Parallel Optimization Of Seismic Wave Simulation Based On E-level Supercomputing Quasi-P Wave Equation

Seismic wave simulation is an integral part of geophysical science research.In geophysics,numerical simulations reveal some physical field information,which are not easily obtained during measurement.Therefore,numerical simulation can better understand the field behavior and underground st ructure.When performing seismic wave simulation calculations,for smaller models,a single CPU can be used for numerical simulation of the wave field.However,when the scale is large,this operation efficiency is far from meeting its industrial applicati on requirements.What's more,when the model is huge,it is often difficult for a single CPU's memory to meet its computational storage requirements.An effective method to solve such a huge amount of calculation or a single CPU memory demand is to use par allel computing.Therefore,it is of great theoretical significance and application value to study the parallel optimization of seismic wave simulation based on elliptic decomposition of quasi-P wave equation.In this paper,the forward simulation algorith m based on the ellipse decomposition method of quasi-P wave equation is carried out on the Dawn E-class prototype.The main work is as follows:(1)An in-node optimization scheme based on CUDA stream and Linux multithreading technology is proposed.CUDA as ynchronous transmission is used instead of synchronous transmission to complete the necessary intermediate data exchange between different CUDA cores while computing,which greatly reduces the idle waiting time of the CUDA cores,and introduces Linux multi threading technology to improve CPU execution efficiency.(2)An inter-node optimization scheme based on multi-node process partition and intra-node parallel acceleration is proposed.The three-dimensional division of the shot data is used to divide the si ngle shot data,and the MPI message transmission mechanism is used to bind the divided sub-domains to each process(computing node)and the data communication between the nodes in the entire calculation process.At the same time,each process(computing no de)uses CUDA streaming and Linux multi-threading technology to effectively improve the computing efficiency of each node's GPU and CPU.Finally,the distributed calculation of each shot data among multiple nodes is realized,and the scalability and portab ility of the algorithm are effectively achieved.(3)Designed a seismic wave simulation web platform based on E-level supercomputing.Hiding the implementation process of the underlying algorithm and complex supercomputing parameter configuration make the application of the algorithm more transparent and easy to operate.