Global Ocean Data Assimilation System Design And Algorithm Acceleration Based On Local Ensemble Transform Kalman Filter

Oceans have played a key role in human existence since ancient times.Numerical prediction and numerical simulation are important and indispensable for marine studies,and data assimilation techniques are one of the most widely used methods.Data assimi-lation is capable not only of optimizing model analysis and observational data to provide more accurate initial fields,but also to producing high spatiotemporal resolution reanalysis datasets.Consequently,it can also contribute to the development of new parameteriza-tion schemes.In addition,with the increase of amount and types of observational data,it is increasingly important and urgent to optimize the computational performance of data assimilation systems.In this paper,we developed a global ocean data assimilation system based on the Lo-cal Ensemble Transform Kalman Filter(LETKF)and the Parallel Ocean Program(POP)of the Community Earth System Model(CESM).Optimum Interpolation Sea Surface Tem-perature(OISST)and Array for Real-time Geostrophic Oceanography(Argo)data were used in the assimilation experiment.The assimilation results compared well with the orig-inal observational data and the SODA(Simple Ocean Data Assimilation)reanalysis data.To enhance the computational performance of the data assimilation system,we optimized the communication,I/O(Input/Output),and load balancing of LETKF.This load bal-ancing strategy can ensure that the system performance is doubled without changing the assimilation result,which greatly saves computation time.Importantly,this optimization strategy has strong scalability and has the same high performance at high ocean model res-olution(0.1~°×0.1~°).The new load balancing optimization scheme is applicable not only to data assimilation methods using localization strategies,but also to the design and devel-opment of data assimilation systems in other fields such as the atmosphere.Furthermore,the optimization in communication and I/O can potentially improve the performance of models at high resolution or with extensive assimilation variables.The research has the following two innovations:(1)A new global ocean data assimilation system with LETKF based on the POP ocean model in CESM was developed.The system is stable and the assimilation perfor-mance is comparable to the state-of-the-art ocean reanalysis data.It provides the founda-tion for development of a coupled data assimilation system.(2)A load balancing strategy based on the“greedy algorithm”is newly developed to solve the problem of uneven distribution of parallel calculation among CPUs in the ocean data assimilation.