Squirrel

Despite significant advances in hardware efficiency, data centers continue increasing their electricity consumption due to rising application demand. Consequently, data center carbon emissions grow. In addition to energy-saving mechanisms, increasing progress in carbon-aware software is necessary to reduce carbon emissions and mitigate climate change. Enhancing workload managers used in data centers with carbon-aware mechanisms counteracts their emissions. State-of-the-art solutions do not cover carbon-aware scheduling for the widely used workload manager Slurm. This thesis develops and evaluates three carbon-aware scheduling algorithms for high-performance computing clusters using the Slurm scheduler. To this end, we elaborate on whether we can reduce operational emissions using spatiotemporal workload shifting within a Slurm cluster in Germany and how the effects differ across energy zones with different characteristics of their electricity grid. We develop and evaluate Squirrel, a carbon-aware batch job scheduler pro- totype that considers the grid carbon intensity and servers’ thermal design power values. We also propose a forecasting algorithm for grid carbon intensity that can be used when there is no access to forecasting data. We measure the energy profiles of different TPCx-AI jobs and use their hourly energy consumptions to simulate workload scheduling for three different cluster configurations. Our simulations compare Squirrel with carbon-agnostic first-in-first-out scheduling in five different energy zones. The comparison is based on the hourly computation of schedules using data throughout 2023. Squirrel achieves average emission savings between 0.3% and 32% in our simulations, demonstrating the practical implications of our research.