Research And Application Of The Algorithm And Parallelization Scheme For Hybrid ETKE-4DVAR Data Assimilation

In the major application field of data assimilation – numerical weather prediction, as the soaring development of observation technology and the research on application of varieties of observation devices, there is an exploding growth in regional and global observation data. Furthermore, the deviations from different devices and observation locations as well as the defects within atmospheric models have caused a more demanding requirement on the precision and computation of data assimilation algorithms. And the development of high performance computers especially SMP clusters also requires a higher degree of parallelization of applications running on them. Thus, data assimilation algorithms have to achieve more parallelization efficiency. For this purpose, we present a parallelization scheme of our chosen data assimilation algorithm, namely, ETKF-4DVAR.Firstly, we give a description of current global researches on the parallelization of data assimilation algorithms and point out their drawbacks and shortcomings. Based on existing data assimilation algorithms, we give an in-depth analysis of theoretical basis and parallelization qualities residing in ETKF-4DVAR. After that we make a brief conclusion of the possibilities of parallelization for each phase of this algorithm as well as the necessity of designing program interfaces which connect model and our data assimilation scheme. Then we present a detailed analysis of the parallelization strategy for the analysis phase of this algorithm so as to offer theoretical basis to subsequent design of our scheme.Subsequently, we also show the implementation and performance requirements of analysis phase based on mode-decomposition(distribute ensembles over processed) and domain-decomposition(divide model domain into sub-domains). Next to that we explain the advantages of domain-decomposition over mode-decomposition both on communication and memory requirements, providing foundations for the choice of adopting domain-decomposition in the experiment schemes.Then we make a further research and analysis on the parallelization of forecast phase. Furthermore, in order to reduce the coupling between model and ETKF-4DVAR scheme we also present a kind of framework structure to define interfaces between these two parts. And accordingly, the detailed design and implementation of our scheme are presented in this thesis, including interface design and MPI-based process configuration in the cases of joint process sets and disjoint process sets as well as the comparison between these two configurations.Finally by presenting the implementation and conducting a series of experiments, we validate the effectiveness and parallelism of our scheme. As the results have shown, the parallelization scheme of ETKF-4DVAR we present in this paper has achieved a satisfactory degree on both data assimilation feasibility and effectiveness. And the parallelization scheme is able to alleviate the negative performance influence introduced by changing data assimilation parameters.