Redundant computation partition on distributed-memory systems

Maximizing parallelism and minimizing communication overheads are important issues on distributed-memory systems. This paper presents a communication optimization technique named redundant computation partitioning, and also gives its implementation. The main idea is to select computation redundancy, represented by a redundant vector, properly for each partitioned loop nest in a parallel loop sequence, so as to acquire a larger parallel region. The technique can eliminate inter-loop communications which will be introduced using ordinary computation partitionings. Some data redistribution points may be eliminated, so the global deduction of data and computation partition can be simplified.