分布式并行矩阵乘算法分析

分布式并行矩阵乘算法包括简单并行分块乘、Cannon、Fox、矩阵乘并行、B-S、B迁移和C迁移算法.并行分块乘法在同行处理器间进行多到多播送.Cannon算法可有目的地在各行和各列施行循环移位.Fox算法的行处理器施行一到多播送,列处理器施行循环单步上移.矩阵乘并行算法采取中子矩阵分配策略对各对应子矩阵相乘.B-S通过消息传递、进行数据迁移.B迁移算法可减少通讯量.C迁移算法可提高并行效率.

Analysis of Distributed and Parallel Matrix Multiple Algorithms

Distributed and parallel matrix multiplication algorithms consists of a simple parallel partitioned algorithm, Cannon algorithm, Fox algorithm, matrix multiplication parallel, A B-S algorithm, B-Shift algorithm and a C-Shift algorithm. For he parallel partitioned algorithm, data is sent from several processors to several processors in the same rows. The shift is moved circularly in all the rows and columns for Cannon algorithm. A row processor of Fox algorithm sends data from a processor to lots of processors, and column processor moves up shift circularly. Matrix multiplication parallel algorithm is that submatrix multiplies corresponding submatrix based on diagonal partition strategy of parallel algorithm takes the strategy of distributed small matrixes. The data is moved through sending messages for B-S algorithm. The communication quantity is effectively reduced with B-Shift algorithm, and the performance is improved for C-Shift algorithm.