计算机工程 ›› 2010, Vol. 36 ›› Issue (06): 12-14.doi: 10.3969/j.issn.1000-3428.2010.06.004

• 博士论文 • 上一篇    下一篇

求解大规模矩阵特征问题的并行算法研究

赵 韬,迟学斌,陆忠华,赵永华   

  1. (中国科学院计算机网络信息中心超级计算中心,北京 100190)
  • 收稿日期:1900-01-01 修回日期:1900-01-01 出版日期:2010-03-20 发布日期:2010-03-20

Study on Parallel Algorithm for Solving Large Matrix Eigenproblem

ZHAO Tao, CHI Xue-bin, LU Zhong-hua, ZHAO Yong-hua   

  1. (Supercomputing Center, Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190)
  • Received:1900-01-01 Revised:1900-01-01 Online:2010-03-20 Published:2010-03-20

摘要: 基于数据并行的重启动Arnoldi并行算法,基于使用数据并行模型的重启动Arnoldi并行算法,提出一个精化重启动Arnoldi并行算法。为了降低弱扩展性对并行性能的负面影响,该算法使用任务图模型并行计算精化向量,减少处理器进程之间的通信次数,有效地实现并行计算。在KD-50-I万亿次机上的测试结果表明,该算法具有较好的可扩展性和并行 效率。

关键词: 矩阵特征值, Arnoldi算法, 并行计算, 精化向量

Abstract: Based on the parallel algorithm for restarted Arnoldi method by using data parallel model, a parallel algorithm for refined restarted Arnoldi method is presented. To impair the negative impact on parallel performance due to weak scalability, Tthe parallel algorithm presented implements parallel computing refined vectors by using task graph model and reduces the communication cost. As a result, the algorithm achieves parallel computing efficiently. Numerical experiments on KD-50-I show that the parallel algorithm presented performs good scalability and efficiency.

Key words: matrix eigenvalue, Arnoldi algorithm, parallel computing, refined vector

中图分类号: