Parallelization And Optimization Of The Global Numerical Weather Prediction Spectral Model Based On Many Integrated Cores(MIC)

With the increase of the resolution of the numerical weather prediction model,the calculation amount is increasing rapidly,and the demand of the super computer performance is also increasing.While the traditional CPU processor architecture has become a bottleneck in the growth of computing performance due to the limitation of Moore's law.The new architecture of MIC core processor is one of the solutions to this problem.Based on this background,the parallel algorithm of global numerical weather prediction model in CPU+MIC hybrid heterogeneous platform is carried out.The main work includes the following two aspects.(1)Using the MIC native mode(native mode)to carry out the study of the global numerical weather prediction model on the single MIC card.Through the research on the process of the model,the research of the thread and the extension of the process,the topology structure of different processor is set up,and the running time of the optimal version is 8.21 s.In the CPU processor as the reference group,the number of processes and the number of threads are also extended to get the optimal running time of 1.7s.This shows that the performance of the sixteen core CPU processor is 4.82 times of the performance of the single block MIC processor,and the performance of the MIC coprocessor cannot be fully demonstrated in the native mode according to the peak performance of the two hardware.But it provides the basis for the application of extended to peer mode.(2)Using the offload model,a numerical weather prediction model is run on the CPU+MIC mixed heterogeneous platform.In this model,the paper proposes a strategy for the selection of hot code segments,and the offload communication domain and the computational domain are separated.For the hot code segment of the MIC core processor,the use of parallelism,cache hit rate,vector and other methods to optimize.After multiple optimization,the time consuming is about 1.8s,compared with the CPU processor,the operation time is 7.6s,the performance is improved.