Combining Dependence and Data-Flow Analyses to Optimize Communication

Reducing communication overhead is crucial for improving the performance of programs on distributed-memory machines. Compilers for data-parallel languages must perform communication optimizations in order to minimize this overhead. In this paper, we show how to combine dependence analysis, traditionally used to optimize regular communication, and a data-flow analysis method originally developed to improve placement of irregular communication. Our approach allows us to perform more extensive optimizations --- message vectorization, elimination of redundant messages, and overlapping communication with computation. We also present preliminary experimental results that demonstrate the benefits of the proposed method. 1 Introduction Distributed-memory machines are becoming widely accepted as the most promising way to achieve high performance computing. They scale to hundreds, and even thousands of processors, thus providing unmatched computing power. However, harnessing this power is an ex..