Research On Optimization Technology Of Parallel Partition For Dynamic Framework Of The Global Circulation Spectral Model

With the increase of the resolution of numerical weather prediction model,the calculation amounts is also increasing,in order to meet the real-time requirements of the actual weather forecast,high performance computing is being applied to weather prediction.For its high forecasting accuracy for global Macro-scale general circulation,global atmospheric spectral model is widely used in the actual prediction.Spectral models can be conceptually divided into two parts: the dynamic framework and the physical process,the function of the dynamic framework is to discretize the model equations without sources and sinks,and to solve them numerically,the dynamic framework is the core of the model operation,if its execution time can be effectively shortened,the operating efficiency of the entire model will improve.In this paper,a new partition scheme is introduced based on the limitations of parallel partitioning for current BCC_AGCM 3.0 dynamic framework.Which means change the original way of division according to the latitude or wavenumber into two-dimensional division.To avoid unnecessary inter-process communication,partition is implemented in the direction of no data relevance and each calculation step uses a different scheme depending on the data relevance.In order to convert between different schemes,data exchange between two calculation steps is required.According to the new partition scheme,the code of dynamic framework is modified.Based on the high-performance computer,the modified dynamic framework is tested,the results show that the modification of the dynamic framework can effectively shorten the running time,the time-consuming can be reduced by 8% on average,and this can reach 15% under some partition schemes.Different modules of the dynamic framework show different improvements,the scandyn module and the scan2 module have a significant improvement when the number of processes is large;the scanslt module shows a better performance in all experiments.Through the comparison of different process settings under the same number of process,different partition schemes have different effects,too many division in the new direction will reduce the improvement brought by the modified framework.In general,the modified dynamic framework shows a shorter run time and a better performance than the current framework.