| 摘要: |
| 划分是一种自动分配计算和数据到各个处理器的编译技术，是分布存储结构下并行编译的核心问题.以往的划分研究较少从生命期的角度考虑数据分解问题，分解在数组的不同生命期中不一致时会产生冗余通信.为解决上述问题，提出了一种数据分解算法，通过定义-引用图来表示数组的数据流信息，并使用分解映射表为数组不同的生命期建立各自的数据分解.对矩阵求逆等9 个实际用例的实验结果表明，与以往不区分生命期的划分研究相比，使用所提算法能够在寻找数据分解时对并行收益做出更准确的评估，减少了通信冗余，从而提升了自动生成的并行代码的加速比. |
| 关键词: 数据分解 数组生命期 自动并行化 分布存储 |
| DOI：10.3724/SP.J.1001.2013.04405 |
| 分类号: |
| 基金项目:国家高技术研究发展计划(863)(2009AA01220)；“核高基”重大专项(2009zx10036-001-001) |
|
| Data Decomposition Algorithm Based on Array Life Cycle |
|
DING Rui1,2, ZHAO Rong-Cai1,2, HAN Lin1
|
|
1.PLA Information Engineering University, Zhengzhou 450002, China;2.State Key Laboratory of Mathematical Engineering and Advanced Computing, Zhengzhou 450001, China
|
| Abstract: |
| Partition is a compiler technique that maps computation and data onto different processors, and is the key issue of automatic parallelization on distributed memory architecture. Array's life cycle has been less considered by previous researches on data decomposition, despite of the fact that the inconsistency of decomposition in different array life cycles often results in communication redundancy. This paper proposes a new data decomposition algorithm which represents data flow information of array by define-use graph, and creates own decomposition for each life cycle of array. The experimental results on Matrix-Inversion and other eight applications show that compared with automatic data decomposition methods that does not distinguish the life cycle of array, the proposed algorithm not only makes more accurate assessment of parallel benefits, but also reduce communication redundancy and rise up the speedup. |
| Key words: data decomposition array life cycle automatic parallelization distributed memory |
