Check out the new USENIX Web site.
Performance VisualizationRelated WorkProfiling Java WorkloadsStatistical Performance Analysis

Statistical Performance Analysis

Recent work uses statistical techniques to analyze performance counter data. Eeckhout et al. [17] analyze the hardware performance of Java programs. They use principal component analysis to reduce the dimensionality of the data from 34 performance counters to 4 principal components. Then they use hierarchical clustering to group workloads with similar behaviors. They gather only aggregate performance counts, and they divide all performance counter values by the number of clock cycles. Ahn and Vetter [1] hand-instrument several code regions in a set of applications. They gather data from 23 performance counters for three benchmarks on two different parallel machines with 16 and 68 nodes. Then they analyze that data using different clustering algorithms and factor analysis, focusing on parallelism and load balancing. We visualize behavior over time, require no instrumentation, and allow the analysis of any kind of derived metric.


Performance VisualizationRelated WorkProfiling Java WorkloadsStatistical Performance Analysis