基于OpenACC的GRAPES_GLOBAL模式长波-中国气象局图书馆.PDF

2 0 1 8 基于OpenACC的GRAPES_GLOBAL模式长 波辐射异构并行化研究 孙晨 王彬 顾文静 魏敏 （国家气象信息中心，北京 100081） 摘要：气象数值模式是天气预报的基本工具和方法，随着技术的发展，模式分辨率有了大幅的提高，分辨率的提升使 计算量呈指数增长，然而气象预报的时效性对并行程序的设计与计算平台的性能都提出了更高的要求。以GRAPES_ GLOBAL数值天气预报模式为研究案例，以“神威·太湖之光”新一代国产异构众核高性能计算系统为试验平台，分析 其程序结构及计算原理，定位影响模式并行效率及扩展性的热点子程序。通过调整程序结构以及添加协处理器加速指示 语句，并针对模式系统消息缓存过大等问题，为长波辐射过程的每个热点子程序分别设计了高效的通讯策略。实现了 “粗粒度MPI并行+细粒度众核OPENACC并行”多级异构并行方案，使其普遍达到3～6倍的加速。本试验充分继承了 原始代码的MPI 级并行，同时利用数量众多的协处理器为其中的热点函数提供加速，有效提升模式的并行效率，节约了 开发成本。 关键词：神威·太湖之光，OpenACC，GRAPES模式，长波辐射过程 DOI：10.3969/j.issn.2095-1973.2018.01.027 A Research About Hybrid Programming and Parallelization of GRAPES_GLOBAL Based on OpenACC Sun Chen, Wang Bin, Gu Wenjing, Wei Min (National Meteorological Information Centre, Beijing 100081) Abstract: Numerical weather model is a basic method and tool of weather forecasting. As the development of technology, the model resolution has been improved greatly, it, however, brings an exponentially-increasing computation cost. The timeliness for the weather forecasting puts forward more advanced requests to the program designing and the performance of computing platform. In this paper, we take GRAPES_GLOBAL as an example, to explore the feasibility of hybrid programming and optimization on the Sunway TaihuLight (new domestic high-performance computing system). By analyzing the program structure and calculation principle, we find hotspot subroutines which are influencing the parallel efficiency and extensibility, then we design an efficient communication strategy for each subroutine in the long-wave radiation by adjusting the structure of program and adding the coprocessor acceleration indicator statements. The implement of the hybrid prog