Research On Migration Efficiency Of Heterogeneous Seismic Data Processing Cluster

Wave equation pre-stack depth migration (WE PSDM) is able to achieve high quality ofmigrated images in complex geological area, and it is one of the important means to find oiland gas. But data amount of WE PSDM is great, and it requires vast computational power,thus its practical application is limited. CPU-GPU heterogeneous clusters in such aspects asperformance, power consumption, cost, cooling has huge advantages, and bring anopportunity for the popularization of WE PSDM. However, compared with the traditionalconsistent, serial, compact CPU model, hybrid CPU-GPU architecture is greatly different insystem organization, processor architecture and programming model. Efficient use ofheterogeneous computing resources is faced with many problems and challenges.This paper first carried on the qualitative analysis and quantitative benchmarking to theeffects of non-uniform access and contention. The result shows that the mismatched data pathand the bus contention and saturation will significantly reduce the performance of datacommunication. It may be the bottleneck for the I/O access demanding migration imaging.Then several strategies of avoiding such bottlenecks are discussed, combined with numericalmethods which are commonly used in migration computation, experiments have been carriedout, and the optimized application in terms of performance and stability is improved.To fully utilize the GPU computing power, this paper analyzes the model of CUDA, andargued that the perspective of considering GPU as a multithreading SIMD processor will helpgrasp the nature of the GPU and the develop more efficient applications. Then throughmicrobenchmarking, some microarchitecture characteristics of Fermi arhcitecutre aredemystified, providing support for in-depth performance optimization. Considering that theFast Fourier Transform is widely used in seismic migration, based on the Fermi architecture,an already highly optimized GPU Fast Fourier Transform program is analyzed. Through dataprefetching and instruction tuning, with increasement to the instruction level parallelism, theperformance is improved by12%with lower occupancy.SIMD branch divergences could lead to a significant performance penalty. In this paper, two kind of code-level optimization strategy are proposed,"convergence" and "extraction".Test results show that the for appropriate branches,"convergence" can increase the share ofeffective results in every SIMD execution step,"extraction" can reduce the length of divergedSIMD branches, and the performance then will be improved.Finally, the real-world migration test results indicate that, reasonable data path planningis most obvious to migration acceleration, deep optimization of the hot spots of GPUcomputing kernel can also bring certain improvements, and contribution of SIMD branchoptimization to the migration speed is relatively small.