Large-scale Distributed Systems and Energy Efficiency: A Holistic View

Chapter 7Energy Efficiency and High-Performance Computing

Pascal Bouvry,¹ Ghislain Landry Tsafack Chetsa,^2,3 Georges Da Costa,³ Emmanuel Jeannot,⁴ Laurent Lefèvre,² Jean-Marc Pierson,³ Frédéric Pinel,¹ Patricia Stolf,³ and Sébastien Varrette¹

¹Computer Science and Communication (CSC) Research Unit, University of Luxembourg, Luxembourg

²INRIA, LIP Laboratory (CNRS, INRIA, ENS Lyon, UCB Lyon1), University of Lyon, Lyon, France

³SEPIA Research Unit, Institute for Research in Informatics of Toulouse (IRIT), University of Toulouse III, Toulouse, France

⁴Runtime Research Unit, INRIA Bordeaux Sud-Ouest, Bordeaux, France

7.1 Introduction

Nowadays, high-performance computing (HPC) infrastructure can consume several megawatts. As of today, the world's fastest supercomputer, according to the Top’5000 list [1], is the Tianhe-2 and consumes more than 17 and 24 with cooling for almost 34 PFlops.

Without any progress in terms of energy efficiency, an Exascale infrastructure (1000 PFlops) would consume around 500 MW or between 10% and 20% of a modern nuclear reactor. To keep the energy consumption of an exascale machine below 30 MW, the energy efficiency of such a system must be multiplied by 16. To perform such progress, tremendous efforts must be carried out in many domains: hardware (e.g., energy-efficient processor or memory) and network, but it is expected that some effort must also be carried out on the software side. It is equally important to improve the energy efficiency of an application ...