A hybrid strategy for auto-scaling of VMs: an approach based on time series and thresholds

Abstract

Demand for performance, availability, and reliability in computational systems has increased lately. Improving these aspects is an important research challenge due to the wide diversity of applications and customers. The elasticity of cloud computing applications plays an important role, and its implementation is based on auto-scaling techniques, which allow to increase and decrease cloud-based application’s capacity. Therefore, it is possible to deal with the workload variation and not interrupt the service. The auto-scaling process can be grouped into two classes: reactive and proactive. In this dissertation, we propose a hybrid auto-scaling approach that uses reactive and proactive solutions. Our proposal is implemented using triggering thresholds and five forecasting models: Drift, Simple Exponential Smoothing, Holt, Holt-Winters and ARIMA. It checks if the CPU utilization has achieved the threshold, if it has, the system is scaled (out/in). On the other hand, if the threshold was not achieved the forecasting phase starts. In this phase, the five forecasting models are trained and tested, so the most suitable one is used to forecast resource utilization. The main goal of this research is to achieve a better QoS related to the cloud computing environment. One of the obtained results shows that our method represents a throughput improvement of 12.11% by using our proposal, instead of only using a threshold-based technique.