Exploiting Parallelism in the Design of Peer-to-Peer Overlays

Abstract

Many peer-to-peer overlay operations are inherently parallel and this parallelism can be exploited by using multi-destination multicast routing, resulting in significant message reduction in the underlying network. We propose criteria for assessing when multicast routing can effectively be used, and compare multi-destination multicast and host group multicast using these criteria. We show that the assumptions underlying the Chuang-Sirbu multicast scaling law are valid in large-scale peer-to-peer overlays, and thus Chuang-Sirbu is suitable for estimating the message reduction when replacing unicast overlay messages with multicast messages. Using simulation, we evaluate message savings in two overlay algorithms when multi-destination multicast routing is used in place of unicast messages. We further describe parallelism in a range of overlay algorithms including multi-hop, variable-hop, load-balancing, random walk, and measurement overlay.