Uncertainty-Aware Decarbonization for Datacenters

Building carbon-free datacenters depends on effective load scheduling, such as suspend-and-resume [1, 12, 18] and wait-and-scale [5, This paper represents the first effort to quantify uncertainty in 16]. The core idea of these scheduling strategies is to adapt to renewable carbon intensity forecasting for datacenter decarbonization. We energy supplies based on carbon intensity forecasts. Inaccurate identify and analyze two types of uncertainty--temporal and spatial--and carbon intensity forecasts can not only fail to reduce carbon discuss their system implications. To address the temporal emissions but may even increase them [4]. While prior work has dynamics in quantifying uncertainty for carbon intensity forecasting, introduced various methods for carbon intensity forecasting such we introduce a conformal prediction-based framework. Evaluation as ARIMA models [3] and neural networks [9, 10], they focus on results show that our technique robustly achieves target point-based estimation, neglecting to account for their uncertainty coverages in uncertainty quantification across various significance levels. As prior studies point out, considering uncertainty is crucial levels. We conduct two case studies using production power traces, for effective scheduling [17].