Abstract:

[Objective] Data assimilation has played an important role in numerical forecasting of the atmosphere and ocean. It can use observation data from different sources to modify initial field data, thereby improving the accuracy of numerical forecasting. The aim is to improve the efficiency of data assimilation calculation based on domestic heterogeneous computing platforms through the improvement of the SVD algorithm. [Methods] This article systematically describes the parallel strategy and implementation method in a large-scale environment, and designs and implements the implementation process and data structure of CPU and GPU-like accelerator collaborative batch SVD solution based on domestic heterogeneous computing platforms. It also provides actual performance improvement test data. At the same time, a complete data assimilation program is implemented using C/C++. [Results] This algorithm fully utilizes the computing resources, CPUs and computing cards of domestic heterogeneous computing platforms, achieving an efficient implementation algorithm for matrix inversion based on singular value decomposition (SVD), and significantly improving the computational efficiency of data assimilation from the basic algorithm. [Conclusions] The efficient implementation algorithm of singular value decomposition based on the collaborative approach of domestic heterogeneous computing platform CPU and computing card can be extended to algorithm implementation in fields such as quantum computing, artificial intelligence, image processing, and signal denoising, etc. It has high application value. The data assimilation application software using C/C++language enriches the application ecosystem of domestic heterogeneous computing platforms.

Key words: Singular Value decomposition, domestic heterogeneous computing platform, data assimilation, numerical forecasting, parallel strategy