Online Learning-Assisted VNF Service Chain Scaling with Network Uncertainties

Abstract

Network function virtualization has emerged as a promising technology to enable rapid network service composition/innovation, energy conservation and cost minimization for network operators. To optimally operate a virtualized network service, it is of key importance to optimally deploy a VNF (virtualized network function) service chain within the provisioning infrastructure (e.g., servers and the network within a cloud datacenter), and dynamically scale it in response to flow traffic changes. Most of the existing work on VNF scaling assume access to precise network bandwidth information for placement decisions, while in reality, network bandwidth typically fluctuates following an unknown pattern and an effective way to adapt to it is to do trials. In this paper, we address dynamic VNF service chain deployment and scaling by a novel combination of an online provisioning algorithm and a multi-armed bandit optimization framework, which exploits online learning of the available bandwidths to enable optimal deployment of a scaled service chain. Specifically, we adopt the online algorithm to minimize the cost for provisioning VNF instances on the go, and a bandit-based online learning algorithm to place the VNF instances which minimizes the congestion in a datacenter network. We demonstrate effectiveness of our algorithms using solid theoretical analysis and trace-driven evaluation.