Exploring multi-core limitations through comparison of contemporary systems

Abstract

Our research aims at evaluating the performance of contemporary multi-core platforms, the quad-core AMD Barcelona and the dual-core Intel Woodcrest, for scientific applications. We used the High Performance Computing Challenge (HPCC) benchmark suite, which contains test cases for quantifying performance of the memory and communication sub-systems. Analyzing the performance of multi-core devices allows us to identify system parameters and configurations that could yield optimal performance. Since these devices are building blocks of the fastest supercomputing systems, the insights into performance behavior provide guidance for designing future high performance computing resources. Our results demonstrate that without multi-core aware optimization the achievable performance decreases as the number of cores per socket increases.