Research On Parallelization Method Of Reverse Time Migration Based On Multi-GPU

The reverse time migration method is currently the theoretically highest seismic migration technology with the highest imaging accuracy.However,as a computationally intensive task,in order to achieve high precision and quasi-phase,the demand for calculation and storage is very huge,especially in recent years In order to meet increasingly sophisticated exploration requirements,the research of reverse time migration has developed from isotropy of underground media to anisotropy,from scalar acoustic wave equation to elastic wave equation,and from two-dimensional model to three-dimensional model,Which brings great challenges to its industrial application.Parallel computing is an effective means to solve such huge calculation and storage problems.When the calculation scale of reverse time migration is large,the storage capacity and computational efficiency of a single GPU often cannot meet the requirements.Therefore,research on the reverse time migration method based on multi-GPU parallel optimization has very important practical significance and application value.In this thesis,the research on multi-GPU parallel optimization of reverse time migration algorithm is carried out on the Sugon advanced computing platform.The main work is as follows:(1)A CUDA parallel method for solving the maximum wave field gradient is proposed,and the GPU optimization of the reverse time migration algorithm is realized.GPU acceleration is performed on the part of the reverse time migration algorithm that is computationally intensive and has no logical dependence.In the part to find the maximum wave field gradient,use CUDA to organize the thread block,and the wave field gradient of the specified span is formed into an iterative pair within the block.Continuously reduce the span to find the maximum value iteratively,which greatly reduces the calculation time and effectively improves the operating efficiency of the reverse time migration algorithm.(2)A method for implementing reverse-time migration for multi-GPU parallel optimization is proposed.Based on the improved reverse time migration algorithm,the parallel optimization method of dual GPU cards is studied,the process of area division and data exchange is elaborated,and the specific implementation process of the dual GPU reverse time migration algorithm is designed.Since the single node of the Sugon computing platform supports up to 4 GPU cards,in order to maximize the use of hardware resources,the algorithm research is extended to four GPU cards.The experimental results show that the multi-GPU parallel optimization method can effectively improve the computational efficiency of the reverse time migration algorithm,and broaden the total memory capacity.The four-GPU card performs better on the problem of larger computing scale.(3)The point-to-point communication technology is proposed to optimize the data exchange process between multi-GPU cards,and the asynchronous transfer function is used to improve the degree of computing overlap between multiple GPUs.According to the point-to-point communication method,the data exchange part of the multi-GPU reverse time migration algorithm has been improved,which effectively accelerates the communication process.At the same time,the asynchronous transfer function is used to avoid the implicit synchronization between the host and the device,and the calculation overlap between multiple GPUs is improved.