| Full waveform inversion(FWI)is famous for its high-precision reflection of geological models in the field of seismic exploration.The FWI algorithm of establishing wave equation in spherical coordinate system can meet the large-scale exploration range and adapt to the real spherical body model.As a computing intensive task,the FWI algorithm in spherical coordinate system needs a great deal of calculation and storage in order to increase the calculation scale and improve the calculation accuracy.This thesis studies the parallelization of the full waveform inversion algorithm in the spherical coordinate system on the Sugon level-E supercomputing platform.The main work is as follows:(1)A method for the multi-node and multi-DCU parallel optimization is proposed.Based on the level-E supercomputing platform,this work divides the spherical body model of the FWI algorithm in the spherical coordinate system,realizes the parallel acceleration of Multiple DCU cards in a node to the grid point calculation of multiple model subdomains,and then uses the RDMA technology based on MPI to expand the multi-DCU optimization algorithm for a node to multiple nodes,so as to fully tap the computing power of the cluster.Finally,several groups of experiments are designed,and the results show that the multi-node and multi-DCU parallel optimization method has a certain acceleration effect on the level-E supercomputing platform.(2)Combined with MPI,HIP,and Open MP,a three-level hybrid parallel optimization strategy is proposed.Based on the parallel optimization method of the multi-node and multi-DCU to realize the parallel of three-dimensional model,continue to use the multiple processes based on MPI to realize the parallel calculation of multi shot data.Based on the specific performance of the algorithm on level-E supercomputing,Open MP technology is introduced to replace the module with weak acceleration effect of DCU cards.Finally,the final optimized algorithm is deployed on the level-E supercomputing,which is comprehensively tested and evaluated from different angles by using the current mainstream parallel performance evaluation theory,and a comprehensive parallel setting scheme of the model parallelization and multi shot parallelization is proposed.The results show that the comprehensive optimization strategy proposed in this work has obvious acceleration effect and good scalability.(3)From the perspective of software development,a "MVTS" software architecture mode for Level-E supercomputing is designed,which realizes the four-layer decoupling among Model,View,Template and SSH-HPC.Using the "MTVS" software architecture,this work develops an energy exploration software prototype system based on the Django development framework with friendly interaction ability.The system not only integrates the software modules of various common algorithm optimization versions for the energy exploration,including three-dimensional FWI software,but also realizes the process construction of FWI algorithm software in spherical coordinate system based on spectral element method.Finally,the source code of this system is released on the Gitees platform.(https://gitee.com/bqf2022/hnuseisoftweb.git)... |
