06221968

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 21, NO. 2, APRIL 2013 469

Distortion-Aware Scalable Video Streamingto Multinetwork Clients

Nikolaos M. Freris, Member, IEEE, Cheng-Hsin Hsu, Member, IEEE, ACM, Jatinder Pal Singh, Member, IEEE,and Xiaoqing Zhu, Member, IEEE

AbstractWe consider the problem of scalable video streamingfrom a server to multinetwork clients over heterogeneous accessnetworks, with the goal of minimizing the distortion of the receivedvideos. This problem has numerous applications including: 1) mo-bile devices connecting to multiple licensed and ISM bands, and2) cognitive multiradio devices employing spectrum bonding. Inthis paper, we ascertain how to optimally determine which videopackets to transmit over each access network. We present modelsto capture the network conditions and video characteristics anddevelop an integer program for deterministic packet scheduling.Solving the integer program exactly is typically not computation-ally tractable, so we develop heuristic algorithms for deterministicpacket scheduling, as well as convex optimization problems forrandomized packet scheduling. We carry out a thorough study ofthe tradeoff between performance and computational complexityand propose a convex programming-based algorithm that yieldsgood performance while being suitable for real-time applications.We conduct extensive trace-driven simulations to evaluate theproposed algorithms using real network conditions and scalablevideo streams. The simulation results show that the proposedconvex programming-based algorithm: 1) outperforms the ratecontrol algorithms defined in the Datagram Congestion ControlProtocol (DCCP) by about 1015 dB higher video quality; 2) re-duces average delivery delay by over 90% compared to DCCP;3) results in higher average video quality of 4.47 and 1.92 dB thanthe two developed heuristics; 4) runs efficiently, up to six timesfaster than the best-performing heuristic; and 5) does indeedprovide service differentiation among users.

Index TermsQuality optimization, rate control, stream adap-tation, video streaming.

I. INTRODUCTION

M ARKET research indicates that mobile data traffic willincrease 39 times over a span of five years, and 66% ofthe increase will be due to mobile videos [4]. In fact, cellularservice providers are having a hard time coping with the huge

Manuscript received September 05, 2011; revised February 22, 2012and April 10, 2012; accepted May 19, 2012; approved by IEEE/ACMTRANSACTIONS ON NETWORKING Editor M. Reisslein. Date of publicationJune 20, 2012; date of current version April 12, 2013. The work of C.-H. Hsuwas supported in part by the National Science Council (NSC) of Taiwan underGrant #100-2218-E-007-015-MY2.N. Freris is with IBM Research, 8803 Rschlikon, Switzerland (e-mail:

[email protected]).C.-H. Hsu is with National Tsing Hua University, Hsin Chu, Taiwan (e-mail:

[email protected]).J. P. Singh is with the Department of Electrical Engineering, Stanford Uni-

versity, Stanford, CA 94305 USA (e-mail: [email protected]).X. Zhu is with Cisco Systems, Inc., San Jose, CA 95134 USA (e-mail:

[email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TNET.2012.2203608

increase in mobile data traffic [1], [3] and will have to care-fully engineer their systems to support high-quality real-timevideo streaming. In wireless networks, one way to achieve thebest possible streaming quality is to leverage all available wire-less spectra by connecting the streaming server to each clientvia multiple access networks. We refer to the clients capableof connecting to multiple access networks as multinetwork ormultihomed clients. Potential application scenarios of multinet-work clients include streaming videos to: 1) multiradio wirelessdevices connected to different Industrial, Scientific, and Med-ical (ISM) bands [37]; 2) cognitive multiradio clients employingspectrum bonding [34]; and 3) multiradio clients connected toboth licensed band (such as 3G cellular network) and ISM band(such as IEEE 802.11 networks) [14]. A streaming server maytransmit a video concurrently over multiple access networksto a multinetwork client, thus aggregating the various wirelessspectra to achieve better streaming quality. We call this setupmultinetwork (multihomed) video streaming, which is particu-larly challenging because access networks are diverse and dy-namic. We note that concurrently activating multiple networkinterfaces may lead to higher energy consumption on mobiledevices. While energy conservation is out of the scope of thispaper, several prior studies [16], [19], [33] propose mechanismsto achieve burst traffic delivery to conserve energy, which canbe used inmultinetwork video systems. Lastly, multihoming canalso be viewed as an alternative to multipath video streaming.Multipath streaming, although studied in the literature, e.g.,[10],is not widely deployed. This is partially due to the additionalrequirements on designated network equipment. In contrast tomultipath video streaming, multihomed video streaming workson the current Internet infrastructure: For example, cellular ser-vice providers can adopt multihomed video streaming to maxi-mize the overall streaming quality without overloading the net-works. Multihomed video streaming however is challenging be-cause of: 1) the heterogeneity and dynamics of access networks,and 2) complicated interdependency among video packets.An approach of arbitrarily splitting a video stream into mul-

tiple substreams and sending each substream over an access net-work may lead to degraded video quality and playout glitches.This is because transmitting a substream at a low rate may un-derutilize the network resources, while transmitting at a rateclose to the available bit rate may lead to network congestion,which in turn causes packet drops and late packet delivery. Tothis end, rate control based on measurements of available bitrate (ABR) and round-trip time (RTT) needs to be performedto achieve a good tradeoff between throughput and delay. Innonscalable video streaming, once the bit rate of each substream

1063-6692/$31.00 2012 IEEE

470 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 21, NO. 2, APRIL 2013

is determined, the video stream must be adapted into the rightformat so that it can be delivered to the client in a timely fashion.We refer to this conversion as stream adaptation, which is typi-cally implemented by means of a computationally demandingoperation called transcoding [29], [36]. In contrast, scalablevideo coding, such as the H.264/SVC standard [29] supports ef-ficient stream adaptation and allows service providers to saveexpenses on deploying streaming servers and transcoders. De-spite a small cost on coding inefficiency, modern H.264/SVCcoders are reported to significantly outperform several scalablecoding schemes and even outperform some nonscalable coderssuch as MPEG-4 Advanced Simple Profile (ASP) [35]. Scal-able video streams feature complex interdependencies amongvideo packets, which stream adaptation must account for. Therate control and stream adaptation problems must be simultane-ously considered for optimal video streaming quality.In this paper, we present a mathematical formulation of the

joint rate control and scalable stream adaptation problem formultiple clients1 concurrently competing for the same accessnetworks (cf. Fig. 12). We abstract the problem of streamingvideos to multinetwork clients and formulate an optimizationproblem to determine, for each client: 1) the streaming rate overeach access network; 2) the video packets to be transmitted; and3) the access network over which each transmitted video packetis sent. Due to the discrete nature of the considered optimizationproblem, and its NP-completeness, we formulate it as an integerprogram in order to derive the global-optimal solutions. Ourcontributions can be summarized as follows. We formulate the joint rate control and packet schedulingproblem as an integer program where the objective is tominimize a cost function of the expected video distortion.We suggest different cost functions in order to provide ser-vice differentiation and address fairness among users.

We propose heuristic algorithms for packet scheduling,analyze their complexity, and study their performancethrough trace-driven simulations.

We consider randomized packet scheduling by relaxing theinteger program into a real-valued optimization problem.We derive convex programming approximations to thisproblem.

We analyze, both analytically and experimentally, the per-formance versus computational complexity tradeoff of theproposed optimization programs and recommend one thatyields good performance while being suitable for real-timeapplications.

Simulation results show that the proposed algorithm:1) outperforms the rate control algorithms defined in theDatagram Congestion Control Protocol (DCCP) stan-dard [22] by about 1015 dB in terms of video quality;2) achieves better balance between performance andruntime; 3) reduces average delivery delay by over 90%compared to DCCP; 4) results in better performance than

1Throughout the paper, we use the terms client and user interchangeably.2This figure shows a sample architecture, in which our algorithm runs on the

streaming server. Before a streaming session starts, each client sends a CON-NECT UDP message to the server using different IP addresses associated witheach client network interface. The different IP addresses allow the streamingserver to direct the video packets over the chosen access networks to the client.

Fig. 1. Sample system architecture of a scalable video streaming system withclients and access networks.

the proposed heuristics, under diverse background trafficload; and 5) indeed provides service differentiation amongusers.

The rest of this paper is organized as follows. We present re-lated work in Section II. In Section III, we expose the problemformulation. In Section IV, we develop deterministic and ran-domized packet scheduling algorithms. We present trace-drivensimulations to evaluate the proposed algorithms in Section V.In Section VI, we discuss some limitations of our approach andpropose future work, while Section VII concludes the paper.

II. RELATED WORKRate control for nonscalable video streams has been exten-

sively investigated [6], [10], [11], [21], [30], [32], [39], [41].Chakareski and Girod [10] propose an algorithm to enable astreaming server to decide which packets to transmit over whichnetwork paths so as to meet the predefined bandwidth con-straints. Szwabe et al. [32] propose an architecture to monitornetwork conditions and control the streaming rate over a singleaccess network. Jurca and Frossard [21] study the problemof rate control for video streaming over a multihop network,assuming known packet loss rates and available bandwidths foreach network link. Chou and Miao [11] propose a video-awareframework to schedule video packets based on their importanceso as to maximize the video quality under given rate constraints.Zhu et al. [41] propose joint routing and rate control algorithmsfor ad-hoc wireless networks, while rate control for clients withmultiple interfaces has been studied in [6], [30], and [39]. Therehas been a wide range of methodologies summoned to addressthe resource allocation problem of video streaming: For ex-ample, Singh et al. [30] propose a solution based on stochasticcontrol of Markov decision processes, Alpcan et al. [6] proposea solution based on -optimal control, and Zhu et al. [39]propose a solution based on convex optimization.Efficient stream adaptation for scalable streams has been

studied in [7], [12], [17], [25], and [31]. Hefeeda and Hsu [17]consider the stream adaptation problem for fine-grained scal-able (FGS) video streaming from multiple senders to a singleclient; they employ a rate-distortion (R-D) function designedfor FGS streams and consider stream adaptation to maximizethe overall video quality. Amonou et al. [7] study the problemof prioritizing video packets of H.264/SVC streams; they em-pirically calculate the distortion impact of dropping each videopacket and give higher priorities to video packets with higherimpact values. Sun et al. [31] propose an R-D model for FGS

FRERIS et al.: DISTORTION-AWARE SCALABLE VIDEO STREAMING TO MULTINETWORK CLIENTS 471

TABLE ILIST OF SYMBOLS USED IN THE PAPER

streams coded by H.264/SVC, based on a generalized Gaussiandistribution source model that captures the drifting error causedby truncating video packets. Mansour et al. [25] study streamadaptation between one base station and multiple clients in asingle-hop wireless network; the clients share a given networkcapacity for receiving FGS streams from the base station.In [12], the authors propose a streaming platform to supportmultihoming, which was tested to reduce video interruptionsand achieve higher and more stable received video quality.This paper builds upon the preliminary results reported

in [13] and [18] and provides more detailed analysis andelaborate evaluation results. To the best of our knowledge, ourwork is the first that simultaneously considers the end-to-endrate control and scalable stream adaptation for multinetworkclients. Previous works either consider nonscalable videostreaming [6], [30], [39] or concentrate on scalable streamadaptation without accounting for heterogeneous and dynamicnetwork conditions [7], [17], [25], [31].

III. PROBLEM FORMULATION

A. System Architecture

Table I summarizes the symbols used in the paper. Amultinetwork scalable streaming system consists of a scalablestreaming server containing a database of scalable videos, andmultinetwork clients, each one having access to hetero-

geneous networks (cf. Fig. 1). When requested by a client, avideo stream is divided into substreams (each transmittedover a distinct network) by a video splitter that further controlsthe rate of each substream to ensure timely delivery of videopackets. For each client, the server sets up a connection overeach access network and transmits substreamover access network . Each client has a dejitter buffer and avideo assembler, which combine the received substreams intoa single scalable video stream. The video stream is then fed toa video decoder.

Access networks are heterogeneous and time-varying, soperiodic measurements of the ABR, , as well as the RTT,, need to be carried out for each access network. In our

implementation, we have opted to use a light-weight tool called[5], although our algorithms are clearly independent

of such a measuring tool. This measurement tool runs onboth server and client sides and monitors end-to-end networkconditions. We develop an algorithm to determine, on theserver side, the streaming rates of individual access networksalong with the video packets to be included in each substream,given information about the network conditions and videocharacteristics.

B. Network ModelFor a given user , we let be the sub-

stream rate over access network and bethe total streaming rate for network . For access network ,we use to denote the packet loss probability, which accountsfor losses due to packets missing their playout deadlines. Weassume that access networks are statistically independent andwrite , where is increasing inand decreasing in . While our analysis can accommodate var-ious queueing models [15] in defining , we adoptthe M/M/1 model that was shown to yield a good approxima-tion in typical streaming applications per previous measurementstudies [39], [40]. We denote the playout deadline3 by anddefine the average one-way delay by . The one-waydelay can be related to the residual bandwidth, , by

, where is a parameter estimated from past ob-servations of via linear regression [39]. Finally, we pe-riodically measure and values and compute the streamingrate in each time window, which in turns allow us to estimatevia

(1)

C. Video ModelWe consider H.264/SVC [29] video streams coded with

medium-grained quality scalability (MGS). Each stream ,, is divided into multiple Network Abstraction Layer

Units (NALUs). For user , each NALU is identifiedby frame number , , and quality layer ,

; NALU corresponds to the base layerof frame , while denote quality enhancementlayers. The H.264/SVC standard imposes decoding depen-dencies among NALUs: depends on all

, while depends on its parent frames asdetermined by the hierarchical prediction structure (cf. Fig. 2).We let4 be the parent frames of frame and useto represent the size of NALU .Let be a boolean decision variable that is equal to 1

if is sent over access network , and is 0 else. We allowfor a packet to be sent over at most one access network; this3For simplicity, we assume in the sequel that the playout deadline is the same

for all users and video packets, whence is a system parameter determined bythe service provider. The general case can be handled by considering playoutdeadlines to depend on users and video packets; in such a case, the loss proba-bility is defined separately for each user and packet via (1).4In this paper, we use bold symbols to represent vectors.


Fig. 2. Dependency among NALUs of H.264/SVC streams. Each square rep-resents an NALU belonging to an MGS layer, and each rounded rectangle rep-resents a video frame.

is because efficient link-layer error control mechanisms, suchas Forward Error Correction (FEC) and Automatic Repeat Re-quest (ARQ), are widely applied in wireless networks to re-duce packet losses [38], hence sending an NALU over multipleaccess networks does not lead to significant improvements onvideo quality, while it increases the network load.We measure video distortion using mean square error (MSE).

We let be the total distortion of frame ,where denotes the truncation distortion and denotesthe drifting distortion. Truncation distortion refers to the qualitydegradation of frame due to dropping NALUs of that frame.Let be the full-quality distortion of frame , achievedwhen all NALUs are received, and bethe additional distortion introduced by dropping NALU .In order to decode , all NALUs must havebeen decoded, thus we have

(2)

Drifting distortion refers to the distortion caused by imperfectreconstruction of parent frames used for interframe pre-diction. In principle, may include all parent frames offrame . Doing so, however, may result in excessive over-head with questionable performance gain. Therefore, in prac-tice, we can either constrain interframe predictions within indi-vidual group-of-pictures (GoPs) or heuristically choose a boundon . Following the discussion in [18] and [24], we pro-pose an affine model

(3)

where are parameters estimated from real data,and each is constrained to be nonnegative,Although our video model is designed for H.264/SVC

streams coded with MGS layers, the model is general and canwork with other types of scalable or nonscalable streams. Forexample, H.264/AVC coders can generate temporal scalablestreams using hierarchical B-frames. The proposed video modelabstracts such H.264/AVC streams as videos with a base layerand no enhancement layer, i.e., , and captures the errorpropagation due to error concealment as the drifting distortion.Such flexibility is due to the fact that nonscalable streams areessentially scalable streams with one quality layer, and thusthey can be captured by our video model.In order to optimize the overall streaming quality, we need

to specify the parameters for the model introduced above.

We have implemented least-squares parameter estimation inMATLAB, which runs offline, as a preprocessing step. We donot assume the parameter estimation is performed online, andits computational complexity is not reported throughout thispaper. Estimating the parameters offline limits the applicationscope mostly to video-on-demand services, in which the pa-rameters can be computed offline and stored as metadata. Suchservices are fairly popular nowadays, e.g., YouTube, Hulu, andNetflix streaming.

D. Optimization ProblemWe denote, by some abuse of notation, the expected

distortion, after accounting for random packet losses, ofthe th frame of user by and define the vectors

. We formulatethe multinetwork scalable video streaming problem as one offinding the values to minimize a convex cost function

, which is nondecreasing in eachargument. One special case of interest is ,where each is convex and nondecreasing in eachargument. We can provide service differentiation amongusers and frames by considering different cost functions, e.g.,

. We can also addressfairness among users, e.g., weighted max-min fairness bysetting .For user , let be the frame rate measured in frames per

second (fps). Then, the average transport stream rate for net-work is given by

(4)

Using the network model (1), the delivery probability ofNALU denoted by is given by

(5)

The expected truncation distortion is still given by (2), and theexpected drifting distortion by (3) if we further assume thatpacket losses are statistically independent.Since NALUs have different sizes, some NALUs may

comprise multiple, say , packets. Typically, is afunction of NALU size ; for example, for a path with max-imum payload length , the streaming server may divide NALU

into packets. We can handle this case byletting be 1, if the th packet of NALU is sentover access network , and 0 else. Then, we may replace (5) and(2) with

(6)

(7)

respectively. In the sequel, we assume fornotational simplicity; extending the optimization program


and the proposed algorithms to handle this general case isstraightforward.The joint rate control and stream adaptation problem, fea-

turing optimization over frames for client , is given bythe integer program

(8a)

s.t. (8b)

(8c)

(8d)

(8e)

(8f)

(8g)

(8h)

(8i)

We consider a recurring scheduling window of duration ,which implies that we have to solve the above optimizationproblem once every seconds. In our implementation, we setas constant and consider variable numbers of packets to be

scheduled within each window, i.e., pick . Usingsuch an approach, the system can adapt to dynamic changes suchas variable network conditions, or new users arriving/leaving,by solving different instances of the optimization problem foreach scheduling window. Finally, note that rate control is per-formed through (8c); this is a form of proactive congestion con-trol, in the sense that it seeks to avoid causing network con-gestion, as opposed to the responsive nature of the widely usedTransmission Control Protocol (TCP).We also consider relaxing to take real values in the

interval . This is a soft decision problem, which can be im-plemented via randomized packet scheduling. Let us define afamily of independent Bernoulli random variables ,for, where takes the value 1 with probability ,

and the value 0 with probability . Given a realiza-tion of the Bernoulli random variables, NALU is sentover network only when ; this event has proba-bility and is independent from the scheduling of otherpackets. The expected truncation distortion is given by (2) ifwe assume that packet losses of access networks are statisti-cally independent from the decision variables . This as-sumption is a gross approximation that can be made fairly accu-rate by considering a two-timescale separation approach: Sup-pose that the optimization window size is large enough for thestochastic process (such as a Markov chain) characterizing thenetwork losses to converge to the stationary distribution. Then,the approximation error in (8c) is negligible in both theory andpractice [20].

E. Properties of the Optimization ProblemThe actual decision variables of the optimization problem (8)

are ; these are only constrained to be either bi-nary (hard decisionsdeterministic scheduling) or lie in

(soft decisionsrandomized scheduling). The distor-tion is a function of the decision variables , how-ever its analytical expression is too complicated to write inclosed form. Instead, we have introduced auxiliary variables

and imposed the equations thatare related with as constraints in the optimization problem (8).This is a technique in optimization usually referred to as up-lifting [8], in which the decision space is increased to yield asimpler objective function accompanied by a set of constraints.The objective function of (8) is increasing in , while

it is decreasing in for each . It is increasing inand for each . Based on these properties, we

can replace the equality constraints in (8c)(8e) with , ,inequality constraints, respectively. This yields an equivalentformulation with no nonlinear equality constraints. The abovemonotonicity properties guarantee that an optimal integer solu-tion for satisfies the property that is sent over somenetwork, only if all are sent over some networkas well. The randomized optimization problem, in which (8i) isreplaced by , is not convex due to multinomialterms in (8d) and (8e). The problem can neither be convertedinto an equivalent convex program by means of exponentialtransformations of the form , nor can it be rendered inthe format of geometric programming [9]. In Section IV-C, wepresent convex approximations to the randomized packet sched-uling problem.

IV. OPTIMIZATION ALGORITHMSIn this section, we propose several deterministic and ran-

domized packet scheduling algorithms. In the sequel, we let, , for all , just for the sake

of notational simplicity but without any loss of generality.

A. Exhaustive SearchThe integer program (8) can be solved by means of exhaus-

tive search; the complexity of a naive exhaustive search is, which can be reduced to in

the light of the fact that each packet is sent over at most onenetwork [c.f (8h)]. If we further exploit the optimality propertyof Section III-E, that is sent over some network, onlyif all are sent over some network, then thismeans that for fixed there are possiblevalues for at optimality. Therefore, the complexity ofan exhaustive search can be further reduced to ,for , or when .

B. Heuristic AlgorithmsWe present heuristic algorithms for deterministic packet

scheduling. The algorithms do not explicitly address servicedifferentiation, i.e., we consider .Simple Rate-Distortion Optimization: The Simple Rate-Dis-

tortion Optimization (SRDO) algorithm takes a maximalallowed packet loss probability as the input and sortsNALUs in descending order of . It sequentially assigns


Fig. 3. Simple Rate Distortion Optimization algorithm (SRDO).

Fig. 4. Progressive Rate-Distortion Optimization algorithm (PRDO).

NALUs to the access network with the smallest until allaccess networks are fully loaded, i.e., right before the smallest

exceeds . SRDO has a worst-case complexity of. The pseudocode is pre-

sented in Fig. 3.Progressive Rate-Distortion Optimization: The Progressive

Rate-Distortion Optimization (PRDO) algorithm considersthe net distortion gain of assigning NALU over accessnetwork , namely , based on the distortion model (cf.Section III-C). Following the video prediction structure, PRDOsequentially schedules the immediately decodable NALU

with the largest nonnegative value to the corre-sponding access network. The algorithm stops when all packetshave been scheduled or when all unscheduled NALUs havenonpositive net distortion gains. PRDO is a greedy algorithmthat relies on evaluating function ; a single function eval-uation has complexity . In the worst case, PRDOmakes such evaluations, so its complexity is

. The pseudocode is presented in Fig. 4.Hybrid Rate-Distortion Optimization: The Hybrid Rate-Dis-

tortion Optimization (HRDO) algorithm uses SRDO to boot-strap a solution, which is consequently used as the initial valuefor PRDO. It has been shown to yield a good tradeoff of perfor-mance versus runtime in simulations [18].

C. Convex ApproximationsIn this section, we derive approximate convex programs

for the randomized packet scheduling problem. The goal

is to approximate the nonconvex constraint set of (8) bya convex superset, by means of convex approximations tothe multilinear product terms in (8d) and (8e). Even thougha solution to an approximate problem might be infeasiblefor the original when considering the augmented space

, this is not an issue becausewe are only interested in obtaining values for the actual deci-sion variables, i.e., the transmission probabilities ;these are only constrained to lie in . We also note here thatour approach can be plainly used to handle the case that someNALUs comprise multiple packets, i.e., ,since (6) and (7) also feature multilinear product terms.In the next lemma, we present a convex programming formu-

lation that approximates multilinear functions in (8d) and (8e)in a term-by-term fashion.Lemma 1 [Term-by-Term Convex Approximation (TTC)]:

The optimization problem

(9a)

s.t. (9b)

(9c)

(9d)

(9e)

(9f)

(9g)

(9h)

(9i)

is a convex program whose optimal value is an underestimateof the optimal value of (8). It consists of deci-sion variables and constraints. Itcan be written as an equivalent smooth convex program by sub-stituting the in (9d) and (9e) with inequality constraints. Ifwe assume that is continuous on , then the convexprogram admits an optimal solution and has the strong dualityproperty.5

Proof: The concave envelope ofon is given by[28]. Applying this to each multinomial

term in (8d) and (8e), we get, by exploiting the monotonicityproperties of Section III-E and the fact that the minimum ofaffine functions is a concave function, convex program (9).The program has a nonempty and compact set of optimalsolutions since the domain of the decision variables isthe compact unit hypercube and since allinequality constraints along with the objective function involvecontinuous functions. The convex program (9) has the strong5Strong duality is important for the performance of numerical methods [9].


duality property as well as a nonempty and bounded set of dualoptimal solutions because it satisfies Slater condition [9]; thereexists a feasible solution for which all inequality constraintsare strictly satisfied. For example, let sufficientlysmall and consider ,

.Remark 1 (Practical Considerations): The approximation

error in TTC may be nonnegligible. The approximation in(9d) does not sufficiently capture the impact of packet losses,which is especially true when the loss probability is small (say5%10%), which further implies that inmost cases, hence . In addition, thegap in approximating the probability that is not receivedwith the minimum of the probabilities that arenot received [c.f (9e)] might not be negligible either.We present another method of approximating the nonconvex

multilinear inequalities (8d) and (8e) by means of their convexenvelopes. This yields the optimal convex approximation of thenonconvex constraint set of (8), but comes at the cost of highcomputational complexity.Lemma 2 [Multilinear Convex Approximation (MC)]: The

optimization problem in (10), shown at the bottom of the page,is a convex program whose optimal value is an underestimate of

the optimal value of (8). If we assume that is continuouson , then the convex program admits an optimal solutionand has the strong duality property.

Proof: Consider a multilinear function on, i.e., a function that is linear in each

argument alone. The convex envelope of is given by [28]

,, where is defined as:

, and. We use this to calculate the concave enve-

lope of the multilinear functions on the right of (8d) and (8e)to obtain convex inequality constraints. The rest of the prooffollows along the same lines as in Lemma 1.Remark 2 [Hybrid Convex Approximation (HC)]: We can re-

place (9d) in TTCwith (10d) for a balance between performanceand computational complexity; we call the resulting problemHC. Note that(10d) is always a better approximation of (8d), asit is the tightest convex approximation of the multilinear func-tion. Therefore, HC is expected to outperform TTC, and weobserved in our experiments that the improvement is signifi-cant in all cases, not only for low-loaded networks, as is imme-diate from Remark 1. However, there is no substantial increase

(10a)

s.t. (10b)

(10c)

(10d)

(10e)

(10f)

(10g)

(10h)

(10i)


in runtime for the case when the number of available networksis low, e.g., . For this reason, we report results using HCin the remainder of this paper. On the contrary, the performancegain of MC over HC was measured to be negligible, particularlyin the face of a runtime that was prohibitive for real-time appli-cations, even for low values of , e.g., .Remark 3 (Computational Complexity): The TTC optimiza-

tion program contains a polynomial number of constraints in, , , . MC requires computing the convex envelope in

(10d) and (10e). The convex envelopes in (10d) can be com-puted offline with tests. This is because the calculationof the convex envelope of on

does not depend on the problem parameters;that is why does not depend on in (10d). However,the convex envelopes in (10e) depend on the problem parame-ters; the computation takes tests, which might takea prohibitively long time. HC contains an exponential numberof constraints in . For fixed , and given , the con-cave envelope in (10d) is given by the minimum of a fixednumber, say of affine functions of (for examplefor , ), while the total number of constraints ispolynomial in . Therefore, for small values of , wepropose using HC for a good tradeoff between performance andruntime.Remark 4: We can improve the convex approxima-

tion in (9e) by replacing parameter with ,where are chosen to satisfy

for a near-optimal solution computedby some heuristic algorithm, like the ones in Section IV-B.Remark 5 (Multiple-Server Extension): We can generalize

the optimization programs to handle the scenario where eachvideo sequence may be streamed from multiple servers, say

, with rate constraints, to the clients . Letbe 1 if the NALU is streamed from server

to client over network , and 0 otherwise. We can consider(8)(10) with the additional constraints

(11)

(12)

where denotes the streaming capacity of server .In this case, server-side scheduling is, by nature, centralized;

different servers need to exchange information to avoid sendingmultiple copies of the same packet.We have implemented the packet scheduling algorithms

based on the aforementioned convex approximations inMATLAB using CVX [2], which is a numerical solver forconvex optimization. However, as we discuss in Remark 2, theHC algorithm gives us a good tradeoff between complexity andperformance. Thus, we only report the performance of HC inSection V.

V. EVALUATION

A. SetupWe use Abing [5] to periodically measure ABR and RTT

values between Deutsche Telekom Laboratories (Berlin,

Fig. 5. Rate increase for different numbers of MGS layers.

Fig. 6. R-D curves of the scalable video streams.

Germany) and Stanford University (Stanford, CA). We collectthe network traces on weekdays with dozens of hosts on eachnetwork generating background traffic. At Deutsche TelekomLaboratories, Abing was run over three access networks: Eth-ernet, 802.11b, and 802.11g. Parts of the network traces wereused in [30] and [39], and further details can be found therein.We consider four 10-s 4CIF (704x576) video sequences:City,

Soccer, Crew, and Harbour, encoded as H.264/SVC scalablestreams using H.264/SVC baseline profile of JSVM ReferenceSoftware. We tested different numbers of MGS layers andfound that does not affect coding efficiency substantially.Fig. 5 illustrates that only results in 5%7.5% higherbit rate than . This shows that additional MGS layers donot lead to severe coding inefficiency. Therefore, each video se-quence is encoded into a scalable stream with eight MGS layersfor higher flexibility. To illustrate the video characteristics ofindividual videos, we plot the R-D curves in Fig. 6.We estimate the video model parameters by extracting and

decoding 32 random substreams from each scalable stream andmeasuring the video quality. Knowing which video packetswere successfully delivered as well as truncation distortion anddrifting distortion, we estimate the model parameters of thevideo model using standard least-squares fitting in MATLAB.To evaluate the accuracy of the video model, we randomlyextracted another 32 substreams from each video stream, com-puted the empirical per-frame video quality, and compared it tothe video quality estimated by the video model. Fig. 7 showsthe actual and estimated video quality of Soccer and Crew: Theproposed video model is quite accurate, and the average abso-lute errors for City, Soccer, Crew, and Harbour were measuredto be 2.82%, 1.38%, 0.74%, and 1.65%, respectively.For comparison, we also encode the same video sequences

into nonscalable streams using the H.264/AVC baseline pro-file of JM Reference Software. We configure the H.264/AVC


Fig. 7. Proposed video model closely followsmeasured quality. Sample resultsfrom (a) Soccer and (b) Crew.

encoder as close to the H.264/SVC configuration as possible,e.g., we use the prediction structure of IPPPPPPP in both cases.Schwarz et al. [29] claim that, compared to nonscalable streams,10%50% bit rate increases of scalable streams are acceptable.To be conservative, we use rate control mechanism to generateH.264/AVC streams with 20% lower bit rates than H.264/SVCstreams; we found that the H.264/AVC streams still achieve0.11 dB better quality than H.264/SVC streams.We implemented a multinetwork streaming server in

NS-2 [27], which supports the SRDO, PRDO, and HC al-gorithms, implemented as MATLAB subroutines. The HCalgorithm uses CVX [2] to numerically solve the convexprogram given in Remark 2. We report runtime values cor-responding to a 2.8-GHz PC with MATLAB R2010a. Forcomparison, we also implemented a multinetwork DCCP [22]streaming server based on an open-source DCCP implemen-tation [26] that supports two standard rate control algorithms:TCP-like and TCP-Friendly Rate Control (TFRC). The DCCPstreaming server sets up a connection over each access net-work and assigns NALUs to each connection from lower- tohigher-quality layers until reaching the rate limit computed bythe rate control algorithms. The DCCP streaming servers withTCP-like and TFRC rate control algorithms are referred to asDCCP-TCP and DCCP-TFRC, respectively.We simulate multinetwork video streaming sessions using

the four videos with random start times in the network traces.We repeat the 10-s video clips throughout the simulations. Weinject background traffic over each network at a rate between20%90% of its available bit rate. We chose , ,

s, s, and . The maximum UDPpacket size is set to 1000 B. If not otherwise specified, we re-port results with 40% background traffic, using average distor-tion as the cost function. We conduct simulations with a singleuser and compare the performance of the proposed al-gorithms and the rate control algorithms defined in DCCP stan-dard. We also run the HC algorithm for three users ofdifferent videos, which is illustrated in Fig. 8. For each setup,we test the algorithms 300 times and consider five performancemetrics: video quality in PSNR, streaming rate, packet deliverydelay, delivery ratio, and runtime.

B. Simulation ResultsNonscalable Versus Scalable Streams: Streaming nonscal-

able videos over a bandwidth-limited channel may lead to un-decodable frames [23], while scalable video streaming systemshave the choice to at least stream the base layer and attain basic

Fig. 8. Simulation setup for multiple user scenarios.

Fig. 9. Delivery ratios with nonscalable and scalable streams: (a) DCCP-TCPand (b) DCCP-TFRC. Sample results from City sequence.

Fig. 10. Video quality achieved by different numbers of access networks:(a) DCCP-TCP and (b) DCCP-TFRC. Sample results from City sequence.

quality. To quantify the benefits of streaming scalable videos,we use delivery ratio as the performance metric, which is de-fined as the fraction of timely delivered frames for nonscalablevideos and the fraction of frames with base layers delivered ontime for scalable videos.We configure a DCCP streaming serverto transmit City (both nonscalable and scalable) over one, two,and three access networks. We plot the resulting delivery ratiowith 95% confidence intervals in Fig. 9. This figure illustratesthat streaming scalable videos results in higher delivery ratio,therefore fewer rebuffering instances and overall a better userexperience. Hence, we only consider scalable streams in the restof this section.Benefits of Multihoming: We instruct a DCCP streaming

server to transmit City over one, two, and three access networksand compute the video quality achieved under 40% back-ground traffic. We plot sample results for a 60-s period usingDCCP-TCP and DCCP-TFRC in Fig. 10; notice that multi-homing can significantly increase video quality and reduce thenumber of quality fluctuations.Video Quality:We compare the video quality achieved by the

proposed algorithms against the DCCP rate control algorithmsunder 40% background traffic. In Fig. 11(a), we plot the videoquality achieved using each algorithm for a 60-s sample period.We observe that both DCCP-TCP and DCCP-TFRC sufferfrom sudden quality drops, unlike the proposed algorithms thatadditionally achieve higher video streaming quality. We report


Fig. 11. Video quality achieved by different algorithms: (a) 60-s sample periodfrom Crew and (b) overall results.

Fig. 12. Streaming rate achieved by the different algorithms: (a) 60-s sampleresults from City and (b) overall results.

the aggregate video quality for different video sequences inFig. 11(b). The proposed algorithms outperform the DCCP ratecontrol algorithms by at least 10 dB in terms of video quality.Streaming Rate: While the proposed algorithms achieve

better video quality, we need to make sure that they do not satu-rate the network resources. We plot the streaming rates achievedby different algorithms in Fig. 12. Fig. 12(a) shows a sampletime period, which reveals that the DCCP rate control algo-rithms (see upper subfigure) result in higher rate fluctuationswhile the proposed algorithms lead to smoother streaming rates(see lower subfigure). This can be attributed to the proactiverate control employed by the proposed algorithms, compared tothe responsive rate control used by DCCP. Fig. 12(b) plots theaverage streaming rates for all videos. This figure indicates thatthe proposed algorithms are conservative in terms of networkresources; they lead to streaming rates comparable to (if notlower than) the DCCP rate control algorithms.Packet Delivery Delay: We calculate the average packet de-

livery delay for the different algorithms. Fig. 13 reveals that, forall videos, DCCP-TCP and DCCP-TFRC lead to average delayof 1.7 and 2.5 s, respectively, while the proposed algorithms re-sult in less than 0.2 s delay, over 90% reduction on average.This shows that schedules produced by the proposed algorithmsdeliver more packets on time, which in turn justifies the bettervideo quality compared to DCCP.Adaptation to Network Dynamics: The proposed algorithms

employ a short scheduling window (in the order of seconds) toadjust to network dynamics. To show the effectiveness of thisapproach, we conduct simulations in which the client gradu-ally gains access to more access networks. More specifically,the video streaming session starts with a single access network,and two additional access networks are activated after 15 and30 s of simulation time, respectively. We plot sample streaming

Fig. 13. Packet delivery delay incurred by different algorithms.

Fig. 14. Streaming rate of individual networks: Network 1 is available for theentire simulation run, while Networks 2 and 3 become available only after 15and 20 s, respectively. Sample results from (a) SRDO and (b) HC.

rates over individual networks in Fig. 14. This figure showsthat our algorithms can quickly adapt to network dynamics bycapitalizing the new access networks shortly after they becomeavailable. We note that short scheduling windows also help toadopt to the access network outages by rescheduling the moreimportant video packets over more reliable access networks.Furthermore, the scheduling window size is a control knob ofthe tradeoff between responsiveness and flexibility, in the sensethat shorter scheduling windows result in faster recovery fromnetwork outages but limit the room for redistributing networkresources among frames in the same scheduling window.Performance Versus Computational Complexity: We com-

pare the performance of the proposed algorithms under differentbackground traffic load, from 20% to 90%. Fig. 15 presentsthe achieved video quality from Harbour and Crew. This figureshows that the HC algorithm outperforms the PRDO algorithm,which in turn outperforms the SRDO algorithm. We also plotthe quality improvement resulted by HC over SRDO and PRDOin Fig. 16; the HC algorithm almost always leads to qualityimprovement, which is more transparent in highly loaded net-works. Specifically, among all videos, the maximum, mean, andminimum quality improvements over SRDO are 7.36, 4.33, and1.19 dB. The maximum, mean, and minimum quality improve-ments over PRDO are 4.71, 1.84, and 0.33 dB.Fig. 17 presents the runtime of the proposed algorithms for

Harbour andCrew; the HC algorithm has an up to 10-fold lowerruntime as compared to PRDO. SRDO runs fast, less than 200ms on average, but it results in lower video quality as illustratedin Figs. 15 and 16. Therefore, we propose using the HC algo-rithm for good performance as well as reasonable runtime. Notethat the runtime of HC is constant independent of backgroundtraffic since it is a convex program that takes the same time to


Fig. 15. Video quality under different background traffic loads. Sample resultsfrom (a) Harbour and (b) Crew.

Fig. 16. Video quality improvement achieved by HC over (a) SRDO and(b) PRDO under different background traffic loads.

Fig. 17. Runtime under different background traffic loads. Sample results from(a) Harbour and (b) Crew.

solve numerically irrespective of background traffic. The run-time of PRDO decreases substantially with background trafficsince there are much fewer packets that can be sent before ca-pacity is reached. The same is true for SRDO, but it is not asapparent because SRDO does not perform the time-costly func-tion evaluation (which PRDO does).Lastly, although our proposed proactive algorithms outper-

form responsive DCCP-TCP and DCCP-TFRC, we need topoint out that DCCP algorithms still have several advantagesover the proposed algorithms. First, DCCP algorithms aresimple and easy to deploy. Second, DCCP algorithms havevery low computational complexity. Third, DCCP works in theconsidered system architecture (Fig. 1) as well as others, whileour proposed algorithms only run on streaming servers. Wewill discuss the last limitation more in Section VI.Multiple Clients and Service Differentiation: We use the

HC algorithm to stream different videos to three clientsunder 40% background traffic load. Three cost functions

, , and are considered, where

Fig. 18. Sample video quality with cost function .

TABLE IIAVERAGE VIDEO QUALITY AND STREAMING RATE WITH DIFFERENT COST

FUNCTIONS

6. We plot the videoquality of individual clients with in Fig. 18, whichshows that the HC algorithm achieves service differentiation:Client 3 (Harbour) has the lowest video quality among allclients.Table II presents the overall video quality under different

weights: Service differentiation can be achieved by a proper se-lection of the cost functions. For example, with , thevideo quality of client 3 (Harbour) is 1015 dB lower than thatof client 1 (Crew), while the gap is reduced to 3 dBwith .Table II gives the average streaming rate under different costfunctions. For , client 1 (Crew) achieves higher videoquality than other clients (cf. Table II), despite receiving lowerrate (cf. Table II); this is because Crew has a steeper R-D curve(cf. Fig. 6).

VI. LIMITATIONS AND FUTURE WORK

This paper considers multihomed, multiple-client videostreaming from a servers perspective (Fig. 1). This results inserver-driven adaptation solutions, which may incur too muchoverhead on the server for many clients. Therefore, each servermight only be able to serve a small number of clients. Whilethe streaming service providers may deploy multiple streamingservers in a server farm, as well as exploit increased computa-tional power via grid or cloud computing, the centralized natureof our solution could still render the proposed algorithms lessefficient in such deployments. For example, probing trafficfrom multiple servers to infer available bit rate and round-trip6The cost functions are application-dependent. A meticulous design of the

cost function to meet some desirable specifications is out of the scope of thispaper.


time could interfere with each other. To tackle this, we cancontrol the number of decision variables by simplifying thescalable stream structures, hence trading streaming optimalityfor shorter running time for the case of many users. Trans-forming our current architecture toward client-driven solutionsis one of our future tasks. Techniques such as Lagrangian de-composition could be used to develop distributed algorithms formore scalable solutions. The resulting distributed algorithmswill be more suitable to client-driven HTTP streaming, such as3GPP/MPEG DASH, which is getting more and more popularnowadays.

VII. CONCLUSION

In this paper, we have addressed various usage scenarios ofvideo streaming from a server to multinetwork clients over het-erogeneous access networks. More precisely, we have formallyabstracted the problem of joint rate control and stream adapta-tion as an optimization problem of minimizing the expected dis-tortion of the received videos subject to constraints based on net-work conditions. We have formulated this problem as an integerprogram for joint rate control and stream adaptation in order todetermine, for each client: 1) the streaming rates over individualaccess networks; 2) the video packets to be transmitted; and 3)the access network each transmitted video packet is sent over, soas to minimize a cost function of the expected distortion at thereceiver side. We have proposed using different cost functionsto account for service differentiation and fairness among users.We have proposed two heuristic algorithms for packet sched-uling, namely SRDO and PRDO. In addition, we have derivedconvex programming approximations to the randomized packetscheduling problem and have studied the tradeoff between per-formance and runtime; one of our randomized algorithms (TTC)has a better runtime at the cost of lower performance, while theother one (MC) has better performance at the cost of exponen-tial complexity. We have proposed a hybrid algorithm (HC) thatyields good performance for a low number of access networkswhile being suitable for real-time applications.We have conducted extensive simulations to compare the per-

formance of HC against SRDO, PRDO, and the rate controlalgorithms defined in the DCCP standard. The simulation re-sults have shown that the HC algorithm: 1) outperforms the ratecontrol algorithms in the DCCP standard by about 1015 dB invideo quality; 2) reduces average delivery delay by over 90%compared to DCCP; 3) results in an average quality improve-ment of 4.33 dB versus SRDO, and 1.84 dB versus PRDO, underdifferent background traffic loads; 4) runs efficiently, up to sixtimes faster than PRDO; and 5) indeed provides service differ-entiation among users.

REFERENCES[1] M. Megna, AT&T faces 5,000 percent surge in traffic, Oct.

2009 [Online]. Available: http://www.internetnews.com/mobility/ar-ticle.php/3843001

[2] CVX: Matlab software for disciplined convex programming, 2009[Online]. Available: http://www.stanford.edu/~boyd/cvx/

[3] K. Fitchard, T-Mobiles growth focusing on 3G, 2009 [Online].Available: http://connectedplanetonline.com/wireless/news/t-mo-bile-3g-growth-0130

[4] Cisco Systems, San Jose, CA, CiscoVisual Networking Index forecastWeb site, 2010 [Online]. Available: http://www.cisco.com/go/vni

[5] Stanford University, Stanford, CA, Abing project page, 2004 [On-line]. Available: http://www-iepm.slac.stanford.edu/tools/abing/

[6] T. Alpcan, J. Singh, and T. Basar, Robust rate control for heteroge-neous network access in multihomed environments, IEEE Trans. Mo-bile Comput., vol. 8, no. 1, pp. 4151, Jan. 2009.

[7] I. Amonou, N. Cammas, S. Kervadec, and S. Pateux, Optimized rate-distortion extraction with quality layers in the scalable extension ofH.264/AVC, IEEE Trans. Circuits Syst. Video Technol., vol. 17, no. 9,pp. 11861193, Sep. 2007.

[8] D. Bertsekas, Convex Optimization Theory. Belmont, MA: AthenaScientific, 2009.

[9] S. Boyd and L. Vandenberghe, Convex Optimization, 1st ed. Cam-bridge, U.K.: Cambridge Univ. Press, 2004.

[10] J. Chakareski and B. Girod, Rate-distortion optimized packet sched-uling and routing for media streaming with path diversity, in Proc.DCC, Snowbird, UT, Mar. 2003, pp. 203212.

[11] P. Chou and Z. Miao, Rate-distortion optimized streaming of pack-etized media, IEEE Trans.n Multimedia, vol. 8, no. 2, pp. 390404,Apr. 2006.

[12] K. Evensen, T. Kupka, D. Kaspar, P. Halvorsen, and C. Griwodz,Quality-adaptive scheduling for live streaming over multiple accessnetworks, in Proc. ACM NOSSDAV, Amsterdam, The Netherlands,Jun. 2010, pp. 2126.

[13] N. Freris, C. Hsu, X. Zhu, and J. Singh, Resource allocation for mul-tihomed scalable video streaming to multiple clients, in Proc. IEEEISM, Taichung, Taiwan, Dec. 2010, pp. 916.

[14] P. Fuxjager, H. Fischer, I. Gojmerac, and P. Reichl, Radio resourceallocation in urban femto-WiFi convergence scenarios, in Proc.Euro-NF NGI, Paris, France, Jun. 2010, pp. 18.

[15] D. Gross, J. Shortle, J. Thompson, and C. Harris, Fundamentals ofQueueing Theory, 4th ed. Hoboken, NJ: Wiley-Interscience, 2008.

[16] Y. He and R. Yuan, A novel scheduled power saving mechanism for802.11 wireless lans, IEEE Trans. Mobile Comput., vol. 8, no. 10, pp.13681383, Oct. 2009.

[17] M. Hefeeda and C. Hsu, Rate-distortion optimized streaming offine-grained scalable video sequences, Trans. Multimedia Comput.,Commun.s, Appl., vol. 4, no. 1, pp. 2:12:28, Jan. 2008.

[18] C. Hsu, N. Freris, J. Singh, and X. Zhu, Rate control and stream adap-tation for scalable video streaming over multiple access networks, inProc. IEEE PV, Hong Kong, Dec. 2010, pp. 3340.

[19] C. Hsu and M. Hefeeda, Broadcasting video streams encodedwith arbitrary bit rates in energy-constrained mobile TV networks,IEEE/ACM Trans. Netw., vol. 18, no. 3, pp. 681694, Jun. 2010.

[20] L. Jiang and J. Walrand, A distributed CSMA algorithm forthroughput and utility maximization in wireless networks, IEEE/ACMTrans. Netw., vol. 18, no. 3, pp. 960972, Jun. 2010.

[21] D. Jurca and P. Frossard, Media-specific rate allocation in heteroge-neous wireless networks, in Proc. IEEE PV, Hangzhou, China, May2006, pp. 713726.

[22] E. Kohler, M. Handley, and S. Floyd, Datagram congestion controlprotocol (DCCP), RFC 4340, 2006.

[23] W. Li, Overview of fine granularity scalability in MPEG-4 video stan-dard, IEEE Trans. Circuits Syst. Video Technol., vol. 11, no. 3, pp.301317, Mar. 2001.

[24] Y. Liang, J. Apostolopoulos, and B. Girod, Analysis of packet loss forcompressed video: Effect of burst losses and correlation between errorframes, IEEE Trans. Circuits Syst. Video Technol., vol. 18, no. 7, pp.861874, Jul. 2008.

[25] H. Mansour, V. Krishnamurthy, and P. Nasiopoulos, Channel awaremultiuser scalable video streaming over lossy under-provisioned chan-nels: Modeling and analysis, IEEE Trans. Multimedia, vol. 10, no. 7,pp. 13661381, Nov. 2008.

[26] N. Mattsson, A DCCP module for NS-2, Masters thesis, Dept.Comput. Sci. Elect. Eng., Lulea Tekniska University, Lulea, Sweden,2004.

[27] The network simulator, 2012 [Online]. Available: http://www.isi.edu/nsnam/ns/

[28] A. D. Rikun, A convex envelope formula for multilinear functions,J. Global Optimiz., vol. 10, no. 4, pp. 425437, 1997.

[29] H. Schwarz, D. Marpe, and T. Wiegand, Overview of the scalablevideo coding extension of the H.264/AVC standard, IEEE Trans. Cir-cuits Syst. Video Technol., vol. 17, no. 9, pp. 11031120, Sep. 2007.

[30] J. Singh, T. Alpcan, P. Agrawal, and V. Sharma, An optimal flow as-signment framework for heterogeneous network access, inProc. IEEEWoWMoM, Helsinki, Finland, Jun. 2007, pp. 112.


[31] J. Sun, W. Gao, D. Zhao, and W. Li, On rate-distortion modeling andextraction of H.264/SVC fine-granular scalable video, IEEE Trans.Circuits Syst. Video Technol., vol. 19, no. 3, pp. 323336, Mar. 2009.

[32] A. Szwabe, A. Schorr, F. Hauck, and A. Kassler, Dynamic multi-media stream adaptation and rate control for heterogeneous networks,in Proc. IEEE PV, Hangzhou, China, May 2006, pp. 6369.

[33] E. Tan, L. Guo, S. Chen, and X. Zhang, PSM-throttling: Minimizingenergy consumption for bulk data communications in WLANs, inProc. IEEE ICNP, Beijing, China, Oct. 2007, pp. 123132.

[34] J. Wang, M. Ghosh, and K. Challapali, Emerging cognitive radio ap-plications: A survey, IEEE Commun. Mag., vol. 49, no. 3, pp. 7481,Mar. 2011.

[35] M. Wien, H. Schwarz, and T. Oelbaum, Performance analysis ofSVC, IEEE Trans. Circuits Syst. Video Technol., vol. 17, no. 9, pp.11941203, Sep. 2007.

[36] J. Xin, C. Lin, and M. Sun, Digital video transcoding, Proc. IEEE,vol. 93, no. 1, pp. 8497, Jan. 2005.

[37] X. Xing, S. Mishra, and X. Liu, ARBOR: Hang together rather thanhang separately in 802.11 WiFi networks, in Proc. IEEE INFOCOM,San Diego, CA, Mar. 2010, pp. 13521360.

[38] Q. Zhang, W. Zhu, and Y. Zhang, End-to-end QoS for video deliveryover wireless Internet, Proc. IEEE, vol. 93, no. 1, pp. 123134, Jan.2005.

[39] X. Zhu, P. Agrawal, J. Singh, T. Alpcan, and B. Girod, Distributed rateallocation policies for multihomed video streaming over heterogeneousaccess networks, IEEE Trans. Multimedia, vol. 11, no. 4, pp. 752764,Jun. 2009.

[40] X. Zhu, E. Setton, and B. Girod, Congestion-distortion optimizedvideo transmission over ad hoc networks, Signal Process., ImageCommun., vol. 20, no. 8, pp. 773783, Sep. 2005.

[41] X. Zhu, J. Singh, and B. Girod, Joint routing and rate allocation formultiple video streams in ad-hoc wireless networks, in Proc. IEEEPV, Hangzhou, China, May 2006, pp. 727736.

Nikolaos M. Freris (M05) received the Diplomain electrical and computer engineering from theNational Technical University of Athens, Athens,Greece, in 2005, and the M.S. degree in elec-trical and computer engineering, M.S. degree inmathematics, and Ph.D. degree in electrical andcomputer engineering from the University of Illinoisat UrbanaChampaign in 2007, 2008, and 2010,respectively.Since 2010, he has been working as a Researcher

with IBM ResearchZrich, Zurich, Switzerland.His research interests lie in wireless and sensor networks as well as data miningwith provable guarantees.Dr. Freris is a member of SIAM and the Technical Chamber of Greece.

Cheng-Hsin Hsu (S09M10) received the B.Sc.degree in mathematics and M.Sc. degree in computerscience and information engineering from NationalChung-Cheng University, Taiwan, in 1996 and2000, respectively, the M.Eng. degree in electricaland computer engineering from the University ofMaryland, College Park, in 2003 and the Ph.D.degree in computing science from Simon FraserUniversity, Burnaby, BC, Canada, in 2009.He is an Assistant Professor with National Tsing

Hua University, Hsin Chu, Taiwan. His research in-terests are in the area of multimedia networking and distributed systems.Dr. Hsu is a member of the Association for Computing Machinery (ACM).

Jatinder Pal Singh (M05) received the B.S. degreefrom the Indian Institute of Technology, Delhi, India,in 2000, and the M.S. and Ph.D. degrees from Stan-ford University, Stanford, CA, in 2002 and 2005, re-spectively, all in electrical engineering.He is the Director of Mobile Innovation Strategy

with the Palo Alto Research Center and Con-sulting Associate Professor with the Department ofElectrical Engineering, Stanford University. He waspreviously Vice President of Research with DeutscheTelekom, Los Altos, CA, one of the worlds largest

ISPs and parent company of T-Mobile.Dr. Singh graduated at the top of his class with the Institute Silver Medal at

the Indian Institute of Technology, Delhi, and was awarded a Stanford GraduateFellowship and Deutsche Telekom Fellowship.

Xiaoqing Zhu (M09) received the B.Eng. degreein electronics engineering from Tsinghua University,Beijing, China, in 2001, and the M.S. and Ph.D. de-grees in electrical engineering from Stanford Univer-sity, Stanford, CA, in 2002 and 2009, respectively.She is currently a member of the Advanced Ar-

chitecture and Research Group, Cisco Systems, Inc.,San Jose, CA. She interned with the IBM AlmadenResearch Center, San Jose, CA, in 2003, and wasat Sharp Labs of America, Camas, WA, during thesummer of 2006. Her research interests lie at the

intersection of multimedia signal processing, wireless communications, andnetworking.Dr. Zhu has served as a reviewer for many journals and magazines,

including the IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS,IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, IEEE TRANSACTIONSON MULTIMEDIA, IEEE Communications Magazine, and IEEE Network. Shehas also helped organize various conferences and workshops, such as IEEEGLOBECOM, IEEE International Conference on Computing, Networkingand Communication (ICNC), and SPIE Visual Communications and ImageProcessing (VCIP). She served as Guest Editor for the IEEE Technical Com-mittee on Multimedia Communications (MMTC) E-Letter, IEEE JOURNALON SELECTED AREAS IN COMMUNICATIONS, and IEEE TRANSACTIONS ONMULTIMEDIA. She was awarded the Stanford Graduate Fellowship from 2001to 2005. She was the recipient of the Best Student Paper Award in ACMMultimedia 2007.

Documents

06221968