Cristian Lumezanu Dave Levin Neil Spring

1. PeerWise Discovery and Negotiation of Faster Paths Cristian Lumezanu Dave Levin Neil Spring

2. Routing Overlays Why are routing overlays not standard practice? Potential BENEFIT is HIGH Deployment COST is LOW PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

3. Scalability? C B A Probing should be done more selectively PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

4. Fairness? cost benefit > 1 cost benefit < 1 Routing overlays should include an incentive mechanism PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

5. PeerWise PeerWise Nodes negotiate and establish pairwise connections to each other Nodes that can help each other find better paths peer Motivation Models autonomous system peerings in the Internet Overlays built on self-interest rather than altruism Cost-benefit ratio known before committing any resources PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

6. PeerWise 50ms D B 40ms 30ms A C 60ms PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

7. PeerWise properties FAIR SCALABLE BENEFICIAL PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

8. PeerWise properties FAIR SCALABLE BENEFICIAL PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

9. Is mutual advantage common? G E B D F A C PEERING SCORE Difference between the number of routes each node uses the other for after all peerings have been established EXPERIMENT SETUP • PeerWise prototype with global knowledge • 256 DNS server data set gathered using the King method • Each node sends data to all other nodes PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

10. Is mutual advantage common? 100 74 % of pairs 74% of the pairs of nodes are happy with existing peerings 0 20 peering score 100   PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

11. PeerWise properties SCALABLE BENEFICIAL FAIR PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

12. Finding shorter detours At least 66% of the pairs of nodes in our datasets are long sides in TIVs B 10ms 10ms A C 100ms Triangle inequality violations indicate the existence of shorter one-hop detours flaws in We use network coordinates to find triangle inequality violations PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

13. Network coordinates and TIVs Internet Metric space B B 38ms 20ms 39ms 42ms C A C 62ms A 26ms error(AC) = -36ms TIVs allowed AC > AB + BC No TIVs AC < AB + BC Long sides shrink Sum of short sides grows PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

14. Embedding errors and TIVs If long sides shrink The more negative the embedding error of an edge, the higher the probability that the edge is a long side in a TIV …and thus has a one-hop shorter detour If sum of short sides grows The more positive the embedding error of an edge, the higher the probability that the edge is a short side in a TIV …and thus is part of a one-hop shorter detour PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

15. PeerWise properties SCALABLE BENEFICIAL FAIR PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

16. Performance best one-hop path direct path 100 PeerWise path % of pairs PeerWise reduces latency by an average of 20% 0 100 Latency (ms) 300 PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

17. Conclusions SCALABLE Selective edge monitoring using network coordinates BENEFICIAL Alternate faster paths discovered from TIVs FAIR Pairwise peerings based on mutual advantage PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

18. Future Work Future Work Why is there mutual advantage? Extensions for low-loss and failure-free paths Deployment THANK YOU! PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

19. Are one-hop detours enough? direct path best one-hop detour path 100 3% of detour paths are longer than 100ms 75 25% of direct paths are longer than 100ms % of pairs 50% improvement in average latency 0 100 300 Latency (ms) PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

20. Connectivity 100 limited knowledge (32 neighbors) 60% of nodes can reach more than 60% of their destinations, with limited knowledge global knowledge % of nodes 40 22 78% of nodes can reach more than 60% of their destinations, with global knowledge 0 detour score 60 100 Percentage of destinations that a node can reach using its peerings, out of all reachable destinations PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

21. Embedding errors and TIVs The more negative the embedding error of an edge, the higher the probability that the edge is a long side in a TIV …and thus has a one-hop shorter detour The more positive the embedding error of an edge, the higher the probability that the edge is a short side in a TIV …and thus is part of a one-hop shorter detour TIV SCORE Percentage of how many times a pair of nodes forms a long side in a TIV out of total number of presences in TIVs PeerWise Discovery and Negotiation of Faster Paths HotNets 2007

22. Embedding errors and TIVs 100 As the estimation error becomes more negative the nodes form more and more long sides more long sides TIV score 50 more short sides 0 -200 0 200 embedding error estimated distance – real distance PeerWise Discovery and Negotiation of Faster Paths HotNets 2007