Parallelizing And Optimizing Seismic Simulations On The New-generation Sunway

Earthquake is a very destructive natural disaster,which threatens the safety of human life and property.With increasing urbanization and growing population,earthquake-induced disasters are receiving increasing attention.Earthquake simulation can be used to improve earthquake theory,restore earthquakes and predict aftershocks,and has high research value in scientific research and people’s livelihood.AWP-ODC is a highly scalable parallel finitedifference application that enables large-scale 3D earthquake simulations.Accurate modeling of seismic models requires a relatively high spatial grid description to reflect high-frequency details.However,the bottleneck of computing power restricts traditional large-scale and highresolution seismic numerical simulation.The new-generation Sunway supercomputer has powerful computing capabilities and can support larger-scale and higher-resolution earthquake simulations,but the new supercomputing architecture and larger-scale earthquake simulations also bring new challenges to parallel optimization work.In this paper,AWP-ODC is transplanted to the new-generation Sunway supercomputer platform,and various parallel optimization methods for earthquake simulation are studied.A series of optimization methods are proposed from three perspectives:multi-level task division,process-level parallel optimization and thread-level parallel optimization,which greatly improves the performance of earthquake simulation.The main work of this paper is as follows:Through in-depth research on the Sunway supercomputing system,a multi-level task division scheme is proposed from multiple dimensions,which decomposes computing tasks into sub-tasks and accurately assigns them to corresponding computing units;Several process-level optimization methods are proposed,and in-depth optimization is carried out from three aspects:increasing communication bandwidth,reducing total communication volume,and computing communication masking.This paper proposes communication reduction and process-level computing communication overlap to reduce process-level communication overhead;Several thread-level parallel optimization methods are proposed.In terms of memory access,DMA data transmission schemes,array merging,data layout transformation,RMA communication schemes and thread-level computing communication overlap are proposed.In computing,a vectorization strategy between multiple points based on data reorganization is proposed.Supported by these optimization strategies,based on the new-generation Sunway supercomputer platform,this paper efficiently expands the earthquake simulation program to 25,000 processes(1,625,000 cores),achieving a peak performance of 4.55Pflops and a parallel efficiency of 98.9%.The earthquake simulation program can achieve a speedup of about 30 times on the Sunway processor.In this paper,large-scale numerical simulation and result analysis of the Tangshan earthquake and Wenchuan earthquake are carried out,and the simulation effect of the earthquake simulation program on real earthquakes is demonstrated.