Prediction-Based VNF Auto-scaling for Energy and QoS Efficiency
Network Function Virtualization; Design of virtual functions; Machine learning; Demand Prediction
Auto-scaling of virtual functions has received attention with proposals that suggest reactive approaches based on monitoring the consumption of resources dedicated to them, and proactive, making decisions ahead of time driven by service demand prediction with the addition of machine learning techniques. Such approaches also enable efficient cost management of the infrastructure, for example, by shutting down unused equipment during periods of low demand for energy saving. However, without accurate knowledge of traffic expectations, wrong decisions can undermine the quality of running services, thus establishing the need to balance the trade-off between cost and performance. Proactive solutions that the literature presents do not evaluate the impact of their decisions on the running services either ways to mitigate their performance losses due to wrong decisions. This work aims to propose a demand prediction-based mechanism for VNF auto-scaling to achieve a reduction in infrastructure power consumption and minimize the impact of decisions made on the Quality of Service of running services. In addition, we intend to assess the impact on service performance and cost under vertical and horizontal scaling schemes in order to guide decisions in this direction as well.