ISCAS OpenIR  > 并行软件与计算科学实验室 
数值计算软件包自适应性能优化若干关键技术及评价标准研究
Alternative TitleKey Techniques in Numerical Software Self-Adapting Optimization and Evaluation of Tuning Process
王宣强
2008-06-05
Degree Grantor中国科学院软件研究所
Degree Level博士
Place of Degree Grantor软件研究所
Keyword数值计算软件 自适应性能优化 评价方法
English Abstract为了能够充分使用计算机资源,使软件运行能够尽可能地接近计算机峰值性能,研究人员 一直在努力。一个思路是为计算机开发优秀的编译器,并使用编译器相关技术对软件进行 性能优化;作为补充,另一个思路是开发能够共用的核心软件包,通过提高核心软件包中 程序的性能提高调用核心库的软件系能。 但科学计算中针对特定的计算机平台和特定的用户问题进行性能调优仍然是一个困难的问 题。速度和可移植性是数值计算软件开发中的一对矛盾。自适应性能优化技术正是为了 解 决数值计算软件中的可移植性问题和进行自动性能优化而提出的,科学家们希望数值计 算 软件能够动态地获悉计算环境的变化以及待求解问题的特征,根据需要改变自身以适应 这些变化以及多种复杂的问题情况,并且在多种求解方法中进行决策以选择最优的解决方 法来求解问题。 本文调研了使用自适应性能优化技术的几个著名软件包:~ATLAS,~SPARSITY,~OSKI以及数字 信号处理领域的快速傅立叶变换软件包FFTW,在此基础上,重点分析和比较了自适应性能 优化的关键技术,着重介绍了矩阵乘计算、MPI通信操作、快速傅立叶变换中涉及的经验搜 索、算法选择和自动代码生成等技术在这些软件包中的应用。 在深入调研的基础上,结合实际应用需求,本文还提出了自适应性能优化过程的新的评价 指标,试图权衡优化效果和优化时间开销,并在不同的实验平台上针对ATLAS的优化过程进 行的实验和过程评价,实验表明,综合优化效果和优化开销能够有效地发现ATLAS自适应优 化过程的特征。将其应用到实际开发和调优过程中,能够在不损失性能的前提下,节省优化时间。 论文最后对HPCC测试软件包在IBM刀片机群上进行了对比测试,发现了测试平台存在的性能 瓶颈,并消除了该瓶颈。表明HPCC软件包确实可以有效的发现被测平台存在的性能瓶颈问 题。
AbstractThis paper investigates the key techniques adopted in the famous numerical computation software as ATLAS, SPARSITY, OSKI as well as the FFTW package in the field of digital signal processing. We nalyzed and compared the application of the self adapting optimization techniques in these software, among which we emphasized on the empirical search, selection of algorithm implementations and auto-generation of codes in matrix computation, MPI communication, fast fourier transformation, respectively. This paper also tried to investigate the evaluation method of the tuning process of self adapting optimization and gave a simple criteria to depict the balance between the optimization result and the time cost. We performed the ATLAS Self-Adapting tuning process on 2 platforms and evaluated this process using the criteria we suggested. we find that through this simple criteria, we can have a good knowledge about the process and can save the tuning time. Finally, we gave the comparative performance testing we performed on the IBM Blade Center Cluster which aims to study the HPCC high performance benchmark and its result evaluation method, we found out the performance bottle neck of the target cluster and removed it through our result evaluation analysis.
Pages63
Language中文
Content Type学位论文
URIhttp://ir.iscas.ac.cn/handle/311060/6270
Collection并行软件与计算科学实验室 
Recommended Citation
GB/T 7714
王宣强. 数值计算软件包自适应性能优化若干关键技术及评价标准研究[D]. 软件研究所. 中国科学院软件研究所,2008.
Files in This Item:
File Name/Size DocType Version Access License
10001_20052801502905(1165KB) 开放获取--Application Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王宣强]'s Articles
Baidu academic
Similar articles in Baidu academic
[王宣强]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王宣强]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.