面向HYCOM的高分辨率海洋数据同化并行算法优化  

Parallel optimization algorithm of high resolution ocean data assimilation for HYCOM

作　　者：曹连雨 张桂赵 肖俊敏 洪学海[2] 谭光明[2] Cao Lianyu;Zhang Guizhao;Xiao Junmin;Hong Xuehai;Tan Guangming(School of Computer&Communication Engineering,University of Science&Technology Beijing,Beijing 100083,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100080,China)

机构地区：[1]北京科技大学计算机与通信工程学院,北京100083 [2]中国科学院计算技术研究所,北京100080

出　　处：《计算机应用研究》2021年第12期3611-3616,共6页Application Research of Computers

基　　金：国家重点研发计划重点专项资助项目(2016YFC1401706);国家自然科学基金资助项目(61802369)。

摘　　要：HYCOM(hybrid coordinate ocean model)海洋数值模式要求较高的吞吐量和相对较小的计算量,这给并行算法设计带来了巨大的挑战。针对具有高吞吐量的海洋数据同化问题,设计了一种基于区域分解的并行优化算法。首先,提出了一种灵活的文件访问方法,可以高效地从磁盘读取大量的数据,避免数据访问冲突,大幅降低磁盘寻址操作的频率。此外,设计了一种避免通信的策略,以一些额外的计算量为代价大幅减少进程间的通信量。最后,提出了一种基于管道流的通信策略,以实现无冲突的消息传递。实验结果表明,该算法与基线算法相比,总体性能提高了5倍,其中文件读取速度提升6倍,进程间的通信性能提升了2.7倍。The HYCOM mode requires higher I/O throughput and less computation,which brings great challenges to the design of high salable parallel algorithm.Therefore,this paper designed a parallel optimization algorithm based on region decomposition for high-throughput ocean data assimilation.First of all,it proposed a flexible file access approach for reading a large number of data from disks.As a result,it avoided the data accessing conflict successfully,and decreased the frequency of disk addressing operations significantly.Next,it designed a communication-avoiding strategy to reduce the communication volume at the cost of some additional computations.Furthermore,it presented a pipe-flow scheme for data exchange to conduct conflict-free message passing.Consequently,it developed a fast data-obtaining algorithm for the data assimilation.The experiments show that the fast data-obtaining algorithm gains a performance of 5x speedup compared with the original one.More specifically,the new approach achieves 6x speedup on average for the file reading process and 2.7x speedup on average for the communication between processors.

关 键 词：数据同化 I/O优化 通信优化 并行计算 区域分解 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]