Parallelization And Optimization Methods For Earthquake Simulation With Complex Topography

Earthquake simulation plays an important role in perfecting seismology theory and earthquake relief,but large-scale earthquake simulation faces severe challenges in terms of calculation and storage.Existing work has performed highly-efficient and high-precision simulation of the Tangshan earthquake using finite difference method on Sunway TaihuLight.However,most earthquakes in China occur in Sichuan and Yunnan regions with complex topography,and are usually beyond the simulation capability of straightforward finite difference method.Based on the Tangshan earthquake simulation work,my thesis introduces curvilinear grid to accurately describe complex topography,and proposes further parallelization and optimization methods for the more complex calculations and data characteristics of the new algorithm,to efficiently scale the Wenchuan earthquake simulation to over ten million cores.The main contributions of my thesis include:· Introduced curvilinear grid features for finite difference method,so as to enable ac-curate description of complex topography in seismic wave propagation;redesigned algorithm to improve the efficiency of earthquake simulation on large-scale parallel heterogeneous systems;proposed multi-level partition scheme,combined with the optimization of network communication and IO communication,and successfully expanded to the whole machine of Sunway Taihu Light.· To further improve the performance,in terms of effective memory utilization,the optimal schemes were explored from two-dimensional and three-dimensional per-spectives respectively;in terms of memory bandwidth,variable fusion methods,grid point rearrangement and collaborative access models were proposed;in terms of computing efficiency,vectorization strategies were proposed for two application scenarios.· Considering the general support for mixed precision of exascale supercomputers,reconfigurable computing platform was used to study and explore mixed-precision computing methods for earthquake simulation;through numerical analysis of simu-lation results,the optimal bitwidth and dynamic scaling factors were determined for different variables,so as to achieve a customized dataflow processor for earthquake simulation.For the case study of the Wenchuan earthquake simulation,parallelization and optimization methods in this paper achieved a sustained performance of 9.07 Pflops using the entire machine of Sunway Taihu Light,and verified the impact of complex topography on the wavefield from the simulation results.On the reconfigurable computing platform,the proposed mixed-precision parallelization and optimization methods in this paper,on the premise of ensuring the correctness of the values and wavefield,achieved an equivalent performance of 13.1 Intel Xeon Gold 6154 processors(18 cores each)or 2.1 SW26010nodes(260 cores each).The above results show methods and software proposed in this thesis can accurately simulate earthquake with complex topography,and gain high computing efficiency on top supercomputing systems.The proposed mixed-precision calculation method lays the foundation for highly efficient earthquake simulations on future exascale systems,and is expected to bring corresponding impetus to the research work in the field of earthquake research.