Abstract

PC-CLUMP (Cluster of Multiprocessor) is one of the most cost-effective commodity-based platforms for HPC applications. The increasing number of CPUs per SMP node realizes very compact system size and very low price on the network interface per processor keeping the number of CPUs in the system. However, the performance of SMP-bus on such an SMPPC node is relatively poor compared with that of SMP-Workstations. Therefore, the application program must achieve a high cache-hit ratio to reduce SMP-bus access for higher performance. In this study, we analyze the performance of PCCLUMP based on the SMP-bus access ratio on several benchmarks as representatives of HPC applications. We measure the ratio of SMP-bus access per total memory access instructions utilizing the performance counters equipped with modern CPUs. From the result of analysis, we can estimate the best numbers of CPUs/node and nodes/system. Moreover, we also confirm the combination of parallel descriptions with OpenMP and MPI is very suitable for programming on PC-CLUMP while there exist several problems on data-process mapping which are related to cache-hit ratio of the applications. 1

Citations