天体聚类应用中的非规则访存优化 - 高性能计算中心 - 上海交通大学.PDF

天体聚类应用中的非规则访存优化 1 1 1 1 1 郝赫 司雨濛 韦建文 文敏华 林新华 1上海交通大学高性能计算中心，上海 200240 (james@) OptimizeIrregularMemoryAccessinAstronomicClusteringApplication 1 1 1 1 1 HeHAO YumengSI JianwenWEI MinhuaWen JamesLIN 1(CenterforHighPerformanceComputing ShanghaiJiaoTongUniversity shanghai200240 China) Abstract Halo-based Galaxy Group Finder(HGGF) implements galaxies category based on dark matter halo and detection on isolated galaxy in the noise so that HGGF plays a very important role in the research of large-scale structureoftheuniverse.Butbecauseofthegrowthofthedatascale it’sextremelynecessarytoshortentherunning time by optimizing the algorithm. After analysis the kernel part of the algorithm was seriously affected by the irregularmemory access sothatweproposed aspecific“DataPre-sorting”approachandanalysedhowitaffectthe process of memory access according to the structure of the algorithm and the irregular memory access pattern. Moreover weoptimizedthememoryaccesswith“DataAlignment”and“LoopFission”andenhancedtheefficiency of OpenMP with “Load Balance” and “Mutex Privatization”. Eventually the HGGF application got 11.6 times speedupon 12threads andgotbetterweak scalability.The followingareouroriginalcontributions:(1)Analysis of the irregular memory access of the HGGF application; (2) Proposed and analysis the “Data Pre-sorting” ; (3) Improvetheparallelperformance ofHGGFapplicationwith anotherfourapproachesincludingdataalignment loop fission loadbalanceandmutexprivatization. Key words astronomic clustering applications; optimize irregular memory access; data pre-sorting; parallel computing 摘要 Halo-basedGalaxyGroupFinder(HGGF)算法实现了基于暗物质光晕的星系分类，以及在噪音状态下对孤 立星系的检测，在研究宇宙大尺度结构及宇宙的演化等领域中占有至关重要的地位。但由于数据规模的增长， 急需对HGGF算法进行优化，以缩短运行时间。经分析，算法的热点部分耗时受到非规则访存的严重影响， 因此我们针对算法的结构和非规则访存模型，提出了数据预排序方法，并分析了该方法如何影响访存过程。在 此基础上，我们还利用数据对齐、循环分解进一步优化访存效率，利用负载均衡和互斥变量私有化的方法提高 了OpenMP 的并行效率，最终将HGGF应用