Reducing Scheduling Regret in Non-Stationary Cloud Environments Through LLM-Augmented Online Convex Optimization

Cloud computing platforms exhibit pronounced workload non-stationarity, rendering static scheduling policies chronically suboptimal. Online Convex Optimization (OCO) offers a principled framework for sequential decision-making under adversarial uncertainty, yet classical OCO algorithms accumulate excessive dynamic regret when the loss landscape drifts rapidly. This paper proposes LLM-AOCO, a scheduling framework that augments OCO with contextual gradient estimates generated by Large Language Models (LLMs). By leveraging LLM semantic interpretation of multi-modal system telemetry—including queue depths, resource utilization vectors, and Service Level Agreement (SLA) violation signals—LLM-AOCO constructs surrogate gradients that substantially improve algorithmic responsiveness to environmental drift. Theoretical analysis establishes dynamic regret bounds of O(√T(1+P_T)) under standard smoothness assumptions, matching the information-theoretic lower bound for non-stationary OCO. Experiments on synthetic non-stationary traces and real-world Azure cluster traces demonstrate that LLM-AOCO reduces cumulative scheduling regret by 34.7% relative to online gradient descent baselines and improves resource utilization by 18.3% compared to state-of-the-art adaptive schedulers.