Publication

Abstract : Understanding the resource demands and managing the resources for future workloads is a challenging task in the cloud. Typically, applications receive dynamic and time-varying workloads. Forecasting future workload and subsequently inferring future resource requirements will aid in better capacity planning and resource utilization. In this paper, we propose an unsupervised approach to predict the number of utilization profiles required to model a workload. An ensemble-based workload forecasting is proposed to predict future workloads. For forecasting, three models, namely ARIMA, XGBoost, and LSTM, are ensembled. Both forecasting and classification are used to predict the number of resources required so that the problem of over-provisioning and under-provisioning of the resources can be addressed during capacity planning and provisioning. The error rate is decreased by 15% with ensemble model when compared with individual models for prediction.