Adaptive Kinetic-Fluid Solvers for Heterogeneous Computing Architectures

TL;DR

This paper demonstrates the feasibility and advantages of porting an adaptive multi-scale kinetic-fluid simulation code to CPU-GPU systems, achieving significant speedups and efficient load balancing for complex flow simulations.

Contribution

The paper introduces a unified flow solver that combines adaptive mesh refinement with cell-by-cell solver selection, optimized for heterogeneous CPU-GPU architectures.

Findings

Achieved 24x24x24 velocity mesh simulations with over a million Boltzmann cells.

Demonstrated double-digit speedups on single GPU and good multi-GPU scaling.

Successfully implemented CUDA kernels for kinetic, DSMC, and mesoscopic solvers on octree mesh.

Abstract

We show feasibility and benefits of porting an adaptive multi-scale kinetic-fluid code to CPU-GPU systems. Challenges are due to the irregular data access for adaptive Cartesian mesh, vast difference of computational cost between kinetic and fluid cells, and desire to evenly load all CPUs and GPUs. Our Unified Flow Solver (UFS) combines Adaptive Mesh Refinement (AMR) with automatic cell-by-cell selection of kinetic or fluid solvers based on continuum breakdown criteria. Using GPUs enables hybrid simulations of mixed rarefied-continuum flows with a million of Boltzmann cells with 24x24x24 velocity mesh. We describe the implementation of CUDA kernels for three modules in UFS: the direct Boltzmann solver using discrete velocity method, the Direct Simulation Monte Carlo (DSMC) solver, and a mesoscopic solver based on Lattice Boltzmann Method, all using octree Cartesian mesh. Double digit speedups on single GPU and good scaling for multi-GPUs have been demonstrated.