This research effort focuses on an integrated compiler-runtime approach to support software distributed shared memory (DSM). The goal is to achieve performance close to hand-tuned message passing programs while still presenting the user with a shared memory programming model. The work is primarily directed towards "networks of workstations", but it is also applicable to dedicated distributed memory machines like the IBM SP-2. Notwithstanding advances in the design and implementation of software DSM systems over the last few years, message passing has remained the most commonly used method for programming distributed memory machines, because message passing systems, in general, perform better than DSM systems. This proposal addresses this issue by building an integrated compiler-runtime DSM system, including the optimization of explicitly parallel shared memory programs through compiler analysis. Specifically, the ability of the compiler to predict access patterns, rather than merely react to them as the runtime system does, will allow prefetching and aggregation of messages, and will avoid unnecessary data movement and consistency actions. The work will be based on the TreadMarks DSM system developed at Rice. The proposed work will contribute to the better utilization of networks of workstations and, more generally, distributed memory machines for parallel computing. Earlier work on TreadMarks has already contributed to that goal, in that it has produced one of the most advanced runtime systems for DSM, that has been used by other researchers for work in computational genetics, mixed integer programming, and seismic modeling. ***
