An Optimal Processor Replacement Scheme for Efficient Communication of Runtime Data Redistribution

AbstractDynamic data distribution is used to enhance data locality and algorithm performance with reducing inter-processor communication in data parallel programs on distributed memory multi-computers. Since the exchange of data is performed at run-time, there is a performance tradeoff between the efficiency of the new data decomposition for a subsequent phase of an algorithm and the cost of exchanging data among processors. In this paper, we present an Optimal Processor Replacement (OPR) scheme to minimize data transmission cost for general BLOCK-CYCLIC data redistribution. The main idea of the proposed techniques is to employ a size oriented greedy matching method or a maximum bipartite matching theory for exploring an ideal one-to-one mapping between logical processors. Based on the matching policy, a realigned sequence of destination processors can be derived and is then used to perform data redistribution in the destination phase. A significant improvement of this approach is that OPR achieves the highest rate of data remain in local space and leading minimum inter-processor communication. The optimal processor replacement scheme can handle array redistribution with arbitrary source and destination distribution in BLOCK-CYCLIC type and can be applied to multidimensional arrays. To evaluate the performance of the proposed technique, we have implemented the OPR method on an SMP Cluster with 24 nodes. The theoretical analysis and experimental results show that our technique provides considerable improvement for runtime data redistribution.