A Comparative Study of ARL Linux Cluster Performance

Since 2003, with the installation of a 256 processor Linux Networx XEON 1686 system (Powell), Army Research Laboratory (ARL) has been providing large- scale Linux cluster production systems for use within the Department of Defense (DoD) High Performance Computing Modernization Program (HPCMP). This initial system was followed in 2004 with the 2,048 processor Linux Networx XEONEM64T (JVN) and 2,304 processor IBM Opteron (Stryker) systems, and in 2007 by the 3,368-core Intel Dempsey (Humvee) and 4,488-core Intel Woodcrest (MJM) systems. These latest systems provide an increased peak performance of over 15 times the original Xeon 1686 system in a four year period. The purpose of this paper is provide a comparative study of the three generations of Linux clusters' capabilities to process some of the most widely used production codes used within the ARL environment. The codes to be benchmarked will include CTH, CFD++, GAMESS, and OVERFLOW. The codes will be run on Powell, JVN and Ping (the Woodcrest testbed machine). The results will focus attention on how architecture enhancements (including CPU speed, memory per node, cache size, and interconnect fabric transfer rates) have affected the overall performance of these systems.