Random graphs as models of hierarchical peer-to-peer networks

This paper proposes the development and application of random graphs-based performance evaluation techniques to understand design trade-offs for hierarchical unstructured peer-to-peer networks. In particular, the connections between lower and higher level peers (that are known as leaves and ultra-peers in the Gnutella jargon) are modeled as a bipartite random graph while the overlay network used by ultra-peers to forward queries is modeled as a generalized random graph. Both the random graph models consider peers of either level as partitioned into classes; this feature is included in the model description to consider the mismatch between the logical topology of the application and the physical deployment of peers throughout the Internet. To assign realistic values to the input model parameters and to validate the model predictions we obtained snapshots of the Gnutella application topology at both levels and conducted simulation experiments on these snapshots. The paper highlights a few exploitations of the modeling technique with a particular focus on the evaluation of the impact of locality awareness on user and network performance measures.