PARALLELIZATION STRATEGY FOR MULTIPLE POINT STATISTICS BASED SIMULATION ON GPU

Open Access
Author:
Malife, Chukwudiuto Udechukwu
Graduate Program:
Energy and Mineral Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
April 01, 2019
Committee Members:
  • Sanjay Srinivasan, Thesis Advisor
Keywords:
  • Reservoir Characterization
  • Reservoir Modeling
  • MPS
  • Geostatistics
  • Geo-statistics
  • Multiple Point Simulation
  • Multiple Point Statistics
  • Multi Point Simulation
  • Multi Point Statistics
  • Spatial Estimation
  • GPU
  • Parallelization
  • Graphics Processing Unit
Abstract:
Although Multiple Point Statistics (MPS) methods can generate realistic reservoir models, they are computationally expensive and time consuming when used for modeling large reservoir. This makes them a poor choice to the reservoir engineer who needs the models on time for analysis and quick decision making. This research aims to develop a parallelization strategy for executing MPS based simulations on a Graphics Processing Unit (GPU). The goal is to show that reservoir modeling with MPS methods can be made faster, hence improving their usefulness as practical choices for modeling large reservoirs. The Single Normal Equation Simulation (SNESIM) algorithm is the MPS method used in this work. The SNESIM algorithm which executes sequentially on a Central Processing Unit (CPU) is first implemented, and subsequently parallelized to be executed on a GPU. The SNESIM algorithm implemented in this work uses a Hash-table as the data structure for storing pattern statistics. To successfully parallelize the SNESIM algorithm, an Iterative Brute-force conflict management strategy is implemented to determine the position of nodes in a simulation grid that can be simulated simultaneously. A Layer Pattern Repetition method is proposed as potential alternative to the Iterative Brute-force strategy for conflict management. The CPU and GPU based SNESIM implementations are used to model the facies distribution of Lobe 10 in the Lobster field which is located in the Gulf of Mexico. The GPU based implementation is evaluated against the CPU based implementation in terms of simulation time and connectivity of features in the generated facies model. The evaluation demonstrates that the GPU based implementation generates a facies model with similar feature connectivity, however the GPU based implementation generates its facies model faster. In other words, parallelization of the SNESIM algorithm does not degrade the quality of realizations generated by the algorithm. In conclusion, the speed of an MPS based simulation can be improved by executing a parallelized implementation of the MPS algorithm on a GPU.