Malleable Techniques and Resource Scheduling to Improve Energy Efficiency in Parallel Applications

Abstract

The high energy consumption of computing platforms has become one of the major problems in high-performance computing (HPC). Computer energy consumption represents a significant percentage of the CO2 emissions that occur each year in the world, therefore, it is crucial to develop energy efficiency techniques in order to reduce the energy consumption in HPC systems. In this work, we present a resource scheduler capable of choosing, using real-time data, the optimal number of processes of running applications. The solution takes advantage of the use the FlexMPI runtime to dynamically reconfigure the application number of processes and DVFS to modify the frequency of cores of the platform. The scheduling algorithms presented in this work include one that minimizes the application energy and another more holistic one that allows the user to balance between energy and execution time minimization. This work presents a description of the methodologies and a experimental evaluation on a real platform.

This work was partially supported by the EuroHPC project “Adaptive multi-tier intelligent data manager for Exascale” under grant 956748 - ADMIRE - H2020-JTI-EuroHPC-2019-1 and by the Agencia Española de Investigación under Grant PCI2021-121966.