Ordonnancement adaptatif en calcul haute performance : modèles prédictifs etmécanismes incitatifs orientés utilisateur pour une allocation des ressources plus sensibleà l'empreinte carbone
Abdessalam Benhari · HAL (Le Centre pour la Communication Scientifique Directe) · 2025
High Performance Computing (HPC) plays a critical role in analyzing and solving complex problems across a wide range of scientific and industrial fields. However, despite its importance, the growing energy consumption associated with HPC raises significant sustainability challenges. This growth, largely driven by increasing computational demands, raises serious concerns about carbon emissions and the related environmental degradation. Various solutions have been developed to mitigate this negative impact, especially through improving the energy efficiency of HPC platforms. However, these measures are generally insufficient when compared to the rapidly evolving landscape of HPC. Thus, there is a need to explore more energy-conscious solutions, particularly through proactive planning methods aimed at reducing the carbon footprint of these systems and by involving users to improve the acceptance of technical solutions.This thesis begins by analyzing the historical trends in HPC systems in terms of performance and energy efficiency, using data from the Top500 and Green500 lists. Leveraging long-term data, we evaluate the carbon footprint of the sector by incorporating information on the energy mix. Since the Top500/Green500 lists provide only partial insight into the environmental impact of HPC systems, we perform a closer analysis of several representative machines from different periods listed in the Top500. This allows us to estimate more precisely the carbon emissions associated with the production and usage of these platforms. Ultimately, we develop a predictive model to estimate the future carbon weight of the HPC sector over the next five years. Our findings indicate that HPC system consumption is on a sharply rising trajectory, while greenhouse gas emissions must be drastically reduced to achieve carbon neutrality. Despite this, HPC can still contribute positively through numerous applications and innovative solutions. According to the IPCC, there is still time to act against climate change. HPC must therefore take part in this effort by ensuring a net positive environmental contribution. In this context, rethinking fundamental mechanisms such as resource management and application scheduling becomes essential.Our second major contribution focuses on new strategies to reduce the carbon footprint of HPC while maintaining optimal performance. These strategies are implemented via scheduling policies that dynamically incorporate both CO₂ rate variations and energy consumption predictions per task. This new information allows the scheduler to adapt its behavior and reduce the system’s carbon footprint. To validate these models, we conducted simulations using real-world data, allowing us to compare them to traditional scheduling policies used on current platforms. The results show that carbon emissions can be reduced by up to 15% without negatively affecting overall resource utilization. Moreover, these scheduling policies remain simple and transparent, facilitating their integration into existing HPC systems and encouraging adoption by users.The third contribution of this thesis broadens our exploration of carbon footprint reduction by focusing on the users of HPC systems. Since users have a substantial impact on system operation, their active involvement is essential. We introduce an incentive mechanism based on a SLURM resource management plugin, which allows for interaction with users. This tool provides real-time feedback on the energy consumption and carbon impact of their submissions. Additionally, users are given the option to run their tasks during periods of lower emissions. This incentive system fosters user responsibility and encourages behavior change aimed at reducing the overall carbon footprint of the system.In conclusion, this thesis presents a comprehensive approach to mitigating the carbon footprint of High Performance Computing by integrating dynamic scheduling policies and user engagement strategies.