Performance Improvement by Overlapping Computation and Communication on SMP Clusters

Clusters of SMPs (Symmetric Mul-tiprocessor Systems) have emerged as important platforms for high performance computing. As a programming scheme on SMP clusters, we proposed hybrid shared memory/distributed memory programming. In order to tolerate inter-node communication , we overlapped the inter-node communication and computation using remote memory based user-level communication primitives NICAM, which we designed on our SMP cluster COMPaS. In this paper, we report on the programming and its performance of COMPaS by overlapping communication and computation. Our experimental results show that communication is almost hidden and the execution time is reduced 25% in the best case.