摘要
近年来,受GPU其高浮点峰值性能的提高和应用领域中大规模科学计算问题的驱动,高性能领域中利用代数多重网格(AMG)求解稀疏线性方程组成为研究热点。针对经典的AMG算法,探究建立阶段(网格粗化)和求解阶段的并行计算结构,提出基于多核异构的AMG并行计算模式。数值实验表明,并行计算模式计算效率相对于串行提高了3~4倍,加速效果明显。
In recent years, the research of Algebraic Multigrid (AMG) for solving sparse linear equations becomes focus in high-performance field, by the driving of improvements of GPU' s high peak floating point performance and application in the field of large-scale scientific computing problem. In this paper, we research the parallel architecture of establishment phase and solution phase for the classical AMG algorithm, and Proposed the parallel computing mode based on Heterogeneous and Multicore. Numerical experiments show that the parallel computing model improves efficiency,it has 3-4 times acceleration relativing to the serial.
出处
《科学技术与工程》
北大核心
2014年第14期120-124,共5页
Science Technology and Engineering
基金
国家自然科学基金(41174099)
山东省自然科学基金(ZR2013DM015)
中央高校基本科研业务费专项项目(11CX04059A)
中央高校基本科研业务费专项项目(13CX06192A)
国家大学生创新性实验项目(111042557)资助
关键词
代数多重网格
并行计算
多核异构
GPU
algebraic muhigrid parallel computing heterogeneous and muhicore GPU