Development And Application Of GPU Accelerated Three-Dimensional Hydrodynamic Numerical Model

Unstructured grid and finite volume hydrodynamic models have been widely used in hydrodynamic numerical simulation of harbor,coastal and offshore engineering.Detailed coastline and topography increases the computation overhead in hydrodynamic numerical simulations.Therefore,hydrodynamic numerical models require methods with high efficiency.The graphics processing unit(GPU)has improved rapidly for scientific computing.Therefore,GPU parallel algorithms have become popular to develop efficient hydrodynamic numerical models.This paper proposes a GPU parallel hydrodynamic numerical model using the compute unified device architecture(CUDA)and based on a three-dimensional unstructured grid and finite volume method hydrodynamic numerical model.The main aspects of this paper are as follows.(1)The process reconstructed by monotonic upstream centered scheme for conservation laws(MUSCL)and Van Albada flux limit methods is shown to fulfill the requirement of mass conservation and enable the model to achieve zero numerical errors.Modification of the water level gradient preserves the well-balanced property.Bottom friction is corrected using an implicit algorithm,according to the velocity direction between time steps.The proposed numerical model improves stability for the shallow water problem solution and dam breaks..(2)The proposed numerical model was applied to several classic examples.Model predictions showed satisfactory agreement with analytical solutions and measured data.This indicates that the proposed model can preserve the well-balanced property,mass conservation,stability,and reliability.(3)A GPU parallel algorithm was developed to realize mode acceleration,and was used to solve the low efficiency problem while simulating a large scale,high resolution problem without requiring high performance computer equipment.First,a kernel function based on the control volume mapping was constructed to handle the control volume cycle calculation partial program within the serial model,which showed that the GPU parallel model could be effectively executed on the GPU.Then the rational distribution of threads model,optimization memory access,and loop fusion were adopted to optimize the GPU parallel model.In contrast to conventional GPU parallel mode,each method can enhance the GPU parallel mode's computing performance.The proposed GPU parallel model was applied to two-and three-dimensional tidal flooding and dam break problems with complex topography,and performances for different calculation conditions were compared.The performance of the GPU parallel model was fully exploited as the number of grids increased.Compared with serial computing,the computing performance achieved 37 times speedup on a Tesla K20 device,providing significantly superior performance.(4)The proposed model was applied to hydrodynamic numerical simulation of Moon Bay and the Xiangshan narrow Bay,analyzing the hydrodynamic characteristics.Time costs corresponding to different models(GPU and MPI parallel mode)were compared.The results suggest that the proposed model is suitable to be applied to many hydrodynamic problems while preserving highly efficient computing performance.