Parallel Black Oil Solvers on GPU

Song Yu

Reservoir simulation for a full field heterogeneous model with millions of grid blocks demands significant computational time; therefore, improving the computational efficiency becomes crucial in designing a reservoir simulator. Graphics Processing Unit (GPU), a new high-profile parallel processor with hundreds of microprocessors, stands out in parallel simulation because of its efficient power utilization and high computational efficiency. Also, its cost is relatively low, making large-scale parallel reservoir simulation possible for most of the desktop users. In this thesis, a GPU-based parallel linear solver package, named Parallel Accelerated Simulation Solvers (PASS), is developed to speed up the reservoir simulation using the GPU. For this new solver package, several different linear solvers and preconditioners have been implemented based on GPU. For solvers, it has GMRES, BICGSTAB and ORTHOMIN, which are commonly used in reservoir simulators. For preconditioners, a group of ILU preconditioners are developed on GPU, including ILU(0), ILUT, block ILU(0) and block ILUT.

In the numerical experiments performed, the SPE10 problem, a 3D heterogeneous model with over one million grid blocks, is selected to test the speedup of the GPU solver. On the state-of-the-art CPU and GPU platforms, the new GPU implementation is able to achieve a speedup of over 8 times in solving linear systems arising from this SPE10 problem compared with the single CPU based sequential solver. Moreover, the GPU solver is successfully coupled with an in-house black-oil reservoir simulator to test the performance of the whole parallel simulation process, with a speedup of around 6 times. The excellent speedup and accurate results demonstrate that the GPU-based parallel linear solver has a great potential in the parallel reservoir simulation.