Green scheduling for LLM workloads with model and data reuse across geo-distributed data centers

作者：Hao Liu, Xiaonyu Hu, Ran Wang, Jie Hao, Qiang Wu, Hongke Zhang · 发表于：Digital Communications and Networks · 年份：2025 · DOI：10.1016/j.dcan.2025.11.006 · 被引用次数：1 · 研究领域：Distributed and Parallel Computing Systems、Cloud Computing and Resource Management、Parallel Computing and Optimization Techniques

The explosive proliferation of Large Language Models (LLMs) imposes significant energy and operational burdens on Geographically Distributed Data Centers (GDDCs), thereby demanding an efficient mechanism for LLMs task scheduling. While prior geo-distributed scheduling methods reduce cost and carbon emissions by exploiting regional heterogeneity, they largely overlook model and data reuse opportunities and the uncertainty of LLM execution times. In this paper, we introduce GCOS, to the best of our knowledge, the first green scheduling framework that incorporates a dual-cache system for both data and models, while jointly optimizing task assignment and cache migration. We firstly propose a dual-cache mechanism that decouples model and data caching to enable fine-grained reuse and minimize redundant transmissions. Subsequently, we propose the Multi-Agent Cache-aware Cooperative Scheduling (MACCS) algorithm, which leverages reinforcement learning to optimize task placement with a focus on minimizing both carbon emissions and cost. Additionally, we design a lightweight execution time predictor, DiPTree, to address the high variability in task execution times. Extensive experiments on real-world datasets demonstrate that GCOS reduces overall cost by up to 92.6 % and carbon emissions by 90.3 %, significantly outperforming existing baselines.