Study And Implementation Of Parallel Computing For IAP9L AGCM

The climate model design and its research is always the critical point during the climate researches. And atmospheric general circulation model is one of the major models, used by many groups around the world for both climate prediction and its researches. Because its numeration needs vast amounts of computation and memory, the history of AGCM is intimately tied with developments in high performance and parallel computing. And it is very important to make parallel computing researches on AGCM.IAP9L AGCM is composed of two components: the"dynamics"component and the"physics"component. And the numerical solution of the dynamics is typically implemented, using a finite difference scheme. To meet its demands for parallel computing, after having an in-depth understanding of the basic principle of the atmospheric general circulation model, we make the data dependence analysis and the time-cost analysis of the serial program of IAP9L AGCM, based on the general parallel computing theory and the idiographic characteristics of the atmosphere model. The process of designing a parallel algorithm generally includes partitioning, communication, agglomeration and mapping. According this process, we develop a parallel algorithm of IAP9L AGCM, based on 1D(latitude) domain decomposition scheme. And we adopt a uniform partition strategy and a load balancing pretreatment method, based on the feedback checkout, to decompose the global domain. And then, we implement the parallel algorithm with MPI and conduct a parallel numerical experiment of IAP9L AGCM on an 8-CPU cluster of workstations. At last, after analyzing and comparing the results of the parallel programs, we conclude three points: (1)The pretreatment method can improve the performance of the parallel program to a certain extent, compared to the uniform partition strategy. (2)For the data-parallel problem of IAP9L AGCM, which is solved by a finite-difference scheme, the parallelization strategy, which is 1D(latitude) domain decomposition scheme, does not scale to very large numbers of processors. When the size of the problem is changeless, in the condition of the suitable communication management, we should choose the right number of CPU to avoid the decline of the parallel speedup. (3)The experiment shows the feasibility and the effectiveness of the algorithm we develop.