海洋环流模式的并行I/O优化研究  

Parallel I/O Optimization Based on Ocean Circulation Model

作　　者：李云龙 祝子杰 孟祥飞 朱小谦[2] 林鹏飞[3] 刘海龙[3] 李庚 LI Yunlong;ZHU Zijie;MENG Xiangfei;ZHU Xiaoqian;LIN Pengfei;LIU Hailong;LI Geng(National Supercomputing Center in Tianjin,Tianjin 300457,China;School of Meteorology and Oceanography,National University of Defense Technology,Changsha 410073,China;The State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics,Chinese Academy of Sciences Institute of Atmospheric Physics,Beijing 100000,China)

机构地区：[1]国家超级计算天津中心,天津300457 [2]国防科技大学气象海洋学院,长沙410073 [3]中国科学院大气物理研究所大气科学和地球流体力学数值模拟国家重点实验室,北京100000

出　　处：《计算机工程与应用》2024年第17期263-271,共9页Computer Engineering and Applications

基　　金：国家重点研发计划项目(2021YFC3101503)。

摘　　要：极端气象日益加剧,严重影响人类的生产生活。如何有效应对未来极端环境变化,更关乎国际竞争。因此,发展高精度高分辨率的地球系统模式推动气候变化研究至关重要。随着模式精度和分辨率的不断提高,I/O性能已成为制约模式计算效率和可扩展性的瓶颈。选取典型的海洋环流模式LICOM3作为研究对象,基于天河HPC系统,详细测试和分析了模式的I/O逻辑过程,发现在3K进程规模下I/O可占整体运行时间一半以上,是限制模式计算性能的关键因素。为此,设计并实现了主要输出过程的MPI-IO并行优化方案,显著提升了模式运行效率,基本解决了模式I/O瓶颈。测试结果表明,MPI-IO并行优化方案将输出时间占比从48.58%降到2.44%,将输出带宽提升到6.21 GB/s,约是原程序的14倍。此外,进一步分析了影响并行I/O性能的因素,以期为相关工作提供参考和指导。Extreme weather is getting worse and worse,which has seriously affected human production and life.How to effectively respond to extreme environmental changes in the future is more related to international competition.Therefore,it is very important to develop high-precision and high-resolution earth system models to promote climate change research.With the continuous improvement of mode accuracy and resolution,I/O performance has become a bottleneck restricting the efficiency and scalability of mode computing.The typical ocean circulation model LICOM3 is selected as the research object.Based on the Tianhe HPC system,the I/O process of the model is tested and analyzed in detail.It is found that under the 3K process,the I/O can account for more than half of the overall running time,which is the key factor limiting the calculation performance of the model.Therefore,the MPI-IO parallel optimization scheme for the main output pro-cesses is designed and implemented,which significantly improves the mode operation efficiency and basically solves the mode I/O bottleneck.The test results show that the MPI-IO parallel optimization scheme reduces the output time from 48.58%to 2.44%and increases the output bandwidth to 6.21 GB/s,about 14 times that of the original program.In addi-tion,the factors affecting the performance of parallel I/O are further analyzed to provide reference and guidance for related work.

关 键 词：并行I/O优化 海洋环流模式 MPI-IO 

分 类 号：TP302.7[自动化与计算机技术—计算机系统结构]