Using sequence mining techniques for performance data

For this project, we attempted to evaluate the effectiveness of rudimentary sequence mining techniques for characterizing I/O trace data. We have taken trace information for a scientific application running on a cluster, and we have attempted to use K-Medoids based clustering algorithms to correlate particular trace sequences with phases of application execution. We also present a novel approach to sequence comparison, where we consider packets in sequences qualitatively based on high-level observations about the distribution of their durations. We show that our approach can reveal correlations between I/O sequences and phases of application execution, and that sequence mining techniques hold promise for adaptive performance monitoring.