A View from inside a Distributed Internet Coordinate System

The Vivaldi algorithm is known to be one of the most interesting approaches for the calculation of Internet coordinates. It is a fully distributed, light-weight and adaptive algorithm. Recent studies show that host coordinates in the Vivaldi system are not stable and are drifting rapidly even when the network delays do not change. In this paper we observe that, despite the instability of Vivaldi coordinates in their absolute values, there is still a stable internal structure that can better reflect the stability of the underlying network. We proceed for this study by extensive simulations and experimentations. In a first stage, we confirm the fact that Vivaldi coordinates oscillate over time because of the adaptive nature of the system. The variations of these coordinates however are most of the time correlated with each other pointing to some stable cluster of nodes seen from inside the network. In a second stage, we present a new clustering algorithm based on the data mining Hierarchical Grouping method to identify this cluster of stable nodes once the network and the host coordinates reach their stationary regime. The metric that we use to cluster nodes in the system is the amount of variation of their Euclidean distances with respect to each other. Our main finding is that such stable cluster of nodes always exists and is grouping most of the nodes. We highlight the utility of such finding with an application that tracks changes in network delays. To this end, we propose to track a simple signal, which is the size of this biggest stable cluster.