THE COST OF EIGENVALUE COMPUTATION ON DISTRIBUTED-MEMORY MIMD MULTIPROCESSORS

作     者：CRIVELLI, S JESSUP, ER 

作者机构：UNIV COLORADO DEPT COMP SCI BOULDER CO 80309 USA 

出 版 物：《PARALLEL COMPUTING》 (Parallel Comput)

年 卷 期：1995年第21卷第3期

页      面：401-422页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Science Foundation, NSF, (CCR-9109785) National Science Foundation, NSF U.S. Department of Energy, USDOE, (DE-FGO2-92ER25122) U.S. Department of Energy, USDOE Oak Ridge National Laboratory, ORNL, (DE-ACO5-840R21400) Oak Ridge National Laboratory, ORNL 

主　　题：LINEAR ALGEBRA EIGENVALUE COMPUTATION DISTRIBUTED-MEMORY MULTIPROCESSOR BISECTION MULTISECTION POLYSECTION INTEL IPSC/2 INTEL TOUCHSTONE DELTA 

摘      要：In [20], Simon proves that bisection is not the optimal method for computing an eigenvalue on a single vector processor. In this paper, we show that his analysis does not extend in a straightforward way to the computation of an eigenvalue on a distributed-memory MIMD multiprocessor. In particular, we show how the optimal number of sections (and processors) to use for multisection depends on variables such as the matrix size and the ratio of communication and computation costs. We also show that parallel multisection outperforms the variant of parallel bisection called polysection proposed by Swarztrauber in [22] for this problem on a distributed-memory MIMD multiprocessor. We present the results of experiments on the 64-processor Intel iPSC/2 hypercube and the 512-processor Intel Touchstone Delta mesh multiprocessor.