随着高性能计算机技术的发展和应用,并行计算已成为保证数值天气预报模式业务运行时效的关键技术之一。目前高性能计算机计算能力已达到每秒千万亿次浮点计算,系统中处理器数量也早已达十万甚至更多,如此巨大的计算资源对应用软件系统...随着高性能计算机技术的发展和应用,并行计算已成为保证数值天气预报模式业务运行时效的关键技术之一。目前高性能计算机计算能力已达到每秒千万亿次浮点计算,系统中处理器数量也早已达十万甚至更多,如此巨大的计算资源对应用软件系统的设计也提出了挑战。数值天气预报软件系统要充分利用高性能计算机提供的计算资源,必须依靠并行计算方法,这包括适合计算问题的可扩展并行算法的设计、合适的数据分配方案以及良好的任务负载平衡方案。作为中国新一代数值天气预报格点模式,GRAPES(Global and Regional Assimilation and PrEdiction System)设计的最终目标是一个科研/业务通用,区域/全球通用模式。作为一个格点模式,GRAPES的并行计算具有与欧洲中期数值预报研究中心谱模式并行计算不同的特点,GRAPES的并行计算采用了经典的水平网格数据划分。但对于全球的GRAPES模式,由于采用拉格朗日差分方案,模式极地及附近区域格点与格点之间距离的减小,使得模式并行计算在采用简单的经纬网格划分方式实现时,必须考虑极地区域并行计算跨越多个处理器时导致的频繁通讯解决途径。本研究提出了利用消息传递组通讯实现全球格点模式并行计算的一种方法,其核心思想是将极点附近一定区域内的处理器按纬向划归不同的处理器组。文中还给出了该实现方法的任务分配算法,提出了改进的任务分配负载平衡方案。在中国气象局高性能计算机IBM-cluster1600上的测试表明,算法具有较好的可扩展性,其负载平衡方案改善了计算的绝对墙钟时间,使并行计算效率提高10%以上。模式的准业务运行结果表明计算墙钟时间基本可以满足数值预报业务的实时性要求。展开更多
The Global/Regional Assimilation and PrEdiction System（GRAPES）is a new-generation operational numerical weather prediction（NWP）model developed by the China Meteorological Administration（CMA）.It is a grid-point m...The Global/Regional Assimilation and PrEdiction System（GRAPES）is a new-generation operational numerical weather prediction（NWP）model developed by the China Meteorological Administration（CMA）.It is a grid-point model with a code structure different from that of spectral models used in other operational NWP centers such as the European Centre for Medium-Range Weather Forecasts（ECMWF）,National Centers for Environmental Prediction（NCEP）,and Japan Meteorological Agency（JMA）,especially in the context of parallel computing.In the GRAPES global model,a semi-implicit semi-Lagrangian scheme is used for the discretization over a sphere,which requires careful planning for the busy communications between the arrays of processors,because the Lagrangian differential scheme results in shortened trajectories interpolated between the grid points at the poles and in the associated adjacent areas.This means that the latitude-longitude partitioning is more complex for the polar processors.Therefore,a parallel strategy with efficient computation,balanced load,and synchronous communication shall be developed.In this paper,a message passing approach based on MPI（Message Passing Interface）group communication is proposed.Its key-point is to group the polar processors in row with matrix-topology during the processor partitioning.A load balance task distribution algorithm is also discussed.Test runs on the IBM-cluster 1600 at CMA show that the new algorithm is of desired scalability,and the readjusted load balance scheme can reduce the absolute wall clock time by 10% or more.The quasi-operational runs of the model demonstrate that the wall clock time secured by the strategy meets the real-time needs of NWP operations.展开更多
文摘随着高性能计算机技术的发展和应用,并行计算已成为保证数值天气预报模式业务运行时效的关键技术之一。目前高性能计算机计算能力已达到每秒千万亿次浮点计算,系统中处理器数量也早已达十万甚至更多,如此巨大的计算资源对应用软件系统的设计也提出了挑战。数值天气预报软件系统要充分利用高性能计算机提供的计算资源,必须依靠并行计算方法,这包括适合计算问题的可扩展并行算法的设计、合适的数据分配方案以及良好的任务负载平衡方案。作为中国新一代数值天气预报格点模式,GRAPES(Global and Regional Assimilation and PrEdiction System)设计的最终目标是一个科研/业务通用,区域/全球通用模式。作为一个格点模式,GRAPES的并行计算具有与欧洲中期数值预报研究中心谱模式并行计算不同的特点,GRAPES的并行计算采用了经典的水平网格数据划分。但对于全球的GRAPES模式,由于采用拉格朗日差分方案,模式极地及附近区域格点与格点之间距离的减小,使得模式并行计算在采用简单的经纬网格划分方式实现时,必须考虑极地区域并行计算跨越多个处理器时导致的频繁通讯解决途径。本研究提出了利用消息传递组通讯实现全球格点模式并行计算的一种方法,其核心思想是将极点附近一定区域内的处理器按纬向划归不同的处理器组。文中还给出了该实现方法的任务分配算法,提出了改进的任务分配负载平衡方案。在中国气象局高性能计算机IBM-cluster1600上的测试表明,算法具有较好的可扩展性,其负载平衡方案改善了计算的绝对墙钟时间,使并行计算效率提高10%以上。模式的准业务运行结果表明计算墙钟时间基本可以满足数值预报业务的实时性要求。
基金Supported by the National S&T Infrastructure Program for the 11th Five-Year Period under Grant No.2006BAC02B00the National Natural Science Foundation of China under Grant Nos.40575050 and 40775073
文摘The Global/Regional Assimilation and PrEdiction System（GRAPES）is a new-generation operational numerical weather prediction（NWP）model developed by the China Meteorological Administration（CMA）.It is a grid-point model with a code structure different from that of spectral models used in other operational NWP centers such as the European Centre for Medium-Range Weather Forecasts（ECMWF）,National Centers for Environmental Prediction（NCEP）,and Japan Meteorological Agency（JMA）,especially in the context of parallel computing.In the GRAPES global model,a semi-implicit semi-Lagrangian scheme is used for the discretization over a sphere,which requires careful planning for the busy communications between the arrays of processors,because the Lagrangian differential scheme results in shortened trajectories interpolated between the grid points at the poles and in the associated adjacent areas.This means that the latitude-longitude partitioning is more complex for the polar processors.Therefore,a parallel strategy with efficient computation,balanced load,and synchronous communication shall be developed.In this paper,a message passing approach based on MPI（Message Passing Interface）group communication is proposed.Its key-point is to group the polar processors in row with matrix-topology during the processor partitioning.A load balance task distribution algorithm is also discussed.Test runs on the IBM-cluster 1600 at CMA show that the new algorithm is of desired scalability,and the readjusted load balance scheme can reduce the absolute wall clock time by 10% or more.The quasi-operational runs of the model demonstrate that the wall clock time secured by the strategy meets the real-time needs of NWP operations.
