Multi-GPU Parallel Computing Method For Thin Shell Structure And Its Application In Car Body Design

The finite element(FE)analysis method based on shell element theory is widely used in CAE analysis of automobile body.In this paper,it is mainly used to solve the nonlinear problem of plate and shell structure and the simulation problem of sheet metal forming.As the FE model becomes more and more complex,a large amount of computation needs to be processed in the simulation process,so the computational efficiency of the simulation is low.Due to its powerful data processing capabilities,parallel computing methods based on GPU have long been used to solve the problem of low computational efficiency in simulation analysis.However,a single GPU is limited by hardware resources and cannot meet the needs of large-scale computing.The current heterogeneous computing platform is capable of configuring 2-4 GPU devices,and the computing platform of the single-machine multi-GPUs breaks through the limitations of hardware resource on a single GPU.The main work of this paper is to apply the parallel technology based on single-machine multi-GPUs to solve the dynamic nonlinear problem of thin-shell structure and the simulation problem of sheet metal forming.The specific work and results are as follows:(1)Before calling multi-GPUs to calculate the FE model,it is quite important to reasonably decompose the FE model.Based on the METIS open source library,a data partition preprocessing system suitable for the algorithm of this paper is developed to realize fast region decomposition of FE mesh data.The system has two advantages: one is that the calculation of each sub-area is balanced,and the other is that the number of boundary points between the regions is small.(2)The multi-GPUs parallel computing method based on FE explicit solution is realized by the hybrid programming model of OpenMP and Compute Unified Device Architecture(CUDA),and the body structure analysis software based on multi-GPUs parallel technology with independent intellectual property rights is developed.The general idea of parallelized implementation is divided into two aspects: on the one hand,CPU thread level parallelism based on OpenMP mode,and on the other hand,GPU lightweight thread parallelism based on CUDA architecture.The former is mainly to open up the number of CPU threads with the same number of GPUs,and one CPU thread controls the startup and suspension of a GPU device.The latter is mainly to establish the abstract mapping relationship between GPU thread and element,node and degree of freedom,so that the explicit iterative solution part based on BT shell element theory can be processed in the whole process on multi-GPUs.(3)In the internal force assembly process of the node based on the pre-indexing strategy,when processing the FE model of the million-level element scale,the calculated circulation amount reaches the trillion level,which is quite time consuming.Therefore,in this paper,the pre-indexing strategy is improved.When the FE model of the million-level element scale is also processed,the calculated circulation quantity is only tens of millions,which reduces the time complexity of the algorithm and significantly shortens the indexing time.The numerical simulation of the nonlinear problem of the plate and shell is carried out.The calculation results show that the parallel computing method based on multi-GPUs can meet the accuracy requirements of engineering problems,and the computational efficiency is obviously improved.When solving the elastic deformation problem of 10,889,568 degrees of freedom on a four-way Titan Xp GPUs workstation,The four-way GPUs has an absolute speedup of 73 times and a relative speedup of 3.4,which is very close to the theoretical value.(4)This paper presents a multi-GPUs parallel computing method suitable for the simulation calculation of sheet metal forming.Aiming at the solving characteristics in the explicit iterative process,a multi-GPUs parallel computing method that is suitable for calculation of element including the constitutive of elastoplastic materials and a GPU parallel computing method of contact force solving are proposed,which realizes the full flow parallel of the explicit iterative process of sheet metal stamping forming simulation.The calculation accuracy and computational efficiency of parallel computing method based on multi-GPUs for sheet metal forming simulation are tested by numerical examples.The results show that the simulation results of the parallel computing method based on multi-GPUs computing platform are consistent with the results of the calculation method based on CPU platform.In terms of computational efficiency,when calculating the finite element model of 490,000 elements,the absolute speedup of nearly 210 times can be achieved based on the four-way Titan Xp GPUs computing platform compared to the calculation speed of the serial program.