Statistical Performance Analysis for Scientific Applications

As high-performance computing (HPC) heads towards the exascale era, application performance analysis becomes more complex and less tractable. It usually requires considerable training, experience, and a good working knowledge of hardware/software interaction to use performance tools effectively, which becomes a barrier for domain scientists. Moreover, instrumentation and profiling activities from a large run can easily generate gigantic data volume, making both data management and characterization another challenge. To cope with these, we develop a statistical method to extract the principal performance features and produce easily interpretable results. This paper introduces a performance analysis methodology based on the combination of Variable Clustering (VarCluster) and Principal Component Analysis (PCA), describes the analysis process, and gives experimental results of scientific applications on a Cray XT5 system. As a visualization aid, we use Voronoi tessellations to map the numerical results into graphical forms to convey the performance information more clearly.