Learning Burst-Aware Early Warning Models for Capacity Stress under AI Workload Surges in Hyperscale Data Centers

The paper introduces a burst-aware early warning framework designed for predicting capacity stress in hyperscale data centers under AI workload surges, particularly from large language models. It employs a lightweight XGBoost model, achieving an ROC AUC of 0.697 and a Recall of 0.914, which enables proactive operational interventions before system degradation occurs. This framework is significant for practitioners as it enhances the resilience of data center operations by integrating predictive analytics into workload management strategies, addressing the unique demands of AI-driven jobs.