Eliminating Redundant Communication of Code Generation for Distributed Memory Machines

How to eliminate redundant communications is a key issue in developing parallelizing compilers for distributed memory machines. An important way to improve the performance of the code generated by parallelizing compiler in distributed memory systems is to eliminate as much communications as possible. This paper investigates algorithms for generating communications code from previously published work and modifies them to eliminate redundant communications. Based on the mathematical model of symbolic coefficient linear inequalities the revisited formal specification of communications sets are presented and the corresponding improved algorithms are given. Experimental results indicate that the new algorithms can reduce communication messages and improve the performance of parallelized code significantly.