Team, Visitors, External Collaborators
Overall Objectives
Research Program
Application Domains
New Software and Platforms
New Results
Bilateral Contracts and Grants with Industry
Partnerships and Cooperations
Dissemination
Bibliography
XML PDF e-pub
PDF e-Pub


Section: New Results

Extraction of Periodic Patterns of Scientific Applications to Identify DVFS Opportunities

Participants : Mathieu Stoffel, Fran├žois Broquedis, Frederic Desprez, Abdelhafid Mazouz [Atos/Bull] , Philippe Rols [Atos/Bull] .

Mathieu Stoffel started his PhD in February 2018 on a CIFRE contract with Atos/Bull. The purpose of this work is to enhance the energy consumption of HPC applications on large-scale platforms. The first phase of the thesis project consists in an in-depth study of the evolution of the metrics characterizing the state of the supercomputer during the execution of a highly parallel application. Indeed, the utilization rates of the different components of the HPC system may demonstrate extreme variations during the execution of the aforementioned application. These variations are sometimes subject to repeat themselves on a regular basis during the application execution. We refer to this phenomena as application "phases". In this context, we developed a tool suite resorting to fine-grain profiling and periodicity analysis to identify optimization opportunities for both performance and power-efficiency. It leverages the fact that a large share of HPC parallel applications are constituted of a restrained set of compute kernels executed a huge number of times to extract periodic patterns representative of the aforementioned kernels. By doing so, our tool offers a simple and condensed proxy to analyze and predict the behavior of complex parallel applications. For instance, we were able to identify and extract periodic patterns for a panel of reference HPC applications such as NAMD and NEMO. Then, as an example of the many ways to exploit the aforementioned extracted periodic patterns, we evaluated the impact of the CPU frequency on the latter. As a result, we were able to identify DVFS opportunities we plan to exploit in a future work.