GPU-Enabled Scalable Multiscale Solver for Reservoir Simulation

As reservoir simulation models continue to grow in their size and complexity, the computational cost of reservoir simulation is constantly increasing. Since most of the reservoir simulation time is typically spent in the linear solver (being in the innermost part and the most challenging to parallelize and scale), efficient linear solvers are of utmost importance for reducing reservoir simulation turnaround times. In this work, we study the scalability of a versatile multiscale linear solver, namely the restriction-smoothed basis multiscale method (MsRSB) (Møyner and Lie (2016)) on the emerging massively parallel GPU architecture, and compare it to its performance on the multi-core CPU architecture. MsRSB, unlike traditional multiscale approaches, uses iterative smoothing to adaptively compute multiscale basis functions, allowing it to handle a wide range of difficult grid orientations seen in real-world industrial applications. While MsRSB can be parallelized directly, its reliance on a smoother to determine the basis of functions results in unusual control and data flow patterns. To achieve effective scalability, these patterns must be carefully designed and implemented on massively parallel systems. We extend Manea et al. (2016) and Manea and Almani (2019) work on parallel multiscale methods to move the MsRSB special kernels to shared-memory parallel multi-core and GPU architectures. Highly heterogeneous multimillion-cell 3D problems, adopted from the SPE10 Benchmark (Christie and Blunt (2001) have been used to illustrate the scalability of our parallel MsRSB development. The GPU implementation is benchmarked on a massively parallel architecture consisting of Nvidia Volta V100 GPUs, while the multi-core implementation is benchmarked on a shared memory multi-core architecture consisting of two packages of Intel's Haswell-EP Xeon(R) CPU E5-2667. For both the setup and solution stages, we compare the multi-core implementation versus the GPU implementation. The GPU-based MsRSB implementation shows great scalability, with over a 4-fold reduction in runtime when compared to the optimized multi-core implementation.