基于长短期记忆网络(LSTM)的数据中心温度预测算法

Data Center Temperature Prediction Algorithm Based on Long Short-term Memory Network

大数据和云计算时代推动数据中心规模迅速扩大,有调查研究显示,国内数据中心年耗电量超过全社会用电量的1.5%,且在数据中心运行时高达10%的机柜运行温度高于设备可靠性的允许范围。温度监测和预测对于防止服务器过热而停机和提高数据中心的能源效率至关重要。文中提出了一种基于长短期记忆网络(LSTM)的温度预测算法,该算法使用数据中心温度监控数据和服务器实际运行参数生成时间序列训练集来训练神经网络模型并预测服务器入口温度。为了降低预测模型的训练时间,基于热局部性原理提出了一种联合建模框架,显著降低了在线温度预测建模的复杂性。在一个有15台服务器的测试台上进行了实验验证,结果表明该方法可以准确地预测动态工作负载的服务器的入口温度演变。

Nowadays,data center has been expanded dramatically by the development of big data and cloud computing.A survey has showed that annual power consumption of data centers in China exceeds 1.5%of the whole society consumption,furthermore,up to 10%of the rack operating temperature is higher than the allowable range of equipment reliability.Temperature monitoring and prediction are crucial for preventing server failures caused by overheating and improving energy efficiency in the data center.We propose a temperature prediction algorithm base on long short-term memory(LSTM)network.In this algorithm,we train the neural network model with temperature monitoring data and time series set composed of parameters of actual runtime and finally obtain the predicted temperature of the server entrance.In order to reduce the training time of the prediction model,we propose a joint modeling framework based on the thermal locality principle,which significantly reduces the complexity of online temperature prediction modeling.The experimental verification on a test bed with 15 servers shows that the proposed method can accurately predict the inlet temperature evolution of servers with dynamic workloads.