Equity-Aware Spatial-Temporal Workload Shifting for Sustainable AI Data Centers (Papers Track)
Mohammad Islam (University of California Riverside); Shaolei Ren (UC Riverside)
Abstract
The escalated demand for hyperscale data centers due to generative AI, has intensified the operational load, leading to increased energy consumption, water usage, and carbon emissions. We propose EquiShift, a novel equitable spatial-temporal workload balancing algorithm that shifts workloads spatially and temporarily across geographically different data centers to minimize the overall energy costs while ensuring fair distribution of water and carbon footprints. Concretely, EquiShift introduces a model predictive control (MPC) framework to solve the equitable load balancing problem, leveraging the predictive capabilities of MPC to optimize load distribution in real-time. Finally, we present comparative evaluations against state-of-the-art load-balancing algorithms to demonstrate the performance of EquiShift which underscores the potential of equitable load balancing as a key strategy for enhancing the sustainability of data centers while achieving fairness in the face of growing computational demands.