Learning-Based Dynamic Resource Provisioning for Network Slicing with Ensured End-to-End Performance Bound

Abstract

To accommodate different sets of network functions with different quality-of-service requirements for different types of applications in 5G networks, network slicing, which dynamically creates virtual networks, was proposed in the literature and IETF. A critical issue for network slicing is to determine the amount of resources for a network slice to ensure the quality-of-service requirement, and as such we need to determine the relationship among traffic demand, amount of resources, and end-to-end delay. This problem is non-trivial in a dynamic, virtualized environment. In this paper, we first use stochastic network calculus (SNC) to study the end-to-end delay bound with given traffic demand and resources. Then, we propose a solution to find the amount of resources that should be allocated with given traffic distribution and end-to-end delay bound. Beyond that, we investigate the range of traffic demands that a network slice can support and design a learning-based dynamic network slice resizing strategy, which can significantly reduce overall resizing cost with quality-of-service guarantee. Our work provides a set of useful tools for network slice tenants to (1) decide the amount of resources to request from physical network providers and (2) cost-effectively adjust the resource amounts that align with the dynamic traffic demand.