Supporting structured shared virtual memory under Mach

Abstract

Distributed Shared Virtual Memory (or DSVM) has been proposed by several researchers as a method for simulating shared memory on a loosely coupled multicomputer. The principle advantage of the DSVM approach is that it provides a shared memory programming model that is familiar to many users, and this simplifies the migration of applications from conventional machines to the loosely coupled environment. However, for use in high-performance parallel processing, the standard DSVM model has several fundamental disadvantages that keep it from performing as well as traditional message-passing models of parallel computation. In this paper, we present a new version of shared virtual memory, called Structured Shared Virtual Memory or SSVM, that overcomes these difficulties. We discuss the implementation of SSVM under the Mach operating system and we present preliminary performance comparisons between this implementation of SSVM and the standard implementation of DSVM under Mach.