Multipath Routing for Video Delivery over Bandwidth-Limited Networks

1. Multipath Routing for Video Delivery over Bandwidth-Limited Networks S.-H. Gary Chan Jiancong Chen Department of Computer Science Hong Kong University of Science and Technology Clear Water Bay, Kowloon

2. Outline Introduction Multipath routing heuristic for point-to-point video delivery Scheduling algorithm at the server to achieve the theoretical minimum start-up delay Extension to point-to-multipoint layered video delivery Conclusion

3. Introduction

4. Research Motivation • Deliver quality video services over bandwidth-limited networks (e.g., the Internet) • Video application requirements • High bandwidth • Low start-up delay or network transmission cost • Traditional routing based on single path approach (e.g., the shortest path routing) is no longer sufficient to meet the bandwidth requirement • QoS routing

5. Negotiating and Guaranteeing QoS in the Internet • Integrated services/Resource Reservation Protocol (RSVP) • Multi-protocol label switching (MPLS) • Differentiated services model (DiffServ) • Traffic engineering • Constraint-based routing

6. Constraint-Based Routing • Compute routes subject to multiple constraints • Distribution of link state information • Route computation • Goals • Select routes that can meet certain QoS requirements • Increase utilization of the network

7. Meeting Bandwidth Requirement with Low Delay: Multipath Routing • The video data is transmitted over multiple paths in the network • Increasing the overall aggregate delivery bandwidth • Routing to meet the bandwidth requirement • The end host needs to do reassembly • Increasing the start up delay • Server scheduling to reduce the delay

8. Previous Work on Multipath Routing • Search multiple paths and select the best one • E.g., selective probing • Find multiple paths for a connection (e.g., disjoint paths routing) • Mainly designed for reliability rather than high aggregate bandwidth

9. Our Work • A multipath heuristics for point-to-point video delivery • Low delay and buffer requirement • Efficient • Given a set of path lengths • The theoretical minimum delay achievable • A scheduling algorithm to achieve that • For point-to-multipoint communication with heterogeneous bandwidth requirement • How the multicast trees should be constructed to minimize the cost of the tree-aggregate • The corresponding number and bandwidth of the video layers

10. Multipath Routing for Point-to-Point Video Delivery

11. A Point-to-Point Video Network

12. Multipath Problem Formulation: Bandwidth-Constrained Delay-Optimized Problem • Given: • A source s • A destination t • Bandwidth requirement B • B less than the max-flow of the network • Find routing and scheduling algorithms to achieve • Bandwidth no less than B • Minimum delay

13. Desirable Properties of Routing Algorithms • Efficient • Similar complexity as the shortest path routing • Fast route convergence • Achieving high end-to-end bandwidth • Preferably the max-flow of the network • Amendable to the current Internet routing

14. A Multipath Routing Heuristics • Find the max-flow sub-graph G’ of the network • Find the shortest-path in the sub-graph G’ • If the aggregated bandwidth of the path(s) found is sufficient, return • Subtract the bandwidth from G’ along the path just found • Repeat steps 2 to 4

15. (20,7) (20,7) v1 v1 v4 v4 (10,12) (10,12) (8,13) (8,13) (15,6) (15,6) (5,13) s s v3 v3 t t (10,5) (10,5) (10,10) (15,7) (15,7) (10,8) (10,8) (15,7) (15,7) (20,6) (20,6) v2 v2 v5 v5 (10,14) An Example

16. Simulation Model • Hierarchical network • 3-hierarchy nodes: backbone routers, border routers and intra-domain routers • Random links • System parameters • Network size • Network density • Connectivity, etc

17. Comparison with the Traditional Approaches • Shortest path • Shortest-feasible path • Remove the links with insufficient bandwidth • Run the shortest path algorithm over the residual network • Performance measures • Success rate in meeting the bandwidth requirement • Bandwidth achieved • End-to-end delay, given by the longest path

18. High Success Rate

19. High Bandwidth Achieved

20. Low Average Delay

21. Hierarchical routing • Logical hierarchical topology as in the Internet • State information • Only full local information is maintained • Remote state information is partially maintained • Compute multiple routes in the regions in parallel • Reduce computation complexity, processing time, and storage

22. s t An example Upper hierarchy Lower hierarchy

23. Server Scheduling Algorithm

24. Problem Formulation • Given a set of path lengths • What is the theoretical minimum start-up delay achievable if video data can be scheduled? • Guarantee continuity • Find a data scheduling algorithm at the server to achieve such minimum delay • No other algorithms can achieve lower delay while maintaining stream continuity

25. A Simple Case • Two paths with the same bandwidth of B/2 but different delays d1and d2 (d1< d2) • Without server scheduling, the start-up delay equals the delay of the longer path, i.e., d2

26. Data Slope=B Slope=B/2 Time d2 d1 0 original delay minimized delay The Theoretical Minimum Delay • Data production and consumption curves • The difference is the buffer requirement • In the example, the minimum start-up delay is (d1+d2)/2

27. The Idea • Don’t indiscriminately multiplex video packets along all the paths • The server sends the video prefixes along the shorter paths • The client plays back the prefixes with stream continuity • Before the data from the longest path arrives

28. Video data To path 1 To path 2 To path 1 The Scheduling Algorithm • The video sequence is partitioned into segments • All the segments are transmitted in parallel over the multiple paths • The earlier segments are transmitted over the shorter paths

29. General Case of Scheduling p K p 2 p 3 . . . p . . . 1 p p 2 1 p 2 p p 1 1 . . . t t t t V i d e o t i m e t 0 1 2 3 K - 1

30. An Exact Solution Solving the Multipath Problem • A network with unit link bandwidth • Multipath is disjoint paths • With scheduling, the problem is to find the shortest-disjoint paths (SDP) • Bandwidth requirement: B units • Find the B-shortest-disjoint paths • The sum of their delays is minimum • The shortest-disjoint paths algorithm is well known

31. Rescheduling Achieves a Delay Comparable to the Shortest Path

32. Extension to Point-to-Multipoint Video Delivery

33. A Video Multicast System • A server and multiple clients • The clients have different bandwidth requirements • A link is characterized by its bandwidth and cost • Find multiple multicast trees spanning the multicast group • Meeting the heterogeneous bandwidth requirements of the members • With minimum cost of the tree-aggregate • Assignment of video layers • A base layer and several enhancement layers • The number of video layers, and • Their respective bandwidths

34. A Simple Case • All the users have the same requirement B • Multiple trees are used to span all the users • With minimum cost of the tree-aggregate • If all the bandwidth requirements are met • A single video layer with bandwidth B • Otherwise, layered video can be used • The higher layers serve users with increasing end-to-end bandwidth

35. Users Base layer tree 1 Base layer tree 2 Enh. layer tree 1 An Example s

36. Problem Formulation: Bandwidth-Constrained Cost-Optimized Problem • Given • A source s • A set of destinations Y (= {y1, y2,…, yn}) • Bandwidth requirement B (= {b1, b2,…, bn} ) • Find multiple trees T to achieve • Bandwidth no less than bifor yi • Minimum cost of the aggregated “mesh” • The corresponding number and bandwidth of the layers, and along which trees a layer transmits • Multiple trees • To find a min-cost tree (Steiner tree) is NP-hard • To construct such multiple trees is even harder

37. Two Heuristics: Multipath Extension • Based on point-to-point multipath heuristic • First meet the bandwidth requirement of each user with the multipath heuristics • Aggregate the paths • Construct trees out of the paths-aggregate • Each tree has a certain bandwidth equal to the bandwidth of the bottleneck link • There is at least one tree spanning all the users • Complexity: O(m|V|3) • Bandwidth-first approach

38. Min-Cost Tree Extension • First find a min-cost multicast tree spanning all the users • Add branches to the tree until all the bandwidth requirements are met • Closest receivers • Forming new trees • Complexity: O(mB|V|2) • Cost-first approach

39. Bandwidth Assignment of Layers • Group the trees spanning the same set of users • Arrange these groups according to decreasing number of users covered • The previous set of users is the superset of the latter • The aggregate bandwidth of the first tree-group is the bandwidth of the base layer • The aggregate bandwidth of the 2nd group is the bandwidth of the enhancement layer 1, and so on

40. Users Base layer tree 1 Base layer tree 2 Enh. layer tree 1 An Example on Layering s

41. Simulation Results • Hierarchical network • Comparing with a single-tree approach (shortest path tree) • Performance measures • Success rate of meeting the bandwidth requirements of the users • Average bandwidth achieved • Cost

42. High Success Rate

43. High Average Bandwidth

44. Slightly Higher Cost

45. Conclusion • Video routing over a bandwidth-limited network • Multi-path heuristic • Achieve high end-to-end bandwidth with low delay • Video scheduling algorithm at the server • Reduce the start-up delay to the theoretical minimum • Extension to multicast environment • Meeting heterogeneous bandwidth requirements • Minimum cost of the tree-aggregate

46. Questions and Answers Thank you!