An Online Model Integration Framework for Server Resource Workload Prediction

Workload forecasting is critical in cloud environments for the reason that cloud service providers must make timely decisions to meet the needs of users. We propose a load prediction method that integrates multiple models to get the advantages of online as well as offline modeling. During the process of forecasting, the data stream is passed through the LSTM model to output the offline prediction results, while the original data stream passes through the cluster and is sent to the bond Hoeffding Tree to output the online prediction results. After that, the outputs of the offline and the online models are integrated according to the integration algorithm, and the obtained prediction results are sent to the online error correction model, the final prediction result is achieved. Experiments show that our method is superior in terms of RMSE and MAE to some offline and online methods.