广州区域数值预报模式并行化计算  

THE PARALLEL COMPUTATION OF GUANGZHOU AREA NUMERICAL PREDICTION MODEL

作　　者：袁金南[1] 王在志[1] 薛纪善[2] 

机构地区：[1]中国气象局广州热带海洋气象研究所,广州510080 [2]中国气象科学研究院,北京100081

出　　处：《应用气象学报》2004年第5期556-563,共8页Journal of Applied Meteorological Science

基　　金："国家智能计算机研究开发中心和国家高性能计算基金"课题;广东省科技攻关项目"珠江三角洲热带气旋暴雨预警信号发布关键技术研究"资助

摘　　要：采用消息传递方式 (MPI)对最近发展的广州区域数值预报模式进行了并行化计算研究。根据模式的结构和计算过程特点 ,模式适合采用水平分区方案进行并行计算处理。在曙光 30 0 0并行计算机上分别采用一维和二维分区并行方案实现了模式的并行化计算 ,并对模式的并行效率、并行加速比和并行通讯时间百分比等做了测试。对测试结果的分析表明 :采用 8个CPU时 ,两种方案都能在 1h内完成 72h的预报 ,一维分区方案的并行效率则保持在 90 %左右 ,可以满足业务运行需要。当模式使用 8个以上CPU时 ,通讯时间迅速增加并超过了计算时间的 5 0 %,模式并行效率明显下降。CPU相同时 。The parallel computation of the recently developed Guangzhou area numerical prediction model was researched by using Message Passing Interface (MPI) method. According to the characters of model structure and model computing process, the model is suited to parallel computation by using horizontal section method. The parallel computation of the model has been realized by using one dimensional and two dimensional section method, respectively, and parallel efficiency, parallel acceleration rate and percent of parallel communication time of the model have been tested on DAWN 3000 parallel computer. Results show that two parallel methods can finish 72 hour forecast in one hour by using 8 CPUs, the parallel efficiency of one dimensional section method keeps about 90%, which can meet the operational need. When more than 8 CPUs are used in the model, the communication time of the model increases rapidly and beyond 50% of the computation time, the parallel efficiency of the model decreases dramatically. One dimensional section parallel method of the model is simpler and more efficient by using the same CPU than two dimensional section parallel method of the model.

关 键 词：并行效率 CPU 分区 并行化 曙光3000 消息传递 并行计算机 数值预报模式 下降 计算时间 

分 类 号：P456.7[天文地球—大气科学及气象学] TP393[自动化与计算机技术—计算机应用技术]