通信学报2025，Vol.46Issue(8)：1-15,15.DOI:10.11959/j.issn.1000-436x.2025135
基于深度强化学习的分离式数据中心存储资源调度优化方法
Storage resource scheduling optimization method for separated data center based on deep reinforcement learning
摘要
Abstract
To address the high overhead of scheduling,deployment,and algorithm execution in disaggregated data center storage systems,a deployment solution was proposed that embeded the storage scheduling mechanism into white-box switches.By enabling the interaction between the switches and data processing unit(DPU),the system achieved auto-mated data migration,thereby freeing up I/O resources occupied by storage access.Through a Bayesian-based data tem-perature sensitivity optimization strategy integrating access volume and access interval,it fed the heat information to a deep reinforcement learning(DRL)scheduler that minimized access latency.Simulations show that,against rule-based LRU and FIFO,latency drops from 38.4%to 67.2%.Against DQN,PPO and DDPG,latency drops from 12.8%to 43.2%.关键词
分离式数据中心/多层级存储/数据迁移/深度强化学习/资源调度Key words
disaggregated data center/multilevel storage/data migration/deep reinforcement learning/resource scheduling分类
信息技术与安全科学引用本文复制引用
袁政利,郭少勇,胡鑫,仝杰,郝佳恺,Michel Kadoch,喻鹏..基于深度强化学习的分离式数据中心存储资源调度优化方法[J].通信学报,2025,46(8):1-15,15.基金项目
国家自然科学基金资助项目(No.62322103) （No.62322103）
中央高校基本科研业务费专项资金资助项目(No.2023ZCTH11) （No.2023ZCTH11）
北京市自然科学基金资助项目(No.4232009)The National Natural Science Foundation of China(No.62322103),The Fundamental Research Funds for Cen-tral Universities(No.2023ZCTH11),The Natural Science Foundation of Beijing(No.4232009) （No.4232009）
