Resource Demand Prediction of Cloud Workloads Using an Attention-based GRU Model

Resources of cloud workloads can be automatically allocated according to the requirements of the application. In the long-term running process, resource requirements change dynamically. Insufficient allocation may lead to the decline of service quality, and excessive allocation will lead to the waste of resources. Therefore, it is crucial to accurately predict resource demand. This paper aims to improve resource utilization in the data center by predicting the resources required for each application. Resource demand forecasting understands and manages future resource needs by mining current and past resource usage patterns. Because we need to analyze time series data with long-term dependence and noise, it is challenging to predict future resource utilization. We designed and implemented an attention-based GRU model. The attention mechanism was added to the GRU model to quickly filter out valuable information from large amounts of data. We used the Azure and Alibaba cluster trace to train our neural network, and used three evaluation indicators RMSE, MAPE and R2 to evaluate our proposed method. The experimental results show that our prediction method has 4.5% improvements in RMSE evaluation criteria and 9.5% improvements in MAPE evaluation criteria compared with single GRU model (without attention mechanism) used. That is, the prediction model with the attention mechanism can improve the accuracy of resource prediction. At the same time, we also studied the influence of the window size on the experimental results, finding that the prediction results are more accurate as the window size increases.