An LSTM‐Based Resource Prediction Model in Google Cloud Data Center

The increasing complexity and dynamic nature of workloads in cloud computing are characterized by various patterns with time‐dependent features. Such variations pose considerable challenges on Cloud Service Providers (CSPs) incurred by the fluctuation of resource demands along with their impact on quality of service (QoS) optimization. The Google cluster trace is an example of such workloads wherein task characteristics significantly differ between applications, which makes the existing approaches insufficient for accurate predictions across various workload scenarios. Such limitations emphasize the need for a comprehensive, domain‐specific consideration of the application of deep learning (DL) models to effectively predict cloud workloads. This paper investigates the design and application of a workload prediction model based on long short‐term memory (LSTM) networks, developed for the unique characteristics of the Google cluster trace. The model leverages the LSTM’s ability in capturing long‐term dependencies within the trace to predict future workload demands, with the extracted features fed into various machine learning (ML) models such as linear regression (LR), K‐nearest neighbor (KNN), and decision tree (DT). The prediction model enables CSPs to proactively allocate sufficient resources for client tasks and optimize workload distribution. It is found that the proposed model combines strengths of feature extraction in LSTM with the prediction’s accuracy of LR to forecast CPU and memory requests of the Google cluster trace. This combination demonstrates superior capabilities compared to convolutional neural network (CNN), autoregressive integrated moving average (ARIMA), and extreme gradient boosting (XGBoost) models.