大规模3D并行分层可扩展矩阵乘法的递阶优化方法

发布时间：2019-04-04 20:49

【摘要】：为提高大规模平台上可扩展矩阵乘法的并行计算效率,提出一种并行分层可扩展矩阵乘法的递阶优化方法。首先,在可扩展矩阵乘法算法(SMM)算法枢轴行和枢轴列通信研究基础上,利用分层方式在更高等级上对网格进行矩形群划分,实现矩阵乘法的二维计算向三维计算转变,并设计对应的集群内通信和集群间通信过程,实现SMM乘法的递阶并行优化(HSMM);其次,对所提HSMM算法进行理论分析,分情况对其通信成本进行分析和预测,推导出最佳计算成本的集群数选取方式;最后,通过在Grid5000和BlueGene/P测试平台实验,显示所提算法在执行时间和通信时间指标上均要优于对比算法,验证了所提算法有效性和理论分析的正确性。
[Abstract]:In order to improve the parallel computing efficiency of scalable matrix multiplication on large-scale platforms, a hierarchical optimization method for parallel hierarchical scalable matrix multiplication is proposed. Firstly, on the basis of the research on the communication between pivot row and pivot row of (SMM) algorithm, the grid is divided into rectangular groups at a higher level by using hierarchical method, and the two-dimensional computation of matrix multiplication is transformed into three-dimensional computation, which is based on the research of the communication between pivot row and pivot row of the extended matrix multiplication algorithm. The corresponding intra-cluster communication and inter-cluster communication process are designed, and the hierarchical parallel optimization (HSMM); for SMM multiplication is realized. Secondly, the proposed HSMM algorithm is analyzed theoretically, and the communication cost is analyzed and predicted according to the situation, and the optimal selection method of cluster number is deduced. Finally, experiments on Grid5000 and BlueGene/P test platform show that the proposed algorithm is superior to the contrast algorithm in terms of execution time and communication time. The validity of the proposed algorithm and the correctness of the theoretical analysis are verified.
【作者单位】：中山火炬职业技术学院信息工程系;广东外语外贸大学思科信息学院;
【分类号】：TP338.6

【参考文献】