Friday 1st of October of 2021
Optimisation at the software level of the energy efficiency associated with the execution of an HPC application on a supercomputer.

PhD thesis summary:

The High Performance Computing (HPC) field is a crucial issue, for both industry and academics: from astrophysics to meteorology, passing by materials science, from Airbus to Total, passing by Pfizer. Computational sciences have become essential, and this dependence implies a never ending urge for more computational power. At the time of writing, all the actors of the HPC field redouble their efforts to reach the ExaScale: 10^18 operations on floating point numbers per second.
Nevertheless, contrary to the previous milestones (e.g. the PetaScale), the computational power achieved by a supercomputer is not the only key performance indicator. Indeed, the first assessment of the electrical power consumed by an exaflopic system were way too high to be acceptable, from both economic and ecological points of view. Consequently, numerous research and development efforts aiming at making supercomputer more energy-efficient were initiated during the last decade. That is precisely the main topic of the work presented in this manuscript, which includes significant contributions to Bull Dynamic Power Optimizer (BDPO), and the conception, development, and experimental validation of Phase - Temporality Analyser (Phase-TA).
BDPO is a dynamic reconfiguration tool, that is to say a daemon executed in parallel of an HPC application, which changes the functioning frequency of the cores of the CPUs to the workload the latter are executing. It has the distinctive feature of being completely agnostic of both the aforementioned executed application and its execution environment, while requiring no specific configuration from the user. Using BDPO induces a 15% decrease of the energy consumption associated with the execution of the two applications NEMO and HPCG, while maintaining the associated performance degradation under 4%. Phase-TA is designed to analyse the profile of an iterative HPC application, notably those produced by BDPO. It detects the locally periodic behaviours, and caracterises them by infering representative patterns for the associated periodicities. What motivated the development of Phase-TA was the possibility to build a relevant and reliable prediction of the future behaviour of the executed application, so as to make the reconfigurations performed by BDPO more efficient. It was experimentally shown that the patterns inferred by Phase-TA are relevant representations of the periodicities featured by HPC applications, and that those periodicities are accountable for a significant part (i.e. more than two thirds) of the execution time of the aformentioned applications. Finally, the performances of Phase-TA make it suitable for on-the-fly analysis of the profile of HPC applications.

Mis à jour le 23 September 2021