Refactoring And Optmizing WRF Model On SW26010 Heterogeneous Multicore Processor

The WRF Model(WeaTher Research and Forecasting Model)is a widely used mesoscale numerical weather forecast model.It is composed of the National Center for Environmental Forecasting(NCEP),the National Center for Atmospheric Research(NCAR),and countless worldwide joint research.The purpose of WRF model is to perform weather forecasting and production data for atmospheric research.Most of the code is written in Fortran,so even on different platforms,as long as you have the applicable Fortran compiler,the source code of the model can be transplanted at the compilation level,which makes it cross-platform.Compared with other climate models,the WRF model has excellent data assimilation technology,powerful physical process modules,and supports nesting.It is very wise to run the WRF model on a supercomputer,you can greatly improve the efficiency of data production and save the waiting time of a large number of researchers.The main work of this paper is refactoring and optimizing WRF model on SW26010 heterogeneous multicore processor and run it on"Sunway·Taihulight"."Sunway·Taihulight" is one of the new generation supercomputer in China,Its peak performance reached 125.4Pflop/s.The basic software on "Sunway·Taihulight" covers custom compilers for C/C++and Fortran.The system support has expanded a lot of syntax OpenACC 2.0 syntax.In addition,the system provides a thread library called Athread,using A thread can achieve fine-grained parallelism.This article introduces the refactoring of the WRF model on "SW2610"heterogeneous many-core processor,and fine-grained optimization of the stencil-type loop through the athread thread library of the domestic platform.For the physical process part,this article introduces the optimization of non-stencil loops by rewriting the function structure and using OpenACC pragma statements to split the loop.For the communication part,we introduced how to use experimental methods to change the process layout and maximize the problem of load imbalance.In addition,we gradually increase the parallelism to expand the entire model on multiple computing cores.In addition,this article introduces the methods of performance analysis and effect evaluation.The instrument performance analysis tool GPTL based on "Sunway·TaihuLight" is introduced,and the performance of WRF mode is tested by using GPTL.By using Athread and OpenACC,the program is optimized.Compared with the pure main core version,the optimized model performance has a speedup of 2.4.