Accelerating Reservoir Simulation with GPUs

Over the past several decades, reservoir simulation has progressed in capability from coarse two-dimensional models to inhomogeneous and detailed three-dimensional models, providing a high degree of fidelity to empirical production rates. Part of this increase can be attributed to advances in solution algorithms and part to advances in computational hardware. In this presentation, we discuss examples of both types. In an IMPES formulation, computational effort is divided into an implicit solve of the discretized pressure equation and explicit updates of the saturation equations. By porting and optimizing saturation updates to run on multiple GPUs, we accelerate their execution speed by a factor of about 400. By replacing the multithreaded direct sparse solver for the pressure equation by an iterative solver preconditioned with algebraic multigrid, we reduce the solution time by a factor of 20. Combining these two enhancements, we improve the overall execution speed of Marathon Oil Corporation Multiscale Simulator (MMS) by a factor of over 100 for the SPE10 Model 2 benchmark problem.