摘要： 基于Householder转换,给出了稠密对称矩阵三对角化的MPI+OpenMP混合并行算法。内容集中在SMP集群系统环境下算法的负载平衡、通信开销和性能评价。OpenMP共享内存并行采用了粗粒度方法,解决了MPI算法中的负载平衡问题,降低了通信开销。在深腾6800上的试验结果表明,MPI+OpenMP版本比纯MPI版本具有更好的性能和可扩展性。
关键词:
对称矩阵,
并行块算法,
三对角矩阵,
加速比,
MPI+OpenMP
Abstract: Based on Householder transformation,MPI+OpenMP hybrid parallel algorithm of tridiagonalizing a dense symmetric matrix is presented.This paper focuses on load balancing,communication overhead and performance evaluation on cluster of SMP’s environments.An OpenMP shared memory parallel approach is coarse grain,this overcomes the load imbalance problem,reduces communication overhead.An experimental analysis on the Deepcomp6800 shows the hybrid MPI+OpenMP algorithm has better performance and scalability than pure MPI algorithm.
中图分类号:
赵永华,迟学斌,陈江. 对称矩阵三对角化的混合并行算法设计[J]. , 2005, 31(22): 49-.
ZHAO Yonghua1,2,3,4,CHI Xuebin1,CHEN Jiang1. Design of Hybrid Parallel Algorithm for Tridiagonalizing Symmetric Matrix[J]. , 2005, 31(22): 49-.