1. Periodic broadcasting with VBR-encoded video Despina Saparilla, Keith W. Ross, and Martin Reisslein 1999 IEEE INFOCOM Hsin-Hua, Lee

2. Objective Develop non-uniform segmentation schemes with VBR-encoded video that significantly reduce the initial start-up latency without appreciably degrading image quality.

3. Introduction(1) VoD (Video on Demand) True VoD: client-centered Arbitrary starting time Waste of network bandwidth Near VoD: data-centered Utilization of network bandwidth and server capacity Start-up latency Batching the same requests before serving Periodic Broadcasting

4. Introduction(2) CBR (Constant Bit Rate) encoding technique Modifying the quantization scale during compression VBR (Variable Bit Rate) encoding technique Quantization level remains constant For the same quality level, Ave. Bit-Rate CBR is typically 2 times or more the Ave. Bit-Rate VBR with VBR video there is potential for increased system efficiency Quality degradation Highly variable bit rate

5. Introduction(3) To obtain dramatic reductions in start-up latency with VBR-encoded video, we must allow for some small fraction of packet loss (due to link buffer overflow). Tradeoff between start-up latency and packet- loss. Proposed Schemes Bufferless multiplexing Smoothing with bufferless multiplexing Server-buffering Client-prefetching

6. Near VoD with VBR-Encoded Video(1) Notations

7. Near VoD with VBR-Encoded Video(2) 12k 12k 12k 12k ˙ ˙ ˙ m k ˙

8. Near VoD with VBR-Encoded Video(3) Each video is divided into K segments according to broadcasting series. General broadcasting series [ e 1,e 2,…,e K-1,e K ] e i : the i th segment consists of e i segmentation units, in general, e 1 =1. N i ( m ): the number of frames in the i th segment of the m th video

9. Near VoD with VBR-Encoded Video(4)

10. Start-up latency vs. Loss Probability(1) Loss of bits occurs when the aggregate bit rate of the traffic (i.e., from all MK streams) exceeds the link’s capacity, C.

11. Start-up latency vs. Loss Probability(2)

12. Numerical Example: Geometric Series(1) Seg. 1 Seg. 2 Seg. 3 Seg. 4 Figure 1: Broadcasting strategy for geometric series with e k =2 k -1. q=K. receiver storage is unlimited. M=10, and N(m)=160,000 frames, about 107 mins. F=25 frames/sec. C: 85~205 Mbps.

13. Numerical Example: Geometric Series(2)

14. Numerical Example: Geometric Series(3) Two performance measures Start-up latency Probability of loss

15. Bufferless Statistical Multiplexing Latency K at least 6 Latency K at least 9 As K increases, prob. of loss also becomes higher. K1K1 K2K2 K3K3 K4K4

16. GoP Smoothing(1) Total start-up latency = max. access time for 1 st video segment + delay introduced due to smoothing over one GOP period. Figure 3: Bufferless multiplexing with smoothing over each GOP period

17. GoP Smoothing(2) We refer to points that correspond to longer total start-up latencies with no further improvement in P loss as dominated. Smoothing over a higher number of GoP periods does not have an adverse effect when low start-up latencies are desirable. Figure 4: Smoothing over many GOP periods (C=145M bps). K=7 K=6 K=5 No significant effect !

18. Buffered Statistical Multiplexing Add in finite size buffer at the server link. Total start-up latency = max. access time for 1 st video segment + B/C. To limit loss it is instead preferable to use a smaller K. K=7 K=6 K=5

19. Join-the-Shortest Queue Prefetching(1) prefetched frames prefetched buffer server client virtual buffer

20. Join-the-Shortest Queue Prefetching(2) The JSQ prefetch policy attempts to balance the number of prefetched frames across all virtual buffers. All the server needs to do is to schedule the broadcast of the frames of the MK video streams as if it were sending them to the MK distinct virtual buffers.

21. C=145 Mbps Join-the-Shortest Queue Prefetching(3) JSQ protocol brings significant improvement over simply multiplexing the video stream onto the bufferless link. 6 x 10 -2 3 x 10 -4 6 x 10 -8 100.7 sec

22. Join-the-Shortest Queue Prefetching with prefetch delay(1)

23. Join-the-Shortest Queue Prefetching with prefetch delay(2) d pre :the prefetch delay in frame periods. Total start-up latency =(N 1 +d pre )/F C=145 Mbps K=7 6.6 x 10 -6 100.4 sec 3 x 10 -4 9 x 10 -5

24. Join-the-Shortest Queue Prefetching with prefetch delay(3)

25. VBR and CBR Compared For the buffered multiplexing, we chose the K value and buffer size combination which gives the lowest delay while having a loss probability less than 10E-7 (essentially a negligible loss probability).