PNNL

Abstract

We propose a technique to optimize the performance of applications using distributed dense arrays and characterized by a nearest-neighbor communication profile by exploiting the topology of SMP clusters. The topological information is exploited for mapping array tiles to processors to reduce the network communication and improve utilization of shared memory for inter-process communication. The potential benefits of using the SMP-aware mapping are demonstrated through a simulation, as well as a real application solving a wind-driven ocean circulation model on an IBM SP. On 128 processors, the execution time was reduced by almost 20 percent without any changes to the original application source code. The proposed mapping approach is applicable to multiple programming models and distributed array management systems.