Distributed Parallel Hybrid CPU-GPGPU Implementation of the Phase-Field Method for Accelerated High-Accuracy Simulations of Pore-Scale Two-Phase Flow

Abstract

Direct numerical simulation of multi-phase pore-scale flow is a computationally demanding task with strong requirements on time-to-solution for the prediction of relative permeabilities. In this paper, we describe the hybrid-parallel implementation of a two-phase two-component incompressible flow simulator using MPI, OpenMP, and general-purpose graphics processing units (GPUs), and we analyze its computational performance. In particular, we evaluate the parallel performance of GPU-based iterative linear solvers for this application, and we compare them to CPU-based implementations of the same solver algorithms. Simulations on real-life Berea sandstone micro-CT images are used to assess the strong scalability and computational performance of the different solver implementations and their effect on time-to-solution. Additionally, we use a Poisson problem to further characterize achievable strong and weak scalability of the GPU-based solvers in reproducible experiments. Our experiments show that GPU-based iterative solvers can greatly reduce time-to-solution in complex pore-scale simulations. On the other hand, strong scalability is currently limited by the unbalanced computing capacities of the host and the GPUs. The experiments with the Poisson problem indicate that GPU-based iterative solvers are efficient when weak scalability is desired. Our findings show that proper utilization of GPUs can help to make our two-phase pore-scale flow simulation computationally feasible in existing workflows.

A pressure-correction and bound-preserving discretization of the phase-field method for variable density two-phase flows

“…Efficient preconditioning of the pressure correction systems is critical for the computational performance of our simulations. If a simple Jacobi preconditioner is used, the solution of pressure correction systems accounts for more than 50% of the total computational cost of large simulations [36]. Our -multigrid implementation is available as open source software [35].…”

On application design for manycore processing systems in the domain of neuroscience

“…More recently, the fundamental problem of rigid body simulation in physics engines for games and animation, has been approached via a solution utilizing both the assets of the CPU and the GPU of desktop computers [189]. In the field of pore-scale modeling, flow dynamics can be simulated and evaluated using distributed, hybrid computational systems of CPU and GPU [190]. Hybrid systems involving FPGAs, as well as multicore CPUs and GPUs, have been utilized in the domain of online security [191], cryptography [192] and deep learning [193].…”