High performance computing and energy efficiency: focus on OpenFOAM

TL;DR

This paper investigates the energy efficiency of high performance computing, focusing on OpenFOAM, by analyzing energy scalability across different architectures and demonstrating that longer computation times can reduce energy consumption.

Contribution

It introduces the concept of energy scalability in high performance computing and compares energy efficiency across architectures using OpenFOAM simulations.

Findings

Most energy-efficient machine varies with calculation time and core count.

Longer calculation times tend to decrease energy consumption.

Energy efficiency depends on architecture and user time preferences.

Abstract

High performance calculation is increasingly used within society. Previously reserved for an elite, based on large computing and storage infrastructures, it is now a core module for many companies. Indeed, high performance calculation makes it possible to design and optimize many elements for a limited cost ,compared to the production of prototypes or tests in situ. It is also widely used in big data and artificial intelligence.It seems essential to ask about theenvironmental impact of these digital practices. A number of actions have already been initiated in this community: GREEN500; European CoC eco-responsibility label for Data centres... but these actions generally look at specific or even idealised situations and/or software.The software qualification process in the field of high performance calculation consists in looking at the scalability of the software. The originality of this study is to focus on energy scalability (calculation return time depending on the power consumed), by considering several architectures (three TOP500 machines and a laboratory cluster).The energy cost of an example calculation could be estimated, which shows that the most efficient machine in terms of calculation time is not necessarily the most energy efficient, and depending on the number of cores/processes chosen, it is not always the same architecture that is the most energy-efficient. It was therefore possible to show that the longer the user is prepared to wait, the less energy is used by the calculation.