Optimizing Communication Using Global Dataflow Analysis

In distributed-memory message passing architectures redu cing communication cost is extremely important. In this paper, we present a technique to improve communicati on globally. Our approach is based on a combination of linear algebra framework and dataflow analys is, and can take arbitrary control flow into account. The distinctive feature of the algorithm is its acc ura y in keeping communication set information, its support for general alignments and distributions inclu ding block-cyclic distributions, and capability of simulating some of the previous approaches by appropriate m odifications. The method is currently being implemented in the PARADIGM compiler. We show how optimizat ions such as message vectorization, message coalescing, redundancy elimination can be supported b y our new framework. Experimental results on an IBM SP-2 show that our technique is effective in reducing b oth the number as well as the volume of the communication.