DAEMON Deliverable 4.2: Refined design of intelligent orchestration and management mechanisms

This deliverable reports the progress made by DAEMON on the design and development of NI-assisted functionalities for Beyond fifth Generation (B5G) mobile networks, with a focus on the area of service and resource management and orchestration. To achieve this overarching goal, we study four key B5G functionalities: (i) energy-aware Virtual Network Function (VNF) orchestration, (ii) capacity forecasting, (iii) automated anomaly detection, and (iv) self-learning management and orchestration. D4.2 builds on the results of D4.1 and presents the latest developments of DAEMON on the identified research threads for each of the above functionalities. In detail, Section 1 introduces the main results of this deliverable and explains their connection to the previous deliverable (D4.1) as well as to the other work packages. The rest of the sections focus on each of the building blocks, and for each solution it is reported which specific requirements from D2.2 are satisfied, how it relates to D2.2 guidelines, which standards are related to it, and how it can be encoded using the MAPE-K representation. This systematic presentation approach allows the reader to understand and compare the proposed solutions in a unified way.
Namely, Section 2 presents the results related to the proposed solutions for Energy-aware VNF orchestration. These solutions are designed to work at different levels of a virtualized next-generation mobile network and specifically are related to (i) placement and autoscaling of virtualized NFs at user/data level (Section 2.1and Section 2.2) and (ii) virtualized NFs at the RAN (Section 2.3 and Section 2.4). The solutions leverage a diverse set of algorithms and AI mechanisms, ranging from Deep Learning to prediction-assisted VNF placement and scaling and to Bayesian learning techniques for the design of meta policies in O-RAN systems. The targeted optimization metrics go beyond energy savings all the way to throughput maximization and service-specific KPIs such as inference accuracy in edge analytic services. Therefore, the proposed solutions not only reduce the energy costs, but are able to trade off energy consumption with performance. The results of this section have been published in scientific conferences and journals and have been verified through full-scale implementation in testbeds and/or data-driven evaluations that are reported in the deliverables of WP5. Section 3 continuous with the solutions related to capacity forecasting methods. It presents the results of anticipatory networking methods where resources are reserved or allocated before the actual demand (or traffic) is realized. In this context, D4.2 presents four solutions: (i) anticipatory allocation of resources, (ii) Virtual Machine (VM) reservation in a network core datacenter, (iii) minimization of video streaming slice OPEX, (iv) and prediction-assisted network resource management. The solutions that are employed to achieve these tasks include time-series forecasting models that merge classical exponential smoothing techniques with Recurrent Neural Networks, meta-learning techniques that identify the most suitable loss function using the problem’s data, and hybrid learning solutions that combine offline-trained predictors and online optimization techniques towards fast and robust resource allocation decisions. The results of this section address several functionality requirements from D2.2 and go well-beyond the state-of-the-art as they reveal and address several limitations of classical AI/ML techniques, such as the slow convergence of online learning algorithms or the suboptimal performance of predefined loss functions. Section 4 presents the progress of DAEMON on the topic of anomaly detection. We present four solutions that tackle anomalies in different parts of the cellular ecosystem, from the RAN and base stations to the interconnection between mobile operators across different countries. In detail, the first solution employs federated learning in order to perform clustering using, without sharing, local data across different agents that are deployed into different network locations. The second proposal integrates advanced deep learning models for anomaly detection in IoT devices for different verticals, using only control plane information. Following that, D4.2 reports results on anticipatory anomaly detection, where an alarm is triggered whenever an abnormal traffic load is expected for a specific Network Slice. The fourth and final solution explores the problem of noisy neighbor in computing environments implementing VNFs of virtualized base stations, and presents experimental evidence of this issue and of its effects on network performance. The assessment of these methods is conducted using a range of datasets and experiments, hence allowing their realistic evaluation. These results are reported in the respective work packages. Finally, Section 5 presents self-learning MANagement and Orchestration (MANO) solutions which are crucial in B5G networks given the range and complexity of the decisions involved in their operation. The first group of solutions studies centralized and distributed Reinforcement Learning (RL) techniques for VNF placements across different network locations toward supporting a newly-deployed network service. This builds upon and extends the respective results reported in D4.1 The second group of solutions considers the problem of automated scaling of allocated network resources in order to meet the time-varying demands of users. To that end, we use two different techniques, namely a control-theoretic solution and an RL-based solution which track the number of consumed resources and increase or decrease the VNF instances accordingly. The evaluation and comparison results are presented in D5.2.

The deliverable concludes in Section 6 by discussing the proposed solutions through the lens of WP2. This way, we provide feedback to WP2 on the applicability of the architecture and in particular we identify two cases where the proposed solutions need to be aligned in order to maximizer the accrued benefits. We identify two such cases, namely the energy-aware VNF placement and vRAN orchestration, and the capacity forecasting and anomaly detection, and we use a detailed MAPE-K representation in order to align them and exploit their complementarity.