一种基于线性代数的计算和数据自动分解算法

An Automatic Computation and Data Decomposition Algorithm Based on Linear Algebra

在针对分布内存体系结构的并行识别技术中,如何对计算和数据进行合理分解,以增加数据引用的本地化、减少处理器间的通信是提高并行程序性能的关键。本文通过对Anderson-lam分解算法完整性的补充,给出了一种可实现无通信的计算划分和数据分布算法,并阐述了对该算法在工程实践中的一些优化考虑。

Increasing the locality of data references and minimizing communication of processors by computation and data decomposition are the key optimization for achieving high performance on large-scale parallel machines. A compiler algorithm that automatically finding computation and data decomposition with no communication based on Andersonlain algorithm is described, and some improvement on engineering implement is also mentioned in this paper.