The Hybrid MPI+OpenMP Parallel Method Of GRAPES-global Model

Clustered SMP systems are gradually becoming more prominence in HPC market,as advances in multi-core technology in recent years. The number ofhigh-performance computers’cores is increasing and computing performanceincreases with a certain annual growth rate. As a typical application ofhigh-performance computing, the accuracy of weather numerical models’forecastis largely due to resolution enhancement.To adapt this hardware architecture that combines both distributed and sharedmemory and solve the problem of computation growth led by increased resolution,utilizing hybrid MPI+OpenMP parallel programming model is a good trial. Thishierarchical programming model can achieve both inter-node and intra-nodeparallelization by using a combination of message passing and thread based sharedmemory parallel paradigms within the same application. MPI is used tocoarse-grained communicate between SMP nodes and OpenMP based on threads is usedto fine-grained compute within a SMP node.In this paper,a hybrid MPI+OpenMP parallel model suitable for GRPAES-Globalmodel is developed with the introduction of horizontal domain decompositionmethod and loop-level parallelization.The preferred method is horizontal domaindecomposition for uniform grid computing and thread-safe procedures, whileloop-level parallelization method is preferred for non-uniform grid computingand the thread unsafe procedures and load imbalance’s physical processes.Experiments with1degree and0.25degree’s dataset are performed and timingon main subroutines of integral computation is compared. Hybrid parallel methodbased on MPI parallelization can improve the degree of parallelism. And hybridparallel method within four threads is superior to pure MPI method underlarge-scale experiment. Hybrid parallel method can also achieve betterscalability than pure MPI method. The hybrid parallel method can keep highparallel efficiency in large parallel scale.The experiment shows hybridMPI+OpenMP parallel method is suitable for GRAPES-Global model.