Predictive Cloud Resource Management by Workload Forecasting with an ensemble of Hybrid ARIMA-LSTM and EL models

Abstract— Due to the dynamic nature of cloud, many modern applications such as IoT, ML and AI based applications depend on cloud for deployment and provisioning. Cloud service providers (CSPs) need to provide the cloud resources requested by the consumer applications with high scalability and less latency. By doing so, the providers can achieve good quality of service (QoS) for their clients. However, meeting QoS with cost-effectiveness is one of the challenging problems for CSPs as the workloads vary dynamically. It is highly necessary to provide an accurate resource estimation and scheduling mechanism at the CSP’s end. Cloud workload prediction helps the CSPs to efficiently manage the cloud resources. To predict the ensuing workload for a CSP, Time series analysis and forecasting techniques may be leveraged. These techniques show how resource requirements change over time and aid in identifying the trends and patterns in the resource requirement and hence determine the optimal amount of resources required in the near future. In this paper, a novel univariate time series forecasting methodology is proposed for modelling the incoming resource requests with respect to time. This paper experiments the ARIMA and LSTM models for building the prediction model. To further improve the prediction accuracy, an ensemble of hybrid of ARIMA and LSTM models and ELM was developed considering the cloud workload’s inherent nature, the volume of the dataset and the technique’s efficiency. The trace-driven experiments based on Google cluster trace dataset demonstrates that the proposed model outperforms the Prophet model in terms of metrics such as MSE, RMSE, NRMSE, MAE, MAPE, R2 Score on incoming CPU request.