Research On The Optimization And Design Of The Initialization Software System For Numerical Weather Forecast

Numerical weather forecast is one of the important applications in high performance computing.It is a technology that solves a set of partial differential equations that describe the evolution of atmospheric physics by computer with the eligible initial value,in order to forecast the weather events in the future.As a forecast problem,the quality of initial value determines the forecast accuracy of the numerical weather forecast to a large extent.The accurate numerical weather forecast must be based on the high quality initial value.Data assimilation?or data initialization?is a method of combining information from different sources to obtain high quality initial values,and is now widely used to provide high quality initial values for numerical weather forecast.In this paper,the existing data assimilation algorithms were deeply researched and analyzed in detail.Aiming at the problems existing in the current data assimilation algorithms,a better four-dimensional variational data assimilation of dimension reduction projection algorithm DRP4Dvar was selected.By means of this algorithm,a set of perturbation collection samples were used to generate the basis vector,the analysis increment of the model space was projected onto the subspace set of basis quantum,and the control variables of the objective function were transformed from the analysis increment to the base vector coefficient.Thus,the projection dimension reduction of computing time was reduced.The optimization effect of projection dimension reduction made the space dimension from the original 10⁸ to no more than 10².Based on this work,the serial version of the four-dimensional variational algorithm of dimension reduction projection was reconstructed and optimized according to the parallel rules of Gridpoint Statistical Interpolation?GSI?.During the optimization process,the analysis and mining the main calculation of algorithm in the minimum of the iterative process was the objective function after bringing in the localized matrix,then the objective function were achieved in parallel through using MPI parallel program model in the 128 multi-core processors,the running time was reduced to about 1/26 of the original,thus solving the technical bottlenecks of the entire assimilation system.In order to test the correctness and performance of data assimilation system in the numerical weather forecast,a reasonable test case was designed and a performance test scheme was proposed.The test was tested on the experimental platform,and the test results were analyzed.The results show that the data assimilation technique proposed in this paper are not only correct,but also improve the initial quality,the performance of the whole system is greatly improved.