JavaDSM is a distributed shared memory (DSM) system that allows Java threads to be executed on distributed nodes. Unlike other approaches, it offers shared memory view in Java virtual machine(JVM) level. This approaches is preferred because it needs no language supports and because it allows the use of Java object-semantics which helps optimize a DSM system for better performance.
In this paper, we present a design, an implementation and a preliminary performance evaluation of JavaDSM. Our objective is to determine the efficiency of a JVM level DSM system. Our result shows satisfactory speedups for the tested workloads like TSP, matrix multiplication and Jacobi relaxation. As compution-to-communication ratio increses we coud observe more speedups, and it means that with suitable networking technology, a DSM at JVM level is a viable technique for parallel processing in a distributed environment.
Other contiributions for the speedups include the replication of shared objects in local memory and the use of the lazy release consistency(LRC) memory model. Exchange of consistency related informations are limited only between a latest lock releaser and the current acquirer of the lock. This optimization entails a significant performance gain compared with other approaches.
