Video Packet Selection and Scheduling for Multipath Streaming

1. Video Packet Selection and Schedulingfor Multipath Streaming IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 3, APRIL 2007 Dan Jurca, Student Member, IEEE, and Pascal Frossard, Senior Member, IEEE 指導老師：童曉儒 教授 學生：許益晨

2. Outline • Introduction • Multipath Video Streaming • General Framework • Streaming Model and Notations • Packet Scheduling Analysis • Unlimited Buffer Nodes • Constrained Buffer Nodes • Distortion Optimized Streaming • Optimal Solution: Depth-First Branch & Bound(B&B) • Heuristic Solution: Load-Balancing Algorithm(LBA) • Simulation Results • Conclusions

3. Introduction

4. Introduction This paper addresses the problem of choosing the best streaming policy for distortion optimal multipath video delivery, under network bandwidth and playback delay constraints.

5. Introduction Internet media streaming applications still suffer from limited and highly varying bandwidth, and from packet loss. • As solution 1： • inter-dependent video packetsto be transmitted • (equivalently the adaptive coding of thevideo sequence) • propose a detailed analysis of timing constraints imposed by • delay sensitive streaming applications. Solution：increase the streaming bandwidth by balancing the load over multiple (disjoint) network paths • propose a heuristic-based approach to the optimization problem, • which leads to a polynomial time algorithm, based on load-balancing • techniques.

6. Multipath Video Streaming

7. Multipath Video StreamingA. General Framework Multipath streaming scenario.The client accesses the streaming serversimultaneously through two different paths, each one composed of two segmentswith intermediate buffers. network segment i instantaneous rate ri(t) instantaneous latency di(t) The network channels between the server and theclient are represented as variable bandwidth, lossless links.

8. Multipath Video StreamingB. Streaming Model and Notations Represented in the last page: • The rate ri(t)is the total bandwidth allocated to the streaming application on segment iat time instant t . • denote the cumulative rate on segment i, up to time instant t, by We first assumethat all segment rates and latencies along with intermediatebuffer capacities are accurately predicted by the server atall time instants, possibly with feedback of the overlay nodes.

9. Multipath Video StreamingB. Streaming Model and Notations • The video sequence is encoded into a bitstream using a scalable • (layered) video encoder. • The general rule stating： • Eachnetwork packet contains data relative to at most one video frame • Encoded video frame can be fragmented into severalnetwork packets

10. Multipath Video StreamingB. Streaming Model and Notations Each packet pn Packet size sn Decoding timestamp tdn Weight wn Directed acyclic dependency graph representation for a typical MPEGlayered-encoded video sequence (one network packet per layer, with IPBPBformat). Thesuccessful decoding of one packet is contingent on thesuccessfuldecoding of some other packets, called ancestors of pn.

11. Multipath Video StreamingB. Streaming Model and Notationst • By transmission policy π =(π1, π2,…, πN) • The policy πN used for packet pn consists in a couple a variables [qn, tsn] that respectively represent the action qnchosen for packet pn, and its sending time tsn.

12. Multipath Video StreamingB. Streaming Model and Notationst That arrival time,tcn Represented is equal to1 if the packet arrives on time at the decoder, and if all its ancestorshave been successfully decoded.

13. Packet Scheduling Analysis

14. Packet Scheduling AnalysisA. Unlimited Buffer Nodes • We consider buffering space in thenetwork nodes and the client is not constrained. • the sending time tsn of each packet • sent on path a

15. Packet Scheduling AnalysisA. Unlimited Buffer Nodes pn entersthe node buffer The arrival time of packet pn

16. Packet Scheduling AnalysisA. Unlimited Buffer Nodes The queuing time bn = transmit bit present in the buffer The minimal playback delay D(π)

17. Packet Scheduling AnalysisA. Unlimited Buffer Nodes • Lemma 1: Given that the streaming server sends the networkpackets in parallel on two paths, and that on each path thepackets are sent sequentially, the playback delay D(π) underthe given policyvector π is a nondecreasing function of n. • Proof:

18. Packet Scheduling Analysis A. Unlimited Buffer Nodes • Lemma 2: Ωnis a nonincreasing function of the packet number.

19. Packet Scheduling AnalysisB. Constrained Buffer Nodes • the buffering space in the intermediate nodes on each path is limited to and respectively. • It tries to avoid buffer overflows by adapting the sending time of each packet to the buffer fullness. Note that it may no longer use the full available bandwidth, without risking loss of packets.

20. Packet Scheduling AnalysisB. Constrained Buffer Nodes it verifies the inequality： also define the maximum buffer occupancy during the wholestreaming process

21. Distortion Optimized Streaming

22. DISTORTION OPTIMIZED STREAMING • Earliest Delivery Path First(EDPF)

23. DISTORTION OPTIMIZED STREAMING • Earliest Delivery Path First(EDPF)

24. DISTORTION OPTIMIZED STREAMING • Earliest Delivery Path First(EDPF)

25. DISTORTION OPTIMIZED STREAMINGA. Optimal Solution: Depth-First Branch & Bound (B&B) At each stage in the tree we can compute D(π), the minimum playback delay and Ωn, the cumulative video quality measure, for a partial scheduling up to packet pn.

26. DISTORTION OPTIMIZED STREAMINGB. Heuristic Solution: Load-Balancing Algorithm If a packet can be scheduled on both network pathswithout interfering with the packetsalready scheduled, the algorithmwill choose the path that offers the shortest arrival time forpacket . If packet can only be scheduled on one path, thealgorithm will insert the packet on that path. Otherwise packetcannot be scheduled on any of the two paths, without interferingwith the already scheduled packets, and the algorithmwill drop packet without transmitting it.

27. Simulation Results

28. Simulation ResultsA. Simulation Setup • Video sequences are compressed with an MPEG4-FGS encoder, at 30 fps with various GOP structures. • simulate network scenarios containing two and three disjoint paths between the server and the client.

29. Simulation ResultsB. Stored Streaming Scenarios

30. Simulation ResultsC. Streaming With Limited Look-Ahead Encoded video frame rate (cumulative) and decoded video frame rates (cumulative) in the case of infinite and constrained intermediate buffers

31. Simulation ResultsD. Streaming With Link Rate Estimation and Channel Losses

32. Simulation ResultsE. Complexity Considerations LBA performance versus complexity (100 frames, average aggregated bandwidth of 450 kbps).

33. Conclusions

34. Conclusions • polynomial time algorithms that still offer close to optimal solutions, in the case of stored videos, and real-time streaming. • Simulation results in both scenarios prove that our proposed heuristic-based solution performs well in terms of final video quality, and is moreover suitable for the case of real-time streaming under strict delay constraints.