Workshop Report for IPTPS'02: 1st International Workshop on Peer-to-Peer Systems

Attendees were welcomed to the Workshop by Frans Kaashoek and Peter Druschel who reported that although the original plan had been to invite 35 people to a 1 1 2 day event, 99 papers had been submitted. The workshop had therefore been scaled up to 50 authors with 32 position papers, 26 of which would be presented over a two day program. Scott Shenker gave a brief overview of the problem space, noting that a great deal of work was being done on Distributed Hash Tables (DHTs) that were location independent routing algorithms. Given a key, they would find the right node. Recent advances in high-dimension routing research guarantee a sub-linear number of hops to locate nodes. However, they assume a uniformity of mesh that does not map well to the Internet. Brocade extends Tapestry to address this by eliminating hops across the wide area and by preventing traffic going through congested stub nodes. " Super-nodes " are spotted within connected domains and a secondary overlay layer is then used to transfer traffic between these super-nodes. Brocade is not useful for intra-domain messages, so it is necessary to classify traffic by destination. Old results for locality are cached and the super-node keeps a " cover net " , an authoritative list of which nodes are local. The local super-nodes are found by snooping peer-to-peer traffic or by consulting the DNS. The super-nodes find each other by using Tapestry, which has built-in proximity metrics. The super-nodes do not need to be incredibly powerful. Simulations show that with 220M hosts, 20K AS's and 10% of the super-nodes " coming and going " then about 9% of the super-nodes' CPU time is spent dealing with changes. Publishing data about the changes uses about 160K bits/sec in bandwidth, and super-nodes need only about 2MB of storage. Simulations also show that the scheme halves delay for short hops and improves overall bandwidth usage.