1. ASAP: an AS-Aware Peer-Relay Protocol for High Quality VoIP Written by Shansi Ren, Lei Guo, and Xiaodong Zhang Department of Computer Science and Engineering The Ohio State University ICDCS ’ 06 Presented by Te-Yuan Huang

2. Outline  Problem Description  VoIP Quality Requirement  Direct vs. Overlay Routing for VoIP

3. Goal  To Prove: Selecting suitable relay peer(s) is critical The mechanism of Skype is not good enough  To Solve: Proposing an AS-Aware Peer-Relay Protocol

4. Outline - Problem Description  Is Internet direct routing enough for VoIP?  Can overlay routing (relay) help improve?  Is existing overlay method including Skype enough?  Can we design an good relay selection method?

5. VoIP Quality Requirement  Mean Opinion Score (MOS) metric MOS > 3.5 is acceptable.  Network factors End-2-End Latency Requirement  one-way latency < 150 ms End-2-End Packet LossRate Requirement  packet loss rate < 5%

6. Problem Description  Is Internet direct routing enough for VoIP?  Can overlay routing (relay) help improve?  Is existing overlay method including Skype enough?  Can we design an good relay selection method?

7. Direct vs. Overlay Routing for VoIP Internet host B host A host C host D host E direct IP routing path 1-hop overlay routing path 2-hop overlay routing path host A communicates with host C CopyRight © Xiaodong Zhang

8. Direct vs. Overlay – End-to-End Latency Measurement Procedures Limewire software modification Gnutella IP crawler BGP tables and updates Gnutella IP probing IP prefix and origin AS extraction online Gnutella IP addresses IP prefix table AS-level cluster Identification and delegate IP selection King tool based prober development cluster delegate IP addresses King prober pairwise IP DNS server latency measurement pairwise delegate IP latency CopyRight © Xiaodong Zhang

9. Direct vs. Overlay – Sessions and their RTTs  A session consists a pair of end host.  Randomly generate 10 5 sessions among cluster delegates.  Session direct RTTs measured by KING.  Optimal 1-hop relay nodes are found by iterated through every possible node.  For delegates a, b, c, relay path a-b-c RTT a-b-c = RTT a-b + RTT b-c + 2*relay delay.

10. Internet e2e RTT Measurement host a host b host c DNS server of host b DNS server of host c Internet IP of host c? IP of host c host a measures RTT b-c via recursive DNS queries CopyRight © Xiaodong Zhang

11. Direct vs 1-Hop Overlay RTT 50% sessions have optimal 1- hop RTT < direct IP RTT 25% sessions whose opt. 1-hop relay can reduce direct IP RTT by more than 50% CopyRight © Xiaodong Zhang

12. Overlay Routing Reduces RTT Sessions whose direct IP RTTs > 300 ms Sessions opt. 1-hop RTTs are always < 300 ms CopyRight © Xiaodong Zhang

13. Problem Description – 2 solved  Is Internet direct routing enough for VoIP? No. 1% - 10% sessions do not meet the RTT quality requirement  Can overlay routing (relay) help improve? Yes. We can always find one-hop relay paths whose RTTs are below the threshold for these sessions.  But why?

14. AS A AS D AS G AS H AS EAS F AS BAS C direct path between AS A and AS B direct path between AS B and AS C direct path between AS A and AS C provider-to-customer edge peer-to-peer edge AS H is congested 1-hop relay path between AS A and AS C via AS B Direct Path Is Congested CopyRight © Xiaodong Zhang

15. AS A AS B AS C AS DAS E AS FAS G AS HAS I direct path between AS A and AS C direct path between AS A and AS Bdirect path between AS B and AS C provider-to-customer edge peer-to-peer edge Multi-homed AS B As 1-hop Relay 1-hop relay path between AS A and AS C via AS BAS B is multi-homed, connects to AS A and AS C CopyRight © Xiaodong Zhang

16. Problem Description  Is Internet direct routing enough for VoIP?  Can overlay routing (relay) help improve?  Is existing overlay method including Skype enough?  Can we design an good relay selection method?

17. Dalian, China Shanghai, China Beijing, China Jingzhou, China Vancouver, Canada Bozeman, MT Austin, TX Jersey City, NJ Reston, VA Williamsburg, VA Baltimore, MD chosen 14 representative Skype sessions Skype Experimental Sites and Sessions CopyRight © Xiaodong Zhang

18. Skype Relay Selection Limits Limit 1: Long latency due to improper relay node selections. Session 4 Session 10 300 ms CopyRight © Xiaodong Zhang Williamsburg 上海 Williamsburg 大連

19. Limit 2: Probing multiple latent nodes in the same AS. Limit 3: Taking a long time to find major relays. relay node DNS zone name relay path RTT 85.64.x.x barak-online.net 360 ms 85.65.x.x barak-online.net 359 ms two probed relay nodes in session 8 Skype Relay Selection Limits CopyRight © Xiaodong Zhang

20. Limit 4: Generating non-negligible overhead. before stabilization after stabilization 10 Skype Relay Selection Limits CopyRight © Xiaodong Zhang

21. Problem Description  Is Internet direct routing enough for VoIP?  Can overlay routing (relay) help improve?  Is existing overlay method including Skype enough? No, there are 4 limits.  Can we design an good relay selection method?

22. ASAP Design Rationale  In general, peer nodes with the same IP prefix are relatively close to each other.  With publicly available BGP tables and updates, an up-to-date annotated AS graph can be built.  Paths with longer AS hops are likely to have longer latencies.  An Internet AS-level direct IP routing path usually has the valley-free property. CopyRight © Xiaodong Zhang

23. bootstrap1 bootstrap2 surrogate SA end host h1 surrogate SB end host h2 Internet AS graph IP prefix to cluster Surrogate IP table IP prefix to ASN table cluster’s close cluster set Internet AS graph cluster’s top node table end host h3 cluster A cluster B cluster C Three Types of ASAP Nodes bootstrap’s data structure cluster surrogate’s data structure Type 3: end hosts Type 2: surrogates Type 1: bootstraps CopyRight © Xiaodong Zhang

24. AS 1 AS 2 AS 3 AS 4 AS 5 AS 6 h1 s1 s2 s4 s5 s6 s3 h4 h3 h6 provider-to-customer edge peer-to-peer edge s1 close cluster ping pong good, 75 ms s2 – 75 ms bad, 350 ms good, 180 ms s5 – 180 ms good, 220 ms s6, h6 – 220 ms good: RTT < 300 ms && loss rate < 5% bad: RTT > 300 ms || loss rate > 5% good, 52 ms s3, h3 – 52 ms Close Clusters Construction Process CopyRight © Xiaodong Zhang

25. AS 1 AS 2 AS 3 AS 4 AS 5 AS 6 h1 s1 s2 s4 s5 s6 s3 h4 h3 h6 provider-to-customer edge peer-to-peer edge h1-h4 Close Relays Selection Process s1 close cluster s2 – 75 ms s5 – 180 ms s6, h6 – 220 ms s3, h3 – 52 ms s4 close cluster s2 – 50 ms s5 – 170 ms RTT h1-s2 + RTT h4-s2 = 125 ms < 300 ms s2 is good relay for h1-h4 VoIP session CopyRight © Xiaodong Zhang

26. bootstrap1 bootstrap2 surrogate SA end host h1 surrogate SB end host h2 Internet AS graph IP prefix to cluster Surrogate IP table IP prefix to ASN table cluster’s close cluster set Internet AS graph cluster’s top node table end host h3 cluster A cluster B cluster C control pkt. voice pkt. surrogate? surrogate IP surrogate? close set? close set close set? close set h2’s close set h2’s close set? voice pkts ASAP Call Session Process bootstrap’s data structure cluster surrogate’s data structure CopyRight © Xiaodong Zhang

27. Evaluation Metrics  Number of quality paths: number of relay paths satisfying the RTT and loss rate requirements  Shortest RTT and highest MOS of these quality paths  Overhead: measured by the number of generated messages to find quality path relay nodes

28. Different Routing Methods  DEDI: uses dedicated relay nodes. (SOSP ’ 01)  RAND: randomly selects relay nodes. (OSDI ’ 04)  MIX: is a combination of RAND and DEDI.  ASAP: selects relay nodes using our AS-aware method.  OPT: always chooses relay nodes that give the shortest overlay routing latency. (Offline method)

29. Number of Quality Paths For 90% sessions, ASAP can find more than 5,000 quality paths DEDI, RAND, and MIX can find no more than 500 quality paths for all sessions

30. Shortest Path RTT 115 ms In ASAP and OPT, all sessions have shorest RTT < 115 ms 1 s In DEDI, RAND, and MIX, more than 5% sessions have shortest RTT > 1s

31. Highest Path MOS 3.85 In ASAP and OPT, all sessions have highest MOSs > 3.85 2.9 In DEDI, RAND, and MIX, about 3% sessions have highest MOSs < 2.9

32. ASAP Is Highly Scalable 23,366 end hosts103,625 end hosts The number of quality paths found by ASAP remains stable under different end host population!

33. ASAP Has Moderate Overhead DEDI, RAND, and MIX all probe fixed number of nodes, i.e., 160, 160, and 200 nodes In ASAP, 85% sessions generate less than 300 messages

34. Problem Description  Is Internet direct routing enough for VoIP? No. 1% - 10% sessions do not meet the RTT quality requirement  Can overlay routing (relay) help improve? Yes. We can always find one-hop relay paths whose RTTs are below the threshold for these sessions.  Is existing overlay method including Skype enough? No. there are 4 limits.  Can we design an good relay selection method? ASAP

35. Conclusion  Their Experiments are fascinating. Intensive & Large Scale  ASAP is still a rough algorithm Details are missing Thank You!