Experimental Evaluation of ML Models for Dynamic VNF Autoscaling
Vasileios Zalokostas-Diplas, Nikos Makris, Virgilios Passas, Thanasis Korakis · 2022
Network Functions Virtualization (NFV) is a key aspect deeply integrated in the latest 5G networks, allowing for the provisioning of elastic resources that adapt in a flexible manner based on the overall network demand. The adoption of NFV architectures is empowered through the evolution of cloud-native and hypervisor tools to support service monitoring, and orchestrate the appropriate decisions for provisioning the scale of the network. Such decisions may directly impact the overall quality of service and experience for users, as well as the energy consumption that the resources use. To this aim, machine learning (ML) - driven optimization for these decisions, relying on inferring the values of future monitored metrics, can assist in deciding proactively on the network scale. In this work, we employ three different candidate solutions (statistical, tree- and CNN-based) for determining the scale of network functions deployed within a cluster of resources, subject to the user demand. We compare and evaluate the different schemes in a real testbed environment, and discuss the benefits of ML-driven optimizations against existing state-of-the-art approaches.