RL-based Virtual Network Embedding using VNF Sharing for Network Slicing in 5G Networks

Abstract

This paper deals with network slicing in 5G networks, where a slice is defined as a set of virtual network function (VNF) instances that collaborate to create an end-toend (E2E) virtual network. A set of slices is implemented on a physical substrate network maintained by an infrastructure provider. The virtual network embedding (VNE) problem deals with the deployment of a network slice’s virtual network request on the substrate. Typically, the resources allocated per slice’s request are not shared with other slices due to privacy, security and performance considerations. However, there are situations in which VNF instances might be aggregated across many slices to further increase the utilization ratio of the substrate infrastructure. Given these shareable VNF nodes, deploying the network slices is effectively the embedding of the numerous virtual network where these slices are linked by the shared VNFs. This paper uses a reinforcement learning (RL) approach for the embedding problem. The approach incorporates sharing based virtual network functions in an existing RL scheme designed for virtual node embedding without much additional computation. The proposed scheme is implemented using a policy based RL method; the performance study shows an increase in the reward ratio by up to 20% compared to the non-sharing case, along with an increase in the acceptance percentage of slices.