基于网络的数据并行计算中分布数组描述研究  

Research on the Distributed Array Description in Parallel Computing Based on Networks

作　　者：胡长军[1] 丁文魁[2] 黄其军[2] 向华[2] 许卓群[2] 

机构地区：[1]清华大学计算机科学与技术系,北京100084 [2]北京大学计算机科学技术系,北京100871

出　　处：《软件学报》2002年第4期608-615,共8页Journal of Software

基　　金：国家自然科学基金资助项目(60173004);国家863高科技发展计划资助项目(863-306-ZT01-02-3)

摘　　要：如何描述分布数组是基于网络数据并行计算的基本问题.从网络并行计算的一般需求出发,讨论了分布数组描述DAD（distributed array descriptor）的内容和结构,具体给出了p-HPF并行编译系统的DAD结构定义.针对分布数据的稀疏存储和紧凑存储模型,给出了数据在Block分布、Cyclic分布和Block_Cyclic（k）分布方式下,全局数组到局部数组转换的计算方法,这些方法已在p-HPF编译器中得到实现并证明了其有效性.最后讨论了分布数据描述的标准化对实现并行计算系统的可移植性和可重用性的意义.How to describe the distributed array is a key point in parallel computing based on networks. Based on the general requirement of the parallel implementation, this paper first discusses the structure and necessary components of DAD (distributed array descriptor). Then a certain DAD structure used by p-HPF parallel compiling system is introduced in detail. Further, for the sparse and the dense storage model, the methods are given out to convert global arrays to local ones in three distribution Blocks, the Cyclic and the Block_Cyclic(k) respectively. These methods are implemented in p-HPF compiler and proved to be effective. Finally, the importance of the distributed array description standard for portability and reuse of the parallel computing systems is discussed.

关 键 词：分布数组描述 网络并行计算 HPF语言 并行编译 数据并行计算 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]