Online VNF scaling with network uncertainties

Abstract

Virtualized Network Functions (VNFs) are gaining a lot of attention from both the industry and the academia, as a promising technology to enable rapid network service composition/innovation, energy reduction and cost minimization for network operators. The VNFs typically run on virtual machine instances in a cloud infrastructure, where the virtualization technology enables dynamic provisioning of VNF instances, to process the fluctuating traffic that needs to go through the network functions in a network service. To optimally operate VNFs, it is of key importance that VNFs be scaled dynamically in response to traffic changes. We target dynamic provisioning of enterprise network services in datacenters, and design efficient online algorithms that would not require any information on future traffic rates. In the meantime, most of the works on VNF scaling assume the availability of precise network information beforehand, whereas in reality, network bandwidth fluctuates and the only way we could gain more accurate information is to do trials. We address the problem by a novel combination of an online algorithm and a bandit optimization framework. Specifically, we adopt an online algorithm to minimize the cost to provision VNF instances and a multi-armed bandit algorithm which makes use of the output of the online algorithm to minimize the congestion of the datacenter network. We demonstrate the effectiveness of our algorithms via solid theoretical analyses and trace-driven simulations.