Carbon Intensity Forecasting with On-Site Renewables for Sustainable AI Workload Scheduling

In recent years, the environmental impact of cloud computing and AI technologies has steadily increased. Data centers currently consume more than 400TWh of electricity each year, and this number is only expected to grow more as companies invest in cloud services and in increasingly compute-hungry Machine Learning (ML) models. While still in an early phase, researchers and industries have started to pay more attention to the rising electricity demand of these technologies, and several computing centers have begun to implement emission reduction measures, such as job execution in low-carbon emissions areas. The allocation of workloads in external regions has, however, led to non-trivial Carbon Intensity (CI) estimations, making it challenging to apply carbon-aware scheduling techniques. Furthermore, accountability measures in these new approaches tend to still be lacking, especially for electricity consumption and emissions, as these metrics are rarely disclosed. This work presents an end-to-end architecture for the minimization of carbon emissions of AI workloads in hybrid cloud environments, based on an ML forecaster, a scheduler, and a provenance-driven accountant component. Experimental results show a 27% reduction in emissions when using the scheduler component against the performance of a carbon-agnostic baseline.