Hypernets -- Good (G)news for Gnutella

Abstract

Criticism of Gnutella network scalability has rested on the bandwidth attributes of the original interconnection topology: a Cayley tree. Trees, in general, are known to have lower aggregate bandwidth than higher dimensional topologies e.g., hypercubes, meshes and tori. Gnutella was intended to support thousands to millions of peers. Studies of interconnection topologies in the literature, however, have focused on hardware implementations which are limited by cost to a few thousand nodes. Since the Gnutella network is virtual, hyper-topologies are relatively unfettered by such constraints. We present performance models for several plausible hyper-topologies and compare their query throughput up to millions of peers. The virtual hypercube and the virtual hypertorus are shown to offer near linear scalability subject to the number of peer TCP/IP connections that can be simultaneously kept open.