Performance Prediction and Network Media Selection for PVM Clusters

Introduction In general, a parallel computer is a computer that has multiple processors connected by a communication network and is capable of using all its processors simultaneously to solve a single problem [34]. Parallel computers have the computational power necessary to solve many of computational science's "grand challenge" problems [27, 34]. However, it can be quite difficult to write programs that perform well on parallel machines. The research proposed here is concerned with performance prediction as a programming and performance debugging tool for parallel computers. 1.1 Parallel Programming Paradigms To solve a problem on a parallel computer, the programmer divides the problem into pieces which are then distributed across the processors of the machine. The problem may be divided according to its functions (control or task parallelism) or its data (data parallelism). Fo