CPU-GPU Heterogeneous Code Acceleration of a Finite Volume Computational Fluid Dynamics Solver

TL;DR

This paper presents a CPU-GPU heterogeneous acceleration approach for a finite-volume CFD solver, demonstrating improved performance through combined CPU and GPU utilization and providing a performance model for estimation.

Contribution

It introduces a performance model for CPU-GPU heterogeneous CFD computing and compares scaling performance with pure CPU or GPU setups.

Findings

Performance improved by combining CPUs and GPUs

Performance model accurately estimates heterogeneous computing benefits

Heterogeneous approach outperforms pure CPU or GPU configurations

Abstract

This work deals with the CPU-GPU heterogeneous code acceleration of a finite-volume CFD solver utilizing multiple CPUs and GPUs at the same time. First, a high-level description of the CFD solver called SENSEI, the discretization of SENSEI, and the CPU-GPU heterogeneous computing workflow in SENSEI leveraging MPI and OpenACC are given. Then, a performance model for CPU-GPU heterogeneous computing requiring ghost cell exchange is proposed to help estimate the performance of the heterogeneous implementation. The scaling performance of the CPU-GPU heterogeneous computing and its comparison with the pure multi-CPU/GPU performance for a supersonic inlet test case is presented to display the advantages of leveraging the computational power of both the CPU and the GPU. Using CPUs and GPUs as workers together, the performance can be improved further compared to using pure CPUs or GPUs, and the advantages can be fairly estimated by the performance model proposed in this work. Finally, conclusions are drawn to provide 1) suggestions for application users who have an interest to leverage the computational power of the CPU and GPU to accelerate their own scientific computing simulations and 2) feedback for hardware architects who have an interest to design a better CPU-GPU heterogeneous system for heterogeneous computing.