Learning Virtual Machine Scheduling in Cloud Computing Through Language Agents

In cloud services, virtual machine scheduling is a typical online dynamic multidimensional bin-packing (ODMBP) problem characterized by large-scale and nonstationary demands. Traditional optimization methods struggle to adapt to dynamic environments, existing learning-based methods often lack generalizability and interpretability, and domain expert–designed heuristic approaches are constrained by rigid strategies. To address these limitations, this paper proposes a hierarchical language agent framework named MiCo, which provides a large language model (LLM)–driven heuristic design paradigm for solving ODMBP. Specifically, ODMBP is formulated as a semi-Markov decision process with options, enabling dynamic scheduling through a micromacro hierarchical architecture, that is, option miner and option composer. Option miner utilizes LLMs to discover context-independent strategies through environment interactions. Option composer uses LLMs to develop a context-aware composing strategy that integrates the context-independent strategies. Extensive experiments on a real-world data set demonstrate that MiCo achieves a 96.9% performance ratio under large-scale and nonstationary scenarios. It maintains high performance even under nonstationary request flows and different configurations. History: Accepted by Ram Ramesh, Area Editor for Data Science & Machine Learning. Funding: The work of J. Luo was supported by the National Natural Science Foundation of China [Grants 72542012, 72571170, and 72031006]. The work of X. Wang was supported in part by a research collaboration with Huawei Cloud [Grant TC20241122017]. The work of W. Li was supported in part by the National Natural Science Foundation of China [Grant 62406270] and the Science and Technology Commission of Shanghai Municipality Shanghai Rising-Star Program [Grant 24YF2748800]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2025.1368 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2025.1368 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .