Research On Parallel Computing Performance Optimization Of Reverse Time Migration Imaging Algorithm

Petroleum is generally buried at 0-10,000 meters underground,passing through the depletion of oil and gas resources in multiple shallow oil and gas fields of 300-5,000 meters.The petroleum exploration industry needs to solve the development of deep oil and gas fields at a depth of 6000-9000 meters.The current main exploration method is seismic exploration.,Magnetic exploration and gravity exploration.Among them,the reverse time migration imaging algorithm is the most commonly used in seismic exploration.It is one of the most accurate imaging methods among the existing alternative methods in petroleum exploration.With the rapid development of computing,high performance,GPU,and other computing technologies,the application of inverse time transform imaging algorithms has also been gradually developed,and significant application effects have been obtained.However,the current Reverse Time Migration imaging algorithm still has problems that need to be further optimized in parallel computing performance.Specifically: due to the insufficient utilization of computing resources such as stand-alone GPUs,the calculation of the algorithm takes a long time,and the utilization of resources is low;the gradual increase,more and more massive data leads to the reverse time migration imaging algorithm in a stand-alone environment Problems such as storage calculations occurred during the next operation.In response to the above problems,this article focuses on the performance of the core computing unit FDTD3 d in the reverse-time migration imaging algorithm in largescale data scenarios,the performance of the reverse-time migration imaging algorithm in a stand-alone environment,and the performance of the reverse-time migration imaging algorithm in a large-scale data environment.To solve the problems of low resource utilization,long calculation time,and storage bottleneck,the research puts forward a series of calculation optimization strategies for large-scale data reverse-time migration imaging algorithms.The main research work and contribution points of this paper are as follows:(1)Aiming at the problem of low utilization of the FDTD3 d Cache,the core computing unit of the reverse time migration imaging algorithm,the research and realization of the pipeline-based parallel FDTD3 d optimization algorithm.Firstly,the problem of the time offset mechanism in the original algorithm is analyzed.Then the FDTD3 d calculation model based on pipeline processing is proposed from the algorithm level.The FDTD3 d calculation optimization method based on data caching is further realized to improve Cache's use and calculation efficiency.(2)In view of the long time-consuming calculation of the reverse time migration imaging algorithm in a stand-alone environment,the research puts forward strategies and methods to make full use of computing resources and improve GPU utilization.Aiming at the problem of low GPU utilization of reverse-time migration imaging algorithm in the single-machine single-GPU scenario,this paper proposes pipelinebased continuation iterative optimization and optimization methods based on data compression and buffering;for the existing reverse-time in a single-machine multiGPU environment,The data communication overhead of the migration imaging algorithm is high.This paper proposes a point-to-point data transmission optimization method.(3)Aiming at the storage and computing bottleneck problem in large-scale data processing scenarios,this paper designs and implements a distributed reverse time migration imaging algorithm based on Spark,using the independence between different shot wavefield data in the reverse time migration imaging algorithm.It realizes the parallel computing method based on the data partition.On this basis,further research and realization of the reverse-time migration imaging algorithm combining Spark and GPU are realized.Finally,through the Pre-Shuffle and Cache optimization strategies and the data load balancing optimization strategy,the performance optimization of the algorithm execution at the Spark system level is realized.(4)The optimization strategy proposed in the research work and contribution point(3)of this article has been deployed in the real production environment of a large stateowned oil exploration company and has been operating stably for nearly one year.Under the premise of ensuring the correctness of the algorithm,the experimental results of performance comparison with the company's original reverse time migration imaging algorithm show that the optimization algorithm proposed in this paper reduces the algorithm running time by more than 43%.