Research On The Gpu-based Parallel Algorithms For Point-based Multiple-point Geostatistical Simulation

The geological structure of the oil and gas reservoir and the reservoir parameters are the key factors which determine the reserves and production capacity of the reservoir. Therefore, the establishment of a three-dimensional digital model which is able to characterize the structure and nature of the reservoir has become an important task in the oil exploration and development. Currently, three-dimensional modeling mainly depends on the geostatistical theory which focuses on the researches of regionalized variable spatial structure and randomness. Due to the limitations for the variogram-based stochastic methods to describe the geometric characteristics of complex geological objects, multiple point geostatistical theory which could characterize the spacial corelations between multiple points came into being. It combines the advantages of object-based and pixel-based stochastic simulation method, and it has been the research focus of the area of3D reservoir modeling over the past decade.With further research, the requirement for more large and more sophistication three-dimensional models increases. In the applications of large-scale and sophisticate geologic modeling, deficiencies become increasingly prominent for the current serial mutiple-point statistical simulation methods. Firstly, the quality of the multiple point geostatistical simulation results is influenced by the input parameters. High-quality simulation realizations are produced by the severe parameters, which require a lot of calculation, leading to a serious decline in computational efficiency. Secondly, when performing geostatistical tasks on geological model, the serial multiple point geostatistical method need a lot of memory space. The limitations in the traditional multiple point simualtion methods have severely affected the practicality of the algorithms.This article focuses on the needs of the large-scale three-dimensional geologic simulation, and the reseaches of the prallel multiple point stochastic simulation algorithms which is based on general-purpose graphics processor(GPGPU) and Compute Unified Device Architecture (CUDA), including the parallel stochastic simulation methods for discrete and continuous variables and the parallel computational methods for the geostatistical model. The work and innovation is mainly reflected in:1. We proposed a parallel SNESIM method for discrete geological variables simulation. i) According to the characteristics of GPU computing architecture, we proposed a parallel decomposition strategy for the SNESIM method, and implemetated thread-level problem partition and highly parallel computing tasks. The parallel strategy significantly improves the computational efficiency, and it does not require a lot of memory space, so it enhanced the adaptability of the algorithmii) In order to avoid the computational bottleneck caused by the communications between the GPU and CPU, we proposed a problem merge strategy based on internal data buffer and two parallel implementations considering the compatibility and memory access conflict. The experimental results shows the effectiveness and efficiency of the parallel consolidation strategyiii) Considering the characteristics of the algorithm and the GPU memory model, we raised further optimization strategies for the parallel implementation of the method. The strategies not only avoid the large number of complex and repetitive computation for the template offset, but also optimize the usage of high-speed shared memory and read-only texture memory.2. We proposed a parallel Direct Sampling method for continous geological variables simulation. i) A parallel search strategy is proposed for the searching tasks in the variable-sized neighborhood. Stable parallel sorting algorithm and auxiliary variables, significantly improves the efficiency of the search of the neighborhood conditions node. ii) A parallel selecting strategy is proposed for the selection tasks of the target node in the training images. Through the improved dual function in the parallel reduction algorithm, the parallelization and certainty of the target point selected on the training images could be achieved,iii) For lack of large scale continuity in the simulations of geological bodies, search ellipsoid is combined with neighborhood seach area. Experimental results show that the continuity in the simulation results can be significantly improved3. For the two-point geostatistical model which characterizes the relationship of regionalized variables, parallel strategies of area-based and statistics-based methods are proposed. The experimental results show that the statistics-based parallel strategy can be much more efficient than the area-based strategy.In conclusion, comparing the parallel multiple-point geostatistics methods proposed in this thesis with the existing serial methods, this thesis put forward new ideas in the combination of the parallel computing and multi-point geostatistical simulation methods, improved the adaptability of multiple-point geostatistics stochastic simulation in the applications of the large-scale three-dimensional modeling, and greatly improve the efficiency of the simulation.