GRAPES全球模式并行优化与众核实现  

Parallel Optimizationand Implementon Manycore Processor of GRAPES Model

作　　者：徐金秀[1] 刘钊[2] 金之雁[3] 陈德训[1] 

机构地区：[1]江南计算技术研究所,无锡214000 [2]上海交通大学计算机系,上海200032 [3]中国气象局国家气象中心,北京100081

出　　处：《高性能计算技术》2014年第1期25-30,共6页

摘　　要：GRAPES（Globaland Regional Assimilation PrEdiction System）全球数值预报模式是中国气象局自主研发的新一代静力／非静力多尺度通用数值预报系统。本文分析了GRAPES全球模式并行计算的特点，尝试在神威蓝光上对GRAPES进行MPI进程级通信优化；对2个耗时较多的核心计算段程序进行众核优化，并在众核FPGA实验平台上进行模拟试验。初步的实验结果表明，本文所设计的通信优化方法和众核并行计算方案可行，不仅提高了GRPAES全球模式的运行效率，而且提高了程序的稳定性和可扩展性。GRAPES （Global and Regional Assimilation PrEdiction System ） is a numerical weather prediction system with hydrostatic/non-hydrostatic and multi-scale general dynamics models developed by CMA（ China Meteorological Administration ）. In this paper, the parallel computing techniques of GRAPES were characterized, the MPI （ Message-Passing Interface ） communication optimizations were carried out on Sunway Bluelight, Two hot subroutines were simulated on a FPGA simulation processor and accelerated. Simulation results varified that the MPI communication strategy and the parallel computing methods presented in this paper are feasible, which are able to highly improve the run-time efficiency of GRAPES and enhance the stability and parallel scalability.

关 键 词：GRAPES 并行计算 通信优化 众核优化 策略 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]