Research On Parallel Computing Of Runoff Generation Based On WEP-L Water Cycle Model

At present,more and more attentions have been paid to the study of watershed hydrology informatization in China,but the simulation speed of watershed hydrology model can not meet the demand of the high-precision study of watershed hydrologic processes due to the huge amount of processing data and complex physical calculation required more time.Parallel computing technology based on multi-core is the mainstream method to improve the simulation speed.However,how to design a parallel algorithm reasonably,which can be applied to the hydrological model and how to efficiently utilize the computing power of CPU and GPU? These have become the key problems to be solved.To solve the above problems,this paper selects three representative parallel technologies,namely OpenMP,MPI and CUDA,based on the runoff generation module of the WEP-L water-cycle model.To realize the parallel design of the model,and to conduct effective analysis and optimization.The main work and achievements are as follows:(1)Design and implementation of parallel computing based on OpenMP.In this paper,the sub-watershed cycle is allocated by exponential scheduling,and the number of sub-watersheds decreases from large to small,and the multi-thread on CPU is allocated dynamically.What's more,the global variables with cyclic dependency are designed by thread privatization to solve the problem of data competition.The parallel model is suitable for a multi-core host.When the number of threads is 20,the total time of the runoff generation module in the Yellow River basin reaches the minimum,which is shortened from 9,556 seconds for a single thread to 1,219 seconds.When the number of threads is 18~20,the peak of speedup ratio is 7.8.(2)Design and implementation of parallel computing based on MPI.In this paper,The whole basin is divided into sub-regions with uniform computation time relatively by using the parallel design mode,and the multi-process on CPU is allocated statically.Then the message communication design adopts the method of combining "point to point" and group communication to carry on the data transmission.The parallel model is suitable for the multi-core host or multi-host cluster.When the increment of communication time is greater than the reduction of calculation time,the parallel computing reaches the bottleneck.At this time,the number of processes is 10,the total time of the runoff generation module in the Yellow River basin reaches the minimum,which is shortened from 5874 seconds for a single process to 1231 seconds,and the peak of speedup ratio is 4.8.(3)Design and implementation of parallel computing based on CUDA.In this paper,To divide the whole basin into sub-regions with the uniform number of sub-basins relatively,and to allocate it statically to thousands of threads on GPU in turn.Then to set the number of thread blocks is twice multiprocessors,which can make full use of the computing resources.As well as limit the number of registers occupied by a thread to exchange the memory access time for more threads to run.The time is shortened from 5251 seconds before the limit to 3422 seconds again.The speedup ratio is between 2 and 3 times the serial computing of the common host.