Full text loading...
-
Realizing the Potential of GPUs for Reservoir Simulation
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, ECMOR XIV - 14th European Conference on the Mathematics of Oil Recovery, Sep 2014, Volume 2014, p.1 - 16
Abstract
Higher stakes from deep-ocean drilling, increasing complexity from unconventional reservoirs, and an overarching desire for a higher-fidelity subsurface description have led to a demand for reservoir simulators capable of modelling many millions of cells in minutes. Recent advances in heterogeneous computing hardware offer the promise of faster simulation times, particularly through the use of GPUs. Thus far, efforts to take advantage of hardware accelerators have been primarily focused on linear solvers and, in particular, simple preconditioners which often sacrifice rapid convergence for the sake of easy parallelism. This relatively weak convergence, the remaining unaccelerated code paths, and communication bottlenecks have prevented dramatic reductions in run time. A comprehensive approach, however, built from the ground up for accelerators, can deliver on the hardware’s promise to meet industry demand for fast, scalable reservoir simulation.
We present the results of our efforts to fully accelerate reservoir simulations on multiple GPUs in an extended black-oil formulation discretized using a fully-implicit finite volume method. We implement all major computational aspects, including property evaluation, Jacobian construction, and robust solvers/ preconditioners on GPUs. The CPR-AMG preconditioner we employ allows low iteration count and near-optimal order(N) scaling of computational effort with system size. This combination of algorithms and hardware enables the simulation of fine-scale models with many millions of cells in minutes on a single workstation without any upscaling of the original problem. We discuss the algorithms and methods we employ, give performance and accuracy results on a range of benchmark problems and real assets, and discuss the strong and weak scaling behavior of performance with model size and GPU count. This work was supported by the Marathon Oil Corporation.