Parallel/Distributed Computing on Clusters of High Performance Risc Workstations

Abstract

Shared—memory vector multiprocessors (VMP) and parallel extensions to FORTRAN offer the most common hardware/software solution to the problem of reducing the elapsed time in large scale computations on numerical grids. Topical kernels in such computations are nested DO—loops in which the outer iterations can be distributed to multiple processors while inner iterations can be dealt with by pipelined vector hardware. An additional level of parallelism can be introduced by grouping VMPs in clusters and by extending FORTRAN to allow a single application to exploit multiple (distributed memory) computers.

The combination in a UNIX environment of high performance RISC machines with high speed interconnect technology constitutes a cost—effective platform of potential interest for application people to explore loosely coupled computing. A short overview of some of the parallel programming environments —mainly based on the message passing programming paradigm—currently in use at ECSEC on such platform is provided.