Doctorate thesis defense of Somlawa Mihia KASSI

Doctorate thesis defense on July 09^th 2024 at 10H00 AM ,in Amphitheater Al-Khawarizmi, SUP'COM 2.

Entitled :ALarge-Scale Virtual Radio Access Networks (V-RAN): Implementation, Resource Optimization and Machine Learning-based Orchestration

Presented by :Somlawa Mihia KASSI

Committee

Abstract

This thesis examined the current state of Radio Access Network (RAN) virtualization and explored the challenges hindering its large-scale implementation in 5G networks and beyond. We began by providing a survey presenting the evolution of the virtualization in radio access networks from Cloud Radio Access Network (C-RAN) to Virtual-RAN up to Open RAN, as well as the various research efforts on V-RAN in the recent years. This survey also detailed the different platforms and tools dedicated to virtualization: Eurecom's OpenAirInterface (OAI), srsLTE, free5GRAN, etc. It also showed the latest V-RAN implementations worldwide, both in research environments and among mobile network operators. Through simulations, we pointed out highlight the limitations and issues preventing the realization of large-scale V-RANs. In view of this assessment, we proposed some solutions, with the main one being the optimization of the OAI protocol stack platform (LTE/5G) of the Eurecom research center of Sophia Antipolis (France). This optimization led to a modular platform with less complex installation and evolution processes, possessing dissociable network functions. We named this platform enhanced-OAI (e-OAI). By using e-OAI as the 5G protocol stack platform, we developed a large-scale V-RAN. This large-scale V-RAN consists of large-scale virtual Base Band Unit pools (vBBU Pools), Remote Radio Heads (RRHs), an Ethernet-based fronthaul network, and a wide range of users. We then added to this V-RAN an orchestrator based on docker swarm technology, providing entity management optimization, supervision, load balancing, and scalability. The results of simulations made with 40 V-RAN entities (vBBUs and RRHs) mainly demonstrated our system's ability to increase the number of entities in the RAN while reducing their overall resource consumption especially CPU (Central Processing Unit) and Memory consumption. Moreover, we integrated network slicing to our e-OAI platform. First, we divided the V-RAN into three different network slices: massive Machine Type Communications (mMTC), enhanced Mobile Broadband (eMBB), Ultra-Reliable and Low Latency Communications (uRLLC). Then, we developed a slice and underlying resource orchestration system composed of Machine Learning (ML) algorithms: Deep Neural Network (DNN) and Gradient Boosting Regressor (GBR). We subsequently proved the efficiency and applicability of these algorithms by testing them on real data from a Tunisian mobile network operator. We further optimized our ML-based orchestration system by integrating a semi-static and adaptive switching scheme. Simulation results proved that our ML orchestration system, coupled with the semi-static and adaptive switching scheme, achieved resource consumption savings (CPU, Memory) from 7 to 12%. In terms of CAPEX (Capital Expenditure) and OPEX (Operation Expenditure), we obtained a significant reduction as well. Finally, we developed an end-to-end virtualized 5G network. It consisted of the previous implemented large-scale V-RAN, a set of connected user equipments, a wireless backhaul network, and a fully virtualized 5G core network. At the heart of this virtual end-to-end 5G network, we also implemented a SDN controller to optimize network administration and resource allocation. Simulation results proved the effectiveness of the system and the correlation between end-to-end intensity traffic and resource consumption. They also demonstrated our network's ability to support high levels of traffic compared to other research efforts.

Keywords

B5G, C-RAN, E2E virtual network, Large-scale virtual BBU Pool, Large-scale V-RAN, Machine Learning (ML), Network slicing, OpenAirInterface, RAN virtualization, RRH, Software Defined Networking (SDN).