Hybrid MPI+UPC parallel programming paradigm on an SMP cluster

摘要

The symmetric multiprocessing (SMP) cluster system, whichconsists of shared memory nodes with several multicore centralprocessing units connected to a high-speed network to form adistributed memory system, is the most widely available hardwarearchitecture for the high-performance computing community. Today, the Message Passing Interface (MPI) is the most widely used parallelprogramming paradigm for SMP clusters, in which the MPI providesprogramming both for an SMP node and among nodes simultaneously.However, Unified Parallel C (UPC) is an emerging alternative thatsupports the partitioned global address space model that can be againemployed within and across the nodes of a cluster.In this paper, we describe a hybrid parallel programming paradigm that was designed to combine MPI and UPC programming models. Thisparadigm's objective is to mix the MPI's data locality control andscalability strengths with UPC's fine-grain parallelism and ease ofprogramming to achieve multiple-level parallelism at the SMP cluster,which itself has multilevel parallel architecture. Utilizing aproposed hybrid model and comparing MPI-only to UPC-onlyimplementations, this paper presents a detailed description ofCannon's algorithm benchmark application with performance results of a random-access benchmark and the Barnes--Hut N-Body simulation.Experiments indicate that the hybrid MPI+UPC model can significantlyprovide performance increases of up to double in comparison withUPC-only implementation, and up to 20% increases in comparison toMPI-only implementation. Furthermore, an optimization was achievedthat improved the hybrid performance by an additional 20%.