Resource management in future mobile networks: from millimetre-wave backhauls to airborne access networks

Abstract

The next generation of mobile networks will connect vast numbers of devices and support services with diverse requirements. Enabling technologies such as millimetre-wave (mm-wave) backhauling and network slicing allow for increased wireless capacities and logical partitioning of physical deployments, yet introduce a number of challenges. These include among others the precise and rapid allocation of network resources among applications, elucidating the interactions between new mobile networking technology and widely used protocols, and the agile control of mobile infrastructure, to provide users with reliable wireless connectivity in extreme scenarios. This thesis presents several original contributions that address these challenges. In particular, I will first describe the design and evaluation of an airtime allocation and scheduling mechanism devised specifically for mm-wave backhauls, explicitly addressing inter-flow fairness and capturing the unique characteristics of mm-wave communications. Simulation results will demonstrate 5x throughput gains and a 5-fold improvement in fairness over recent mm-wave scheduling solutions. Second, I will introduce a utility optimisation framework targeting virtually sliced mm-wave backhauls that are shared by a number of applications with distinct requirements. Based on this framework, I will present a deep learning solution that can be trained within minutes, following which it computes rate allocations that match those obtained with state-of-the-art global optimisation algorithms. The proposed solution outperforms a baseline greedy approach by up to 62%, in terms of network utility, while running orders of magnitude faster. Third, the thesis investigates the behaviour of the Transport Control Protocol (TCP) in Long-Term Evolution (LTE) networks and discusses the implications of employing Radio Link Control (RLC) acknowledgements under different link qualities, on the performance of transport protocols. Fourth, I will introduce a reinforcement learning approach to optimising the performance of airborne cellular networks serving users in emergency settings, demonstrating rapid convergence (approx. 2.5 hours on a desktop machine) and a 5dB improvement of the median Signal-to-Noise-plus-Interference-Ratio (SINR) perceived by users, over a heuristic based benchmark solution. Finally, the thesis discusses promising future research directions that follow from the results obtained throughout this PhD project.