The Research On Hybrid Parallel Algorithm For Forward Modeling And Inversion Of2.5-Dimensional CSAMT Based On MPI And OpenMP

With the continuous development of magnetotelluric explorationmethod, the performance requirements of traditional algorithm isincreasingly higher. Two dimensional algorithm is inadequate insimulating the actual terrain and three dimensional algorithm is tootime-consuming to putting into practical application, so2.5dimensionalforward and inversion is causing more and more attention. On the otherhand, parallel computing technology improves rapidly and is used in alarge number of domains. Thus, this paper combines the parallel computingwith2.5dimensinonal forward and inversion.Aiming at the problem of the low efficiency of the computing of2.5-dimensional controlled source audio magnetotelluric (CSAMT) serialforward and inversion, this paper implements a hybrid parallel algorithmbased on distributed MPI and multi-core multi-thread OpenMP. Besides,this paper implements a kind dynamic programming to achieve loadbalancing and enhance parallel efficiency. Forward parallel algorithmcan reach7.8times speedup and17.19times speedup with dynamicprogramming at most.Inversion parallel algorithm can reach2.83timesspeedup at most.In the forward, the algorithm is decomposed into two parallel layers.The outer parallel layer is frequency domain loop and is decomposed withMPI method. Main process assign tasks to the other collaborativeprocesses which decompose the inter parallel of wave domain with OpenMP. During the OpenMP computing in each process, several threads of CPUcalculating the filed values in different waves through setting sharedand private values. Then processes passing the result to the main processand main process will finish surplus job.In the inversion, the parallel algorithm decomposes the frequencydomain with OpenMP method. A group of threads calculates the field valuesof different frequencies. Then the main thread integrates the values tooutput the final data.To improve the parallel efficiency and exploit the parallelcapability of multicore and multithread, this paper implements a kindof dynamic task programming while assigning tasks. According to each PCprocess’s parallel computing capability, the tasks will be assignedLegitimately. This method is able to make processes’s computing timealmost the same and shorten the waiting time of processes which largelyimproves the parallel efficiency.Finally, this paper proves the accuracy and efficiency of the parallelalgorithm though a group of tests in different modelings and PC clusters.The result indicates that this algorithm can reach linear speedup on thepremise of accuracy. If using the dynamic task programming to assign tasks,this algorithm is able to finish in a shorter time and higher efficiency.