Some Performance Issues Associated with CEWESMSRC Distributed - Memory ArchitecturesS

Views, opinions, and/or ndings contained in this report are those of the au-thor(s) and should not be construed as an oocial Department of the Defense position, policy, or decision unless so designated by other oocial documentation. Abstract Some typical performance issues associated with sparse matrix codes which use the Message Passing Interface and distributed memory machines will be discussed. In particular, average, sustained, oating point performance will be examined as a function of problem size and number of processors for an unstructured uid mechanics solver. These issues will be examined within the context of the iterative solution of nonsymmetric, linear systems of equations, such as those that arise from nite volume approximations to convection-diiusion problems. A brief description of the parallel CGSTAB (stabilized bi-conjugate gradient) iterative method will be followed by comparisons of its performance on the IBM SP, SGI Origin 2000, and Cray T3E at the Corps of Engineers Waterways Experiment Station Major Shared Resource Center (CEWES MSRC). In general, these machines have chips that are rated at several hundred MMops, but actual performance obtained in practice is much less. Sustained performance is a strong function of cache and communication performance. The aim of this report is not to determine which machine is \better" for certain classes of problems, but rather to illustrate performance trade-oos that users can expect to address in most problems run on these machines.