Learning energy performance of parallel algorithms on an HPC infrastructure

In recent years, the High-Performance Computing (HPC) community has placed a growing emphasis on energy efficiency, recognizing the escalating energy consumption of non-computational components within these systems. This shift has prompted discussions on environmental sustainability in the Information Technology (IT) sector, particularly in HPC, where the term "green" has become a recurring theme. This context sets the stage for the exploration of artificial intelligence (AI) as a key player in predicting and optimizing energy parameters within HPC systems. The coexistence of AI with traditional optimization techniques is highlighted as a promising approach. The thesis, with its central focus on two primary objectives, unfolds within this landscape. Firstly, the thesis aims to compile a detailed dataset encompassing the energy consumption of nodes in an HPC system, specifically tailored to various configurations of jobs executed on the system. Secondly, it seeks to assess the efficacy of an AI model in providing realistic predictions of energy consumption for unknown cases. The specific research focus hones in on parallel algorithms designed for solving linear systems, both in fault-tolerant and non-fault-tolerant versions. This choice is rooted in the widespread applications of linear systems across diverse sectors, adding practical relevance to the undertaken research.