A Heuristic Rule of Partitioning Irregular Loop for Parallelizing Compilers

Abstract

For irregular applications on distributed-memory systems, computation partition is an important issue on parallel compiling techniques of parallelizing compilers. In this paper, we propose a local optimal solution, called heuristic computes rule (HCR), which could be used for irregular loop partitioning. This rule considers both the iteration being partitioned and the iterations partitioned, which ensures that iterations are assigned so as to produce less communication costs. And HCR rule proposes that irregular loop partitioning should trade off the maximum message degrees of processors, the number of messages, the message sizes, and workload balance. In our experiments, we compare HCR with almost owner computes rule and least communication computes rule. The results show that the executing of irregular loop partitioned by HCR rule has much less communication cost and achieve better performance.