Elastic Function Chain Control for Edge Networks under Reconfiguration Delay and QoS Requirements

Abstract

Network Function Virtualization (NFV) allows network providers to reconfigure their edge processing infrastructure in an online fashion, to adapt it to the changing traffic demands (intensity and type of computation requests). In this work, we consider the Virtual Network Function Placement and Chaining (VNFPC) problem, whose aim is to elastically deploy services (i.e., chains of multiple Virtual Functions, VFs) through three phases: function placement, assignment, and chaining. For this problem, a predictive control framework is proposed to solve these three phases jointly by horizontally scaling VF instances, adapting their number to current and predicted demands, while ensuring that flows' Quality of Service (QoS) requirements (latency) are met. Our technique accounts for the delays and costs incurred in reconfiguration operations and uses a Gaussian Mixture Model, trained with real traces collected by base stations across the city of Milan (Italy), to estimate future computing demands. The proposed predictive control method is tested against a heuristic policy for several meaningful metrics, achieving up to 99% less edge control overhead and allowing a reduction of the blocking probability by 95% with respect to the heuristic (for the same energy consumption).