The objective of this project is to develop algorithms for implementing a software distributed shared memory that reduce both the number of messages and the amount of data transferred between processors. A software distributed shared memory (DSM) enables processes to share memory even when they execute on processors that only communicate via message-passing. In recent years, the developers of tightly-coupled multiprocessors have adopted relaxed memory consistency models to reduce the latency of remote memory access. However, in a distributed memory multiprocessor, it is equally important to reduce the number of messages used for remote memory access. The principal investigator has developed an algorithm that implements a well known form of relaxed memory called Release Consistency. This algorithm, called Lazy Release Consistency, differs from most previous work because it is designed specifically for software DSM, rather then being an imitation of hardware implementations. Our preliminary results using trace-driven simulation indicate that Lazy Release Consistency significantly reduces both the number of messages and the amount of data transferred between processors. In order to reduce communication, Lazy Release Consistency performs more computation. To assess the impact of this trade-off, Lazy Release Consistency is being implemented in the Munin DSM and application program performance compared to previous work.
