Roch Glitho

Known as one of the key enablers of 5G networks and beyond, network function virtualization provides flexibility and scalability by enabling the deployment of network services in the form of chained Virtual Network Functions (VNF) known as VNF Forwarding Graphs (VNF-FGs). However, the already deployed and running VNF-FGs could become inefficient as the services' demand increases and network conditions fluctuate. Thus, not only does the number of VNF instances need to be adjusted (i.e., function scaling), but also the topology of the VNF-FG and its mapping to the physical network may need proper adjustments via so-called topology adaptation. Since function scaling and topology adaptation may have a mutual correlation, a disjoint approach could lead to inefficient results. In this paper, we propose a deep reinforcement learning-based joint function scaling and topology adaptation framework that leverages a deep double Q-network architecture enhanced with action filtering to update the function and topology of VNF-FGs already deployed. Our evaluation results show that the proposed framework improves up to a 28% success rate compared to our benchmarks while incurring a low modification cost and maintaining competitive inference time among learning-based approaches.

Network function virtualization is a key enabler for 5G and beyond, allowing network functions to run on general-purpose hardware for flexibility and resource efficiency. User plane network functions (UP-NFs) in the Radio Access Network (Distributed and Centralized Units) and the core (User Plane Function) are deployed across heterogeneous platforms, including virtual machines, containers, and bare metal nodes. These functions are typically implemented as virtual network functions (VNFs) and interconnected as VNF forwarding graphs (VNF-FGs), capturing interdependencies and traffic flows. As 5G infrastructures scale, energy efficiency becomes a critical challenge, especially when embedding VNF-FGs onto platforms with varying overhead, performance, and power consumption. Inefficient embedding leads to excessive energy use and degraded service quality. This paper addresses the green embedding of UP-NF-based VNF-FGs to minimize power consumption while meeting strict latency and bandwidth requirements. We propose a novel graph attention network-enhanced deep reinforcement learning (GAT-DRL) approach. The GAT component assigns dynamic attention weights, effectively modeling the relationships within the VNF-FG and the substrate network, enabling the DRL agent to make context-aware, energy-efficient embedding decisions. Simulation results show that our approach reduces power consumption by up to 38% and improves acceptance rates by up to 49%, demonstrating strong potential for scalable, energy-efficient 5G deployments.

This paper addresses the challenge of delivering low-latency, scalable immersive experiences by exploiting a hybrid continuum of cloud, edge, and In-Network Computing (INC) resources. Indeed, delivering low-latency, scalable immersive experiences requires the transfer of a large amount of digital assets of different sizes, many of them consisting of large, static scene elements corresponding to service-specific and user-specific components. We argue in this paper that such elements could be separated within an in-network rendering farm while dynamically caching popular assets and synchronizing rapidly changing, user-centric data at INC, Edge or Cloud nodes. Still all theses need to be orchestrated efficiently. To efficiently orchestrate these heterogeneous resources, we formulate in this paper a multi-objective optimization problem—maximizing resource efficiency, minimizing end-to-end latency, and maximizing user request acceptance. This optimization problem is then solved via a deep reinforcement learning (DRL) framework that adaptively assigns functions across all layers in real time. The purpose of our proposed popularity-based replication and pre-caching is to further reduce latency for the most frequently accessed assets, while we offload lightweight rendering operations directly onto programmable switches to cut down on round-trip delays. Extensive simulations, benchmarked against multiple baselines, demonstrate that our approach consistently maintains sub-20ms end-to-end delays and achieves superior resource utilization efficiency under dynamic workloads. These results validate the potential of integrating INC into the Compute Continuum and use a DRL-driven orchestration, both together allowing to meet the stringent Quality of Service (QoS) and Quality of Experience (QoE) requirements of next-generation immersive applications.

The growing demand for computational power in cloud computing has made Graphics Processing Units (GPUs) essential for providing substantial computational capacity. Efficiently allocating GPU resources is crucial due to their high cost. Additionally, it’s necessary to considercloud environment characteristics, such as dynamic workloads, multi-tenancy, and requirements like isolation. One key challenge is efficiently allocating GPU resources while maintaining isolation and adapting to dynamic workload fluctuations. Another challenge is ensuring scheduling maintains fairness between tenants while meeting task requirements (e.g., completion deadlines). While existing approaches have addressed each challenge individually, none have tackled both challenges simultaneously. This is especially important in dynamic environments where applications continuously request and release GPU resources. This paper introduces a new dynamic GPU resource allocation method, incorporating fair and requirement-aware task scheduling. We present a novel algorithm that leverages the multitasking capabilities of GPUs supported by both hardware and software. The algorithm schedules tasks and continuously reassesses resource allocation as new tasks arrive to ensure fairness. Simultaneously, it adjusts allocations to maintain isolation and satisfy task requirements. Experimental results indicate that our proposed algorithm offers several advantages over existing state-of-the-art solutions. It reduces GPU resource usage by 88% and significantly decreases task completion times.

Logs are prevalent in modern cloud systems and serve as a valuable source of information for system maintenance. Over the years, many supervised, semi-supervised, and unsupervised log analysis methods have been proposed to detect system anomalies. In particular, semi-supervised methods have garnered increasing attention as they balance reduced labeled data requirements and optimal detection performance, contrasting with their supervised and unsupervised counterparts. However, existing semi-supervised log analysis methods often suffer from practical challenges, such as log instability, imbalanced class data, and labeling dependency, which are pervasive issues in real-world systems. To address these challenges, we propose ALogSCAN, a self-supervised method to detect anomalies at the host level of cloud systems. ALogSCAN introduces the Dynamic Frequency-based Log Filtering (DFLF) technique to mitigate the potential bias introduced by highly frequent log messages, thereby focusing more on infrequent yet critical log messages. Moreover, the self-supervised nature of ALogSCAN eliminates the need for time-consuming manual labeling of log data, and enables the DFLF technique to continuously adapt to evolving log sequences, maintaining robustness against unstable log data. We have evaluated ALogSCAN on two widely used public datasets and one private dataset from Ericsson Research, and the experimental results demonstrate its effectiveness, consistently outperforming existing methods in various scenarios.

In-network machine learning (ML) offers a cutting-edge approach for promptly detecting malicious traffic. Existing methods often rely on one-size-fits-all ML models that fail to adapt to evolving attack traffic patterns, leading to a time-consuming and labor-intensive process for updating ML model from the control to the data plane. To address these limitations, we propose an automated, data-driven method for identifying novel malicious traffic patterns and updating ML model seamlessly in programmable networks. The proposed method sets drift detection thresholds based on baseline performance from historical (i.e., training) data and uses these thresholds to detect anomalies in unseen (i.e., testing) data. We continuously adjust the thresholds to accommodate data distribution changes and in-network inference results while minimizing sensitivity to minor fluctuations. We evaluate the proposed method using two intrusion detection datasets, CICIDS2017 and UNSW-NB15. The experimental results demonstrate its efficacy in safeguarding against evolving volumetric attacks. Additionally, we compare the conventional model performance-based drift detection method with an adaptive monitoring window-based approach, highlighting the latter’s advantage in balancing drift detection efficacy and minimizing its adaptation impact, i.e., disruptions to normal network traffic are reduced by an average of 20%. The adaptive method dynamically adjusts the drift monitoring window size to adapt to the characteristics of the unseen traffic patterns.

Applications deployed in clouds are susceptible to performance degradation due to diverse underlying causes such as infrastructure faults. To maintain the expected availability of these applications, Machine Learning (ML) models can be used to predict the impending application performance degradations to take preventive measures. However, the prediction accuracy of these ML models, which is a key indicator of their performance, is influenced by several factors, including training data size, data sampling intervals, input window and prediction horizon. To optimize these data-related parameters, in this paper, we propose a surrogate-assisted multi-objective optimization algorithm with the objective to maximize prediction model accuracy while minimizing the resources consumed for data collection and storage. We evaluated the proposed algorithm through two use cases focusing on the prediction of Key Performance Indicators (KPIs) for a 5 G core network and a web application deployed in two Kubernetes-based cloud testbeds. It is demonstrated that the proposed algorithm can achieve a normalized hypervolume of 99.5% relative to the optimal Pareto front and reduce search time for the optimal solution by 0.6 hours compared to other surrogates and by 3.58 hours compared to using no surrogates.

Providing quick and reliable emergency communication in situations of natural disasters or unforeseen incidents may be crucial. In such situations, traditional communication infrastructure, such as ground-based wireless base stations, may become temporarily damaged or unavailable to support emergency teleoperations. Considered a promising solution, unmanned aerial vehicles (UAVs) can be deployed as flying base stations or relays to provide fast and reliable communication between physicians and remote robots in both uplink and downlink directions, while meeting strict transmission requirements. This paper addresses the joint optimization problem of UAV positioning and resource allocation in UAV-assisted wireless networks to minimize the number of deployed UAVs, all while satisfying stringent transmission quality demands. The formulated problem is a non-convex mixed-integer programming problem, which we prove to be NP-hard. We first develop efficient greedy and metaheuristic genetic algorithms. Then, we propose an efficient centralized deep reinforcement learning solution based on the deep deterministic policy gradient (DDPG), where the agent learns optimal UAV positions and resource allocation. Simulation results demonstrate that the greedy solution closely matches the performance of both the genetic and deep reinforcement learning approaches, with a significant reduction in computational complexity. Furthermore, the results highlight the effectiveness of the deep reinforcement learning solution in minimizing the number of UAVs required to fully satisfy the transmission requirements of all users in both uplink and downlink directions.

The advent of 5G networks has facilitated various Industry 4.0 applications requiring stringent Quality-of-Service (QoS) demands, notably Ultra-Reliable Low-Latency Communication (URLLC). Multi-Access Edge Computing (MEC) has emerged as a key technology to support these URLLC applications by bringing computational resources closer to the user, thus reducing latency. Meanwhile, Network Function Virtualization (NFV) supports 5G networks by offering flexibility and scalability in service provisioning across various applications. Despite their benefits, MEC networks must adapt to dynamically fluctuating user demands and varying workloads, which can create challenges in maintaining QoS. This paper addresses the Virtual Network Function (VNF) placement problem in MEC networks, focusing on minimizing costs while ensuring QoS through VNF reuse. We propose a novel solution based on the Deep Transformer Q-network (DTQN) algorithm, leveraging reinforcement learning to optimize VNF placement and redeployment. Extensive simulations demonstrate that our DTQN algorithm outperforms baseline approaches, achieving up to a 9% improvement over the D3T-based method and up to 56% over the DQN-based method in terms of average rewards under specific scenarios. This results in significant improvements in cost efficiency, resource utilization, and QoS maintenance.

The complexity and dynamicity of microservice architectures in cloud environments present substantial challenges to the reliability and availability of the services built on these architectures. Therefore, effective anomaly detection is crucial to prevent impending failures and resolve them promptly. Distributed data analysis techniques based on machine learning (ML) have recently gained attention in detecting anomalies in microservice systems. ML-based anomaly detection techniques mostly require centralized data collection and processing, which may raise scalability and computational issues in practice. In this paper, we propose an Asynchronous Real-Time Federated Learning (ART-FL) approach for anomaly detection in cloud-based microservice systems. In our approach, edge clients perform real-time learning with continuous streaming local data. At the edge clients, we model intra-service behaviors and inter-service dependencies in multi-source distributed data based on a Span Causal Graph (SCG) representation and train a model through a combination of Graph Neural Network (GNN) and Positive and Unlabeled (PU) learning. Our FL approach updates the global model in an asynchronous manner to achieve accurate and efficient anomaly detection, addressing computational overhead across diverse edge clients, including those that experience delays. Our trace-driven evaluations indicate that the proposed method outperforms the state-of-the-art anomaly detection methods by 4% in terms of F1-score while meeting the given time efficiency and scalability requirements.