摘要
并行计算软件库,如scalapack和plapack,通过提供大量并行计算子过程来提高并行计算软件开发效率。为方便用户使用并行软件库,该文提出了一种基于性能预测技术来开发并行计算软件库的方法,这是一种隐藏并行计算细节的有效途径。并就构成该软件库的神威超级计算机系统并行软件开发工具,并行线性代数软件库的数据自动分布的方法和供用户调用并行软件库的接口设计技术进行了讨论。实验结果表明,这种技术能有效提高并行计算软件效率。
Parallel computing software library,such as scalapack and plapack,can provide many parallel computing subroutines to assist in improving development efficiency.In this paper,a design method of parallel computing software library based on performance predict technology is provided,it is a promising way to hide the detail of parallel computing.Several critical techniques in constructing parallel computing software library are also discussed,such as parallel software development tools of″SFENWEI″super computer,the method of data distribution and interface design technology.The experimental results show that this technique can get good result in improving the efficiency of parallel computing software.
出处
《计算机工程与应用》
CSCD
北大核心
2004年第29期6-8,37,共4页
Computer Engineering and Applications
基金
国家部委开发项目资助
关键词
并行软件库
性能预测
数据划分
神威计算机
并行计算
parallel software library,performance predict,data distribution,SFENWEI computer,parallel computing