优化和并行一个油藏数值模拟软件中的解法器

OPTIMIZING AND PARALLELING A SPARSE LINEAR EQUATIONS SOLVER PACKAGE FOR RESERVOIR SIMULATING SOFTWARE

在当前共享存储对称多处理 ( SMP)并行机上 ,基于指导语句的并行程序设计模式 ,讨论了多功能油藏数值模拟软件中求解超过百万节点规模的解法器 ( MFS)的并行和优化技术。首先 ,结合当前微处理器的高性能特征 ,为了提高 Cache命中率 ,改进了数据和循环结构 ,并组织了 MFS的性能优化 ,在 R50 0 0上获得了 2 0 %的性能提高 ,并消除了并行化将可能引入的 Cache一致性冲突 ;然后 ,基于循环合并、区域分解和大粒度流水线并行技术 ,实现了 MFS的并行化 ;最后 ,在 POW-ER CHALL ENGE R80 0 0的 6台处理机和 R1 0 0 0 0的 8台处理机上 ,对三维三相 50万和 1 0 0万节点规模问题 ,分别组织了数值实验 ,并取得了超过 60 %的并行效率。

Based on the contemporary shared memory Symmetric Multi Processing parallel machines (SMP), and the directive parallel programming platform, this paper discusses optimization and parallelization of a sparse linear equations solver package MFS for reservoir simulation software with more than one million nodes.Firstly, with the high performance characteristics of contemporary microprocessors, we organize the optimization of MFS by adjusting both data and cycle structures for the Cache hit ratio, and improve the performance by 20% for R5000, and moreover, eliminate the potential Cache coherence collision during parallelization.Secondly, we parallelize the solver package MFS with the effective techniques such as cycle combination, domain decomposition and large granularity parallel pipeline.Finally, under POWER CHALLENGE R8000 (6 CPUs) and R10000 (8 CPUs), we organize the numerical experiments for some three dimensional three phase problem with half and one million nodes respectively, and the parallel efficiencies are all higher than 60%.