A Decentralized Network Coordinate System for Robust Internet Distance Prediction

Network coordinate systems have recently been developed as a scalable mechanism to predict latencies among arbitrary Internet hosts. Our research addresses several design challenges of a large-scale decentralized network coordinate system that were not fully addressed in prior work. In particular, we examine the design issues of a decentralized network coordinate system operating in a peer-to-peer network with high churn, high fractions of faulty or misbehaving peers, and high degrees of network path anomalies. This paper presents a fully decentralized network coordinate system, PCoord, for robust and fault-tolerant Internet distance prediction. Through extensive simulations, we examine the convergence behavior and prediction accuracy of PCoord under a variety of scenarios, and compare its performance with an existing network coordinate system, Vivaldi. Our results indicate that PCoord is robust under high churn, and degrades gracefully even under high fractions of faulty nodes, and high degrees of triangle inequality violations in the underlying network distances. Finally, our results indicate that even under an extremely challenging flash-crowd scenario where 1740 hosts simultaneously join the system, PCoord is able to converge to 12% median relative prediction error within 10 seconds.