Virtual Radios, Real Services: Enabling End-to-End Network Slicing through Radio Virtualisation

Abstract

This work aims to resolve some resource management problems that arise due to changes in the mobile network market and business models, brought by recent trends in commercial mobile networks, such as the widening range of services offered by Mobile Network Operators (MNOs), and the increasing importance of vertical industries and service providers as sources of revenue. These changes led the 3GPP to introduce network slicing, proposing a logical division of the mobile network infrastructure into independent virtual networks, known as Network Slices (NSs). Each NS acts as a completely isolated End-to-End (E2E) network, tailored with different resources and functionality to support applications with specific Quality of Service (QoS) requirements. Similarly to the separation of mobile networks into Core Network (CN) and Radio Access Network (RAN) segments, the offer of NSs can be split into CN as a Service (CNaaS) and RAN as a Service (RANaaS), allowing MNOs to provide their clients with custom CN slices, custom RAN slices, or a combination of both, forming fully custom E2E NSs. In particular, this thesis studies the resource management functionality for enabling E2E network slicing through radio virtualisation. The problems we address in this thesis can be summarised in the form of the following research questions: - What are the requirements for using radio hypervisors to support RANaaS? - How can radio hypervisors provision and instantiate custom virtual radios on demand? - How to model the embedding of heterogeneous RAN slices on RANaaS platforms? - Could separate specialised orchestrators be used to provide fine-grained resource allocation on E2E networks? - How to coordinate multiple network segments and orchestrators to guarantee a consistent QoS for E2E NSs? To addresses these questions, we apply a range of experimental approaches, analytical methods, and new architectural proposals, including the qualitative review of the vast literature on radio virtualisation, radio slicing and RANaaS; the implementation of a software layer that provides radio hypervisors with resource management features for supporting RANaaS; the discussion and formalisation of the radio processing isolation problem; the modelling of the resource allocation of heterogeneous RAN slices on RANaaS platforms; and the proposal of a hierarchical orchestration architecture for E2E networks as well as its implementation as a proof-of-concept prototype. First, we investigate the functionality and challenges for enabling RANaaS through radio virtualisation. We identify the key requirements for using radio hypervisors to support RANaaS, evaluate how radio hypervisors in the literature meet these requirements, and observe that most of the existing radio hypervisors lack some of the essential features for RANaaS. Then, we introduce a modular software layer that can sit on top of existing radio hypervisors and provide them with the missing slicing functionality to support RANaaS. We integrate our software layer with a radio hypervisor and validate its ability to create and deploy virtual radios with different Radio Access Technologies (RATs) or numerologies on-demand, capable of displaying performance comparable to bare metal. Next, we focus on the resource management of RANaaS platforms regarding the mapping of real radio resources to realise virtual radios, also known as the virtual wireless network embedding problem. We observe that current models and solutions for the embedding of RAN slices are tied to particular radio virtualisation mechanisms, not being suitable for virtual radios with different RATs or numerologies. Then, we introduce a technology-agnostic modelling approach for embedding heterogeneous RAN slices, which is transparent to the characteristics of the underlying virtual radios. We also propose a resource management optimisation problem that is solved at the MNO, for determining the optimal allocation of RAN slices to minimise the total isolation overhead. Finally, we look into the coordination of the resource management across multiple network segments to deploy E2E NSs with consistent QoS. We propose a hierarchical orchestration architecture that addresses the limited support and oversimplified resource allocation on different network segments of existing E2E orchestration solutions, by enabling the independent management of each network segment using existing distributed specialised orchestrators used to manage a real network infrastructure. Then, we present a prototype higher-level orchestrator, the hyperstrator, for coordinating the resource management of the existing specialised orchestrators and the deployment of E2E NSs across network segments. We also verify that the distributed nature of our orchestration architecture introduces a negligible overhead for instantiating E2E NSs.