High Performance Manycore Solvers for Reservoir Simulation

SUMMARY The forthcoming generation of many-core architectures compels a paradigm shift in algorithmic design to effectively unlock its full potential for maximum performance. In this paper, we discuss a novel approach for solving large sparse linear systems arising in realistic black oil and compositional flow simulations. A flexible variant of GMRES (FGMRES) is implemented using the CUDA programming model on the GPU platform using the Single Instruction Multiple Threads (SIMT) paradigm by taking advantage of thousands of threads simultaneously executing instructions. The implementation on the GPU is optimized to reduce memory overhead per floating point operations, given the sparsity of the linear system. FGMRES relies on a suite of different preconditioners such as BILU, BILUT and multicoloring SSOR. Additionally, the solver strategy relies on reordering/partitioning strategies algorithms to exploit further performance. Computational experiments on a wide range of realistic reservoir cases show a competitive edge when compared to conventional CPU implementations. The encouraging results demonstrate the potential that many-core solvers have to offer in improving the performance of near future reservoir simulations. Abstract With the advent of the processor technology revolution, many-core computing with Graphics Processing Units (GPUs) provides enormous processing capability suitable for large scale engineering and scientific applications. The forthcoming generation of many-core architectures compels a paradigm shift in algorithmic design to effectively unlock its full potential for maximum performance. In this paper, we discuss a novel approach for solving large sparse linear systems arising in realistic black oil and compositional flow simulations. A flexible variant of GMRES (FGMRES) is implemented using the CUDA programming model on the GPU platform using the Single Instruction Multiple Threads (SIMT) paradigm by taking advantage of thousands of threads simultaneously executing instructions. The implementation on the GPU is optimized to reduce memory overhead per floating point operations, given the sparsity of the linear system. FGMRES relies on a suite of different preconditioners such as BILU, BILUT and multicoloring SSOR. Additionally, the solver strategy relies on reordering/partitioning strategies algorithms to exploit further performance. Computational experiments on a wide range of realistic reservoir cases show a competitive edge when compared to conventional CPU implementations. The encouraging results demonstrate the potential that many-core solvers have to offer in improving the performance of near future reservoir simulations.