Topology-Aware Overlay Networks for Group Communication

1. Topology-Aware Overlay Networksfor Group Communication Minseok Kwon and Sonia Fahmy Department of Computer Sciences Purdue University {kwonm, fahmy}@cs.purdue.edu http://www.cs.purdue.edu/~fahmy/

2. Overlay Multicast Switzerland Russia Austria Germany USA Japan Korea

3. End System Multicast • Proposed by Y. Chu, S. Rao, S. Seshan and H. Zhang (CMU) in 2000. • One of the early application-level multicast protocols • Easy to deploy • Self-organizing • Target application: small sparse groups (audio/video conferencing groups)

4. Challenges • Efficiency • How to reduce delay penalty (relative to unicast) and reduce number of duplicate packets (link stress)? • Scalability • How to reduce the amount of data maintained by each group member, and reduce routing information exchange?

5. Our Primary Focus • Can we exploit underlying network topology data which routing protocols anyway establish (together with measurements) for constructing efficient low-overhead overlays for application level multicast?

6. Benefiting from overlapping shortest paths Minimize the delay penalty and the number of duplicate packets. Source Shortest path for the relay node Shortest path for the new member Other group member (now serving as relay Node) New member

7. TAG (Topology Aware Grouping) • Mostly single source or core-based multicast • Overlay tree is organized based on path overlap information • Delay (primary) and bandwidth (secondary) are considered as metrics • Target application: latency-based applications (limited bandwidth streaming applications, multi-player online game)

8. TAG Features • Reduces delay penalty • Reduces the number of duplicate packets • Low space complexity: a small amount of information (IP addresses and paths of only parent and children) is maintained at each group member • Low average time complexity: member join and leave require O((log n)2) and O(log n), respectively, where n is the total number of group members

9. Difficulties with TAG • When to re-organize? • Overhead at (or near) the sender/core if many members join/re-join at the same time • Bandwidth is only considered as a secondary metric (as a loose constraint, and to break ties among equal delay paths)

10. TAG: Definitions • Path from A to B • A sequence of routers comprising the shortest path from A to B • Spath of A • Path from the sender/core to A • Length of path P • The number of routers in the path P • A  B • if spath of A is a prefix of spath of B

11. FIND JOIN FIND TAG: Member Join Root Path Matching Member1 New Member Member2 • A new member finds the parent and children by recursively applying the path matching algorithm

12. FT FT FT D1: (R1,R3) D2: (R1,R6) D4: (R1,R2,R4) D7: (R1,R2,R7) D1: (R1,R3) D2: (R1,R6) D4: (R1,R2,R4) D7: (R1,R2,R7) D1: (R1,R3) D2: (R1,R6) D4: (R1,R2,R4) D7: (R1,R2,R7) TAG: Complete Path Matching New member: D8: (R1,R2,R4,R5) D8: (R1,R2) D8: (R1,R5) Family Table D8  D4, D8  D7: New member subscribes here D4 and D7 are the children of new member D8. New member subscribes here D4  D8: D4 is the best candidate to Proceed with

13. FT D1: (R0,R1) Router End host FT FT R5 R6 D2: (R0,R1,R2,R4) D4: (R0,R1,R2) FT D2: (R0,R1,R2,R4) D5: (R0,R1,R2,R3) FT D2: (R0,R1,R2,R4) D2 D1 D3 D4 D5 FT FT D3:(R0,R1,R2,R4) D3:(R0,R1,R2,R4) TAG: Member Join Example Source S R0 R1 R2 R3 R4

14. FT D1: (R0,R1) Router End host R6 FT D2: (R0,R1,R2,R4) D5: (R0,R1,R2,R3) D2 D3 D5 D4 D1 FT D3:(R0,R1,R2,R4) TAG: Member Leave Example Source S R0 R5 R1 R2 R3 D4 is leaving. R4

15. TAG: Partial Path Matching • Complete path matching does not consider available bandwidth • Minus-k (or partial) path matching • Node A can be the parent of node B if A has a common spath prefix of length (spath of A) – k with B • Example S R0 D1: (R0,R1,R3) R1 R3 D1 R2 D4: (R0,R1,R2,R4) R4 k=1 D4

16. TAG: Partial Path Matching • Mitigates possibly high link stress and limited bandwidth near a constrained node • When is partial path matching activated? • Partial path matching is activated when the available bandwidth < bwthresh (loose constraint) • With partial matching, a new member examines several delay-based paths and selects the path which maximizes bandwidth (tie breaker) • k may be dynamically increased depending on the available bottleneck bandwidth and other constraints • Last hop(s) delay bounds, etc. can also be used as (loose) constraints, in addition to available bandwidth

17. How do we obtain topology and bandwidth data? • Topology: • Traceroute (experiments show that 10 % of the routers do not respond) • Topology server (e.g., OSPF topology server, AS-level maps) • Comparing common subsequences can be used instead of matching paths when complete information is not easily available • Bandwidth/delay: • Bandwidth estimation tools (e.g., pathchar, nettimer) • In-band measurements

18. Ongoing work • More intelligent path matching with multiple tight or loose constraints and incomplete topology data • Fault resilience • Periodic probing of parent and children • Adaptivity to changes • An intermediate node probes paths to its children • Path-based aggregation of destinations • A change in spath affects members which overlap with the spath

19. Economies of Scale Factor • Two important questions to answer about an overlay multicast tree • How much bandwidth does TAG save compared to unicast (1) ? • How much additional bandwidth does TAG consume compared to IP multicast (0.8+) ? 1. J. Chuang and M. Sirbu, “Pricing Multicast Communications: A Cost-based Approach.” Proc. of Internet Society INET, 1998. 2. G. Phillips, S. Shenker, and H. Tangmunarunkit, “Scaling of multicast trees: Comments on the Chuang-Sirbu scaling law.”, ACM SIGCOMM, 1999.

20. A Simple Model • An end-host can be attached to any router • A router can have more than one end host attached to it Primary Source Router End host k k-ary tree

21. TAG Model • Case 2 • No host connected to A • Case 1 • At least one host connected to A B B A C(k) A C(1) C(2) A single packet hop over the link B

22. Economies of Scale Factor • Modeling results • Simulation results • Can we develop a more realistic model? (e.g., unary nodes representing transit routers added to the tree)

23. Performance Evaluation • Simulations • Session-level simulations for TAG and ESM • TAG • Minus-k partial matching: fixed k=1, loose bwthresh=200 KB • ESM • Degree bounds of a member in mesh: lower bound = 3, upper bound = 6

24. Performance Evaluation • Topologies • Transit-Stub model: GT-ITM • TS1 (492 nodes), TS2 (984 nodes), TS3 (1640 nodes) • Random symmetric link delays from 1 to 55 ms in transits and 1 – 3 ms in stubs • 100 MB to 500 MB backbone bandwidth and 500 KB to 1 MB for the bandwidth of edge links • AS-level: AS maps from NLANR, Inet • AS97, Inet97 (3015 nodes) • AS98, Inet98 (3878 nodes) • AS99, Inet99 (4872 nodes)

25. Performance Evaluation • Performance Metrics • Relative Delay Penalty (RDP): The relative delay increase between two nodes in TAG against unicast delay between the same two nodes • Link Stress: (Total or maximum) Number of duplicate copies of a packet over a physical link • Mean Available Bandwidth: The mean available end-to-end bandwidth between every two nodes

26. Results: Mean RDP ESM performance significantly improves when upper degree bound is increased to 12

27. Results: Total Stress

28. Results: Maximum Stress Partial path matching helps reduce the stress near highly constrained nodes.

29. Results: Mean Bandwidth

30. Results: ASMap and Inet

31. Related Work • End System Multicast • ScatterCast, Yoid, ALMI, Overcast, Bayeux, SCRIBE, CAN-based multicast • Overlay networks • RON, Detour • Unicast-based multicast protocol • REUNITE • Theoretical studies • Node degree constraints and diameter bounds in overlay multicast networks

32. Conclusions • Network topology information is used to construct an overlay multicast network: low delay penalty and a small number of duplicate packets • Delay (primary metric) and bandwidth (secondary) are considered as metrics • Economies of scale factor is 0.94+ for TAG • Simulation results indicate the effectiveness of TAG in building efficient trees for a large number of group members, with appropriate parameter values

33. Ongoing and Future Work • Two-tiered TAG • Core receivers should meet given requirements (latency or bandwidth) • More scalable • More adaptive to dynamic changes • Implementation and experiments