Latest revision as of 16:06, 25 November 2022

Summary

Preparing legacy codes for the upcoming exascale systems is a timely topic since the unveiling of the Frontier system in June 2022. In this work we describe the steps taken to prepare the AVBP code for this new step in computing ressources. AVBP [6] is a parallel CFD code that solves the three-dimensional compressible Navier-Stokes equations on unstructured and hybrid grids. AVBP is a cutting-edge software when it comes to distributed memory CPUs, scaling efficiently up to 200.000's of cores on Bluegene or AMD Epyc2 systems. However, other types of architectures such as ARM processors and accelerators are gaining popularity and play a significant role in the exascale era. We first explore the usage of ARM processors, then GPU accelerators through OpenACC[2] directives. This work highlights the difficulties of porting a legacy code to those architectures and solutions implememented so far for performance.

Abstract

Full Paper