Techniques for efficient execution of fine-grained concurrent programs

Abstract

Concurrent object-oriented programming languages are an attractive approach for programming massively-parallel machines. However, exploiting object-level concurrency is problematic as the linkage and communication overhead can overwhelm the benefits of the fine-grained concurrency. Our approach achieves efficient execution by tuning the grain size, matching the execution grain size to that efficiently supportable by the architecture. To verify the feasibility of grain-size tuning, we study the invocation locality of a collection of object-oriented programs. The results suggest that local constraints on placement combined with code specialization can produce a significant increase in execution grain size. We describe several compile-time analyses which identify opportunities to increase grain size. These analyses identify static relationships between objects and enable transformations to reduce invocation cost. Some initial measurements are presented.