Design, implementation and experimental evaluation of a network-slicing aware mobile protocol stack

  1. García Avilés, Ginés
Supervised by:
  1. Pablo Serrano Yáñez-Mingot Director
  2. Marco Gramaglia Co-director

Defence university: Universidad Carlos III de Madrid

Fecha de defensa: 11 June 2021

Committee:
  1. Francisco Valera Pintor Chair
  2. Vincenzo Sciancalepore Secretary
  3. Xenofon Foukas Committee member

Type: Thesis

Abstract

With the arrival of new generation mobile networks, we currently observe a paradigm shift, where monolithic network functions running on dedicated hardware are now implemented as software pieces that can be virtualized on general purpose hardware platforms. This paradigm shift stands on the softwarization of network functions and the adoption of virtualization techniques. Network Function Virtualization (NFV) comprises softwarization of network elements and virtualization of these components. It brings multiple advantages: (i) Flexibility, allowing an easy management of the virtual network functions (VNFs) (deploy, start, stop or update); (ii) efficiency, resources can be adequately consumed due to the increased flexibility of the network infrastructure; and (iii) reduced costs, due to the ability of sharing hardware resources. To this end, multiple challenges must be addressed to effectively leverage of all these benefits. Network Function Virtualization envisioned the concept of virtual network, resulting in a key enabler of 5G networks flexibility, Network Slicing. This new paradigm represents a new way to operate mobile networks where the underlying infrastructure is "sliced" into logically separated networks that can be customized to the specific needs of the tenant. This approach also enables the ability of instantiate VNFs at different locations of the infrastructure, choosing their optimal placement based on parameters such as the requirements of the service traversing the slice or the available resources. This decision process is called orchestration and involves all the VNFs withing the same network slice. The orchestrator is the entity in charge of managing network slices. Hands-on experiments on network slicing are essential to understand its benefits and limits, and to validate the design and deployment choices. While some network slicing prototypes have been built for Radio Access Networks (RANs), leveraging on the wide availability of radio hardware and open-source software, there is no currently open-source suite for end-to-end network slicing available to the research community. Similarly, orchestration mechanisms must be evaluated as well to properly validate theoretical solutions addressing diverse aspects such as resource assignment or service composition. This thesis contributes on the study of the mobile networks evolution regarding its softwarization and cloudification. We identify software patterns for network function virtualization, including the definition of a novel mobile architecture that squeezes the virtualization architecture by splitting functionality in atomic functions. Then, we effectively design, implement and evaluate of an open-source network slicing implementation. Our results show a per-slice customization without paying the price in terms of performance, also providing a slicing implementation to the research community. Moreover, we propose a framework to flexibly re-orchestrate a virtualized network, allowing on-the-fly re-orchestration without disrupting ongoing services. This framework can greatly improve performance under changing conditions. We evaluate the resulting performance in a realistic network slicing setup, showing the feasibility and advantages of flexible re-orchestration. Lastly and following the required re-design of network functions envisioned during the study of the evolution of mobile networks, we present a novel pipeline architecture specifically engineered for 4G/5G Physical Layers virtualized over clouds. The proposed design follows two objectives, resiliency upon unpredictable computing and parallelization to increase efficiency in multi-core clouds. To this end, we employ techniques such as tight deadline control, jitter-absorbing buffers, predictive Hybrid Automatic Repeat Request, and congestion control. Our experimental results show that our cloud-native approach attains >95\% of the theoretical spectrum efficiency in hostile environments where state-of-the-art architectures collapse.