Performance analysis of the ARL Linux Networx Cluster

Within the past year, a 256-processor I686 Linux Cluster was installed at the Army Research Laboratory (ARL) Major Shared Resource Center (MSRC) to augment the center's current unclassified scientific application processing capabilities. The Purpose of this paper is to provide a comparative analysis of wall-clock-time performance of this system and the other unclassified HPC platforms currently installed at the ARL MSRC. A suite of vendor applications currently receiving significant utilization on the ARL platforms will be used to perform this analysis. The other existing unclassified HPC platforms at ARL include: an SGI 3800 with 512 processors, an IBM SP3 with 1024 processors, and an IBM SP4 with 128 processors. The following application codes will be used: CTH, CFD++, OVERFLOW, GAMESS, COBALT, LS_DYNA and FLUENT. Each code will be run using 16, 32, 48, 64, 80, 96, 112, and 128 processors. Using these timing metrics we will analyze the appropriateness of the 1686 architecture for use as a large-scale distributed computing platform for each of these scientific application codes.