程序自动并行化过程中的任务划分  

On Task Partitioning in Automatic Parallelization

作　　者：冯百明[1] 丁晓宁[1] 肖骊[1] 况正谦[1] 康继昌[1] 

机构地区：[1]西北工业大学计算机科学与工程系,陕西西安710072

出　　处：《西北工业大学学报》2000年第2期298-301,共4页Journal of Northwestern Polytechnical University

摘　　要：任务划分是程序并行化过程中的重要一环 ,它的优劣直接影响并行程序的正确性。本文以基于数组元素局部相关的计算为基础 ,讨论了将串行程序所描述的任务划分为多个子任务的方法 ,提出了编译程序静态识别与并行程序动态判断相结合的方法 ,成功地实现了对多个串行程序的正确并行化 ,该方法降低了对串行程序书写格式的要求 。Task partitioning has great influence on the performance of the resulting parallel programs in automatic parallelization. We developed a software system——NPUPAR, for converting serial programs to parallel ones in CFD. In this paper, we present the method for task partitioning in NPUPAR. In section 1, we described loop partitioning and non loop partitioning in detail. We proposed a method to take advantage of both static analysis and dynamic decision when loop domain or array subscripts is uncertain. Such a method greatly extended the scope of serial programs to be converted. An example is given in section 2. It shows that our method for task partitioning is effective.

关 键 词：串行程序 并行程序 任务划分 并行计算机 

分 类 号：TP338.6[自动化与计算机技术—计算机系统结构]