rate Adaptation

Next, we evaluate the impact of rate adaptation on theperformance of QDM. In this simulation, 30 nodes are


Fig. 7. Impact of node distribution.

uniformly distributed in a 140 m by 140 m field. We letlocation of clients and background traffic keep static.Figs. 8a and 8b compare the performance of QDM withadaptation and that of QDM without adaptation in terms ofthe average PSNR of a multicast group and the PSNR of theworst node, respectively. Adaptation means that QDMadapts the rate dynamically for each sampling interval; noadaptation means that AP uses the fixed rate selected in thestart-up stage without updating the rate afterward. We plotthe results of first 20 seconds of a video stream, andillustrate how video quality varies with time.

We note that even though the environment is stable(i.e., static nodes and background traffic), the average PSNRin Fig. 8a and the PSNR of the worst node in Fig. 8b stilldegrade because the video bit rate may vary with time. Therate selected in the initial sampling intervals is only suitablefor the video bit rate in that period, but may not provide anadequate video quality when the video bit rate changes.Specifically, when the video bit rate increases, the schedulermay need to increase the number of essential frames (ne) inorder to provide a minimum video quality PSNRmin for

every member. Hence, QDM can produce a better videoquality than no adaptation and provide the worst node atleast an adequate quality because it adapts the rate tovariable video bit rates even when the wireless environmentis static. We note that sometimes the PSNR of the worstnode is still lower than PSNRmin. This is because the videobit rate in some periods is burst so that the availablebandwidth is not enough to deliver all essential frames Feused to satisfy PSNRmin. Sometimes, adaptation mightperform worse than no adaptation, e.g., at 12th second,because it is unnecessary to waste the sampling overheadwhen the initial rate happens to be suitable for the currentvideo rate. However, we note that, for most of the time,adaptation can gain higher visual quality than no adaptationbecause the bit-rate of a video stream usually fluctuates.

5.3 Impact of Reclustering

We next examine whether the system can gain fromreclustering. We consider a scenario that clients are mobilebased on three different mobility patterns. First, wedistribute nodes uniformly in the topology, and let each


Fig. 9. Impact of reclustering.

Fig. 8. Impact of rate adaptation.

node move according to the random waypoint model [32].Fig. 9a presents the performance of various mean mobilityvelocities, with the standard deviation 0.5. The figure doesnot include the results of the dynamic-programming-basedsolution because it can only be solved in preprocess, and,thus, cannot adapt to dynamic scenarios. In this simulation,we tend to evaluate the effect of reclustering clients as thetopology changes due to the client movement. The figureshows an interesting phenomenon that video quality doesnot decline much if QDM does not recluster the clientswhen they move to the new locations. This is because, inthe random waypoint model, the node distribution duringthe movement is similar to the original one. That is, eachnode moves to a new location selected randomly from thetopology, so the locations of clients during/after movementstill follow the uniform distribution. Thus, the number ofmembers in each cluster Ci 2 C does not change much. Thisis the reason why AP can still estimate ~PSNRtotal correctlybased on the original cluster sizes jCij, even if it does notrecluster the clients.

We consider another mobility pattern that the networksize expands from 60 m� 60 m to 150 m� 150 m. We let eachnode move away from AP so that the number of clients ineach cluster could change with time. When nodes move awayfrom AP, there may be more members that can only receive ata lower rate. Thus, without reclustering, AP may get anincorrect estimate of ~PSNRtotal by

PCi2C PSNRðCHiÞ � jCij

because jCij has changed when the network size grows. TheCDF of PSNR perceived by each member in Fig. 9b indicatesthat, without reclustering, the AP keeps using the out-of-daterate such that only few nodes staying nearby AP can getsatisfactory PSNR. Those members moving away from APcannot receive data sent at the original rate anymore, andthereby perceive a degraded video quality. However, byreclustering the clients, AP can characterize the up-to-dateheterogeneity of clients after movement, and adapt the rate toproducing an adequate quality under mobile environments.Besides, QDM with reclustering achieves the average PSNR33.18 dB, which is higher than that without reclustering.

Finally, we consider an extreme scenario where theclients are originally uniformly distributed with a distanceto the AP smaller than 50 m. During the streaming period,the cluster heads do not move, while the other clients moveaway from the AP. Such a scenario occurs in a conference ora hotspot setting when attendees move from one place toanother. We let the distribution of the mobile clients’ speedfollow the uniform distribution. We then check the gain ofreclustering when the mean moving speed of the mobileclients varies from 1 to 20 m/s. The speeds are selected tosimulate different mobile scenarios, such as 1 m/s for thewalking speed, 5-10 m/s for the running speed, and 20 m/sfor the driving speed. We notice that 20 m/s might not be areasonable setting for the conference scenario, but is onlyused to test the limitation of our protocol. Moreover, toprevent a user from moving quickly to the area out of thetransmission range, we let all the users keep static for thefirst 10 seconds, and start moving at the 10th second. Ifthe users move to the boundary of the transmission range,they stop moving and keep staying in the same location.

The PSNR value is computed over the entire videostreaming duration.

Fig. 9c shows that enabling reclustering produces abetter visual quality because, even if the cluster heads arestable, the mobile clients can trigger reclustering when theirSNRs change significantly. The figure also shows that thegain of reclustering is not obvious when the moving speedis slow because the structure of clusters does not changemuch during the streaming period. Reclustering howeverplays a key factor that improves the visual quality when themoving speed increases to 10 m/s, resulting in a smoothchange in the cluster structure. When the speed keepsincreasing up to 20 m/s, the gain of reclustering is limited.The reason of this limitation is two-fold: 1) when thestructure of clusters changes too fast, the cost of recluster-ing could offset the gain of reclustering; 2) more impor-tantly, when the moving speed becomes 20 m/s, all clientscould move to the region where only the lowest rate isavailable. As a result, there is no need to trigger reclusteringbecause most of clients, except cluster heads, can onlyreceive at 1 Mb/s. In this case, the system cannot gain fromreclustering.

We next investigate how the value of the triggeringthreshold (�) affects the overhead of reclustering in theconference scenario, i.e., the third mobile scenario, with themean speed 10 m/s. We then change the value of � from 0.1to 0.7. Fig. 9d indicates that when the threshold is set to atoo small value, the system cannot quickly adapt to networkchanges. On the contrary, when the threshold increases, thecost of performing the probing procedure for reclusteringcould occupy a few more airtime and, thus, slightly sacrificethe transmission opportunities for the useful video stream.To adapt to network dynamics, setting � to a value between0.4 and 0.7 is a reasonable setting.

5.4 Impact of Two-State Adaptation

Table 1 shows the overhead of QDM based on the two-statemachine and periodical sampling (called one-state for shortdue to always in the active state). We mentioned in Section 4that the sampling overhead is much higher than thefeedback overhead. Hence, by applying the two-statemachine, QDM can adapt the size of the sampling intervaldynamically, and reduce the sampling overhead from 16.21to 6.9 percent. Fig. 10a also shows that PSNR in the one-statescheme sometimes drops, e.g., at 26th and 44th seconds,because the scheduler may waste some unnecessary sampleGOPs to test unsuitable rates. On the contrary, the proposedtwo-state mechanism can avoid this situation by letting thesystem switch to the static state when the environment isstable. Thus, in the two-state scheme, AP can send thestream at the previous selected rate, which is suitable for thecurrent environment, instead of testing the wrong rates forsample GOPs. Besides, for most of the time, the two-statescheme outperforms the periodical sampling scheme. This


TABLE 1Overhead Comparison (Video Size: 8.8 Mb)

means that the proposed two-state machine can assist thescheduler in distinguishing whether the environment isdynamic, and allow the system to switch to the correct state.Otherwise, the two-state mechanism may produce a worsequality than periodical sampling if it keeps the systemstaying in a wrong state.

We next examine the effect of two-state adaptation in amobile scenario, where clients have the same uniformmobility pattern with the scenario shown in Fig. 9a. Fig. 10bplots the comparison between two-state adaptation andperiodical sampling in such a mobile scenario. The figureshows that, when the environment is dynamic, the sendershould stay in the active state to keep tracking the besttransmission bit-rate. Our two-state adaptation scheme candetect the network dynamics automatically and tune itselfto operate similar to periodical sampling. As a result, itcan perform as well as periodical sampling in a dynamicenvironment.

6 EXPERIMENTAL EVALUATION

We use real traces to examine the performance of QDM.Experimental setup. We implement a trace-based test

environment to evaluate the performance of QDM in realwireless channels. To do so, we deploy a sender at thelocation marked as “X,” as shown in Fig. 11, and put areceiver at one of the 30 marked locations in Fig. 11. Foreach receiving location, we make the sender transmit astream of 1,000-byte packets. The packet stream cyclesbetween transmitting at each of the 802.11b rates, i.e., 1, 2,5.5, and 11 Mb/s. Each packet is marked with a sequencenumber. The receiver logs the packets as a trace file usingWireshark [33]. We offline process the traces, and produce acombo-mask that represents the actual packet loss patternof using each different rate on the air. For each location, werepeat the experiment to collect 10 traces. We then use each

trace to compute the best bit-rate that can achieve thehighest throughput for each marked location. Given thebest rate of each different location, we feed the correspond-ing combo-mask to our program that performs ourclustering-based algorithm and compute the total PSNRbased on the packet loss pattern of the real traces.

Results. For each trace-based evaluation, we randomlyassign k multicast clients to k randomly selected locationsmarked in Fig. 11, where k varies from 5 to 25. For eachselected location, we run our trace-based evaluation 10 timesusing 10 different traces. The results are the average over20 random location assignments. Fig. 12a shows the PSNRcomparison between our QDM and the previous schemes,such as ARSM and H-ARSM. The figure shows that ourQDM enables heterogeneous visual quality and thereforeimprove the average PSNR. Fig. 12b plots the CDFs of PSNRwhen there are 25 clients. The figure shows a trend similarto our simulation results. This figure also indicates that, inH-ARSM, most of clients that cannot receive at the highestrate cannot receive part of video frames and thereby cannotget a satisfactory visual quality. This is also why H-ARSMproduces a lower average PSNR than ARSM.


Fig. 11. Testbed. Dots refer to client locations. Mark “X” refers to senderlocation. Fig. 12. Trace-based evaluation.

Fig. 10. Impact of two-state rate adaptation.

7 CONCLUSIONS

This paper investigated the rate scheduling problem of

video multicast for heterogeneous clients over wireless

environments. We have proposed a rate scheduling model

that solves the theoretical optimal solution by dynamic

programming. A practical protocol, called QDM, was

further presented for real-time video streaming even with-

out preprocess on computing the rate-distortion function

and estimating the loss probability of each wireless link. In

QDM, we exploit a cluster-based structure to provide

differentiated qualities for heterogeneous clients. Based on

the information reported by cluster heads, the sender can

estimate the total video quality and explore a suitable rate

for each video frame based on a sample-based scheme. The

performance evaluation shows that QDM can produce a

gain of 2-5 dB in terms of the average video quality as

compared to the leader-based scheme, while also guarantee

that each client perceives at least a minimum video quality.

In addition, the two-state rate-adaptation scheme in QDM

can adapt the rate to network dynamics and variable video

bit rates with a reduced sampling overhead.

ACKNOWLEDGMENTS

This work was partially supported by the National Science

Council of ROC under contract No. NSC99-2218-E-001-005

and NSC100-2221-E-001-005-MY2.

REFERENCES

[1] K.A. Hua and F. Xie, “A Dynamic Stream Merging Technique forVideo-on-Demand Services over Wireless Mesh Access Net-works,” Proc. IEEE CS Seventh Ann. Conf. Sensor Mesh andAd Hoc Comm. and Networks (SECON), 2010.

[2] A. Kamerman and L. Monteban, “WaveLAN-II: A High-Perfor-mance Wireless LAN for the Unlicensed Band,” Bell Labs TechnicalJ., vol. 2, no. 3, pp. 118-133, 1997.

[3] G. Holland, N. Vaidya, and P. Bahl, “A Rate-Adaptive MACProtocol for Multi-Hop Wireless Networks,” Proc. ACM MobiCom,2001.

[4] B. Sadeghi, V. Kanodia, A. Sabharwal, and E. Knightly, “OAR: AnOpportunistic Auto-Rate Media Access Protocol for Ad HocNetworks,” Proc. ACM MobiCom, 2002.

[5] J. Bicket, “Bit-Rate Selection in Wireless Networks,” PhDdissertation, Massachusetts Inst. of Technology, 2005.

[6] J. Camp and E. Knightly, “Modulation Rate Adaptation in Urbanand Vehicular Environments: Cross-Layer Implementation andExperimental Evaluation,” Proc. ACM MobiCom, 2008.

[7] G. Judd, X. Wang, and P. Steenkiste, “Efficient Channel-AwareRate Adaptation in Dynamic Environments,” Proc. ACM MobiSys,pp. 118-131, 2008.

[8] M. Vutukuru, H. Balakrishnan, and K. Jamieson, “Cross-LayerWireless Bit Rate Adaptation,” Proc. ACM SIGCOMM, 2009.

[9] Z. Lei and N.D. Georganas, “Rate Adaptation Transcoding forVideo Streaming over Wireless Channels,” Proc. IEEE Int’l Conf.Multimedia and Expo (ICME), 2003.

[10] S. Pal, S.R. Kundu, A.R. Mazloom, and S.K. Das, “Video RateAdaptation and Scheduling in Multi-Rate Wireless Networks,”Networking: Proc. Sixth Int’l IFIP-TC6 Conf. Ad Hoc and SensorNetworks, Wireless Networks, Next Generation Internet, pp. 475-487,2007.

[11] S. Lee and K. Chung, “Combining the Rate Adaptation andQuality Adaptation Schemes for Wireless Video Streaming,”J. Visual Comm. and Image Representation, vol. 19, no. 8, pp. 508-519, 2008.

[12] J. Kuri and S. Kasera, “Reliable Multicast in Multi-Access WirelessLANs,” Proc. IEEE INFOCOM, Mar. 1999.

[13] A. Basalamah, H. Sugimoto, and T. Sato, “Rate Adaptive ReliableMulticast MAC Protocol for WLANs,” Proc. IEEE VehicularTechnology Conf. (VTC Spring), May 2006.

[14] Y. Park, Y. Seok, N. Choi, Y. Choi, and J.-M. Bonnin, “Rate-Adaptive Multimedia Multicasting over IEEE 802.11 WirelessLANs,” Proc. IEEE Consumer Comm. and Networking Conf., 2006.

[15] J. Villalon, P. Cuenca, L. Orozco-Barbosa, Y. Seok, and T. Turletti,“Cross-Layer Architecture for Adaptive Video Multicast Stream-ing over Multirate Wireless LANs,” IEEE J. Selected Areas inComm., vol. 25, no. 4, pp. 699-711, May 2007.

[16] O. Alay, T. Korakis, Y. Wang, and S. Panwar, “Dynamic Rate andFEC Adaptation for Video Multicast in Multi-Rate WirelessNetworks,” Mobile Networks and Applications (Online Version),vol. 15, pp. 425-434, 2009.

[17] “Objective Perceptual Multimedia Video Quality Measurementin the Presence of a Full Reference,” ITU recommendation J.247(08/08), http://www.itu.int/rec/T-REC-J.247/en, 2012.

[18] S. Gringeri, R. Egorov, K. Shuaib, A. Lewis, and B. Basch, “RobustCompression and Transmission of MPEG-4 Video,” Proc. ACMSeventh ACM Int’l Conf. Multimedia, pp. 113-120. 1999,

[19] D.-N. Yang and M.-S. Chen, “Bandwidth Efficient Video Multi-casting in Multiradio Multicellular Wireless Networks,” IEEETrans. Mobile Computing, vol. 7, no. 2, pp. 275 -288, Feb. 2008.

[20] C. Hsu and M. Hefeeda, “Flexible Broadcasting of Scalable VideoStreams to Heterogeneous Mobile Devices,” IEEE Trans. MobileComputing, vol. 10, no. 3, pp. 406-418, Mar. 2011.

[21] A. Chen, D. Lee, and P. Sinha, “Optimizing Multicast Performancein Large-Scale WLANs,” Proc. IEEE 27th Int’l Conf. DistributedComputing Systems (ICDCS), 2007.

[22] R. Chandra, S. Karanth, T. Moscibroda, V. Navda, J. Padhye, R.Ramjee, and L. Ravindranath, “DirCast: A Practical and EfficientWi-Fi Multicast System,” Proc. IEEE 17th Int’l Conf. NetworkProtocols (ICNP), 2009.

[23] S. Deb, S. Jaiswal, and K. Nagaraj, “Real-Time Video Multicast inWiMAX Networks,” Proc. IEEE INFOCOM, 2008.

[24] K. Leung, M. Clark, B. McNair, Z. Kostic, L. Cimini, and J.Winters, “Outdoor IEEE 802.11 Cellular Networks: Radio andMAC Design and Their Performance,” IEEE Trans. VehicularTechnology, vol. 56, no. 5, pp. 2673-2684, Sept. 2007.

[25] D. Cox, “Delay Doppler Characteristics of Multipath Propagationat 910 MHz in a Suburban Mobile Radio Environment,” IEEETrans. Antennas and Propagation, vol. AP-20, no. 5, pp. 625-635,Sept. 1972.

[26] S. Sen, S. Gilani, S. Srinath, S. Schmitt, and S. Banerjee, “Designand Implementation of an ‘Approximate’ Communication Systemfor Wireless Media Applications,” Proc. ACM SIGCOMM, 2010.

[27] S. Jakubczak, H. Rahul, and D. Katabi, “One-Size-Fits-All WirelessVideo,” Proc. ACM SIGCOMM HotNets Workshop, 2009.

[28] J. Chakareski, J. Apostolopoulos, W.-T. Tan, S. Wee, and B. Girod,“Distortion Chains for Predicting the Video Distortion for GeneralPacket Loss Patterns,” Proc. IEEE Int’l Conf. Acoustics, Speech, andSignal Processing (ICASSP ’04), 2004.

[29] A. Wyner and J. Ziv, “The Rate-Distortion Function for SourceCoding with Side Information at the Decoder,” IEEE Trans.Information Theory, vol. IT-22, no. 1, pp. 1-10, Jan. 1976.

[30] “The Network Simulator NS-2,” http://www.isi.edu/nsnam/ns,2012.

[31] “802.11b PHY Simulink Model,” http://www.mathworks.com/matlabcentral/fileexchange/2262, 2012.

[32] D.B. Johnson and D.A. Maltz, “Dynamic Source Routing inAd Hoc Wireless Networks,” Mobile Computing, vol. 353, pp. 153-181, 1996.

[33] “Wireshark,” http://www.wireshark.org, 2012.


Kate Ching-Ju Lin received the PhD degreefrom National Taiwan University in 2009. Shewas a visiting scholar in the Computer Scienceand Artificial Intelligence Laboratory (CSAIL) atMIT in 2007. After her graduation, she joined theResearch Center for Information TechnologyInnovation at Academia Sinica, Taiwan. She isnow an assistant research fellow. Her currentresearch interests include wireless systems,wireless mesh networks, and sensor networks.

She is a member of the IEEE.

Wei-Liang Shen received the BS degree fromNational Central University, Taiwan, in 2008 andthe MS degree from National Taiwan Universityin 2010. He is also working toward the PhDdegree in the Department of Electrical Engineer-ing, National Taiwan University. He is currently aresearch assistant in the Research Center forInformation Technology Innovation at AcademiaSinica, Taiwan. His current research interestsinclude multiuser MIMO systems and peer-to-

peer streaming.

Chih-Cheng Hsu received the MS degree fromNational Cheng Kung University, Taiwan, in2006 and the PhD degree from National TaiwanUniversity in 2012. After his graduation, hejoined the Computer Science and InformationEngineering Department at the National TaiwanUniversity and he is now a postdoctoral fellow.His research interests include wireless sensornetworks, underwater sensor networks, P2Pcomputing, and cloud networks.

Cheng-Fu Chou received the PhD degree fromthe University of Maryland, College Park. Afterhis graduation, he joined the Computer Scienceand Information Engineering Department at theNational Taiwan University, where he is now anassociate professor. He was a visiting scholar atthe University of Southern California. His currentresearch interests are in peer-to-peer networks,distributed multimedia systems, multi-hop wire-less networks, sensor networks, and their

performance evaluation. He is a member of the IEEE.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

Networks