基于GPU的多点地质统计逐点模拟并行算法的研究

发布时间：2018-03-24 23:30

本文选题：储层表征　切入点：随机模拟　出处：《中国科学技术大学》2013年博士论文

【摘要】：油气储层的地质结构及储层参数是决定着储层的储量及产能的关键因素。因此,建立能够表征储层结构及性质的三维数字化模型已经成为油田勘探开发中的一项重要工作。目前,进行三维建模主要依赖于研究区域化变量空间结构性和随机性的地质统计学理论。由于传统的基于变差模型的随机模拟方法在描述复杂地质对象空间几何特征方面的局限,能够表征空间多点之间相关关系的多点地质统计理论应运而生。它结合了基于对象和基于像素随机模拟方法优点,近十年来一直是储层三维建模领域研究的热点。随着研究的深入,对于地质三维模型的规模和精细程度的要求日益提高。在进行大规模精细地质模拟时,逐点模拟的串行多点地质统计模拟方法中存在的不足日益突出。首先,串行多点地质统计方法的模拟效果受输入参数的影响。高质量的模拟效果受到严格参数的限制,从而需要大量计算,导致运行效率严重下降。其次,串行多点地质统计方法在对地质模式进行建模时,需要大量的内存空间。有限的物理内存会使大规模的模拟失败。逐点模拟的串行多点地质统计模拟方法中的局限性使算法的实用性受到了严重影响。本文从大规模三维精细地质模拟的需要出发,重点研究了基于通用图形处理器(GPGPU)和计算统一设备架构上(CUDA)的多点地质统计领域的并行计算方法,包括应用于离散变量和连续变量的并行随机模拟方法及地质统计模型的并行计算方法。本文的工作和创新主要体现在： 1.提出了应用于离散地质变量的并行SNESIM方法 ⅰ)根据GPU计算架构的特点,提出SNESIM并行方法的问题分解策略,实现了线程级别的问题划分和计算任务的高度并行,显著提高了计算效率。并且该并行策略不需要大量的内存空间,增强了算法的适应性。 ⅱ)为了减少GPU和CPU之间通信交互而造成的计算瓶颈,提出了基于GPU内部数据缓冲的合并方案,并给出了两种子问题并行合并策略,考虑了兼容性和内存访问冲突问题。实验结果表明了并行问题合并策略的有效性和计算效率。 ⅲ)考虑算法的特性和GPU的内存模型,进一步改进了并行方法的GPU实现。不仅避免了对于模板偏移的大量复杂和重复的计算量,并且对高速共享内存和只读纹理内存的使用进行了优化。 2.提出了应用于连续地质变量的并行Direct Sampling方法ⅰ)针对可变大小邻域节点的搜索问题,改进了对并行邻域条件节点的搜索策略。通过稳定的并行排序算法和辅助变量,显著提高了邻域条件节点的搜索效率。 ⅱ)针对对训练图像上目标点的选择问题,提出了一种并行选取策略。通过改进的并行规约二元函数算法,实现了训练图像上目标点选择的并行化和确定性。 ⅲ)针对模拟大范围地质体的连续性的不足,提出了将搜索椭球概率与搜索领域结合的改进方法,实验结果表明了模拟结果的连续性得到显著改善。 3.对于表征区域化变量相关关系的两点地质统计模型,提出了基于研究区域和统计量的两种并行策略。实验结果表明了基于统计量的并行策略在计算效率方面优于基于研究区域的策略。总而言之,多点地质统计逐点并行模拟方法与现有的串行方法相比,在并行计算与多点地质统计模拟方法的结合中提出了新的思路,改进了多点地质统计随机模拟方法在大规模三维建模中的适应性,且大幅度提高了模拟效率。
[Abstract]:The geological structure and reservoir parameters of oil and gas reservoir is a critical factor of reserves and productivity of the reservoir. Therefore, to establish a three-dimensional digital model of reservoir structure and properties has become an important task for the exploration and development of oilfield. At present, the three-dimensional modeling mainly depends on the research on the regional spatial structure variables and the randomness of the geostatistics theory. Because the traditional stochastic simulation method based on variation model in describing complex geological object spatial geometry limitations to multi-point geostatistical theory came into being the relationship between characterization of multi point. It combines the advantages of object based and pixel based on the method of stochastic simulation, in recent ten years is a hot research field of 3D modeling of the storage layer.
With the in-depth study, the scale and sophistication of 3D geological model of the increasing demand. In large-scale fine geological modeling, simulation of serial point-to-point multiple point geostatistics modeling method are increasingly prominent. First, simulation results of serial multiple point geostatistics method by simulation of input parameters. The effect of high quality is strictly limits for parameters which need a large amount of calculation, resulting in a serious decline in efficiency. Secondly, geological statistics method in multi-point serial modeling of geological model, requires a large amount of memory space. Physical memory is limited and makes simulation failure. By simulation of large-scale serial multiple point geostatistics limitations of the method in the simulation of the practicability of the algorithm have been seriously affected.
From the needs of large-scale 3D geological simulation of focusing on the GPU based (GPGPU) and Compute Unified Device Architecture (CUDA) on the parallel computing method of multiple point geostatistics in the field of parallel computing method including parallel stochastic simulation method and the geological statistics model is applied to the discrete and continuous variables in this paper. The main work and innovation in:
1. a parallel SNESIM method applied to discrete geological variables is proposed.
I) according to the characteristics of GPU computing architecture, proposed a parallel SNESIM method of problem decomposition strategy, realize the division and computing task thread level parallelism, significantly improve the computational efficiency. And the parallel strategy does not require a lot of memory space, enhance the adaptability of the algorithm.
II) in order to reduce the computational bottleneck of communication between GPU and CPU caused by the proposed GPU merging scheme based on internal data buffer, and gives the parallel merging strategy of two seed, considering the compatibility and memory access conflicts. The experimental results show the effectiveness and efficiency of parallel computing problems with strategy.
Considering the characteristics of algorithm and the memory model of GPU, the GPU implementation of parallel method is further improved. It not only avoids a lot of complex and repetitive computation for template offset, but also optimizes the use of high-speed shared memory and read-only texture memory.
2. proposed parallel Direct Sampling method I applied to continuous geological variables) to the problem of searching for variable size neighborhood nodes, improved the parallel neighborhood conditions node search strategy. Through the parallel sorting algorithm and the auxiliary variables are stable, improve neighborhood conditions node search efficiency.
II) aiming at the selection of target points on the training images, we propose a parallel selection strategy. Through the parallel protocol two yuan function to improve the algorithm, parallelization and deterministic target selection of training images.
In order to solve the problem of continuity in simulating large scale geological bodies, an improved method combining search ellipsoid probability with search field is put forward. Experimental results show that the continuity of simulation results has been significantly improved.
3., for the two geostatistical models representing the correlation between regionalized variables, two parallel strategies based on the research area and statistics are proposed. The experimental results show that the parallel strategy based on statistics is superior to the strategy based on the research area in terms of computation efficiency.
In a word, multiple point geostatistics by compared with the existing methods of serial parallel simulation method in combination with parallel computing and multi point geostatistics simulation method put forward new ideas, improve the geostatistical stochastic simulation method in large-scale 3D modeling of adaptability, and greatly improves the simulation efficiency.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP338.6

【参考文献】