Autoscaling prediction models for cloud resource provisioning

Abstract

The elasticity mechanism of cloud computing can auto scale cloud resources to meet users' need. Elastic adding or removing virtual machines is the most common method to achieve the auto scaling. But the elastic scaling often takes tens of minutes, which is inefficient for the running workload. To reduce the latency and improve the quality of service (QoS), the new virtual machine should be provisioned when the request arrives. In this paper, we present a prediction framework for virtual machines provisioning. This prediction framework includes three main modules: monitor, filter and predictor. This framework aims to predict the upcoming workload and provision the virtual machines in advance. To get the reasonable monitored metrics, we propose the Kalman filter method to preprocess the raw data. Moreover, we present five different prediction models as the based predictor. These prediction models include moving average (MA), auto regression (AR), auto regression integrated moving average (ARIMA), neural networks (NN) and support vector machine (SVM). Meanwhile, we propose four evaluation metrics, including the prediction error, the time saving, under-prediction resource and over-prediction resource, to evaluate the performance of prediction framework. In addition, we use Alicloud as the experimental infrastructure. Experimental results demonstrate that the prediction framework can reduce the latency of provisioning cloud resource and improve the cloud service quality.