Full-GPU Simulation of Coupled Multi-Reservoir Models: Implementation and Field Examples

Abstract

Abstract Graphics Processing Units (GPUs) have been successful in providing up to an order of magnitude performance benefit for reservoir simulation; this has been demonstrated for both standalone black-oil and compositional models. To our knowledge however, a demonstration of the applicability of GPUs to the simulation of models coupled at the asset level is still missing. In this paper, we present how a full-GPU reservoir simulator has been turned into a field simulator whereby multiple reservoirs can be coupled by a field scheduler, using the Message Passing Interface (MPI). The default scenario is to use one (or possibly multiple) GPUs per reservoir. As an alternative, it is possible to leverage NVIDIA's Multi Process Service (MPS) to run the full asset on a single GPU when the reservoir models aggregate can fit in memory, resulting in optimal usage of available computing resources. The field scheduler may integrate reservoir models using three different logic in the definition of the coupling frequency: (a) periodic, (b) at each time-step, or (c) at each Newton iteration. Smoothness of production and injection profiles typically improves from (a) to (c), while the simulation cost can either increase or decrease depending on the model specifics. Using a set of case studies based on real industrial problems, we demonstrate the robustness and performance of the implementation. We also discuss the impact on resources utilization in a typical computational environment used by reservoir engineers, in particular the possibility to run probabilistic forecast workflows on a limited number of cluster nodes. This paper presents for the first time the effectiveness and efficiency of a GPU based reservoir simulator for integrated asset models consisting of many reservoirs connected to a common surface gathering system. The implementation offers the capability to fully exploit the memory bandwidth and the fine grain parallelism offered by modern GPUs to perform complex simulations, providing further evidence that leading-edge high-performance systems, typically designed by energy companies for seismic processing, are also suitable for reservoir simulation.