Distributed Hierarchy of Clusters in the Presence of Topological Changes

Abstract

We propose an algorithm that builds a hierarchical clustering in a network, in the presence of topological changes. Clusters are built and maintained by random walks, that collect and dispatch information to ensure the consistency of clusters. We implement distributed communication primitives allowing clusters to emulate nodes of an overlay distributed system. Each cluster behaves like a virtual node, and executes the upper level algorithm. Those primitives ensure that messages sent by a cluster are received and treated atomically only once by their recipient, even in the presence of topological changes. Decisions concerning the behavior of the cluster (virtual node for the higher level algorithm) are taken by the node that owns the random walk at this time. Based on this abstraction layer and the overlay network it defines, we present a distributed hierarchical clustering algorithm, aimed at clustering large-scale dynamic networks.