区域分解对气象模式并行计算速度的影响  

Influence of Domain Decomposition on Parallel Computation Speed of Meteorological Model

作　　者：臧增亮[1] 饶宣锐 潘晓滨[1] 张理论[3] 王春明[1] 何宏让[1] 

机构地区：[1]解放军理工大学气象学院 [2]二炮装备研究院,北京100085 [3]国防科技大学计算机学院,长沙410073

出　　处：《计算机工程》2008年第17期4-6,共3页Computer Engineering

基　　金：国家自然科学基金资助项目(40505023,40705020);中国科学院大气物理研究所LASG实验室开放课题基金资助项目

摘　　要：通过数值试验分析了区域分解策略对ARPS气象模式并行计算速度的影响,发现无论是否使用编译优化技术,均以分解后数据区域近似为正方形时具有最大的加速比和并行效率。在二级编译优化的情况下,并行速度还和分解方向有关,在y方向上的分解比在x方向上的分解更有利于提高并行效率,而在无优化情况下,并行速度和分解方向几乎无关。并从通信量和编译优化的角度对试验结果进行了讨论和分析。The influence of domain decomposition on the parallel computation speed of ARPS model is analyzed on the base of some experiments. Results show that the subdomalns similar to square possess the biggest acceleration rate and efficiency of parallel whether the optimization lever -o2 is applied to or not. Under the circumstances of optimization lever -o2, parallel speed also relates to decomposition direction. The decomposition along y direction is preferable to that along x direction for improving in parallel efficiency. However, under the circumstances without optimization, parallel speed is almost independent of decomposition direction. Furthermore, experimental results are discussed in the paper from the angles of communications traffic and optimization.

关 键 词：区域分解 数值模式 并行计算 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]