Optimized BLAS and Its Effect on Performance of Parallel Programs
-
摘要: 利用SMP结构的多处理器结点通过高速网络构造高性能并行计算系统是当前的一种发展趋势.为了使BLAS在SMP结点上具有更高的效率,本文讨论了BLAS3的多线程化问题;同时以SUMMA作为并行计算的实例,说明提高结点机程序的性能对并行计算性能的影响.试验结果显示,在一定的条件下,多线程BLAS3在SMP平台上可以获得超线性加速比;结点计算性能的提高的同时必须提高网络有效带宽,才能充分发挥并行计算的效能.Abstract: It is the trend of using SMP board as the compute node of a high performance system. The benefits of multithreading is discussed firstly and the BLAS3 is rewritten to get higher performance on a Dual Pentium II system. To investigate the relation between the performance of a single compute node and the entire parallel system, the SUMMA(Scalable Universal Matrix Multiplication Algorithm) is taken as an instance of our research works afterwards. The result demonstrates that the higher the performance of a SMP compute node is, the more sensitive to the capability of the SAN(System Area Network) the performance of the whole parallel system is.
-
Key words:
- parallel processing /
- linear algebra /
- optimization /
- multithread /
- BLAS /
- SUMMA
-
[1] Dongarra J J, Gustavson F G, Karp A. Implementing linear algebra algorithms for dense matrices on a Vecto pipeline machine[J].SIAM Rev, 1984,26:91~112. [2] Chandrika Kamath, Roy Ho, Dwight P Manley.DXML:A high-performance scientific subroutine library. http://www.digital.com/info/DTJF04/ DTJF04SC.TXT. [3] 李忠泽,陈 瑾,龙 翔,等.基于Pentium Pro的高性能BLAS的设计与实现[J].北京航空航天大学学报,1998,24(4):454~457. [4] Golub G H, Van Loan C F.Matrix computations[M]. 2nd ed. Baltimore:Johns Hopkins University Press, 1989. [5] Cannon L E.A cellular computer to implement the Kalman filter algorithm. Bozeman:Montana State University, 1969. [6] Fox G C, Johnson M A, Lyzenga G A, et al.Solving problems on concurrent processors[M]. Englewood Cliffs:Prentice Hall, 1988. [7] Agarwal R C, Gustavson F, Zubair M. A high-performance matrix multiplication algorithm on a distributed memory parallel computer using overlapped communication[J]. IBM Journal of Research and Development, 1994, 38(6):673~681. [8] Geijin R Van de, Watts J. SUMMA:scalable universal matrix multiplication algorithm. Technical Report of The University of Texas,TR-95-13, 1995. -
点击查看大图
计量
- 文章访问数: 2928
- HTML全文浏览量: 154
- PDF下载量: 6
- 被引次数: 0
下载:
