REAL-TIME STATISTICAL CLUSTERING FOR EVENT TRACE REDUCTION

Event tracing provides the detailed data needed to understand the dyna mics of interactions among application resource demands and system res ponses. However, capturing the large volume of dynamic performance dat a inherent in detailed tracing can perturb program execution and stres s secondary storage systems. Moreover, it can overwhelm a user or perf ormance analyst with potentially irrelevant data. Using the Pablo perf ormance environment's support for real-time data analysis, we show tha t dynamic statistical data clustering can dramatically reduce the volu me of captured performance data by identifying and recording event tra ces only from representative processors. In turn, this makes possible low overhead, interactive visualization, and performance tuning.