Generative Coflow Scheduling for Cross-Silo Federated Large Language Models Synchronization in Computing Power Networks

作者：Xuening Shang, Deyun Gao, Dong Yang, Weiting Zhang, Hongke Zhang · 发表于：IEEE Transactions on Network Science and Engineering · 年份：2025 · DOI：10.1109/tnse.2025.3627637 · 被引用次数：2 · 研究领域：Software-Defined Networks and 5G、Parallel Computing and Optimization Techniques、Big Data and Digital Economy

To efficiently transmit bandwidth-intensive synchronization flows of cross-silo federated large language models in computing power networks, coflow scheduling emerges as a promising approach to minimize waiting time by balancing bandwidth allocation and flow priority within each coflow group. However, due to heterogeneous computing resource and dynamic network environment, traditional coflow scheduling leaves a performance gap in iteration completion time (ICT), comprising different computing completion time (CCT) and untraceable flow completion time (FCT) within a coflow group. To fill this gap, we first formulate an ICT-oriented coflow scheduling system model that integrates a CCT-aware scheduling mechanism to collect complicated coflow group relation, and an ICT-oriented Markov Decision Process (MDP) to model the relationship between FCT and dynamic link load by a policy network. To leverage coflow group relation during policy training, we propose a group-aware deep reinforcement learning framework, which incorporates a group-aware trajectory collection mechanism and a group-aware reward design. Furthermore, we replace the traditional single-step policy model with a multi-step generative diffusion model, enabling coflow scheduling policy generation based on the dynamic load information of flows within each coflow group. Simulation results on real-world traces demonstrate that the proposed method outperforms baselines in terms of ICT.