Network Function Virtualization (NFV) is one of the main enablers behind the promised improvements in the Fifth Generation (5G) networking era. Thanks to this concept, Network Functions (NFs) are evolving into software components (e.g., Virtual Network Functions (VNFs)) that can be deployed in general-purpose servers following a cloud-based approach. In this way, NFs can be deployed at scale, fulfilling a great variety of service requirements. Unfortunately, the complexity in the management and orchestration of NFV-based networks has increased due to the diverse demands from a growing number of network services. Such complexity calls for an automated and autonomous solution that self adapts to the needs of those network services. In this paper, we propose and compare a Deep Reinforcement Learning (DRL) agent, a classical Proportional–Integral–Derivative (PID) controller, and a Threshold (THD)-based algorithm for autonomously determining the amount of VNF instances to fulfill a service latency requirement without knowing or predicting the expected demand. Finally, we present a comparison of the three approaches in terms of created VNFs and peak latency performed in a discrete event simulator.

This research was partially funded by Ctrl App, a research project of the Fund for Scientific Research Flanders (FWO) under grant agreement No. G055619N, and by the European Union's Horizon 2020 research and innovation programme under Grant Agreement No. 101017109 (DAEMON).