A UNIFIED FRAMEWORK FOR OPTIMIZING COMMUNICATION IN DATA-PARALLEL PROGRAMS

This paper presents a framework, based on global array data-flow analy sis, to reduce communication costs in a program being compiled for a d istributed memory machine. We introduce available section descriptor, a novel representation of communication involving array sections. This representation allows us to apply techniques for partial redundancy e limination to obtain powerful communication optimizations. With a sing le framework, we are able to capture optimizations like 1) vectorizing communication, 2) eliminating communication that is redundant on any control flow path, 3) reducing the amount of data being communicated, 4) reducing the number-of processors to which data must be communicate d, and (5) moving communication earlier to hide latency, and to subsum e previous communication. We show that the bidirectional problem of el iminating partial redundancies can be decomposed into simpler unidirec tional problems even in the context of an array section representation , which makes the analysis procedure more efficient. We present result s from a preliminary implementation of this framework, which are extre mely encouraging, and demonstrate the effectiveness of this analysis i n improving the performance of programs.