Research On Computing Optimization Strategies Of CPU-GPU Seismic Data Processing Platform

Seismic data processing is a nexus key link in the oil and gas exploration, and it is also an important application field of high performance computing. As the refinement of the exploration, it puts forward higher requirements on the computing capability of the system. CPU-GPU heterogeneous computing systems have a significant advantage in terms of cost and performance. For the ability of better meeting the needs of seismic data processing, they are gradually to be concerned about. However, compared with the mature CPU environment, GPU is still insufficient in terms of programming models and optimization methods, which also bring some difficulties for developers to its efficient use. This paper will focus on optimizing seismic data processing on CPU-GPU heterogeneous computing platform, and then expand research on the following four aspects.First is the GPU parallelization of the most time-consuming migration in the seismic data processing, then on this basis, the CPU and GPU co-processing strategy in seismic migration has been proposed to fully tap the computing power of the system. Then the effectiveness of the strategy is verified by testing, and several factors may affect the processing efficiency have also been discussed.For the problem of CUDA has no explicit and efficient GPU global synchronization mechanism and needs kernel reboot, in this paper, two GPU global synchronization strategies without kernel reboot have been designed and implemented under the existing architecture and take use of the GPU synchronization mechanism that CUDA provided. Then using the widely used FFT in seismic data processing to test the strategies shows that they are significantly better than the existing GPU global synchronization methods.For the problem of conditional branch leads to GPU efficiency declining, in this paper, two conditional branch optimization strategies based on the thought of convergence are proposed at the software level and utilizing existing hardware and architecture. These two strategies perform optimization by compressing the number of efficient SIMD operation. The test results show that, these strategies achieve expected speedup under the premise to meet the conditions of use.Finally, in this paper, the above optimization strategies are tested by applying them to real seismic data processing. The results show that the GPU parallelization and CPU-GPU co-processing of the seismic migration achieve the significant acceleration effect, the GPU global synchronization strategies further reduce the total processing time, and the contribution made by GPU branch optimization strategies is relatively small.The works of this paper significantly improve the efficiency of seismic data processing and shorten the processing period and have great economic significance in the actual production.