A Novel Container Workload Prediction Method Based on Automatic Classification and Attention-based LSTM

With the rapid development of cloud native technology, a growing number of applications run in containers. For cloud service providers, accurate container workload prediction can contribute to the improvement of resource utilization and service level agreement. However, containers that support different sorts of businesses have different preferences on resource requirements, and thus the prediction methods via one general model cannot work for all containers. In this paper, we propose a novel workload prediction method to adapt to different classes of containers and obtain more accurate prediction results. We first develop an automatic classifier to classify the containers according to their resource usage records. For each type of container, we train an attention mechanism-based Long Short-Term Memory(LSTM) model to predict the workload. Then a weight optimizer and the trained LSTM models are assembled into an ensemble predictor. The proposed weight optimizer can dynamically update weights according to the real-time container workload. Experiments on the Google cluster-data dataset show that the accuracy of our ensemble predictor is 7% higher than that of other two competitive methods.