通信学报2025，Vol.46Issue(11)：147-161,15.DOI:10.11959/j.issn.1000-436x.2025195
基于DDPG强化学习的多集群算力资源调度算法
Multi-cluster computing power resource scheduling algorithm based on DDPG reinforcement learning
摘要
Abstract
Addressing algorithm deficiencies and severe resource fragmentation,particularly regarding heterogeneous re-sources like GPU,in multi-cluster scheduling,the problem was modeled as a Markov decision process.An expert aug-mented dual experience deep deterministic policy gradient(EADE-DDPG)algorithm was proposed.EADE-DDPG inte-grated a self-attention mechanism for dynamic state extraction and emploied a dual experience buffer(combining expert/offline and real-time/online knowledge)to ensure high sample efficiency and stability.The model also featured an ex-panded delayed deployment action for handling resource scarcity.Validation using a real-world Alibaba dataset shows EADE-DDPG yields optimal GPU resource utilization,boosting maximum task deployment by 14.83%,cutting resource fragmentation by up to 14.50%,and decreasing cross-cluster GPU load variance by up to 89.01%,thus significantly im-proving scheduling efficiency and load balancing.关键词
多集群/算力资源/资源调度/深度强化学习/深度确定性策略梯度Key words
multi-cluster/computing power resource/resource scheduling/deep reinforcement learning/DDPG分类
信息技术与安全科学引用本文复制引用
胡亚辉,王越嶙,张宸康,洪雨琛,范鹏飞,宋俊平,周旭,鲍震..基于DDPG强化学习的多集群算力资源调度算法[J].通信学报,2025,46(11):147-161,15.基金项目
国家重点研发计划基金资助项目(No.2024YFB2908700) （No.2024YFB2908700）
高校基本科研业务费专项资金资助项目(No.2025ZKPYZN02) （No.2025ZKPYZN02）
国家能源集团科技环保有限公司开放课题基金资助项目(No.YZ-2025-101)The National Key Research and Development Program of China(No.2024YFB2908700),The Fundamental Re-search Funds for the Central Universities(No.2025ZKPYZN02),Open Research Topics of CHN Energy Technology&Environment Limited(No.YZ-2025-101) （No.YZ-2025-101）
