Design and Implementation of a Dependable Backhaul Solution based on the Multi-Armed Bandit problem

Sammanfattning

5th Generation (5G) Networks were designed to satisfy our modern requirements for connectivity. Because of them, in the upcoming years millions of new devices will get connected. Not only that, but 5G networks will help us build Smart Cities, Smart Factories, allow us to immerse ourselves in the digital world through Virtual Reality (VR) and much more. With the push towards a modular and open ecosystem, enterprises can now run a 5G network on premises. At this moment, radio access technology and core network technology have advanced well enough to provide the low delays and high speeds. However, the backhaul connections still face challenges, which could undermine the overall performance. This is because high-performing backhauls can not always be deployed, due to financial or technology-specific reasons. Therefore, the utilization of multiple low-cost backhauls has the potential to solve this issue. Multipath TCP (MPTCP) is the most popular protocol for multipath communication. However, due to its poor performance in heterogeneous environments, another solution is needed. This degree project investigates the utilization of multiple cost-efficient backhaul connections to provide the needed performance and resilience. To efficiently exploit these connections an approach is proposed, which considers the live performance characteristics of the links. Furthermore, to fairly distribute the traffic among the paths a path scheduler is developed based on the Multi-Armed Bandit (MAB) problem. The proposed solution was evaluated in a real heterogeneous environment over a Low Earth Orbit (LEO) satellite link and an optical fiber link, where our path scheduler outperformed Multipath TCP (MPTCP). Furthermore, the solution was tested in a simulated scenario with homogeneous links, where it doubled the overall throughput. The experimental analysis yields that the Sliding-Window Upper Confidence Bound (UCB) algorithm is a promising solution for the existing issues in multi-path scheduling.