基于强化学习的数据中心智能机架级制冷系统研究

Intelligent Rack-level Cooling System in Data Center Based on Reinforcement Learning

在数据中心内,冷气的供应量与需求量不匹配是导致机架热点和能源利用率低的主要原因。数据中心现有的机架级制冷方案总是提供过量的冷气,而缺少对冷气供应的控制研究。过量的冷气不仅会加剧冷资源的浪费,而且依旧无法消除机架热点。针对上述问题,提出数据中心智能机架级制冷系统。该系统利用马尔科夫决策过程(MDP)建立问题模型。首先,通过LoRa传感器节点采集机架各层的入风口温度;然后,将LoRa网关接收到的温度数据上传到控制模块,控制模块采用基于深度强化学习(DQN)的算法求解问题模型;最后,通过驱动模块控制主动通风地板(AVT)的风扇转速,为机架各层存放的服务器提供适量的冷气。实验结果表明:在使用智能机架级制冷系统后,实验机架各层的入风口温度均降低到22℃以下,且实验机架各层的温度均方差相较于使用前降低至0.109 5。

In the data center,the mismatch between the supply and demand of the air-conditioning is the main reason causing rack hot spots and the low energy utilization. Existing rack-level cooling solutions in data centers always provide excess cooling air,not only aggravating the waste of cold resources,but also still cannot eliminate rack hot spots. However,there was a lack of research on the control of cooling air supply. In view of the above problems,the intelligent rack-level refrigeration system was proposed in this paper. The Markov Decision Process(MDP)was used to build a problem model. First,the air inlet temperature of each layer of the rack was collected through the LoRa(Long Range)sensor node. Then,the temperature data received by the LoRa gateway were uploaded to the control module,and the control module used an algorithm based on the deep reinforcement learning(DQN)to solve the problem model. Finally,an appropriate amount of cooling for the servers stored on each floor of the rack was provided through controlling the fan speed of the Active Ventilation Floor(AVT)by the drive module. The experimental results showed that after using the intelligent rack-level refrigeration system,the temperature of the air inlets of each layer of the experimental rack fell to below22℃,and the temperature mean square error of each layer of the experimental rack decreased to 0. 109 5 compared to that before using the system.