Large scale geostatistics with locally varying anisotropy
Classical geostatistical methods are based on the hypothesis of stationarity, which allows to apply repetitive sampling in different locations of the spatial domain, in order to obtain enough information to infer cumulative distributions. In case of non stationarity, anisotropy is observed in the un...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/674552 |
| Acceso en línea: | http://hdl.handle.net/10803/674552 https://dx.doi.org/10.5821/dissertation-2117-369087 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Informàtica 004 311 |
| Sumario: | Classical geostatistical methods are based on the hypothesis of stationarity, which allows to apply repetitive sampling in different locations of the spatial domain, in order to obtain enough information to infer cumulative distributions. In case of non stationarity, anisotropy is observed in the underlying physical phenomena. This feature manifest itself as preferential directions of continuity in the phenomena, i.e. properties are more continuous in one orientation than in another. In the case of local anisotropy, each location of the domain in study presents different preferential directions of continuity. The locally varying anisotropy (LVA) approach in geostatistics allows to incorporate a field of local anisotropy parameters defined for each domain point. With this additional input, more realistic spatial simulations can be generated, including geological features to the computational model such as folds, veins, faults, among others. Since the seminal article published by Boisvert and Deutsch (2011), to the best of the author's knowledge, no further analysis or public code improvements were developed. This is in part because acceleration and parallelization techniques must be applied to the inner kernels of the baseline LVA codes. Large execution time is needed to generate small-scale domain simulations, making large-scale domain simulations a prohibitive task. The contributions of this thesis are accelerating and parallelizing classical and LVA-based geostatistical simulation methods, particularly sequential simulation, which is one of the most common and computationally intensive methods in the field. This fact was recently remarked by some of the main authors in the field, Gómez-Hernández and Srivastava (2021), which shows the relevance of this work today. Two main parallel algorithms and an optimized version of a kd-tree search implementation are presented, all of them applied to both classical and LVA-based sequential simulation implementations. The first parallel algorithm is related to the parallel simulation of different domain points, after rearranging the order of simulation but preserving the exact results of a single-thread execution. The second parallel algorithm is related to the parallel search of neighbour points in the domain, which will be used to build data dependencies for the parallel simulation of points. The optimized kd-tree search was used in each test case in order to reduce the computational complexity of neighbour search tasks. Its modified implementation reduces the number of branching instructions and introduces specialized code sections to accelerate the execution. The main focus is on multi-core architectures using OpenMP and optimization techniques applied to Fortran and C++ codes. Additionally, acceleration and parallelization techniques were also applied to auxiliary applications, such as shortest path and variogram calculation on hybrid CPU/GPU architectures using Fortran, C++ and CUDA codes. In the last application, an analytical and heuristic model was developed to estimate the optimal workload distribution between CPU and GPU in the hybrid context. The overall results of this work are a set of applications that will allow researchers and practitioners to accelerate dramatically the execution of their experiments and simulations, being sgsim, sisim, sgs-lva and sisim-lva the accelerated codes presented. Final speedup results of 11x and 50x are obtained for non-LVA codes using 16 threads, and 56x and 1822x are obtained for LVA codes using 20 threads. These tools can be combined with other geostatistical tools, in order to improve the existing landscape of open source codes that can be used in practical scenarios. |
|---|