中文摘要：

为了研究GPU的通用计算能力和适合SMP集群的编程模型,首次提出MPI＋CUDA多粒度混合并行编程的新方法,节点间采用MPI实现粗粒度并行,节点内采用CUDA实现细粒度并行的混合编程方式。利用此方法在搭建的3节点SMP集群环境中,测试了大规模矩阵乘问题的并行计算能力。实验结果表明,该方法能够显著提升并行效率,同时证明MPI＋CUDA混合编程模型能够充分发挥SMP集群节点间分布式存储和节点内共享内存的优势,为装有CUDA-enabled GPU的SMP集群提供了一种有效的并行策略。

英文摘要：

In order to study the general computing ability of GPU and programming model suitable for SMP clusters,a method of MPI＋ CUDA multi-grain hybrid parallel programming is proposed,which implemented the coarse grain parallel programming between the nodes and the fine grain parallel programming in the node using MPI and CUDA respectively.A SMP clusters environment is built with 3 computers,and the large-scale matrix multiplication with multi-grain hybrid parallel programming is achieved.It is indicated that MPI＋ CUDA hybrid parallel programming is an effective parallel strategy for the parallel computing in the SMP clusters with CUDA-enabled GPU,which exerts the advantages of the share memory in the node and the distributional memory between the nodes in the SMP clusters.This method promote parallel efficiency remarkably.