Global partial replicate computation partitioning
Yiran Wang, Li Chen, Zhao-Qing Zhang · Proceedings of the International Conference on Parallel Processing · 2004
Early parallelizing compilers use the owner-computes rule to partition computation. Partial replication is then introduced to eliminate near-neighbor communication at the cost of some replicated computation, hence improves the performance and scalability. Current exploration of partial replicate computation partitioning is limited within a single loop nest. We present a formal description of the global partial replicate computation partitioning problem, a simplified cost model and a heuristic solution. Experimental results show that the solution is superior to local approaches.