Gossip-based Unstructured Overlay Networks: An Experimental Evaluation

Gossip-based protocols offer a scalable and reliable approach to a number of large-scale distributed applications. The basic idea is for each node to periodically select a random peer node to exchange information with. Analytical studies reveal a high reliability of gossip-based protocols. However, a usual assumption of these studies is that the peer is chosen uniformly at random from the set of all nodes. In practice—instead of requiring all nodes to know all the peer nodes— a scalable way to implement random peer selection is by constructing dynamic unstructured overlays through gossiping membership information itself. In this paper we generalize existing gossip-based overlays by introducing a general scheme in which existing overlays as well as novel protocols can be implemented. The central theme of this paper is exploring and comparing several implementations of our abstract scheme. Through extensive experimental analysis, we show that all of them lead to different communication topologies none of which is uniformly random. This clearly renders traditional theoretical approaches invalid. Our observations help explain important differences between design choices of a gossipbased protocol and how these impact the reliability of the overlay. Understanding these differences poses new interesting theoretical problems. . This work was partially supported by the Future & Emerging Technologies unit of the European Commission through Project BISON (IST-2001-38923). . University of Bologna, Italy . EPFL, Lausanne, Switzerland . Microsoft Research, Cambridge, UK . Vrije Universiteit, Amsterdam, The Netherlands