Guanyu Ding, Shiyu Yang, Han Lin, Z Chen, J. C. Yang
Cloud resource scheduling presents a fundamental challenge in modern distributed computing, where heterogeneous workloads, complex task dependencies, and multi-objective optimization requirements exceed the capabilities of traditional rule-based and small-scale machine learning approaches. Existing schedulers struggle to dynamically adapt to evolving workload patterns while simultaneously satisfying Service Level Agreement (SLA) constraints, resource efficiency targets, and fairness policies. This paper introducesLLMSched, a novel Large Language Model (LLM)-driven adaptive scheduling framework that synergistically combines the contextual reasoning capabilities of foundation models with the execution guarantees of classical optimization techniques. Our approach transforms heterogeneous cluster states-including task dependency graphs, real-time resource utilization metrics, and SLA specifications-into a unified structured-textual representation that leverages LLMs' few-shot learning and causal reasoning abilities to generate intelligent scheduling candidates. These candidates are subsequently refined through a lightweight Integer Linear Programming (ILP) module that ensures feasibility and optimality under resource constraints. We evaluate LLMSched on Google's production cluster trace dataset, demonstrating significant improvements over state-of-the-art baselines: 23.7% reduction in average job completion time, 18.4% improvement in resource utilization efficiency, and 31.2% decrease in SLA violations across diverse workload scenarios. Extensive ablation studies validate the contributions of each architectural component, while robustness analysis under workload perturbations confirms the framework's practical viability. Our work establishes a new paradigm for intelligent resource management that bridges the gap between neural reasoning and algorithmic precision, opening avenues for LLM applications in systems optimization domains.