摘要
未来计算性能的提升依赖的是更高的并发度,而不再是更快的时钟频率,这将致使传统的时间步进算法成为数值模拟非定常问题的一个瓶颈.该文实验性地探究求解二维非定常4-Laplacian问题的具有高并发度的并行解法器,其中全离散格式为向后Euler格式和双线性矩形元,时间并行策略为通信器和进程分组下基于完全近似格式的多重网格规约时间算法.数值对比实验表明:基于F-FCF松弛、细/粗时间网格层的粗化因子为16/4的MGRIT算法具有更高的并发度,相对文献[Falgout,et al.SIAM J Sci Comput,2017,39:S298-S322]中的最优MGRIT解法器,它可提速2.4倍.
Future performance increases will be provided through greater concurrency,not faster clock speeds.Thus,traditional time-stepping algorithms are becoming a huge bottleneck in time integration simulations.The goal of this work is the high concurrency solution of two-dimensional unsteady 4-Laplacian problems with MGRIT by experimental investigations,where we utilize the backward Euler to discretize in time and bilinear quadrilateral elements in space,as well as the full approximation scheme(FAS)multigridreduction-in-time(MGRIT)algorithm based on communicators and process topologies.Numerical experiments indicate that MGRIT algorithm with F-FCF relaxation and the fine/coarse coarsening factor as 16/4 achieves higher concurrency,and runs 2.4 times faster than the fastest MGRIT solver in[Falgout,et al.SIAM J Sci Comput,2017,39:S298-S322].
作者
岳孝强
刘一寅
瞿创成
YUE Xiao-qiang;LIU Yi-yin;QU Chuang-cheng(School of Mathematics and Computational Science,Xiangtan University,Xiangtan 411105 China)
出处
《湘潭大学学报（自然科学版）》
CAS
2019年第1期49-54,共6页
Journal of Xiangtan University(Natural Science Edition)
基金
国家自然科学基金项目(11601462)
湖南省军民融合产业发展专项资金"自适应多水平解法器及其在ICF数值模拟中的应用"
湖南省自然科学基金项目(2018JJ3494)
