A multi-GPU parallel computing method for 3D random vibration of train-track-soil dynamic interaction

In this paper, an efficient computation method based on a multi-GPU parallel algorithm is proposed to overcome the low efficiency in random calculation of the train-track-soil coupled system (TTSCS). Firstly, for the large time consumption caused by solving multiple independent equations of TTSCS at different frequency points in serially random vibration analysis, the multi-GPU parallel algorithm is proposed and programmed based on the OpenMP-CUDA algorithm. The tasks of solving multiple linear equations for random vibration analysis of the TTSCS are distributed to different GPUs for parallel execution. On each GPU, the large sparse linear equations of TTSCS are solved by the CUDA-based parallel preconditioned conjugate gradient (PCG) method, and the large sparse matrix is stored in the compressed sparse row (CSR) format to save memory space. Then, the parallel computing program is implemented on the MATLAB-CUDA hybrid platform. Finally, numerical examples show that the efficiency of solving large sparse linear equations based on the multi-GPU parallel algorithm implemented on a 4-GPU node and the GPU-accelerated PCG algorithm implemented on a personal computer with a single GPU is 22.59 times and 3.75 times that of the multi-point synchronization algorithm (MPSA), respectively.