基于GPU的地震资料处理并行加速研究

发布时间：2018-03-22 11:03

本文选题：地震资料处理　切入点：正演模拟　出处：《西南石油大学》2015年硕士论文　论文类型：学位论文

【摘要】：在地震资料处理环节中,随着处理对象及算法的复杂程度越来越高,作为高精度反演、逆时偏移等地震资料处理的重要组成部分,地震波正演模拟对计算速度的要求越来越高,传统的以CPU为主的串行算法已经不能满足日益复杂的地震勘探对计算效率的需求,而目前实际工作中使用的计算机集群并行计算面临系统能耗和维护成本过高的问题,而GPU拥有自身的并行架构特性,且功耗低、成本小,进行GPU并行计算正可以解决这些问题。有限差分法求解地震波动方程时,数值频散问题对模拟结果影响重大,通过实验对比,分析得到处理数值频散的几种方法。继而而研究了PML(完全匹配层)吸收边界条件处理边界反射问题,通过实验对比,说明了边界条件对正演模拟的重要性,并推导出了相应的声波方程的离散差分形式,有限差分方法和吸收边界处理都很适合在GPU平台做并行计算。 NVIDIA公司开发的基于GPU的统一设备架构平台CUDA, CUDA首先是在硬件支持的基础上,对GPU进行多线程并行计算的软件平台,处理地震波场有限差分计算的频散问题时,因为差分阶数和网格规模的增加造成计算效率很低,而且在剖分网格上做地震波场的有限差分计算,网格上各个节点的数值计算存在很大的并行性,所以正演模拟有限差分非常适合在GPU平台做多线程并行计算。本文基于CUDA软件结构,实现了二维声波方程有限差分数值解法的具体实现方法和流程,并通过不同配置硬件环境下的运行效率对比分析,充分证明,基于CUDA软件平台可明显提高地震波正演模拟的计算效率,为地震资料处理环节提供了技术实现的可能。
[Abstract]:In seismic data processing link, as the processing object and the algorithm complexity is increasing. As an important part of high precision inversion, reverse time migration of seismic data processing, seismic forward simulation of the computing speed of the increasingly high demand, mainly CPU the traditional serial algorithm can not meet the increasingly seismic exploration the computational efficiency of the complex needs, and computer cluster currently used in the practical work of parallel computing system facing the problem of high energy consumption and maintenance cost, while GPU has a parallel architecture characteristics of itself, and the power consumption is low, the small, GPU parallel computing is to solve these problems.
Finite difference method for solving seismic wave equations, the numerical dispersion effect on the simulation results, through the experimental comparison, analysis of several treatment methods. Then the numerical dispersion of PML (perfectly matched layer) absorbing boundary condition boundary reflection problem, through the experimental comparison, illustrates the importance of boundary conditions on the forward the simulation, and deduces the discrete differential equations corresponding to the wave form, the finite difference method and the absorbing boundary treatment is very suitable for parallel computing on the GPU platform.
GPU unified device architecture based on CUDA developed by NVIDIA company, is the first CUDA based on hardware support, GPU multi-threaded parallel computing software platform, processing of seismic wave field finite difference calculation of the dispersion problem, because the difference order and the scale of the grid increases caused by the calculation efficiency is very low but, in the mesh on the seismic wave field finite difference calculation, calculation of the inherent parallelism of each node of the large grid numerical modeling, so the finite difference is very suitable for parallel computing on the GPU platform to do more threads.
In this paper, based on the CUDA architecture, implementation methods and processes to achieve the two-dimensional acoustic wave equation finite difference numerical solution, and proved by comparing the efficiency of different hardware configuration environment analysis, the computing efficiency of CUDA software platform can significantly improve the seismic simulation based on technology offers the possibility for seismic data processing link.

【学位授予单位】：西南石油大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P631.44

【参考文献】