Critical analysis of HPL performance under different process distribution patterns

Abstract

High Performance Linpack (HPL) benchmark is a standard tool to measure the performance of High Performance Computing Clusters (HPCC). There are several parameters which can be optimized to maximize performance delivered by HPCC. HPL provides many such parameters for evaluation of performance. P & Q are two such values which determine the distribution pattern of processes for executing HPL. In this paper HPL runs with different values of P and Q have been carried out for three different values of N while keeping the block size (NB) constant. Value of N represents the problem size of task being executed. Results obtained are included in tables and are also depicted in form of graph and are analytically discussed. These results also provide the basis for deeper analytical approaches for HPCC performance evaluations. This critical analysis also reveals that maximum performance is delivered if the task is evenly distributed over all the available cores and nodes in the HPCC.