Optimization And Parallel Implementation Of Oceanic Variational Data Assimilation

High performance parallel computation is closely related to the development of numerical oceanic prediction, and it prompts the innovation of prediction methods and the improvement of prediction resolution and precision. OVALS is developed by the Institute of Atmospheric Physics, which provides initial fields to ENSO(El Nino Southern Oscillation) prediction system by using the 3D-Var method to assimilate various temperature and salt data. In order to take advantage of the computation power of HPC and improve the in-time property of prediction, the thesis focuses on the serial performance optimization and parallelism of OVALS system.Based on the research of various optimization methods and basic principal of OVALS, the thesis uses the TAU, which is a powerful performance analysis tool, to analyze the OVALS system. Loop optimization and high performance math library methods are introduced to optimize the hotspots and performance bottlenecks of the application. To the weakness of I/O performance, redundant I/O operation elimination, observation data buffers creation methods are used to improve the I/O performance. The performance of optimized OVALS is triple as much as the original.After deeply analyzing the data flow of OVALS system, the thesis parallelizes OVALS systme. In the process of parallelizing the ASSIMI part, the Layer-First-Processor-Allocation algorithm is proposed and the global communication operations and parallel I/O are implemented based on the LFPA algorithm. In the process of parallelizing the ASSssh part, two-dimentional irregular domain decomposition technique is used to make the computation load of parallel OVALS system balance. When the OVALS is running with 36 processes, the speed-up gets up to 17 and a high parallel performance is achived.