Automation mechanisms for improved energy efficiency in the 5G core network

Energy efficiency has quickly become one of the main priorities in the telecommunications sector, with many mobile network operators setting the goal to become net-zero from carbon emissions by the year 2050. The path towards greener mobile networks is only in its infancy and will require significant efforts, in particular towards enhanced observability mechanisms and energy-aware autonomous orchestration procedures. This dissertation aims to advance the state-of-the-art in green 5G and future 6G mobile network design and operation through a combination of empirical measurements, experimental deployments, and machine learning modeling. First, we review the standardized deployment models for private 5G networks, which serve as our primary use case, as well as the main challenges concerning their autonomous management and orchestration. We then focus on studying the enabling mechanisms allowing network automation to be used as an energy-saving tool. To this end, we design and implement a methodology and an experimental testbed which we use to conduct a data collection campaign measuring the power consumption of different 5G core network deployments. Analyzing the collected data, we discuss the advantages and disadvantages of alternative virtualization technologies and compare software and hardware-based power metering solutions. The collected metrics are also used to create a digital twin model of an edge host capable of predicting its behavior and performance under different workloads. We test different machine learning and interpolation techniques, training using complete and incomplete datasets and comparing the accuracies. Finally, we demonstrate one possible integration between the power consumption monitoring and an autonomous network orchestration task. We chose the task of user plane function selection in a distributed 5G deployment over an edge-cloud environment, which is a common problem in private mobile network deployments. We developed a simulated environment using the previously derived performance and power computation models. In this environment, we trained a deep reinforcement learning algorithm to minimize the overall energy consumption while satisfying the users' latency and bandwidth requirements.