Self-similarity in Message Passing Parallel Processing Communication

Communication performance analysis of message passing parallel programs relies on accurate modeling of cluster network traffic patterns. To this end we have developed a tool for collecting traffic samples by simulating parallel processing environment with statistically tractable network traffic. Our results from experiments show network packet traffic generated by messagepassing parallel programs in a Beowulf cluster setting exhibits strong self-similarity, which was tested using Hurst parameters. As most known Hurst parameter estimators are heuristics, we propose a novel approach for statistical inference via parametric bootstrapping, and show confidence intervals of the estimated parameters. The empirical results demonstrate the usefulness of the traffic generation tool as well as the Hurst parameter interval estimation method in the investigation of parallel processing communication performance.