Reinforcement learning based dynamic resource migration for virtual networks

Abstract

Network virtualization techniques enable network operators to implement and provide multiple virtual networks (VNs) on a common substrate network infrastructure. The network resources should be able to be dynamically reallocated among VNs or migrated from a place to another one in various situations, such as to construct a new urgent VN in case of occurrence of some urgent contingency, or to satisfy quality of service (QoS) requirements of non-urgent VNs in case of time-varying network traffic conditions as much as possible. In this paper, we propose methods to automatically and dynamically select and migrate resources of non-urgent VNs. In particular, our proposal applies reinforcement learning for the selection of resources from alternate places to satisfy the QoS requirements. We evaluate the performances in terms of the satisfaction level of QoS requirement with various frequency of network traffic variation in a given duration and learning parameter configurations. Simulation results show that the proposed dynamic resource migration method can increase the number of times that non-urgent VNs' QoS requirements are satisfied in comparison with a static resource assignment method. Moreover, we show that, depending on traffic variation and parameter configuration, applying reinforcement learning can increase the number of times the QoS requirement is satisfied compared to the dynamic method with completely random resource selection.