Holistic Hardware Counter Performance Analysis of Parallel Programs

Automatic Performance Analysis

计算机科学;物理学

Brian J. N. Wylie ; Bernd Mohr ; Felix Wolf

Others  :  http://drops.dagstuhl.de/opus/volltexte/2006/507/pdf/05501.WylieBrian.Paper.507.pdf
PID  :  6652

学科分类：计算机科学（综合）

来源: CEUR

【 摘 要 】

The KOJAK toolkit has been augmented with refined hardware performance counter support, including more convenient measurement specification, additional metric derivations and hierarchicalstructuring, and an extended algebra for integrating multiple experiments. Comprehensive automated analysis of a hybrid OpenMP/MPI parallel program, the ASC Purple sPPM benchmark, is demonstrated with performance experiments on equisized POWER4-II-based IBM Regatta p690+cluster, Opteron-based Cray XD1 cluster and UltraSPARC-IV-based Sun Fire E25000 systems. Automatically assessed communication and synchronisation performance properties, combined with a rich set of measured and derived counter metrics, provide a holistic analysis context and facilitate multi-platform comparison.