Fifth Generation (5G) networks and beyond are envisioned to provide user-focused communications, supporting diverse services with enhanced Quality of Service (QoS). Pivotal to this evolution is user equipment, which is increasingly performing advanced computational tasks beyond the edge of the network, known as the Extreme-Edge. Seamless integration of Extreme-Edge devices (EEDs) into the 5 G framework is however hindered, due to challenges in terms of device management, resource restrictions and interoperability issues. To address these barriers, we realize the Extreme-Edge Orchestrator (EEO), a management and orchestration framework enabling the extension of the 5 G cloud-to-edge continuum towards the Extreme-Edge. The EEO enables real-time resource monitoring and lifecycle management of network applications, including Artificial Intelligence/Machine Learning (AI/ML) tasks, deployed on EEDs. Unlike existing theoretical studies, our solution is deployed on an operational research-center-wide 5 G testbed and evaluated using an AI/ML-based network QoS prediction application, in the automotive domain. Our results show that the EEO supports efficient resource utilization, dynamic EED selection under device mobility scenarios and maintains robust service performance under computational stress-demonstrating its capability to support next-generation network services.

The advent of 5G and beyond systems is expected to shape the automotive vertical, as safety-critical vehicular applications rely on the network to meet their stringent Quality of Service (QoS) requirements. Predictive QoS (pQoS) has been proposed as a mechanism that allows automotive applications to proactively adapt in view of forthcoming QoS changes. Although pQoS is typically facilitated via classical (centralized) Machine Learning (ML) methods, the demand for data privacy has led to the emergence of distributed ML schemes. Efficient training of ML models however requires large volumes of (kinematic-state and connectivity) QoS data, so as to capture the involved spatio-temporal effects.To that end we hereby present and publicly share NordicDat, a QoS dataset collected during a two-week measurement campaign, driving across three European countries. NordicDat contains over 90K samples of physical layer, network and mobility-related features. Contrary to prior works, it includes multiple instances of cross-boarder roaming, diverse vehicle speed profiles and radio access technologies (generations). Further, we provide a thorough NordicDat data analysis, highlighting the dependencies between the NordicDat’s features and the resulting QoS values (throughput, delay). To showcase its broad usability, we train pQoS ML models over NordicDat in classical and distributed fashion. Our results demonstrate for the first time the viability of distributed pQoS with real-word data, which achieves similar (within a margin of 10%) accuracy to that of classical ML, cropping privacy-preserving benefits.

In this paper we propose a DC network (DCN) architecture that interconnects servers in the intra-rack and inter-rack domain, utilizing optical switching at each domain. The proposed interconnection techniques are studied as an intermediate step before migrating the entire DCN to all-optical schemes. Unlike other studies, we study the server-to-server communication across the whole DCN. For the performance evaluation we produce numerical results for throughput and end-to-end delay for three traffic classes co-existing in DCN s. The numerical analysis reveals that bandwidth utilization reaches 90% and 100% in the intra- and inter- domain respectively. Meanwhile, the maximum end-to-end delay for the highest priority packets under congested load is lower than 0.56 and 0.41 µs for the two examined intra-rack capacity scenarios of 400 and 600 Gbps respectively. A comparative study shows that our solution can effectively interconnect up to 10000 servers with lower environmental footprint and end-to-end delay than other DCN s.

The convergence of Machine Learning (ML) with the edge computing paradigm has paved the way for distributing processing-heavy ML tasks to the network's extremes. As the edge deployment details still remain an open issue, distributed ML schemes tend to be network-agnostic; thus, their effect on the underlying network's resource consumption is largely ignored.In our work, assuming a network tree structure of varying size and edge computing characteristics, we introduce an analytical system model based on credible real-world measurements to capture the end-to-end consumption of ML schemes. In this context, we employ an edge-based (EL) and a federated (FL) ML scheme and in-depth compare their bandwidth needs and energy footprint against a cloud-based (CL) baseline approach. Our numerical evaluation suggests that EL exhibits a minimum of 25% bandwidth-efficiency compared to CL and FL, if employed by a few nodes higher in the edge network, while halving the network's energy costs.

The proliferation of machine learning (ML) applications has lately witnessed a considerable shift to more distributed settings, even reaching hand-held mobile devices; there, contrary to typical Centralized learning (CL) whereby the involved (large amounts of) training data are centrally gathered to train models, the load of training tasks is distributed across a set of capable mobile learners at the expense of their own energy. The idea of Federated learning (FL) has emerged as a privacy-preserving mechanism suggesting that the ML model parameters rather than data, are sent over the network to a central point of aggregation. However, when relaxing the privacy concerns, the debate strongly relates to the available network resources. Interestingly, the sofar theoretical or even experimental comparison of the two approaches overlooks network conditions and remains of low realism. In this work we rely on past measurement studies to introduce a realistic system model that accounts for all involved mobile network conditions such as bandwidth and data availability (af-fecting training accuracy and model aggregation) as well as user mobility patterns (affecting data loss). A dedicated simulation framework we have developed replays rich mobile-traces allowing for a comprehensive comparison of the two ML approaches over a large set of training data shedding light on network-resources utilization, energy efficiency and training convergence. Intuitively, our results suggest that the ratio between the employed raw data and the corresponding ML model shapes the conditions under which FL acts as a network-efficient alternative to CL. Interestingly enough, asymmetry in data availability across users as well as their varying number are shown to hardly affect the FL approach in traffic and energy needs, pointing both to its promising potential and the need for further research.