1. Breaking the Single-Path Barrier Brad Smith Jack Baskin SoE Research Review Day 10/20/2011

2. Outline Research Corporate Partnership Open Source Network Lab 10/20/11Jack Baskin SoE Research Review Day 2

3. The Internet is Single-Path Compute best path per destination Destination-based (“hop-by-hop”) forwarding 2 problems – Quality-of-Service – Congestion S 10/20/11 3 Jack Baskin SoE Research Review Day

4. Problem – Quality-of-Service Requires multiple paths per destination Example – 2 paths 2Mbps, 200ms 100Kbps, 20ms Depends on application! – Video streaming – Voice over IP (VoIP) b/w = 100Kbps latency = 20ms b/w = 2Mbps latency = 200ms Next hop for D? S a b D 10/20/11 4 Jack Baskin SoE Research Review Day

5. Problem – Congestion 45 16 32 16 18 32 16 18 16 32 7 7 74 47 4 7 4 7 7 7 10/20/11 5 Jack Baskin SoE Research Review Day Strong tendency for paths to share links…

6. Previous Work Circuit switch – special path per flow Solve part of the problem – Congestion – only need a small number (≤ 4) – Minimum delay – Partial solutions - disjoint widest and shortest In practice - over-provision a single path Challenge – what are enough paths? 10/20/11 6 Jack Baskin SoE Research Review Day

7. Best Set of Paths Paths as points in multi-dimensional space Some paths are “better” than others Best set of paths are those with none better Paths that provide the full range of performance 100Kbps, 20ms 10/20/11 7 Jack Baskin SoE Research Review Day

8. Using the Best Set of Paths Assign flows to paths that satisfy QoS In general, there is more than one… …choose one that minimizes congestion 10/20/11 8 Jack Baskin SoE Research Review Day 100Kbps, 20ms

9. Simulations Generate random networks Compute routing tables at all nodes – Total bandwidth – Delay Generate random stream of flows Use oracle to assign flows to paths – Satisfies QoS – Has bandwidth 10/20/11 9 Jack Baskin SoE Research Review Day

10. Simulations (cont) Measure Call Acceptance Ratio – % flow requests successfully routed Across range of networks – Size – number of vertices – Connectivity – average degree (# neighbors) 10/20/11Jack Baskin SoE Research Review Day 10

11. CAR: 350 Vertices, Degree 32 10/20/11 11 Jack Baskin SoE Research Review Day

12. CAR: 350 Vertices, Degree 16 10/20/11 12 Jack Baskin SoE Research Review Day

13. CAR: 350 Vertices, Degree 4 10/20/11 13 Jack Baskin SoE Research Review Day

14. Converge when lightly loaded and over-loaded Multipath does better in-between Fewer resources  poorer and less distinct performance 10/20/11Jack Baskin SoE Research Review Day 14

15. Target CAR Flow Rate 10/20/11Jack Baskin SoE Research Review Day 15 95% 600 Load supported by given infrastructure & routing architecture.

16. 95% CAR Rate: Degree 32 10/20/11 16 Jack Baskin SoE Research Review Day Multipath provides dramatic capacity increase with same infrastructure.

17. MP @ 95% CAR Rate – # Vertices 10/20/11 17 Jack Baskin SoE Research Review Day Solid gains with increasing infrastructure.

18. Ratio MP:SP 95% CAR – # Vertices 10/20/11 18 Jack Baskin SoE Research Review Day 4 to 11x gains… with opportunities for improvement(!).

19. From 1 to many layers… 10/20/11Jack Baskin SoE Research Review Day 19

20. Future Work Routing protocols (path computation) – Link state – Distance vector Congestion management (path selection) – Routing – Network feedback 10/20/11 20 Jack Baskin SoE Research Review Day

21. Thank you!