Abdessalam Benhari

Thesis Summary :

High-Performance Computing (HPC) plays an essential role in the analysis and resolution of complex problems in various scientific and industrial fields. However, despite its importance, the environmental consequences of its growing energy use pose significant sustainability challenges. This increase, largely driven by rising computing needs, raises significant concerns about the associated carbon emissions and environmental degradation. Several solutions have been developed to limit this negative impact, notably linked to increasing the energy efficiency of HPC platforms. However, these solutions generally prove insufficient in the face of environmental challenges, given the rapid evolution of the field. It is therefore necessary to think about more sober solutions, in particular scheduling methods oriented towards reducing the carbon footprint of these systems, and solutions involving users to ensure greater acceptance and encourage frugal behavior.
Top500/Green500 lists. Using this information, which is available over a long period of time, we study the carbon footprint of the field, including energy mix data. Since the Top500/Green500 lists only provide part of the information needed to assess the environmental impact of HPC systems, we take a closer look at some examples of machines listed in the Top500 (from different periods) to estimate more precisely the Carbon emissions linked to the production and use of these platforms. Finally, we derive a predictive model to estimate the weight of the HPC domain over a five-year horizon. We deduce from this study that the consumption of HPC systems is following a significant upward trajectory, while greenhouse gas emissions need to be drastically reduced to achieve carbon neutrality. However, many applications and solutions on these systems can help to limit the negative effects. Today, IPCC experts believe that there is still time to take action against climate change. The HPC field must therefore play its part in this effort by ensuring a positive overall balance. With this in mind, it is essential to rethink certain fundamental mechanisms of HPC operation, in particular resource management policies and application allocation/scheduling.
This is precisely the aim of our second contribution, which proposes new strategies to reduce the carbon footprint of HPC while maintaining optimal performance. These strategies are implemented through scheduling policies that dynamically integrate the variation of the CO2 rate and a prediction of the energy consumed per task. This new information is used to adjust the scheduler's behavior in order to reduce the system's carbon footprint. To validate these models, we carried out simulations on real data, enabling us to compare them with conventional scheduling policies. The results show a reduction in carbon emissions of up to 15%.
The third contribution of this thesis is to broaden the scope of our exploration of carbon footprint reduction solutions by targeting the users of HPC systems. The latter have a considerable impact on the way these systems operate, hence the need to take them into account by creating active participation. We are introducing an incentive mechanism based on a plugin in the SLURM resource management software, enabling interaction with users. This tool provides users with real-time feedback on the energy consumption of their submissions and the associated carbon impact. What's more, the tool enables users to choose to execute their tasks during periods of lower emissions. This incentive mechanism empowers users by encouraging behavioral changes aimed at reducing the system's overall carbon footprint.
Overall, this thesis proposes a comprehensive approach to mitigating the Carbon footprint of High-Performance Computing (HPC), by integrating dynamic scheduling policies and user engagement strategies to promote more sober and sustainable computing practices.