ANU Computer Science Technical Reports

TR-CS-08-03

Jie Cai, Alistair P. Rendell, Peter E. Strazdins, and H'sien Jin Wong.
Predicting Performance of Intel Cluster OpenMP with Code Analysis Method.
November 2008.

[POSTSCRIPT (276423 bytes)] [PDF (283214 bytes)]

Abstract: Intel Cluster OpenMP (CLOMP) extends OpenMP programs onto clusters by using page-based software Distributed Shared Memory (sDSM) system. With the commercial release of CLOMP, the interest in cluster-enabled OpenMP systems is likely to increase. Page faults detection and servicing are the major overheads of such systems. This paper presents a code analysis method to estimate the number of page faults caused when running an OpenMP Program with CLOMP. Then, utilizing SDP model to predict performance of CLOMP. On a 4-node Intel cluster connected via both InfiniBand (IB) and Giga-Ethernet (Giga-Eth) interconnects, some NAS Parallel Benchmarks (NPB) are used to evaluate and predict CLOMP. The estimates show less than 10% prediction error on most benchmarks, which indicates that the code analysis method-SDP model is an effective approach to predict performance of CLOMP.

Technical Reports <Technical-DOT-Reports-AT-cs-DOT-anu.edu.au>

Last modified: Tue May 31 12:56:01 EST 2011