Deterministic fog computing architecture for 5G applications in underserved communities

Abstract

The need to cope with the continuously growing number of connected users and the increased demand for mobile broadband services in the Internet of Things has led to the notion of introducing the fog computing paradigm in fifth generation (5G) mobile networks in the form of fog radio access network (F-RAN). The F-RAN approach emphasises bringing the computation capability to the edge of the network so as to reduce network bottlenecks and improve latency. In addition, the F-RAN method is feasible enough for deployments in underserved regions. However, despite the potential, the management of computational resources remains a challenge in F-RAN architectures. Thus, this dissertation aims to overcome the shortcomings of conventional approaches to computational resource allocation in F-RANs. This research first investigates applications of machine learning (ML) techniques in fog computing and 5G networks, as well as present approaches to the resource allocation problem. The potential of ML in future wireless networks is highlighted along with the limitations of current resource allocation methods. Consequently, two resource allocation techniques are presented as a solution-a reactive algorithm based on the auto-scaling method in cloud virtualisation and a proactive algorithm based on Q-learning in reinforcement learning (RL) -along with their respective architectures for implementation. The effectiveness of the proposed resource management techniques is demonstrated through simulation modelling. The proposed reactive auto-scaling algorithm yields favourable performance results in terms of latency and throughput, compared with other systems in literature. The proposed Q-learning algorithm is more efficient than popular RL algorithms in literature and outperforms the reactive method in terms of CPU utilisation and virtual link utilisation, which demonstrates the potential of ML in 5G F-RAN architectures for resource management