Multi-threaded Design for a Distributed Shared Memory System

Abstract

this paper fescribes the design and implementation of a multi-threaded Distributed Shared Memory (DSM) system, called Cohesion, which provides high programming flexibility and latency masking, and supports load redistribution. Cohesion differs from many other DSM systems by providing a variable shared address space, instead of a shared variable model, in which user and system data can be located. This address space is partitioned into regions, each maintained by a distinct mermory coherent protocol. User and system data with similar behaviors can placed in the same regions and maintained by a protocols that suits their nature, thereby improving the system performance. The design issues such as synchronization and thread management have been reconsidered to support multi-threading. The results of Successive-Over-Relaxation and Quick Short on Cohesion show that multi-threading has better performance than single-threading since communication can be overlapped with computation.