Research On High Performance Of GRAPES Tangent/Adjoint Model With The MPI/OpenMP

Our independent research and development of the numerical Prediction System GRAPES(Global/Regional Assimilation and Prediction System) based on the principle of scientific research and business general, parallel computation model was designed and implemented, now has made significant progress. But it still exists problems: GRAPES have multiple mode at the same time, and has a large amount of calculation and storage, running time is also very long.In order to meet the demand of GRAPES calculation, in addition to improve the clustering performance, are also very good implementation of the parallel computing application ontology for improve the computing power. As a cluster of the most commonly used message passing interface MPI has both high performance, portable, powerful wait for a characteristic, is widely used in distributed storage system, has become one of the standard of parallel programming. But there are also instances show that the use of pure MPI message-passing programming model does not give full play to the cluster there are both distributed storage and Shared storage hardware resources advantage. And using MPI for messaging between the nodes, the nodes use OpenMP Shared storage MPI + OpenMP hybrid programming model can take full advantage of the characteristics of the cluster, in the optimization of node communication, which can obtain more ideal resultOn how to improve the GRAPES model on in the code, to improve the speedup and scalability, is key and key point in the research of this paper. Based on detail GRAPES model itself characteristics, this paper studies the mechanism of OpenMP and MPI parallel programming, on this basis to realize the tangent with pattern of MPI + OpenMP hybrid programming model. Article also testing the tangent mode and the adjoint model takes out time consuming larger subroutine, the subroutine in such aspects as communication processes and threads are optimized, and the GRAPES and redesign the communication between. For the above work is effective for GRAPES tangent with pattern, also has carried on the pure MPI and MPI + OpenMP hybrid model testing contrast, as you can see the results can be as pure MPI parallel program parallel efficiency drop to below 90%, the use of hybrid parallel, the parallel efficiency is higher than the former 5% to 5%, the result data validation in the using MPI on GRAEPS + OpenMP hybrid model of rationality, and mixed programming model can more effectively improve the efficiency of GRAPES mode of operation and speed than the conclusion. Finally also I/O bottleneck problem of the traditional model of GRAPES were analyzed, and the use of IBMcluster1600 disk system of its own resources advantage, using MPI standard defined in communication operation MPI_FILE_WRITE and MPI_FILE_READ parallel file system for parallel, speaking, reading and writing, speaking, reading and writing, and through the contrast test of parallel can be seen from the results, speaking, reading and writing can improve the GRAPES of the tangent with extensibility model.