基于深度学习的主机负载在线预测模型研究

数据中心主机负载预测对于数据中心的资源调度和节能具有重要意义，但是目前缺乏一个通用模型以准确预测所有类型数据中心的主机负载情况。为了使主机负载预测模型具有一定的自适应性，提出一种基于深度循环神经网络编码器-解码器的多步在线预测模型。通过线上实时采集的能耗数据进行在线训练，同时设计一个在线监控模块，对模型的预测准确性进行实时监控和调整，使得该模型在不同数据中心中均能获得较准确的预测值。利用Google开源的时长为29天的数据中心主机负载数据集进行实验，结果表明，该模型的预测准确性接近离线训练，其预测性能优于ESN和LSTM模型。

Research on Online Prediction Model of Host Load Based on Deep Learning

Host load prediction plays a key role in data center resource scheduling and energy saving, but there is no universal model that can accurately predict the host load of all types of data centers.In order to improve the adaptability of host load prediction, this paper proposes a multi-step online prediction model based on a deep cyclic neural network encoder-decoder.The model is trained online with energy consumption data collected in real time.In addition, an online monitoring module is designed to monitor and adjust the prediction accuracy of the model in real time, so that the model can obtain more accurate prediction results in different data centers.Experiments are carried out on a data set open-sourced by Google, which contains host loads of a data center spanning 29 days.The experimental results show that the proposed model displays similar prediction accuracy to the models trained offline, and its prediction performance is better than that of ESN and LSTM.