基于MPI的分布式水循环模型并行计算性能研究

Study on the Parallel Calculating Performance of the Distributed Water Cycle Model Based on MPI

为满足分布式水文模型快速模拟的需要,引入并行计算技术。基于MPI编程模型实现了WEP-L分布式水循环模型产流模块的并行编程,采取子流域任务划分和对等模式实现了模型的并行设计。为了减少进程间的通信时间,在产流计算之前,数据按随年、月、日变化以及不随时间变化分成四类,分批进行通信;产流计算完成之后,采用聚合通信方式中的数据收集,快速统计结果。模型并行化后应用于黄河流域,结果如下:(1)随着参与计算的进程数增加,并行计算的加速比呈先增加后减少的趋势,并行效率随进程数增加呈线性下降趋势。(2)模型并行性能受通信开销制约,当通信开销增量大于产流计算时间减少量时,加速比达到峰值4.8。

In order to meet the needs of fast simulation in the distributed hydrological models,the parallel computing technology is introduced.This study achieved parallel programming of runoff-module in the distributed water cycle model(WEP-L)based on MPI and applied task division of sub-basin and Peer-to-Peer mode to fulfill the parallel design.The way to reduce the communication time between processes is that:divide the data into four categories by the variation with years,months,days or time-invariant,then communicate them in batches before the runoff calculation.After completing runoff calculation,the statistics of the results were quickly gathered by the mode of collective communication.The model was applied in the Yellow River basin after parallelization,and the results can be as follows:(1)as the number of processes increases,the speedup ratio of parallel computing shows a trend of increasing first and then decreasing,and the parallel efficiency decreases linearly;(2)the parallel effect of the model is restricted by the communication overhead,and when the increase of communication overhead is greater than the decrease of runoff calculation time,the speedup ratio of the parallel computing reaches a peak of 4.8.