| 摘要: |
| 本文提出了一种面向分布存储器多机系统的并行化编译方法．针对分布存储并行系统的特点，作者采用的基本优化策略是：折衷并行性与数据引用局部性；减少和隐藏通信开销．通过对基于仿射函数的程序分解方式所导致的数据通信性质的分析，得到了适合分布存储结构特殊要求的并行性开发方法．为了在保持并行性的前提下最小化通信数据总量，提出了基于齐次线性方程组求解的程序全局优化分解方法．为了优化数据通信的组织，提高结点代码的效率，又提出了一种以线性不等式组作为工具的更加实用的通信优化和结点代码生成方法. |
| 关键词: 并行化编译 分布存储 优化编译 消息传送 任务划分 单程序多数据流 |
| DOI： |
| 分类号: |
| 基金项目:本文研究得到国家攀登计划资金资助. |
|
| AN INTEGRATED OPTIMIZATION SCHEME IN PARALLELIZING COMPILERS |
|
Sun Tong,Li Sanli,Li Xiaoming
|
| Abstract: |
| This paper presents a complete suit of systematic optimizing methods which may be used in parallelizing compilers for multicomputers or computer clusters. In the compilation scheme, two strategy are adopted. One is trading off parallelism and communication cost and the other is reducing and hiding communication overhead. Through analyzing the properties of data communication required for the program partition approach based on affine functions, the authors find a method to exploit parallelism in serial programs satisfying the special requirements of distributed memory machines. In order to minimize the total of data needing to be communicated, they invent a global optimization program partition method based on solving linear equations. In order to optimize the organization of communication codes and generate more efficient node programs, they invent a more practical method based on linear inequalities to perform communication optimization and node programs generation. |
| Key words: Parallelizing compiler distributed memory optimizing compiler message passing task partition SPMD. |
