基于MPI/OPENMP混合编程的三维粒子模拟并行优化

针对MPI、OPENMP并行程序各自存在的缺陷,将MPI与OPENMP结合起来,实现了MPI/OPENMP混合并行编程;通过实验对MPI、OPENMP并行程序及MPI/OPENMP混合并行编程进行了分析.实验结果表明:MPI/OPENMP混合并行编程可以大大减少通信量,其效率和加速比均优于纯MPI并行程序,克服了MPI并行程序中因粒子分布不均匀使负载不均衡而导致的程序性能下降的问题,使得可以利用集群中的更多结点来进行计算,缓解了MPI并行程序的通信延迟问题;同时,MPI/OPENMP混合并行编程克服了OPENMP并行程序依赖于单台计算机处理能力和存储空间的问题,大幅度提高了模拟规模.

Parallel Optimization of Three-Dimension Particle Simulation Based on Mixed MPI/OPENMP Programming

In order to remedy the shortcomings of the parallel MPI program and the parallel OPENMP program,a mixed parallel programming combining MPI and OPENMP is implemented.Then,the parallel processes of the MPI program,the OPENMP program and the mixed parallel MPI/OPENMP programming are respectively analyzed through experiments.The results show that the mixed parallel MPI/OPENMP programming greatly reduces the communication traffic,achieves an efficiency and a speedup rate higher than those of the parallel MPI program,and avoids the performance degradation of parallel MPI program due to the load imbalance caused by uneven particle distribution.Thus,more nodes in the cluster can be used to perform a computation and the communication delay can be successfully decreased.Moreover,the mixed parallel programming significantly enlarges the simulation scale because it overcomes the dependency of parallel OPENMP program on the processing capacity and the storage space of a single computer.