An improved generalized conjugate residual squared algorithm suitable for distributed parallel computing
An improved generalized conjugate residual squared algorithm suitable for distributed parallel computing
复制标题
DOI：
10.1016/j.cam.2014.04.009
复制
发表时间：
2014-12
期刊：
影响因子：
--
通讯作者：
Xian-yu Zuo;Li-tao Zhang;T. Gu
中科院分区：
文献类型：
--
作者：
Xian-yu Zuo;Li-tao Zhang;T. Gu
本文在Zhang and Zhao（2010）的GCRS算法和Gu et al.（2007）的思想基础上，提出了一种适用于分布式并行环境的改进广义共轭残差平方（IGCRS）算法。该算法通过改变GCRS算法的计算顺序，使每次迭代的内积相互独立，从而使内积所需的通信时间与有效计算重叠，从而将两个全局同步点减少为一个.理论分析和等效率分析的数值比较表明，IGCRS方法比GCRS方法具有更好的并行性和可扩展性，并行性能可提高约2倍。最后，数值实验表明，IGCRS方法比GCRS方法具有更好的并行性能和更高的可扩展性，通信量平均提高了52.19%，与理论分析相符。
In this paper, based on GCRS algorithm in Zhang and Zhao (2010) and the ideas in Gu et al. (2007), we present an improved generalized conjugate residual squared (IGCRS) algorithm that is designed for distributed parallel environments. The new improved algorithm reduces two global synchronization points to one by changing the computation sequence in the GCRS algorithm in such a way that all inner products per iteration are independent so that communication time required for inner products can be overlapped with useful computation. Theoretical analysis and numerical comparison of isoefficiency analysis show that the IGCRS method has better parallelism and scalability than the GCRS method, and the parallel performance can be improved by a factor of about 2. Finally, some numerical experiments clearly show that the IGCRS method can achieve better parallel performance with a higher scalability than the GCRS method and the improvement percentage of communication is up to 52.19% averagely, which meets our theoretical analysis.