Study Of GRAPES Numerical Weather Prediction System Optimization On Domestic High Performance Computers

Numerical weather prediction is a comprehensive subject which is combined with high performance computing technology and numerical weather forecast technology. On the one hand, numerical weather forecast technology is an important driving force of high performance computer development; On the other hand, high performance computers provide numerical weather prediction with software and hardware to use more complicated and accurate models. In recent years, the performance of super computers has increased incredibly. Heterogeneous processors have become the constituent part of heterogeneous high performance computers. As the storage system and memory access method of heterogeneous computers are different from the computers made up of single core or multi-core processors, it is necessary to study new parallel computing methods that can take full advantage of heterogeneous high performance computer architecture.GRAPES is a new generation numerical weather prediction system developed by China, the communication mode and calculation process of which are very complicated. There is urgent need to figure out how to make good use of heterogeneous high performance computer architecture advantage and improve the scalability and computing efficiency.The thesis introduces the scientific computing principle of GRAPES, finds the hotspot subroutine of the program. To solve the communication hotspot concentration and large system message buffer problems, the thesis designed highly effective communication strategies for each subroutine respectively. The thesis also designed a multi-level heterogeneous parallel scheme and several optimization strategies for the iteration integration part of GRAPES model so that the program can make good use of the co-processors to accelerate calculation and obtain ideal parallel efficiency.By using different resolution of input data, the thesis tested the program with different amount of processors and obtained elapsed time and speedup. The results show that the optimization schemes can effectively improve the efficiency of GRAPES model and scalability, making good use of the heterogeneous many-core processor. In addition, the multi-level heterogeneous parallel computing scheme is also applicable to other high performance computing applications with multiple cycles.