Power Measurements and Saving Techniques in Beyond 5G Networks

The rise of 5G systems and Beyond 5G (B5G), along with state-of-the-art cloud and software engineering solutions, have brought to a radical shift of networking technologies with respect to the previous generations. However, the increased performance and flexibility have come at the cost of reduced sustainability, mainly from an environmental standpoint. Infrastructure owners can introduce power management solutions in their datacenters, but the improvement is negligible unless tenants operating on top of the physical infrastructure are enabled and incentivized to customize their slices and their applications towards more power efficient and carbon neutral profiles. For these reasons, the thesis addresses the power efficiency issue. In further detail, herein, I present a flexible Observability, Monitoring and Analytics framework whose main targets revolve around measuring the power consumption of B5GNetwork Functions (NFs) and the application of algorithms (also involving Artificial Intelligence (AI)) to promote power efficiency. The crucial part of the aforementioned framework is the Management Data Analytics Function (MDAF) whose main task is to provide a power measurement of B5G NFs which takes into account the role virtualization plays (i.e., the sharing of physical resources). Tests are performed in a 5G network composed of opensource softwares and commercial products. Through a statistical analysis, Scaphandre, a well-known power monitoring tool in virtualized environments, is found to be the most appropriate tool of comparison for the MDAF. Results show that the MDAF is able to detect a higher power consumption than Scaphandre which could become significant in case of NFs deployed in multiple instances. Moreover, the difference in power is mainly ascribable to operations related to the container management orchestrator and the scheduler of the Operating System (OS). One of the two power saving techniques presented in this thesis concerns the exploitation of the well known power saving technique commonly used in general-purpose server: Low Power Idle (LPI), which consists in putting to (various degrees of) “sleep” Central Processing Units (CPUs) when not in use. Tests are performed on a 5G User Plane Function (UPF) and its power consumption and performance (latency) are evaluated. Results show that, in most cases the deeper the “sleep” states are enabled, the more power is saved. Furthermore, results show that this power saving does not come at a performance cost since no packet losses are detected in any test. Finally, concerning the Analytics part of the aforementioned framework, an Artificial Neural Network (ANN) -based algorithm is presented. Its main goal is to, via a multiclass classification problem, drive the NF scaling decisions. This algorithm uses a blend of both “black-box” Key Performance Indicators (KPIs) (e.g., CPU and memory utilization) and “white-box” KPIs (e.g., NF metrics, such as the number of Packet Data Unit (PDU) sessions, etc.). The model proposed herein is application- and implementation-independent since the white-box KPIs are extracted from 5G standardized interfaces. The proposed model is tested in a 5G environment and compared with a standard threshold-based one. Results show a 97∼98% accuracy in the training and validation phases and an overall accuracy of more than 97% on never seen before samples and, thus, a decrease in computational resource usage.