An Effective and Efficient Traffic Smoothing Scheme for Delivery of Online VBR Media Streams

Transcription

1 An Effective and Efficient Traffic Smoothing Scheme for Delivery of Online VBR Media Streams Ray-I Chang, Meng-Chang Chen, Jan-Ming Ho and Ming-Tat Ko Institute of Information Science Academia Sinica, Taipei, Taiwan, ROC {william, mcc, hoho, Abstract Traffic smoothing for delivery of online VBR media streams is one of the most important problems in designing multimedia systems. Given available client buffer and a window-sliding size, conventional approaches try to reduce bandwidth allocated in each window. However, they can not lead to the minimization of bandwidth allocated for transmitting entire stream. Although a window-sliding approach was introduced to further reduce the bandwidth allocated recently [21], it was computational costly. In this paper, an effective and efficient online traffic-smoothing scheme is proposed. Different from the conventional static window-sliding approaches, our approach dynamically decides the suitable window-sliding size to online smooth the bursty traffic. Then, an aggressive workahead scheme is applied in transmitting entire stream. By examining different media streams, our approach has small bandwidth, high bandwidth utilization and small computation cost. Considering the online transmission of a Star War movie, our approach result is 13% less for the bandwidth and 4% less for the network idle rate than SLWIN(1). Comparing the number of window sliding, our approach is 75% less than SLWIN(1). The relations between the characteristic of input traffic and the behavior of obtained scheduling results are discussed. Finally, an extension of the proposed approach to resolve the latency and quality tolerance applications is also introduced. I. Introduction In a distributed multimedia system, VBR media streams should be smoothed and transmitted from servers across network to clients by following a transmission schedule. To guarantee media QoS (quality-of-service), the transmission schedule should satisfy real-time requirements for jitter-free playback. Besides, the allocated resources (such as network bandwidth) should be minimized and fully utilized to support as many requests as possible. Generally, a network transmission schedule can be categorized as either client-controlled [1] ÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁÁ * This work was partially supported by NSC under grants NSC E-1-11, NSC E-1-12, and NSC E or server-controlled [2-9]. In client-controlled schemes, the occupancy of client buffer can be monitored with better knowledge by deciding the playback status. However, the client needs to send feedback messages to servers. As the network is possibly congested, this clientcontrolled scheme has the drawback of feedback overhead and the responses may be delayed. Recently, servercontrolled schemes have received great attentions for delivery of VBR media streams to guarantee deterministic network services. If the media stream is pre-recorded, the entire traffic can be accurately characterized to allocate appropriate amount of resources and to minimize the specified cost functions (i.e. critical/peak bandwidth or rate variability) [2-15]. The smoothed traffic usually has smaller transmission cost than the original bursty traffic. Notably, if the traffic is generated online, the characteristics of online traffic can not be analyzed offline as presented in [16, 21]. Although some prediction methods are presented to estimate the sizes of future frames based on their previous frames [19-2], the prediction may not be correct. In 1997, a window-based traffic-smoothing approach was introduced for delivery of online media streams [21]. Given a suitable delay as the time window size (called delay window), media frames in this time window are known. Thus, the smoothing algorithms presented for prerecorded media streams can be applied to smooth the traffic in this time window. In [21], a variation of the offline smoothing algorithm (proposed in [5] to minimize the allocated bandwidth for a pre-recorded media stream) is applied in each time window to smooth online bursty traffic. They smoothed the traffic between different time windows independently. Note that, for traffic smoothing, the large-size frames should be pre-transmitted ahead of their playback time to reduce required bandwidth. Let s consider a bursty media stream, in which, a sequence of small-size frames (called window w i ) is followed by an action-scene with large-size frames (called window w i+1 ). This kind of bursty traffic can be found in any media streams. If the hopping-window scheme [21] is applied (w i and w i+1 are mutually exclusive), the large-size frames in w 2 can not be pre-transmitted to smooth the bursty traffic. Although bandwidths allocated in w i and w i+1 are minimized separately, they do not lead to the

2 minimization of bandwidth required for transmitting the entire stream. Recently, a window-sliding approach with static window-sliding size [21] is introduced to further reduce the peak bandwidth allocated by overlapping time windows. However, it is hard to decide the best windowsliding size to smooth the peak bandwidth. Generally, if the window-sliding size is large, the obtained improvements will be limited. If the sliding size is too small, the computation will be time-consuming. Although the increasing of time window size may reduce the required peak bandwidth, the delay time for media playback would be increased. It is not acceptable for some real-time applications. To alleviate these drawbacks, in this paper, a dynamic window-sliding size scheme is proposed. Our approach can dynamically select the suitable window-sliding size to smooth bursty traffic. Then, an aggressive workahead scheme that operates for consecutive windows in a dependent manner is applied in transmitting entire stream. Thus, the large-size frames could be pre-transmitted as early as possible to reduce the required bandwidth. In this paper, we have examined our proposed approach by several benchmark streams. They are chosen because of their variety in different burst traffics. Experiments show that our proposed method is effective and efficient. Given the client buffer and the delay time window, our proposed approach requires smaller network bandwidth and obtains higher network occupancy than the conventional static window-sliding (SWS) approaches. The required computation cost (the window sliding times) is also smaller than that of SLWIN(1) (an 1-frame windowsliding approach). Our first contribution is the use of aggressive workahead scheme to increase transmission performance. The second is the use of dynamic window-sliding size selection so as to reduce computation costs. The proposed approach can be integrated with other approaches to improve the playback quality. If network congestion or buffer overflow is detected, we can decrease the encoding speed and the encoding quality. Based on the transmission deadline for each frame (can be computed by [2]), the suitable frames can be dropped or deferred for playback to fit the available bandwidth. This bounded bandwidth allocation technique is empirically useful when considering real-world applications with latency- and quality-tolerance [1]. II. Delivery of Online VBR Media Streams At client sites, media frames are played periodically. The playback schedule can be described by the cumulative media frame sizes called cumulative playback function (CPF) [2-3]. Define a media stream as V = { f f 1... f n-1 ; T f Á where f i is the i-th media frame size, T f is the period time for frame playback and n is the number of frames in this media stream. Without loss of generality, we assume that the client starts playback at time throughout this paper. The cumulative playback function F(t) at time t is defined as follows. F(t) =, if t < =F i, if i*t f t < (i+1)*t f, where i < (n-1) = F n-1, if (n-1)*t f t F i = F i-1 + f i is the i-th cumulative media size (where i < n and F -1 = ). As F(t) is a stair function, we can define F(i*T f ) - and F(i*T f ) + as the lower and the upper corner values at time i*t f. Assume that the server starts transmission at time -d*t f and d*t f is called the delay time. Assume that the transmission rate r i is applied to transmit the media stream between time i*t f to time (i+1)*t f. A transmission schedule can be identified by a sequence of allocated transmission rates r -d r -d+1... r r 1... r n-2. The peak bandwidth allocated for media transmission is defined as max{ r i : for all i }. A. Jitter-Free Transmission Schedule Based on the above definition of CPF, we define the cumulative transmission function G(i*T f ) = G((i-1)*T f ) + r i *T f as the amount of data sent by the transmitter. Notably, a jitter-free transmission schedule should ensure that a complete data frame is transmitted before its display without buffer overflow and underflow. As the consumed data F(t) is a stair function and the buffer size b is bounded, the cumulative data transmitted before time t should not be larger than H(t) = min{ F n, F(t) + b Á. It can be easily proved that, for all i, f i b. Besides, G(t) should not be smaller than F(t) for continuous playback. A jitterfree transmission schedule G(t) should satisfy F(t) + G(t) H(t) - as shown in Fig. 1. buffer b d delay H(t) G(t) F(t) t time Fig. 1. A jitter-free transmission schedule for the given client buffer size and initial delay. For a jitter-free transmission schedule, its performance can be measured by the required resource and the resource utilization. Generally, the resource considered in a multimedia system includes the memory buffer and the

3 network bandwidth. As shown in previous studies [2-9], large initial delay and client buffer can act as a good reservoir to regulate the difference between transmission and playback rates. However, in a real-world system, the available buffer size and the acceptable delay time are bounded and highly dependent on the service provided. In [2], we presented a linear-time transmission schedule algorithm for delivery of pre-recorded media stream. An O(nlogn) algorithm is presented to offline compute the relation functions between the required resources [18]. Based on these relation functions, the session set-up protocol is as simple as a table lookup which is a constant time computation. Although this approach has been proved optimal on the required resource and the resource utilization, it is not designed for delivery of online generated media streams. B. Online-Generated VBR Media Stream In a distributed multimedia system, media data can be pre-recorded or online-generated. The pre-recorded media data are generally stored in disks [24]. Based on the data layout, a disk retrieval scheduler [22-23, 28] is applied to decide the retrieving time of data blocks. Then, a network transmission schedule is designed to reduce the required resource to support as many requests as possible. In a prerecorded media stream, as all the frame sizes are known, the transmission schedule can be computed offline [2-9]. However, in an online generated media stream, the frame sizes are not pre-specified before session setup. At any time t, the previous frames may be transmitted and the upcoming frames are not generated. Recently, the window-based traffic model was applied to smooth the online media stream by introducing a small playback delay [21]. The transmission schedule is decided by a time window under the constraints of F(t) and H(t) [2-21]. A simple example to illustrate this window-based online media stream is shown in Fig. 2. frames before f i are already transmitted and frames after f i+w-1 are not generated, only frames in this window can be applied to smooth the traffic. We can simply treat the media frames in each time window as pre-recorded. Thus, the traffic smoothing methods presented for the prerecorded streams (such as CBA [6], MVBA [5], LA [2] and CRTT [9]) can be applied in these time windows. In conventional SWS (static window-sliding) approaches, traffic smoothing for different time windows is operated independently [21]. They are also called the IWS (independent window smoothing) approaches in this paper. A simple example to demonstrate its drawback is given in Fig. 3. Suppose that the window w contains only smallsize frames and the next window w + contains large-size frames. The window w - contains the average-size frames. If the traffic in different windows were smoothed independently, the required bandwidth in w would be minimized without considering the allocated bandwidth in w - and the bursty traffic in w +. Thus, the peak bandwidth allocated in w - would not be used to smooth the bursty traffic in w +. In contrast, if the bandwidth allocated in w - is considered in w and w +, a better schedule with smaller bandwidth can be obtained (as shown in Fig. 3). Our proposed algorithm is based on this aggressive concept to improve the required peak bandwidth and network idle rate. conventional schedule a better schedule w w+ W D d server client Fig. 2. The window-based online media stream with window size W and initial delay D. In a window-based model with window size W, media frames f i f i+1... f i+w-1 are generated for scheduling with delay time W*T f. The initial delay for playback is D = (d+w)*t f (without loss of generality d = 1). Notably, as w- Fig. 3. The drawbacks of independent bandwidth minimization. In [21], a sliding window-based approach SLWIN(k) is proposed to reshape the media traffic once k new frames are generated (where k is called the windowsliding size). Given a suitable window size W, SLWIN(k) first considers the window w with frames f -1 f f 1... f W-1. The frame size of f -1 is. As shown in Fig. 2, the scheduling algorithm is executed and starts the transmission schedule at time W*T f. After d*t f preloading time, the client receives frame f and starts the playback. At time (W+k)*T f, frames f W f W+1... f W+k-1 are generated. The scheduling algorithm is executed again to shape the remainding traffic of f f 1... f W-1 and these k new added frames. This k-frame window-sliding approach

4 should run the O(W) traffic smoothing algorithm O(n/k) times (called window-sliding times). The computation cost is proportional to the window-sliding times. Sometimes, the over-determined bandwidth reduction may defeat the pre-transmission of media frames. Although the traffic of the current window is smoothed, it may result in larger peak bandwidth and lower resource utilization for the entire media stream. III. Our Proposed Approach In this paper, an online traffic-smoothing method is proposed to dynamically select the suitable windowsliding size. In addition, our bandwidth minimization procedures for different windows are operated dependently and aggressively. Notably, in each window, the peak bandwidth allocated in the previous windows (called r max ) is considered as an input threshold parameter. The initial value of r max is given as a suitable non-negative number, i.e. the average rate, for transmitting the first window. Assume that the current considered window is f i f i+1... f i+w-1 and f j (i j i+w-1) is the first un-transmitted frame (f i f i+1... f j-1 were transmitted in the previous time window). Let t i (p i = G(t i )) as the start time (point) that applies the transmission rate r i. The lower corner of H(t i ) p i the upper corner of F(t i ). For the transmission segment <p i, p i+1, r i >, the transmission starts from time t i *T f and ends at time t i+1 *T f with transmission rate r i. As shown in Fig. 4, there are three transmission segments <p, p 1, r >, <p 1, p 2, r 1 > and <p 2, p 3, r 2 >. By applying conventional IWS (independent window smoothing) approaches, r 2 is smaller than r max as the allocated bandwidths in these two windows are minimized independently. Thus, the large frames would not be pre-transmitted to reduce the peak bandwidth allocated. By our proposed approach, the transmission schedule is constructed to consider the bandwidth allocated in the previous windows. Based on the parameter r max, we do not only smooth the traffic but also pre-transmit the bursty traffic as early as possible to reduce the required bandwidth and its idle rate. p r r1 r2 p1 p2 p3 r r 1 r 2 p p 1 p 2 conventional proposed Fig. 4. Comparisons of the peak bandwidth allocated by conventional independent window smoothing (IWS) approaches and our proposed approach. p 3 A. Required Bandwidth and its Idle Rate Before describing the proposed algorithm, we first define (s, G(s)) as the start-point of the segment and the increasing variable t would be tested as the possible endpoint of the segment. Assume that the media frames in the current window are f i f i+1... f i+w-1 where f j (i j i+w-1) is the first un-transmitted frame (f i f i+1... f j-1 are transmitted in the previous window). Define Q = F(j*T f ) - - F(i*T f ) as the pre-transmitted traffic size by the previous transmission schedule. In this paper, we compute R F (t) = (F(t) - G(s) - Q) / (t - s) as the test rate that started from time s and ended at time t of F(x). R H (t) = (H(t) - - G(s) - Q) / (t - s) is the test rate that started from time s and ended at time t of H(x). The detail description of the proposed algorithm in each window is shown as follows. ALGORITHM: Online Traffic Smoothing if (i = j) i = j - 1 // we have i j s = i*t f, t = t H = t F = j*t f, and G(s) = F(j*T f ) - // Initial the segment points r F = R F (s) and r H = R H (s) // Initialize the transmission rates repeat t = t + T f if (r H < R F (t )) Á // up-bounded by H(x) output segment: < G(s), H(t H ) -, r H > r max = max{ r max, r H Á, G(t H ) = H(t H ) -, Q=, s = t H, t = s + T f, r F = R F (s) and r H = R H (s) Á else if (r F > R H (t ))  // low-bounded by F(x) output segment: < G(s), F(t F ), r F > r max = max{ r max, r F Á, G(t F ) = F(t F ), Q=, s= t F, t = s + T f, r F = R F (s) and r H = R H (s) Á else if (t = i+w-1)  // final open area if (r H < r max )  output segment: < G(s), H(t H ) -, r H > G(t H ) = H(t H ) -, Q=, s = t H, t = s + T f, r F = R F (s) and r H = R H (s) Á else { // up-bounded by r max r max = max{ r max, r F Á, t F = (F(i+W-1)-G(s)) / (r max *T f ) *T f output segment: < G(s), F(i+W-1), r max > s = t Á Á else  // try the next frame t = t if (r H R H (t))  r H = R H (t), t H =t Á if (r F R F (t))  r F = R F (t), t F =t Á Á until (s = i+w-1) This presented procedure is based on an offline smoothing algorithm [5] with additional aggressive transmission

5 scheme [2]. Note that, in the first window w, the index i and j are initialized as -d and, respectively. You can easily prove that i j in other windows. B. Required Computation Cost In conventional SWS (static window-sliding) approaches, the window-sliding size is a static value k W. Given the sliding size k, the traffic-smoothing procedure will be executed periodically after k new frames are generated (as shown in Fig. 5). Although a small sliding size could lead to a small peak bandwidth, the required computation cost increases proportionally. Besides, it is hard to decide the suitable sliding size to minimize both the computation time and the peak bandwidth allocated. rate rate conventional proposed (static sliding 5 windows) (dynamic sliding 5 windows) Fig. 5. Our proposed approach can dynamically decides the suitable window-sliding size to shape the bursty traffic. Considering SLWIN(1) with k = 1, it requires n sliding-times and O(W n) computation. In our proposed approach, the smoothing procedure is executed only when the previous scheduled frames are all transmitted. The window-sliding sizes at different time window would be different. Generally, the applied window is hopping (window-sliding size = W) for CBR media traffic. When the bursty traffic is presented, the siding-size can automatically decrease to make a better control of the required bandwidth. Thus, the bandwidth requirement can be reduced. Notably, the suitable window-sliding size can be automatically decided in O(1). Since the select window-sliding size is usually larger than 1, the required sliding-times would be much smaller than n. Besides, as our smoothing procedure considers the un-transmitted frames only, we evaluate each frame just once. It totally takes only O(n) computation time. As suggested in [21], the window size W is selected to be an integer multiple of the size of GOP (group-ofpicture). In this paper, we assign W as the number of frames in a GOP. If the smoothing window is started by a large I-frame, the required peak bandwidth may be higher. In our algorithm, whenever the next window will start with an I-frame, we adjust the window size (for example, extend the window size to W+1) to include this I-frame. C. Other Extension Problems In conventional online traffic smoothing approaches, the available bandwidth is assumed to be infinite. Besides, they assume that media frames are not dropped or deferred during transmission. There are no additional control scheme provided. However, these assumptions are not guaranteed in current networks. In many real-world applications, the available bandwidth is bounded. Besides, the online media streams are latency- or quality-tolerance. The required QoS is characterized by a bounded delay in playback or a bounded decay in quality. Thus, given a bounded network bandwidth, they want to admit as many requests as possible under required QoS. Whenever network congestion is detected or the available bandwidth is decreased, we can slow-down the encoding speed or decrease the encoding quality to support more requests. The proposed algorithm can be extended to resolve these bounded bandwidth problems. space time drop frame defer display Fig. 6. We can drop frames or defer the frame display to reduce the required bandwidth. Define the bounded bandwidth as r available, we can modified the value of r max = min{r available, r max } in our proposed algorithm to design a new transmission schedule. If the available bandwidth is too small, some frames may miss their transmission deadlines for continuous playback. Table I. Statistics of the test VBR-encoded MPEG media streams. (fps: frame-per-second, AVG: average bit-rate, STD: standard deviation of frame size) Stream Frame # fps Max Frame AVG STD Burst Traffic Star War KB.36 Mb/s 2.3 Convex-and-Concave Princess Bride KB 1.17 Mb/s 4.8 None (large STD) CNN News KB 1.17 Mb/s 3.7 None Lecture KB.33 Mb/s 1.6 Convex Advertisements KB.45 Mb/s 1.9 Convex-and-Concave

6 Window-Sliding Times Star War Peak Bandwidth Allocated (Mb/s) Star War offline Network Idle Rate (%) Star War Fig. 7. The window-sliding times, the required network bandwidth and the bandwidth idle rate to online transmit the Star War movie stream. Window-Sliding Times Advertisements Peak Bandwidth Allocated (Mb/s) Advertisements offline Network Idle Rate (%) Advertisements As shown in Fig. 6, given the available bandwidth, the transmission deadline for each media frame can be easily computed by the lazy-aggressive scheme proposed in [2]. Considering a MPEG stream, we can directly drop P- and B-frames if their deadlines are missed. If an I-frame has missed its transmission deadline, we can try to drop other B/P-frames or to defer the stream display as shown in Fig. 6. Thus, the modified media traffic can always fit the available bandwidth. IV. Experiment Results In this paper, we have examined the proposed online transmission method by many benchmark streams shown in Table I [25-26]. Without loss of generality, we evaluate the proposed approach by the peak bandwidth allocated, the required sliding-times and the network idle rate (the bandwidth occupancy the actual bandwidth consumed divides the peak bandwidth allocated). Comparisons are made with an optimal offline scheduler proposed in [2] and the conventional window-based online schedulers presented in [21]. Our first benchmark is an over two hours long MPEG-encoded Star War movie. It has large media frames and high variation in frame sizes as many real-world media streams. The next two media traces examined are nearly 9 minutes long CNN News and Princess Bride video streams those were encoded by hardware MPEG encoder. As the hardware encoder uses variable distortion coding to maintain its target rate, the Fig. 8. The required network bandwidth, the window-sliding times and the bandwidth idle rate to online transmit the Advertisements video stream. variations in frame sizes of these two video streams are small. The last two test data are Advertisements and Lecture video traces encoded by the UCB software MPEG encoder [25] with constant distortion coding. Their variations in frame sizes are different due to the different video contents. In Lecture, the same speaker and his slides are shown along with only zooming and panning. Since the frame contents are very similar, the variation in frame sizes is small. However, in Advertisements, the frame contents change from one scene to another scene in a sequence of advertisements. Although the mean value of frame sizes is small, the frame size variation is very high. Table I shows the detail statistics of test media streams. A. Comparisons with Conventional Approaches When comparing the required bandwidth and network idle rate, experiments to the Star War movie show that the required bandwidth of SLWIN(1) is over 13% higher than that of our proposed approach (as shown in Fig. 7 with a reasonable buffer size > 9 KB). Besides, the obtained network idle rate of SLWIN(1) is over 4% higher than that of our proposed approach. Comparing to the conventional SLWIN(1) and SLWIN(W) approaches, our proposed approach can utilize the allocated network bandwidth more efficiently. Under the same constraints of initial delay and buffer size, our obtained network idle rate is smaller than those obtained by SLWIN(1) and SLWIN(W). Furthermore, the obtained improvement is

7 increased dramatically whenever a sufficient buffer is allocated. Comparing the required window-sliding times, our approach is 75% less than SLWIN(1). Since the windowsliding size used in our algorithm is larger or equal to that of SLWIN(1), our proposed approach requires less sliding-times than that of SLWIN(1). Although we have introduced a fast algorithm to speedup SLWIN(1) by computing only the first transmission segment in each window, the consumed CPU time is still very large. Our experiments show that the proposed approach can compute almost two times faster than this fast SLWIN(1) algorithm. Our proposed approach is shown to be efficient and effective. Although the SLWIN(W) approach with a W- frame hopping-window could compute faster, however, it requires higher bandwidth and network idle rate. The requirements in large system resources make SLWIN(W) impractical to play back burst VBR media streams. KBytes Cumulative Playback Function: Star War Frame Index Fig. 9. Star War movie stream contains a long-term concave burstiness. KBytes Cumulative Playback Function: Advertisements Frame Index Fig.1. Advertisements video stream contains a longterm convex burstiness. bursty traffic. The obtained results are comparable to the optimal offline scheduler. Comparing to SLWIN(W), we achieve over 22% improvements for the required bandwidth and 5% improvements for the network idle rate. V. Conclusion This paper presents a traffic smoothing method for delivery of online media stream. The proposed approach can dynamically decide the suitable window-sliding size to smooth the VBR traffic. We have explored the proposed algorithm by transmitting several benchmark video streams. Given the client buffer size and the playback delay, our proposed approach would require smaller network bandwidth and computation cost than conventional approaches [21]. Besides, the occupancy of network obtained is also better than those obtained by the conventional SLWIN(1) and SLWIN(W) approaches. Our proposed approach is shown effective and efficient. It significantly reduces the computation cost without significantly degrading the performance. In this paper, the bandwidth minimization procedure that operates for consecutive windows in a dependent manner. Thus, the selection of initial bandwidth would highly affect the obtained results. In this paper, the initial bandwidth is simply decided by the frame sizes in the first window. However, these frames may not be suitable for characterizing the entire stream. Our future work is to design a more intelligent method to select the suitable initial bandwidth. Although SLWIN(1) coupled with our aggressive workahead scheme will allocate as small bandwidth as other policy, it is really time-consuming as SLWIN(1) (the 1-frame window-sliding approach) is applied (see our experiments in Fig. 7 and 8). More effort should be taken to reduce its computation cost. It is necessary to design a new algorithm to further minimize the peak bandwidth allocated for delivery of online media streams with low time complexity and high resource utilization. An optimal algorithm is necessary for traffic smoothing of online VBR media streams [3]. We need to prove that the allocated bandwidth is minimized with the minimum required sliding-times and the optimal network idle rate. Fig. 8 shows the required window-sliding times, the required network bandwidth and idle rate to online transmit the Advertisements video stream. Comparisons are made with the optimal offline scheduler (offline), the SLWIN(1) approach and the SLWIN(W) approach. Our obtained result is better than that of SLWIN(1). When a small buffer (< 2 KB) is provided, our approach requires nearly the same sliding-times as SLWIN(W) to shape the Acknowledgements The authors would like to thank the anonymous referees for their valuable suggestions about this paper. In addition, thanks also go to Miss Yu-Lin Chiu for early reading this paper.

8 Reference [1] J.Y. Hui, J. Zhang and J. Li, "Quality of service control in GRAMS for ATM local area network," IEEE JSAC, [2] R.I Chang, M. Chen, M.T. Ko and J.M. Ho, "Optimizations of stored VBR video transmission on CBR channel," SPIE VVDC, pp , [3] R.I Chang, M. Chen, M.T. Ko and J.M. Ho, "Designing the on-off CBR Transmission Schedule for Jitter-Free VBR Media Playback in Real-Time Networks," IEEE RTCSA, pp. 1-9, [4] J.M. McManus and K.W. Ross, "Dynamic programming methodology for managing prerecorded VBR sources in packet-switched networks," SPIE VVDC, [5] J.D. Salehi, Z.L. Zhang, J.F. Kurose and D. Towsley, "Supporting stored video: reducing rate variability and end-to-end resource requirements through optimal smoothing," ACM SIGMETRICS, pp , [6] W. Feng and S. Sechrest, "Smoothing and buffering for delivery of prerecorded compressed video," IS&T/SPIE MMCN, pp , [7] W. Feng, F. Jahanian and S. Sechrest, "Optimal buffering for the delivery of compressed prerecorded video," IASTED International Conference on Networks, Jan [8] W. Feng, "Rate-constrained bandwidth smoothing for the delivery of stored video," IS&T/SPIE MMCN, [9] J.M. McManus and K.W. Ross, "Video On Demand over ATM: Constant-Rate Transmission and Transport," IEEE INFOCOM, March [1] M. Grossglauser, S. Keshav and D. Tse, "RCBR: a simple and efficient service for multiple time-scale traffic," ACM SIGCOMM, August [11] H. Zhang and E.W. Knightly, "A new approach to support delay-sensitive VBR video in packetswitched networks," NOSSDAV, [12] E.W. Knightly, D.E. Wrege, J. Liebeherr and H. Zhang, "Fundamental limits and tradeoffs of providing deterministic guarantees to VBR video traffic," ACM SIGMETRICS, pp , [13] S. Radhakrishnan and S.V. Raghavan, "A flexible traffic shaper for high-speed networks: design and comparative study with leaky bucket," CS-TR-3495, University of Maryland, [14] E.W. Knightly and P. Rossaro, "Smoothing and multiplexing tradeoffs for deterministic performance guarantees to VBR video," Int. Symp. Multimedia Communication and Video Coding, [15] J. Rexford and D. Towsley, "Smoothing variable-bitrate video in an internetwork," SPIE VVDC, [16] A.R. Reibman and A.W. Berger, "Traffic descriptors for VBR video teleconferencing over ATM networks," IEEE/ACM ToN, June [17] M. Grossglauser and S. Keshav, "On CBR Service," IEEE INFOCOM, March [18] R.I. Chang, M.C. Chen, J.M. Ho and M.T. Ko, "Characterize the minimum required resources for admission control of pre-recorded VBR video transmission by an O(nlogn) algorithm," IEEE IC3N, [19] S.S. Lam, S. Chow and D.K.Y. Yau, "An algorithm for lossless smoothing of MPEG video," ACM SIGCOMM, [2] T. Ott, T.V. Lakshman and A. Tabatabai, "A scheme for smoothing delay-sensitive traffic offered to ATM networks," IEEE INFOCOM, [21] J. Rexford, S. Sen, W. Feng, K. Kurose, J. Stankovic and D. Towsley, "Online Smoothing of Live, Varibale-Bit-Rate Video," NOSSDAV, [22] M. Chen, D.D. Kandlur and P.S. Yu, "Optimization of the grouped sweeping scheduling (GSS) with heterogeneous multimedia streams," ACM Multimedia, [23] E. Chang and A. Zakhor, "Scalable video data placement on parallel disk arrays," IS&T/SPIE Symposium on Electronic Imaging Science and Technology, [24] Y.C. Wang, S.L. Tsao, R.I. Chang, M.C. Chen, J.M. Ho and M.T. Ko, "A fast data placement scheme for video server with zoned-disks," SPIE VVDC, [25] [26] ftp::/thumper.bellcore.com [27] J. Zhang, "Optimal buffering algorithms for clientserver VBR video retrievals," Ph.D. Thesis, The State University of New Jersey, [28] R.I. Chang, W.K. Shih and R.C. Chang, "A new real-time disk scheduling algorithm and its application to multimedia systems," in Lecture Notes on Computer Science: 5th IEEE IDMS, [29] R.I. Chang, M.C. Chen, J.M. Ho and M.T. Ko, "Traffic-Smoothing for Delivery of Online VBR Media Streams by a Dynamic Window-Based Approach," TR-IIS9813, Institute of Information Science, Academia Sinica, [3] R.I. Chang, "Optimal Window-Based Traffic- Smoothing for Delivery of Online VBR Media Streams," Tech. Report, Institute of Information Science, Academia Sinica (under preparing).