Distributed shared virtual memory on RISC System/6000 clusters and large scale computations: Two case studies

Abstract

Many of today's moderately parallel and massively parallel multicomputers are non-shared memory systems, using the message passing programming paradigm for parallel applications. This is significant, since the shared memory paradigm is a well established programming paradigm on high performance computers, like the vector multiprocessors (vmp). Distributed shared virtual memory (dsvm) provides a way of using the shared memory programming paradigm on a set of loosely coupled computer systems. The work to date suggests that one can implement some moderately parallel applications on dsvm loosely coupled systems and achieve performances comparable to those on message passing non-shared memory systems. This paper reports first on DSVM6K, a dsvm system implemented in the AIX version 3 operating system on IBM's RISC System/6000, and then describes MTF6K, a prototypal implementation of MultiTasking Fortran on a cluster of RISC/6000 engines: MTF is a component of the IBM VS FORTRAN library which provides FORTRAN programmers with coarse grain, explicit multitasking primitives. In the last two sections of the paper two large scale applications are analyzed, a large scale 2D seismic modeling application and a 3D aerodynamic simulation. © 1992.