Hierarchical clustering and k-means analysis of HPC application kernels performance characteristics

In this work, we present the characterization of a set of scientific kernels which are representative of the behavior of fundamental and applied physics applications across a wide range of fields. We collect performance attributes in the form of micro-operation mix and off-chip memory bandwidth measurements for these kernels. Using these measurements, we use two clustering methodologies to show which applications behave similarly and to identify unexpected behaviors, without the need to examine individual numeric results for all application runs. We define a methodology to combine metrics from various tools into a single clustering visualization. We show that some kernels experience significant changes in behavior at varying thread counts due to system features, and that these behavioral changes appear in the clustering analysis. We further show that application phases can be analyzed using clustering to determine which section of an application is the largest contributor to behavioral differences.