基于J-Earth领域编程框架的高效并行大气环流模式(H-GAMIL)构建

高性能并行计算,一直是气候系统模式研究中的重要科学前沿。然而,随着计算机技术的发展,计算机体系结构越来越复杂,以及气候系统模式本身的复杂性使得开发高性能气候系统模式成为一个难题。采用了一种全新的并行程序设计方法:基于针对气象科学中的计算特点发展出的领域编程框架J-Earth,构建了高性能并行大气环流模式H-GAMIL。在对原模式GAMIL分析和解构以后,结合J-Earth的特点,采用面相对象的编程方式,对H-GAMIL进行重构。重构后的H-GAMIL具有现代软件标准化模块化的特点,并具有自动二维剖分、MPI(Message Passing Interface)和Open MP(Open Multi-Processing)混合并行实现、并行输出等新功能。解决了原模式一维剖分受到处理器限制,串行IO等问题。在此基础上,本文对H-GAMIL进行了并行效率测试,结果显示,模式水平分辨率为1°(纬度)×1°(经度)时,可使用上千处理器核,并行效率达40%以上且负载平衡达70%。测试结果表明了H-GAMIL具有较好的并行效率和可扩展性;同时对H-GAMIL模拟性能进行对比分析,其结果表明H-GAMIL保持了原模式的守恒性,并且具有与原模式相当的模拟性能,能够满足气候研究的计算需求,达到实用程度。

The Development of High Performance Atmospheric General Circulation Model (H-GAMIL) Base on Programming Framework J-Earth

High-performance parallel computing has always been an important scientific issue in the study on climate system models. However, with the development of computer technology, computer architecture is becoming more and more complicated while the complexity of the climate system has made it a challenge to develop high-performance climate system models. This paper adopted a completely new parallel programming method. Based on the programming framework J-Earth, a high-performance parallel general circulation model, H-GAMIL, was developed. With detailed analysis and deconstruction of the original GAMIL model, the characteristics of J-Earth was combined with GAMIL to reconstruct the H-GAMIL in this paper by using the object-oriented programming method. The reconstructed H-GAMIL possesses the features of standardization and modularization of modern software and also has new functions of automatic 2D subdivision, hybrid parallel implementation of MPI (Message Passing Interface) and OpenMP (Open MultiProcessing), as well as parallel output, thus solving the problem that the one-dimensional subdivision of the original model was subjected to the restriction of the processor and the serial IO problem in the original model. Based on the above work, in the present study we conducted a parallel efficiency test on the H-GAMIL. The result shows that when the model''s horizontal resolution was 1°×1°, hundreds of processors could be used and the parallel efficiency could reach above 40%, and the load balance totaled more than 70%. The test result suggests that the H-GAMIL possesses a relatively high parallel efficiency and scalability. Meanwhile, a comparative analysis of the simulation performance of H-GAMIL was conducted in this study. The result shows that the H-GAMIL maintains the conservativeness of the original model and has the same simulation performance as the original one. Thus it meets the computing requirements for climate study and can be applied for practical use.