天体大规模数值模拟软件性能优化

发布时间：2018-08-12 13:53

【摘要】：在百万亿次超级计算机深腾7000上进行了基于Aztec库和PETSc库的天体大规模数值模拟程序对比测试,并对性能测试结果进行了分析.实验结果表明,当8核/节点时,PETSc程序运行时间是Aztec程序的23.34%～41.10%,平均为29.31%;当4核/节点时,PETSc程序运行时间是Aztec程序的19.72%～31.34%,平均为24.97%.在测试过程中,还观察到同样的问题规模和处理器核数,每节点采用4个处理器核比采用8个处理器核时的运行时间,在基于Aztec库的情况下平均下降29%,在基于PETSc软件包的情况下平均下降36%.此外,还给出了对PETSc程序进一步对角线存储优化的结果.实验结果体现了深腾7000提供的有效支持,对软件下一步的优化研究有指导意义.
[Abstract]:A large scale numerical simulation program for celestial bodies based on Aztec library and PETSc library was carried out on a hundred trillion times supercomputer Shenteng 7000, and the performance test results were analyzed. The experimental results show that the running time of the PETSc program is 23.34 of that of the Aztec program when the core / node is 8, with an average of 29.31, and the running time of the PETSc program with 4 cores / nodes is 19.72and 31.34 of that of the Aztec program, with an average of 24.97. During the test, the same problem size and number of processor cores were observed, with four processor cores per node compared to eight processor cores. In the case of Aztec library, the average drop is 29 parts, and in the case of PETSc software package, the average drop is 36%. In addition, the results of further diagonal storage optimization for PETSc programs are given. The experimental results reflect the effective support provided by Shenteng 7000, and have guiding significance for the further optimization research of software.
【作者单位】：中国科学院软件研究所;中国科学院计算机科学国家重点实验室;中国科学院研究生院;中国科学院上海天文台;
【基金】：国家高技术研究发展计划资助项目(2006AA01A125,2009AA01A134,2009AA01A129) 国家科技重大专项资助项目(2009ZX01036-001-002) 中国科学院知识创新工程领域前沿资助项目(ISCAS2009-DR05)
【分类号】：TP319;P1

【参考文献】