Implementation and Performance Analysis of a Parallel Oil Reservoir Simulator Tool Using a CG Method on a GPU-Based System

Abstract

An oil reservoir simulator is a crucial tool used by petroleum engineering to analyze reservoir conditions. To increase its performance, we implement a parallel version of the tool on a Graphic Processing Unit (GPU), using Computer Unified Device Architecture (CUDA) programming model and the Single Instruction Multiple Threads (SIMT). This paper presents our parallel implementation and performance analysis for 1-D, 2-D, and 3-D oil-phase reservoirs. The implementation and the performance evaluation reveal the gains and the losses achieved by the parallelization of a reservoir simulator on a Graphics Processing Unite (GPU) system. The performance results show that despite the interdependency between the different computational parts of the Conjugate Gradient (CG) method used as a linear solver in the parallel reservoirs, a speedup of 26 can be easily obtained for an oil reservoir simulator using 15 streaming multiprocessors (SMs), compared to a sequential CPU execution. The parallel execution scales well with grid dimensionality.