454IEEE Transactions on Consumer Electronics, Vol. 53, No. 2, MAY 2007Minimum Cost Scheduling of Stored Video in Dynamic BandwidthAllocation NetworksMyeong-jin Lee, Member, IEEEAbstract — In this paper, a minimum cost schedulingalgorithm is proposed for high-quality video streaming indynamic bandwidth allocation networks. It targets theproblem of service cost minimization in optimal smoothingalgorithms. Based on the bandwidth usage and the number ofbandwidth renegotiations, service cost functions are definedfor each transmission segment and a group of segments. Therelative cost flags are also defined for each segment, whichindicate the direction in which the segment is merged withless service cost. By inspecting the relative cost flags on theboundaries of segments, it is determined whether neighboringsegments should be merged, until there is no reduction in theoverall service cost. From the simulation results, theproposed algorithm is shown to merge neighboring segmentsin the direction minimizing the service cost. Compared withthe optimal smoothing algorithm, the overall service cost isgreatly reduced with a similar level of network resourceusage and less transmission rate variability, especially forlarger renegotiation cost parameters 1 .Index Terms — minimum cost scheduling, RCBR service,video streaming, segment merging.I. INTRODUCTIONMost current and future video applications will require theplayback of stored video over a high-speed network. Also,with the development of quality of service (QoS) provisioningtechnologies, there is increasing demand for high-qualityvideo services for stored video contents, e.g. high qualityvideo-on-demand or broadcasting over premium networkservices. For high-quality and constant-quality video services,studies on dynamic bandwidth allocation methods haveaddressed the inherent multiple-timescale burstiness ofvariable bit-rate (VBR) video , , , , , . Forrenegotiated network services, the cost may depend on theservice policies of networks with different pricing parametersand models. For renegotiated constant bit-rate (RCBR)service, the service cost can be calculated from the overallbandwidth usage and the number of renegotiations during aservice connection . Thus, for high quality videostreaming, an appropriate transmission schedule to minimizethe service cost while keeping the service quality unchangedshould be determined.1 This research was supported by the MIC (Ministry of Information andCommunication), Korea, under the ITRC (Information Technology ResearchCenter) support program supervised by the IITA (Institute of InformationTechnology Advancement). (IITA-2006-(C1090-0603-0002))Myeong-jin Lee is with the Korea Aerospace University, Goyang-City,Gyonggi-do, 412-791, KOREA (e-mail: email@example.com).Contributed PaperManuscript received April 11, 20070098 3063/07/$20.00 © 2007 IEEEFor video streaming, finding an optimal transmissionschedule has been the main focus of the majority of research.Optimization criteria are set to increase the multiplexing gain, to lessen the burden of networks for bandwidthrenegotiations , or to increase the user utility consideringboth the quality and the service cost . Salehi  proposedan optimal smoothing algorithm for stored video to minimizethe rate variability and evaluated the impact of the optimalsmoothing on network resources needed for video transportunder RCBR service. However, the resulting schedule is notoptimal in terms of service cost. Jiang  proposed a generalmethodology for obtaining an optimum transmission schedulewith different requirements. Using dynamic programming,optimum schedules were determined with discrete streamingrates and different cost functions. However, the algorithmcannot be applied to a real video streaming environmentwhere there are numerous frames to transmit and thetransmission rate is continuous. Thus, conventional optimalalgorithms are not optimal with respect to service cost northey are not practical enough to be applied to real videostreaming environments. Also, in order to meet theoptimization criteria, the resulting schedules of the optimalsmoothing algorithms sometimes have a larger number ofshort segments, which means frequent renegotiations. Lee reported that the portion of the short segments is notsmall and it may cause renegotiation failures or an increase inthe service cost. Although algorithms to lessen the burden ofthe networks induced by frequent renegotiations have beendeveloped , , they cannot minimize the overall servicecost for general video streaming environments.In this paper, we address the problem of service costminimization in video streaming under RCBR service. First,we state the problem of short renegotiation intervals and theiraffect on the overall service cost in an optimal smoothingalgorithm . Second, we propose a service costminimization algorithm for video streaming. Theoretically, itis possible to find the minimum cost schedule using Jiang’smethodology . However, this approach is not feasible for acontinuous transmission rate, because the computationalcomplexity increases to infinity. To solve the problem ofcomputational complexity, a two-pass approach is usedwherein an optimal schedule with some criteria is firstgenerated, and then the minimum cost schedule is found withthe resulting optimal schedule as the input.The remainder of this paper is organized as follows. Insection II, we state the service cost minimization frameworkfor video streaming. In section III, we discuss the need forAuthorized licensed use limited to: National Cheng Kung University. Downloaded on March 2, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
M.-j. Lee: Minimum Cost Scheduling of Stored Video in Dynamic Bandwidth Allocation Networks 455service cost minimization in conventional optimal videosmoothing algorithms. In section IV, we propose a servicecost minimization algorithm. In sections V and VI, we presentsimulation results and our conclusion, respectively.II. SERVICE COST MINIMIZATION FRAMEWORK FORSTORED VIDEO TRANSMISSIONIn this section, we describe the service cost minimizationframework for stored video transmission. A network servicemodel, a transmission unit for pricing, and the correspondingpricing model, which will be considered throughout the paper,are explained.A. Dynamic Bandwidth Allocation for Video StreamingCompressed video sources show burstiness over multipletime-scales, periods of milliseconds to several seconds.Because of the characteristics of multiple time scaleburstiness and long-term correlation in compressed video, thepossible multiplexing gain could be low in preventive calladmission control mechanisms for guaranteeing Quality ofService (QoS). Thus, dynamic bandwidth allocation during aconnection has been considered -, where the serverrequests bandwidth dynamically according to its currentbandwidth demand. In this paper, a renegotiated constant bitrate(RCBR)  service model is considered for videostreaming. This model has the key advantage that thesignaling cost and the network function for renegotiations areas simple as those of the constant bit-rate (CBR) service.For efficient transmission of video over the RCBR service,numerous studies on the traffic smoothing of live and storedvideo transmission have been reported , , . Thesmoothed rates are constrained by the lower and the upperbounds of the accumulated transmission rate to meet the endto-enddelay constraint of video transmission systems.Notably, differences exist in the objectives of optimization.Jiang  focused on the cost of video streaming and solvedthe optimal scheduling problem by using dynamicprogramming under the assumption of discrete transmissionrates. However, the algorithm cannot be used for realenvironments where the transmission rate is continuous.Salehi  proposed an algorithm to minimize the ratevariability but did not consider the service cost in depth.B. Cost for RCBR Service of Stored VideoThe scheduled result S * of a compressed video sequenceconsists of successive transmission segments, where eachsegment is described by its transmission rate and the segmentlength. For convenience, the i th transmission segment S i isrepresented bySi ≡ ( Ri, Li), (1)where R i and L i are the transmission rate and the length ofthe i th segment, respectively. Also, the transmission rates ofthe neighboring segments should be different from that of thethi segment.To find the minimum service cost for video streamingunder RCBR, we use the network pricing function  of theformN*CS ( ) = α∑ RL i i + β ⋅N(2)i=1where N , α , and β are the number of transmissionsegments in S * , the cost parameter for bandwidth usage, andthe cost per renegotiation, respectively. The functionconsiders both the cost for bandwidth usage and the signalingcost for renegotiations.III. PROBLEMS IN CONVENTIONAL OPTIMAL VIDEOSTREAMING ALGORITHMSIn this section, we define a minimum cost schedulingproblem for general video streaming environments where thenetwork pricing model considers both the bandwidth usageand the renegotiation cost. Conventional optimal smoothingalgorithms are not optimal in terms of service cost ,  northey are not practical enough  to be applied to real videostreaming environments. Although Jiang’s approach presentsa general optimization framework, the computational loadgrows to infinity for a continuous transmission rate. Also,while Salehi’s approach can handle continuous transmissionrates and minimizes the rate variability with lowcomputational complexity, it does not consider the servicecost induced by the bandwidth usage and signaling forrenegotiations.A. Cost Functions for Transmission SegmentsIt is generally acceptable that the transmission cost forRCBR service is the sum of the costs for the bandwidth usageand the signaling for renegotiations. In this section, anisolated transmission segment is defined as a single segmentwith a transmission rate different from those of theneighboring segments. The transmission cost of a singleisolated segment can then be represented byCS ( k)= α RL k k + β(3)where α and β are the cost parameter for bandwidthusage and the cost per renegotiation, respectively. The reasonfor including the single renegotiation cost β is that theisolated segment must have one transmission raterenegotiation at its end.Also, the transmission cost of a group of successivesegments from S k to k ncan be represented as follows.k+nS +CS ( k; Sk+n) = α∑ RL i i + β ⋅ ( n+1)(4)i=kTo show the effect of short transmission segments on theoverall cost, the cumulative cost ratio is defined as follows.Nk= 1, L ≤L∑ CS ( k )kCCR( L)= (5)CS ( ; S )1NThis represents the ratio of the transmission cost for shortsegments to the overall cost for the whole video sequence.Authorized licensed use limited to: National Cheng Kung University. Downloaded on March 2, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
458renegotiation cost parameter β , the code first calculates therelative cost flags for each segments (lines 4~6). By scanningthe segments from the first to the last, the relative cost flagsfor each segment boundary are then checked for segmentmerging (line 9). If all the flags are set, the segments aremerged (lines 10~12). Otherwise, the current segment staysisolated (line 14). Finally, the number of segments is updated(line 17) and the same procedure is repeated until there is nosegment to be merged (lines 3 and 18).Fig. 5. Service cost reduction during execution of the proposedminimum cost scheduling algorithm. Bond sequence, D = 40 .Fig. 6. Scheduled results of the proposed algorithm. Bond sequence,D = 20 .V. EXPERIMENTAL RESULTSIn this section, we present some experimental results forthe proposed service cost minimization algorithm. We usedfour MPEG video traces in . They are 40,000 frames longeach and the GOP pattern is IBBPBBPBBPBB. Because wefocus on the service cost of video streaming, it is assumedthat renegotiation requests are always accepted by networks.The scheduled result of the MVS algorithm is used as theinput to the proposed algorithm. Also, all the performanceparameters are evaluated with reference to the MVSalgorithm. For the calculation of the overall service cost,IEEE Transactions on Consumer Electronics, Vol. 53, No. 2, MAY 2007because we are focusing on the service cost reduction, not theabsolute service cost, it is sufficient to consider the ratio ofthe cost parameters β to α . Thus, α is fixed at 0.005 andonly β is allowed to vary over a given range.Fig. 5 shows how much the overall service cost is reducedfor different renegotiation cost parameters during theexecution of the proposed algorithm. The left-most pointscorrespond to the service costs of the MVS algorithm, whichare the initial costs to the proposed algorithm. The larger therenegotiation cost parameter, the more the proposedalgorithm is repeated and the overall service cost is reduced.The algorithm executes rapidly because it only handles butthe transmission rate and the length of each segment ratherthan the bit-rate information of every frames. For example,although the bond sequence is 40,000 frames long, there are291 transmission segments in the MVS scheduled result withan end-to-end delay of 40 ( D = 40 ). Thus, by using theproposed algorithm, the minimum cost schedule can becalculated in real-time for various network service pricingpolicies.Fig. 6 shows the transmission segments generated by theproposed algorithm with different renegotiation costparameters. Although there are frequent transmission raterenegotiations for some frame intervals in the MVSalgorithm, the proposed algorithm yields much longertransmission segments while avoiding high costrenegotiations. As the renegotiation cost parameter isdecreased, the transmission patterns of the proposedalgorithm become increasingly similar to those of the MVSalgorithm.Fig. 7 compares the performance between the proposed andthe reference MVS algorithm for different renegotiation costparameters. The bandwidth usage and the number ofrenegotiations in the MVS algorithm are fixed. In theproposed algorithm, as the renegotiation cost parameterincreases, the bandwidth usage increases, which causes anincrease in the overall service cost. However, the number ofrenegotiations tends to decrease substantially, which causes adecrease in the overall service cost. These two tendencies arereflected in the overall service cost, as given by (2), andfinally determine the minimum service cost. As therenegotiation cost parameter increases, while the service costincreases very steeply in the MVS algorithm, the service costincrement is not as large in the proposed algorithm. Thus, theproposed algorithm minimizes the service cost for differentservice pricing environments. This can be achieved bymerging neighboring segments in the direction minimizing theoverall service cost, which results in a decrease in the numberof renegotiations and a slight increase in the bandwidth usage.In order to demonstrate the impact of the proposed trafficsmoothing approach on the network resource requirements,the effective bandwidth, the coefficient of variance (COV),and the peak to average ratio (PAR) are calculated for thesmoothed traffic of the proposed and the MVS algorithms andAuthorized licensed use limited to: National Cheng Kung University. Downloaded on March 2, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
M.-j. Lee: Minimum Cost Scheduling of Stored Video in Dynamic Bandwidth Allocation Networks 459also for raw traffic without smoothing. The effectivebandwidth statistically estimates the network throughputrequired to transmit the video through a b -byte buffer with atolerable loss rate of γ . For a video with transmission ratesc1, c2,..., c N , the effective bandwidth is computed asNlog( ∑ exp( θ ⋅ c / )) /i 1 i N θ where θ =− log( γ / b). The=coefficient of variance is the ratio of the standard deviation tothe mean of the given traffic. In , the MVS algorithm isshown to minimize the empirical effective bandwidth, whichmeans that the network could allow more service connectionswith certain levels of guaranteed QoS. As shown in Fig. 8(a),the effective bandwidth of the proposed algorithm is quiteclose to that of the MVS algorithm. Although the effectivebandwidth increases for larger renegotiation cost parameters,the amount is rather slight. Also, from Figs. 8(b) and (c), it isfound that the COV and the PAR are equal or less than thoseof the MVS algorithm. This means that the transmission ratevariability of the proposed algorithm is less than that of theMVS algorithm, especially for service environments withlarger renegotiation cost parameters.In summary, compared to the performance of the MVSalgorithm, the proposed algorithm is shown to reduce theoverall service cost greatly with a similar level of networkresources usage and with less transmission rate variability forvarious service pricing environments.VI. CONCLUSIONIn this paper, a minimum cost scheduling algorithm isproposed for high-quality video streaming in dynamicbandwidth allocation networks. It targets the problem ofservice cost minimization in optimal smoothingalgorithms. Based on the bandwidth usage and the numberof renegotiations, service cost functions are defined foreach segment and a group of segments. The relative costflags are defined for each segment in order to indicate thedirection in which the segment is merged with less servicecost. In the proposed algorithm, by inspecting the relativecost flags of neighboring segments until all the flags arezero, it is determined whether segments should be merged.From the simulation results, the proposed algorithm isshown to merge neighboring segments in the directionminimizing the service cost. Also, compared with theMVS algorithm, the service cost is greatly reduced with asimilar level of network resources usage and with lesstransmission rate variability for larger renegotiation costparameters. The proposed algorithm can be used tominimize the service cost for RCBR service and can beeasily extended to other service environments withdifferent service pricing models.Fig. 7. Overall service cost of the proposed algorithm. Bond and news sequence, D = 10 .Fig. 8. Characteristics of the scheduled traffic. Bond and news sequence, D = 40 . Switch buffer size=5kbits, cell loss prob.=10 -3Authorized licensed use limited to: National Cheng Kung University. Downloaded on March 2, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
460REFERENCES T. V. Lakshman, A. Ortega, and A. R. Reibman, “VBR video: tradeoffsand potentials,” Proceedings of the IEEE, Vol. 86, No. 5, pp. 952-973,May. 1998. M. Grossglauser, et al, “RCBR: a simple and efficient service formultiple time-scale traffic,” IEEE/ACM Trans. Networking, Vol. 5, No.6, pp. 741-755, 1997. H. Zhang and E. W. Knightly, “RED-VBR: a renegotiation-basedapproach to support delay-sensitive VBR video,” ACM/Springer-VerlagMultimedia System Journal, Vol. 5, No. 3, 1997. D. Reininger, D. Raychaudhuri and J. Hui, “Bandwidth renegotiationfor VBR video over ATM networks,” IEEE J. Select. Areas Commun.,Vol. 14, No. 6, pp. 1076-1086, Aug. 1996. O. Rose, “Statistical properties of MPEG video traffic and their impacton traffic modeling in ATM systems,” 20th Conference on LocalComputer Networks, Oct. 1995. Z. Jiang, L. Kleinrock, “A general optimal video smoothing algorithm,”IEEE Infocomm, 1998. J. Salehi, Z. Zhang, J. Kurose and D. Towsley, “Supporting storedvideo: reducing rate variability and end-to-end resource requirementsthrough optimal smoothing,” IEEE/ACM Trans. Networking, Vol. 6,No. 4, Aug. 1998. C. E. Luna, L. P. Kondi, and A. K. Katsaggelos, “Maximizing userutility in video streaming applications,” IEEE Trans. Circuits andSystems for Video Technology, Vol. 13, No. 2, Feb. 2003.IEEE Transactions on Consumer Electronics, Vol. 53, No. 2, MAY 2007 M. Lee, “Video traffic prediction based on source information andpreventive channel rate decision for RCBR,” IEEE Trans.Broadcasting, Vol. 52, No. 2, pp. 173-183, June 2006. M. Lee, K. Yoo, and D. Lee, “Dynamic bandwidth allocation for storedvideo under renegotiation frequency constraint,” MMNS 2006, LNCS4267, pp. 98-109, Oct. 2006.Myeong-jin Lee (M’01) received the B. S., M. S., andPh. D. degrees, all in electrical engineering from KoreaInstitute of Science and Technology (KAIST), Daejon,Korea, in 1994, 1996, and 2001, respectively. From 2001to 2004, he was a Senior Engineer with the System LSIBiz., Samsung Electronics, Gyeonggi, Korea. From 2004to 2007, he was an Assistant professor of the Departmentof Electrical Engineering, Kyungsung University, Busan, Korea. In 2007, hejoined the School of Electronics, Telecommunications, and ComputerEngineering at Korea Aerospace University (KAU), Gyeonggi, Korea, wherehe is an Assistant Professor. His current research interests are in the areas ofvideo coding and multimedia communication systems. Prof. Lee received theSamsung Human Tech Thesis Prize in 2001. He is a member of the IEEE andthe Korean Information Science Society.Authorized licensed use limited to: National Cheng Kung University. Downloaded on March 2, 2009 at 22:29 from IEEE Xplore. Restrictions apply.