采用曙光超级计算机的海冰模型并行性能调优研究

Research on Parallel Performance Tuning of Sea Ice Model Based on“Dawning”Supercomputer

海冰对全球气候变化非常敏感,海冰变化具有全球气候变化指示和放大器作用,海冰模型是地球系统模型中最重要的组成部分之一。由于海冰模型应用本身具有计算密集型特点,所以海冰模型的模拟计算过程一般需要数小时至几天不等。因此,如何提升海冰模型的运算效率,探索海冰模型的最佳性能成为业界研究热点。将海冰模型应用程序移植到曙光超级计算机上,通过大气的T62原始网格数据转化生成海冰模型所需要的1度输入数据,使海冰模型模拟结果更精确。基于海冰模型的7种网格划分方式,通过改变开启的MPI进程数目、调整网格划分方式及相关参数配置,对海冰模型的并行性能进行调优。通过性能调优研究发现,在slenderx2划分方式下,320个MPI进程与16个MPI进程相比,性能提升了8.32倍。该成果能帮助研究人员更高效地使用海冰模型,也为海冰模型的并行优化工作提供借鉴。

Sea ice is one of the most sensitive parts of global climate change.Sea ice changes have the role of an indicator and amplifier of global climate change.The sea ice model is also one of the most important components of the earth system model.Since the sea ice model application itself is computationally intensive,the simulation calculation process of the sea ice model generally takes several hours to several days.Therefore,how to improve the computational efficiency of the sea ice model and explore the best performance of the sea ice model has become a hot issue in the industry.This research first transplanted the sea ice model application program to the“Dawning”supercomputer;secondly,in order to make the simulation results of the sea ice model more accurate,the 1 degree input data required for the sea ice model was generated through the transformation of the original T62 grid data of the atmosphere.;Finally,based on the seven meshing methods of the sea ice model,the parallel performance of the sea ice model can be tuned by changing the number of open MPI processes,adjusting the meshing method and the configuration of related parameters.Through performance tuning research,it is found that in the slenderx2 partitioning method,the performance of 320 MPI processes is increased by 8.32 times compared with 16 MPI processes.It provides help for related researchers to use the sea ice model more efficiently,and also provides reference for the parallel optimization work of the sea ice model.