A Parallel Algorithm Of E Grid-based Numerical Prediction Model AREM

The AREM mode (Advanced Regional Eta-coordinate Model) is established byInstitute of Atmospheric Physics, Chinese Academy of Sciences. It is a mesoscaleregional numerical prediction model with independent intellectual property rights forChina rainstorm forecast. This mode is widely used at domestic and foreign inmeteorology, hydrology, environmental and military research and business units.However, the computation growth rapidly with the improved mode resolution, theoriginal parallel programs have been increasingly unable to meet its huge demand forparallel computing. In this case, the research on the AREM mode high efficient parallelalgorithms is of great significance.The AREM modes latest parallel version is a MPI parallel version established byPu Ye, etc. This version uses a one-dimensional data partitioning strategy. Along themeridional direction divide forecast mesh evenly, and assigned each part to therespective processors for parallel computing, the assignments difference of eachprocessor up to one zonal line.The one-dimensional data partitioning strategy has the advantage of easyprogramming and simple communication. But divided in one direction, the AREMprogram can only do coarse-grained parallelism. When the number of processorsincreases, the AREM one-dimensional data partitioning strategy load imbalancesituation intensifies and sharply increase in total communication. So the parallelscalability and the parallel efficiency of this version are not good. The run tests showthis version can only be effectively extended to about36processors, can not meet thedemand of the current model of parallel computation.In order to solve the poor parallel scalability of one-dimensional data partitioningstrategy, this paper designs and implements an AREM mode two-dimensional datapartitioning strategy and corresponding communication algorithms. Two-dimensionalsubdivision strategy can be achieved to the program’s fine-grained parallelism, so it hasbetter load balancing and less total communication, which leads to a good parallelscalability. Finally, we tested the performance of the new subdivision strategy. Theresults show: The new implementation can significantly improve program parallelscalability and parallel efficiency.