三维有限差分瞬变电磁正演的OpenAcc并行研究

发布时间：2018-07-31 09:04

【摘要】：近年来,随着瞬变电磁技术的不断发展,瞬变电磁仪器的不断进步,我们可以采集到越来越精确可靠的瞬变电磁野外数据。然而我们队瞬变电磁法的研究仍不深入,没能很好地发挥作用。目前瞬变电磁反演,主要以一维为主,二维三维反演仍不成熟。为了更加深入地研究瞬变电磁法,我们对时间域有限差分三维瞬变电磁正演进行深入研究,但是常规地CPU上的计算耗时太长,模型太大的话计算时间长到无法忍受。随着近些年GPU并行计算的深入发展,形成了包括CUDA、OpenGL、OpenAcc 等一大批 GPU 并行技术。但是由于 CUDA、OpenGL等传统的并行技术要与硬件结构高度结合,因而程序编写难度较大,并且对于已经写好的程序移植性较差,所以技术普及应用受到限制。而OpenAcc技术解决了传统并行技术在这些方面难题,并取得了较大的进步。本文利用新型的GPU并行编程技术OpenAcc对时域有限差分三维瞬变电磁正演算法进行改进,极大缩短了三维正演的计算时间,增大了计算效率,并针对其进行了如下研究:首先,应用Yee (1966)提出的交错网格方案和Wang Hohmann (1993)提出的DuFort-Frankel方案对均匀网格剖分下的麦克斯韦方程组的FDTD离散形式进行推倒,并采用软源直接注入的方式施加激励源,推导得到有源媒质中的MaxWell有限差分方程组。对GPU并行技术的发展与优势进行简要介绍,重点介绍了 OpenAcc并行技术的执行模式与优势,应用实际实例证明OpenAcc并行技术相较于传统CUDA技术的优势以及选取其作为本文研究工具的原因;其次,对FDTD三维正演算法进行分析,找出程序最耗时的部分,应用OpenAcc并行技术对其进行并行化处理。并应用经典的均匀半空间模型、层状模型以及三维异常体模型对并行程序的准确性进行验证;最后,对并行程序的加速性能进行评价,通过试验,并行程序的最大加速比可以达到20x左右。通过大量的模型试验得到在本文所用的计算环境下并行程序的计算上限,分析网格剖分方式对并行程序加速性能的影响,寻找最适合的模型剖分方法。最后在更加实际的有覆盖的地质模型、复杂的地质模型进行计算,验证了并行程序在绝大多数地质模型下准确性极高,加速性性能也达到满意效果。
[Abstract]:In recent years, with the continuous development of transient electromagnetic technology and the continuous progress of transient electromagnetic instruments, we can collect more and more accurate and reliable transient electromagnetic field data. However, the study of transient electromagnetic method in our team is still not thorough and has not worked well. At present, transient electromagnetic inversion is mainly one-dimensional, two-dimensional inversion is still immature. In order to study the transient electromagnetic method more deeply, we study the 3-D transient electromagnetic forward modeling in finite difference time domain, but the computation time on the conventional CPU is too long, and if the model is too large, the calculation time is too long to bear. With the development of GPU parallel computing in recent years, a large number of GPU parallel technologies have been formed, such as CUDAN OpenGL OpenAcc and so on. However, because the traditional parallel technology such as CUDA-OpenGL is highly integrated with the hardware structure, it is difficult to program writing, and the portability of the written program is poor, so the popularization and application of the technology is limited. OpenAcc technology solves these problems in traditional parallel technology and has made great progress. In this paper, a new GPU parallel programming technique, OpenAcc, is used to improve the 3-D transient electromagnetic forward algorithm of finite-difference time-domain, which greatly shortens the computing time and increases the computational efficiency. Using the staggered grid scheme proposed by Yee (1966) and the DuFort-Frankel scheme proposed by Wang Hohmann (1993), the FDTD discrete form of Maxwell equations under uniform mesh division is pushed down, and the excitation source is applied by direct injection of soft source. The MaxWell finite difference equations in active media are derived. The development and advantages of GPU parallel technology are briefly introduced, and the execution mode and advantages of OpenAcc parallel technology are emphatically introduced. A practical example is used to prove the advantage of OpenAcc parallel technology compared with traditional CUDA technology and the reason why it is chosen as the research tool in this paper. Secondly, the 3D forward algorithm of FDTD is analyzed to find out the most time-consuming part of the program. OpenAcc parallel technology is applied to parallelize it. The accuracy of the parallel program is verified by the classical uniform half-space model, layered model and 3D abnormal volume model. Finally, the acceleration performance of the parallel program is evaluated and tested. The maximum speedup of parallel programs can reach about 20x. Through a large number of model tests, the upper limit of parallel programs is obtained in the computing environment used in this paper. The effect of grid generation on the acceleration performance of parallel programs is analyzed, and the most suitable model generation method is found. Finally, in the more practical overlay geological model and the complex geological model, it is verified that the parallel program is highly accurate under most geological models, and the speedup performance is satisfactory.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P631.325

【参考文献】