IMEXLBM 1.0: A Proxy Application based on the Lattice Boltzmann Method for solving Computational Fluid Dynamic problems on GPUs

TL;DR

IMEXLBM 1.0 is an open-source proxy application based on the lattice Boltzmann method, designed for efficient CFD simulations on GPUs, aiding high-performance computing development and scientific discovery.

Contribution

The paper introduces IMEXLBM 1.0, a minimal, portable, GPU-capable proxy app that captures key features of large CFD applications for HPC benchmarking and development.

Findings

Successfully solves CFD benchmark problems on GPU systems

Demonstrates effective CPU-GPU hybrid computation with MPI and OpenMP

Verifies accuracy by comparing CPU-only and CPU+GPU results

Abstract

The US Department of Energy launched the Exascale Computing Project (ECP) in 2016 as part of a coordinated effort to achieve the next generation of high-performance computing (HPC) and to accelerate scientific discovery. The Exascale Proxy Applications Project began within the ECP to: (1) improve the quality of proxies created by the ECP (2) provide small, simplified codes which share important features of large applications and (3) capture programming methods and styles that drive requirements for compilers and other elements of the tool chain. This article describes one Proxy Application (or "proxy app") suite called IMEXLBM which is an open-source, self-contained code unit, with minimal dependencies, that is capable of running on heterogeneous platforms like those with graphic processing units (GPU) for accelerating the calculation. In particular, we demonstrate functionality by solving a benchmark problem in computational fluid dynamics (CFD) on the ThetaGPU machine at the Argonne Leadership Computing Facility (ALCF). Our method makes use of a domain-decomposition technique in conjunction with the message-passing interface (MPI) standard for distributed memory systems. The OpenMP application programming interface (API) is employed for shared-memory multi-processing and offloading critical kernels to the device (i.e. GPU). We also verify our effort by comparing data generated via CPU-only calculations with data generated with CPU+GPU calculations. While we demonstrate efficacy for single-phase fluid problems, the code-unit is designed to be versatile and enable new physical models that can capture complex phenomena such as two-phase flow with interface capture.