The Parallel Computing Of Adjoint Models Of Variational Data Assimilation In Numerical Weather Forecasting

The combination of scaling parallel computing technologies and numerical weather forecasting research results is of great significance to develop the high performance parallel computer application, to improve the numerical weather forecasting level, and to advance the parallel computer study. Based on the principles of scaling parallel computing and variational data assimilation, this dissertation systematically studies scaling parallel algorithms and parallel implementing technologies of adjoint models, and designs high performance scaling parallel computing adjoint models. The researches are aimed at the 'Military Global Numerical Weather Advanced Prediction System' and the 'National Global and Regional Multiple Scale Advanced Prediction System', and the researches closely follow international parallel computing and numerical weather forecasting technology studies. The main achievements in this dissertation are summarized as follows.1) The decomposition technology for establishing adjoint equations is put forward. The adjoint models are derived from continuous and discrete two dimension shallow-water models.2) The parallel 'Arakawa C -grid' algorithm of adjoint model was designed. The parallel algorithm holds parallel computing high performance scalability.3) Based on 3-time data distribution and 4-time two-dimension matrix transposition, the parallel spectral algorithm of adjoint models is put forward, achieves good load balance, decreases the communication overhead, and holds good application and high performance scalability. The first national scaling high performance parallel medium-range numerical weather forecasting spectral model system PT106L19 is designed. Compared with CCM3, the system holds higher performance and betterscalability. Compared with T L 213L31, it holds the same performance.4) The circle data partition method of spectral adjoint model is put forward. Themethod increases 30.7% of parallel computing ratio under distributed-memoryenvironments with 32-processors. The butterfly-net data redistribution of spectral adjoint models is put forward. In this method, the communication overhead is decreased.The communication latency hiding technology of spectral adjoint models is put forward, which includes computation and communication overlapping and structure data transposition. The technology increases 12.9% of the parallel computing ratio under distributed-memory environments with 32-processors. 5) The fact that the parallel computing designing principles of spectral adjoint model and forecasting modes are equivalent is shown clearly. With the combination continuous and discrete methods, the adjoint model of Semi-Lagrangian spectralmodel T L 213L31 is designed.