Toward Full GPU Implementation of Fluid-Structure Interaction

• Published in: 2019 18th International Symposium on Parallel and Distributed Computing (ISPDC) pp. 16 - 22
• Main Authors: Beny, Joel, Kotsalos, Christos, Latt, Jonas
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2019
• Subjects: CUDA; Force; GPU; Graphics processing units; immersed boundary; Instruction sets; Kernel; lattice Boltzmann; Lattices; nodal projective finite elements; red blood cells; Sociology; Statistics
• DOI: 10.1109/ISPDC.2019.000-2

Fluid-structure interaction (FSI) is a notoriously difficult topic in the fields of Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEA), as it requires the deployment of a coupling framework between two different numerical methodologies. Theoretically, the coupling may suffer from numerical instabilities or other types of incompatibilities. Practically, the strategies used to optimize and parallelize the respective solvers need to be put into a common framework for the run-time coupling to exhibit acceptable performance. In this article, we present a strategy for deploying a FSI system in the many-thread framework of general purpose Graphics Processing Units (gpGPUs). The system uses the lattice Boltzmann method (LBM) to simulate fluid flow, a novel finite element (FE) solver for the non-linear structural analysis of deformable bodies, and the immersed boundary method (IBM) to impose the noslip boundary condition on the fluid-structure interface. In the literature, GPU implementations of LBM-IBM codes are restricted to situations where the immersed surfaces are very small compared to the total number of fluid cells [10], as it is often the case for exterior flow simulations around obstacles. This article investigates two situations in which the density of immersed objects is inherently high. In the first case, we mimic the physics of a rigid (non-deformable) multi-rotor propeller with imposed motion and a solid-vertex to fluid-node ratio of less than 1%. In the second case, we simulate deformable red blood cells (RBCs), whose motion is resolved by the FEM part, flowing in the blood plasma, with a vertex-to-node ratio of 1%-10%.