Time series forecasting of cloud data center workloads for dynamic resource provisioning

Cloud computing offers on-demand, elastic resource provisioning that allows an enterprise to provide services to their customers at an acceptable quality while consuming only the requisite computing resources as a utility. Since cloud computing resources scale elastically, utilizing cloud computing reduces the risk of over-provisioning, wasting resources during non-peak hours, and reduces the risk of under-provisioning, missing potential customers. By using an automated resource scaling algorithm, a system implemented using cloud services can fully exploit the benefits of on-demand elasticity. A simple reactive scaling algorithm, resource scaling is triggered after some monitored metric has crossed a threshold, suffers from the fact that cloud computing workloads can vary widely over time and a scalable application needs time to perform the triggered scaling action. Instead, resources can be proactively requested by forecasting future resource demand values based on demand history. Obtaining accurate prediction results is crucial to the efficient operation of an automated resource scaling algorithm. In this work, several forecasting models are evaluated for their applicability in forecasting cloud computing workloads. These forecasting methods were compared for their ability to forecast real cloud computing workloads including Google cluster data and Intel Netbatch logs. Three tests are performed to evaluate the accuracy of each forecasting model: out-of-sample forecasting, rolling forecast origin cross-validation, training set length analysis.