Summary Massively parallel single-instruction, multiple-data (SIMD) computers have shown much promise in solving numerically intensive problems ranging from molecular modeling to computational fluid dynamics. Much of the research on the use of parallel computing for reservoir simulation, however, has been parallel computing for reservoir simulation, however, has been limited to coarse-grain, shared memory computers or medium-grain hypercubes. We report here on research performed on a massively parallel SIMD computer with 65,536 processors. This work addresses issues in machine architecture, programming environment, and formulation of a reservoir simulator programming environment, and formulation of a reservoir simulator on SIMD computers. Towards this end, a three phase, three dimensional, IMPES compositional simulator was developed in Fortran 8X. Problems associated with reservoir simulation on SIMD computers such as an appropriate data structure for the SIMD architecture, the treatment of multiple fluid phases, data communication, and the matrix solution are addressed in this paper. In addition, table lookup poses a unique problem for SIMD computers because storage of large tables on each processor is impractical, as is the use of the front-end computer to perform this function. The problem is overcome through a novel use of local memory access. We found that massively parallel computers can be used to run very large reservoir models and that the cost of data transfer between processors need not be prohibitive. Reservoir models with two hydrocarbon components and up to 2,097,152 grid blocks were successfully run using this simulator. Computational speeds on the order of one giga FLOPS were achieved for the generation of the Jacobian mix and the matrix solution using 65,536 processors. Introduction The current trend for improving supercomputer performance is to increase the intrinsic speed of the individual processors and/or to add more processors. As clock speed approaches the maximum physical limit for semiconductors, the addition of more processors is physical limit for semiconductors, the addition of more processors is the more promising direction. Parallel computers, in particular those with distributed memory, have generated much interest for numerically intensive applications. These parallel computers hold the promise of being more cost effective than supercomputers with shared memory, and have the potential to solve much larger problems.
A general purpose reservoir simulator was developed as a vehicle to investigate vectorization and parallel processing on the Cray XMP/48. The simulator can model black oil, compositional and steam injection processes. It includes both fully implicit and IMPES formulations. Microtasking is used for parallel processing in order to maintain the portability of the code. The simulator calculations other than linear equation solution were structured to facilitate vectorization and parallel processing. Several issues concerning parallel processing are discussed. These include task granularity, load balance, synchronization, memory contention, and balance between vectorization and arallel processing. Based on three test problems, the speedup of the calculations due to vectorization ranges from 5.3 to 10. The speedup can be augmented by as much as 3.3 times through parallel processing. The scheme proposed here is general enough to be applicable to other vector and parallel processing computers. Introduction Two factors which are important in simulator utilization are speed and maintainability. Vectorization and parallel processing can significantly improve speed. A general purpose reservoir simulator can improve maintainability by elimimating the need to duplicate routines which are common to many simulators. Although much attention has been paid to applying vectorization in reservoir simulation, little work was done in parallel processing. This work is intended to test the feasibility of using both vectorization and parallel processing in a general purpose reservoir simulator. Although the simulator was developed for a Cray multiprocessor machine, the methodology used in the development is applicable to other multiple and vector processors. This paper is divided into four parts. The first part describes the mathematical formulation and numerical solution procedure for the general purpose simulator. The second part discusses the fundamentals of vectorization and parallel processing. The third part shows how vectorization and parallel processing is applied to the simulator. Finally, the fourth part describes the performance tests and discusses the results of these tests. MATHEMATICAL MODEL AND NUMERICAL SOLUTION PROCEDURE Mathematical Model The mathematical model is formulated to simulate black oil, compositional, and thermal recovery processes. It includes both fully implcit and IMPES formulation. A compositional approach is used to model black oil processes. Darcy's law and instantaneous phase equilibrium are assumed valid. In order to simulate different processes using a single code, mutual solubilities of the components in the mixture have to be distinguished. Mutual solubilities of water and hydrocarbon components are considered unimportant in black oil or compositional processes, but are important for steam injection processes. P. 329^
This paper describes the formulation of a thermal simulation model in a vectorized, general purpose reservoir simulator. The thermal simulation model, an option in this general purpose simulator, is equivalent to a three-phase, three-dimensional, thermal compositional simulator. It has a variety of options for modeling fluid properties and phase behavior. These include the EOS method (the three parameter Peng-Robinson equation of state) and the non-EOS methods (table look-ups and correlations). The model also considers the solubility of water in the oil phase. Furthermore, all calculations are extensively vectorized.Test problems including problem No. 3 of the Fourth SPE Comparative Solution Project are used to evaluate the effectiveness of vectorization. Results show that the vectorized calculations outperform the nonvectorized ones by about an order of magnitude in computational speed on the Cray X-MP /48. Overall, this thermal model runs about 5 times faster than a previously reported conventional, partially vectorized thermal simulator.Finally, the use of this thermal simulation option to study the sensitivity of simulation results to various fluid property and phase behavior calculation methods is illustrated with a hypothetical problem. Results show that simulation results can be used to determine the importance of different Kvalue and volume calculation methods, and the inclusion of water-in-oil solubility.References and illustrations at end of paper.123
A 2-dimensional cellular automaton model is developed to simulate Darcy's flow in complex, heterogeneous porous media. Abstract This paper reports on the development of a 2-dimensional cellular automaton (CA) model to simulate fluid flow in porous media. In CA, simple rules of particle interactions at a lattice are used to simulate complex flow phenomena. Since the numerical operations involved are largely bit manipulation, CA can be potentially more efficient in memory usage than conventional potentially more efficient in memory usage than conventional methods, such as finite difference or finite element methods. While other research on the use of CA to model porous media flow primarily focuses on Stokes flow at the pore level, with the intent of understanding flow at the core scale, this work concentrates on modeling Darcy flow in large heterogeneous fields. Permeability variations and anisotropy are modeled by a Permeability variations and anisotropy are modeled by a distribution of ‘soft scatterers’ and 'directional scatterers'. Collision of incoming particles with scatterers is a stochastic process. Anisotropy is modeled via a preferred direction of process. Anisotropy is modeled via a preferred direction of collision between the directional scatterers and incoming particles The CA model is verified through the study of single-phase flow in porous media. Simulation results obtained from this CA model show that particle flow rates are linearly proportional to pressure gradient (pressure is modeled through particle density in pressure gradient (pressure is modeled through particle density in CA), and permeability is linearly correlated with scatterer density over a certain domain. Therefore, the CA model can be used to predict Darcy-like flow. It is shown that heterogeneity is modeled predict Darcy-like flow. It is shown that heterogeneity is modeled properly by varying the density of scatterers. In addition to the single fluid model, a model with two fluids with identical physical properties has also been developed to study dispersion characteristics of a CA model. In the numerical simulation results, the transverse diffusivity converges to the theoretical constant, but the longitudinal diffusivity increases with time. Finally, extension of the current model to a more general case (i.e., three-dimensional, multiphase flow) is also discussed. Introduction Flow in porous media is generally described by Darcy's law, which linearly relates fluid velocity to pressure gradient. Although the governing equations are linear in reservoir simulation, we have not fully resolved difficulties associated with multiscale phenomena, complicated permeability distribution in a reservoir, and nonlinear functional dependencies description of a reservoir includes large uncertainties due to the difficulty of field measurements at the reservoir conditions. Thus, they are often described by a statistical method rather than by a deterministic one.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.