Decentralized communication is gaining increasing importance in modern computer networks, particularly as edge computing systems evolve towards greater autonomy and resilience. Many emerging applications, from collaborative Machine Learning (ML) to distributed sensing and decision-making, rely on local device-to-device communication rather than centralized coordination. However, existing networking solutions are primarily designed for resource-rich edge nodes and are often unsuitable for highly constrained devices, which lack the memory, energy, and processing capacity to support conventional communication frameworks. To address this limitation, we present μ swim , a lightweight gossip-based communication protocol tailored for constrained edge devices. Built with portability in mind, μ swim extends the Scalable Weakly-consistent Infection-style process group Membership (SWIM) protocol design to operate efficiently in heterogeneous environments, enabling devices with limited capabilities to participate in decentralized edge networks. The implementation provides a minimal Application Programming Interface (API), supports customizable event messages, and offers both Concise Binary Object Representation (CBOR) and JavaScript Object Notation (JSON) message encoding for flexibility across platforms. Through extensive evaluation on real constrained devices, we show that μ swim matches the energy footprint of Constrained Application Protocol (CoAP) and Message Queuing Telemetry Transport (MQTT) while reducing the per-device packet load by 45%–50%, resulting in substantial operational cost savings when scaled to large deployments. Alongside these results, we evaluate convergence behavior, dissemination latency, and encoding overhead, demonstrating that μ swim provides efficient and scalable decentralized communication without sacrificing responsiveness in constrained environments.

Heterogeneous communication environments are expected to be pivotal in sixth-generation (6G) networks, enabling devices to seamlessly utilize various coexisting connectivity options, including cellular technologies (e.g., 4G/5G) and complementary access technologies like Wi-Fi, Bluetooth, and other short-range wireless systems. The challenge lies not just in device diversity but in effectively managing environments where nodes can dynamically switch between different network interfaces to optimize performance, reliability, and quality of service. This paper explores the role of Digital Twins (DTs) as an enabling framework for managing these multi-connectivity environments, where both mobile and non-cellular access technologies collectively enhance overall network functionality. We highlight key limitations in current DT implementations for these connectivity-rich scenarios and propose a scalable DT framework tailored explicitly for integrated cellular and auxiliary wireless access. Developed with commercial mobile devices as endpoints and a widely adopted twinning platform, our solution enhances management efficiency, optimizes resource allocation, and improves Quality of Service (QoS) in dynamic connectivity settings. We evaluate two architectural designs for DT deployment with respect to workload distribution and scalability. Additionally, we present a generative AI-driven analytics pipeline that consolidates descriptive, diagnostic, predictive, and prescriptive analytics into a unified closed-loop system. Experimental results from a real-world smart campus environment show that the offloaded architecture keeps CPU usage below 6% on mobile devices with up to 40 nodes while maintaining low latency and stable performance. The analytics pipeline achieves near-interactive response times, ensuring real-time network awareness and adaptive decision-making in multi-connectivity network environments.

The rapid rise of Language Models (LMs) has expanded the capabilities of natural language processing, powering applications from text generation to complex decision-making. While state-of-the-art LMs often boast hundreds of billions of parameters and are primarily deployed in data centers, recent trends show a growing focus on compact models—typically under 10 billion parameters–enabled by techniques such as quantization and other model compression techniques. This shift paves the way for LMs on edge devices, offering potential benefits such as enhanced privacy, reduced latency, and improved data sovereignty. However, the inherent complexity of even these smaller models, combined with the limited computing resources of edge hardware, raises critical questions about the practical trade-offs in executing LM inference outside the cloud. To address these challenges, we present a comprehensive evaluation of generative LM inference on representative CPU-based and GPU-accelerated edge devices. Our study measures key performance indicators—including memory usage, inference speed, and energy consumption—across various device configurations. Additionally, we examine throughput-energy trade-offs, cost considerations, and usability, alongside an assessment of qualitative model performance. While quantization helps mitigate memory overhead, it does not fully eliminate resource bottlenecks, especially for larger models. Our findings quantify the memory and energy constraints that must be considered for practical real-world deployments, offering concrete insights into the trade-offs between model size, inference performance, and efficiency. The exploration of LMs at the edge is still in its early stages. We hope this study provides a foundation for future research, guiding the refinement of models, the enhancement of inference efficiency, and the advancement of edge-centric AI systems.

On-device inference offers significant benefits in edge ML systems, such as improved energy efficiency, responsiveness, and privacy, compared to traditional centralized approaches. However, the resource constraints of embedded devices limit their use to simple inference tasks, creating a trade-off between efficiency and capability. In this context, the Hierarchical Inference (HI) system has emerged as a promising solution that augments the capabilities of the local ML by offloading selected samples to an edge server/cloud for remote ML inference. Existing works, primarily based on simulations, demonstrate that HI improves accuracy. However, they fail to account for the latency and energy consumption in real-world deployments, nor do they consider three key heterogeneous components that characterize ML-enabled IoT systems: hardware, network connectivity, and models. To bridge this gap, this paper systematically evaluates HI against standalone on-device inference by analyzing accuracy, latency, and energy trade-offs across five devices and three image classification datasets. Our findings show that, for a given accuracy requirement, the HI approach we designed achieved up to 73% lower latency and up to 77% lower device energy consumption than an on-device inference system. Despite these gains, HI introduces a fixed energy and latency overhead from on-device inference for all samples. To address this, we propose a hybrid system called Early Exit with HI (EE-HI) and demonstrate that, compared to HI, EE-HI reduces the latency up to 59.7% and lowers the device’s energy consumption up to 60.4%. These findings demonstrate the potential of HI and EE-HI to enable more efficient ML in IoT systems.

Smart spaces, physical spaces that are integrated with sensor-enabled IoT devices, are a powerful paradigm for optimizing the operations of the space and improving its quality for the occupants. Managing the applications and services running in the space is a complex task as the operations of the devices and services are dependent on the physical characteristics of the space, the occupants of the space, and the technologies that are being integrated. Digital twinning, the combination of physical representations with a virtual counterpart, is a potential technology for facilitating the management of smart space devices and services. While digital twins are increasingly adopted in industry, their use in everyday environments remains low due to difficulties in creating and linking the virtual representation with the physical environment. In this paper, we propose our vision for the adoption of digital twinning as a pathway to improve the functions of smart spaces. We derive a generic reference architecture that comprises four layers, covering the physical space, the sensing infrastructure, the network interfaces, and the underlying computational infrastructure. Next, we identify and address key requirements for the uptake of digital twins in smart space and assess their benefits using the ascendancy model of business analytics. Finally, to demonstrate the practicality of digital twinning, we present a proof-of-concept digital twin for the TellUs smart space at the University of Oulu in Finland and use it to highlight the potential benefits of different ascendancy levels.

The conventional federated learning (FedL) architecture distributes machine learning (ML) across worker devices by having them train local models that are periodically aggregated by a server. FedL ignores two important characteristics of contemporary wireless networks, however: (i) the network may contain heterogeneous communication/computation resources, and (ii) there may be significant overlaps in devices’ local data distributions. In this work, we develop a novel optimization methodology that jointly accounts for these factors via intelligent device sampling complemented by device-to-device (D2D) offloading. Our optimization methodology aims to select the best combination of sampled nodes and data offloading configuration to maximize FedL training accuracy while minimizing data processing and D2D communication resource consumption subject to realistic constraints on the network topology and device capabilities. Theoretical analysis of the D2D offloading subproblem leads to new FedL convergence bounds and an efficient sequential convex optimizer. Using these results, we develop a sampling methodology based on graph convolutional networks (GCNs) which learns the relationship between network attributes, sampled nodes, and D2D data offloading to maximize FedL accuracy. Through evaluation on popular datasets and real-world network measurements from our edge testbed, we find that our methodology outperforms popular device sampling methodologies from literature in terms of ML model performance, data processing overhead, and energy consumption.

The Hierarchical Inference (HI) paradigm has recently emerged as an effective method for balancing inference accuracy, data processing, transmission throughput, and offloading cost. This approach proves particularly efficient in scenarios involving resource-constrained edge devices like micro controller units (MCUs), tasked with executing tinyML inference. Notably, it outperforms strategies such as local inference execution, inference offloading, and split inference (i.e., inference execution distributed between two endpoints). Building upon the HI paradigm, this work explores different techniques aimed at further optimizing inference task execution. We propose three distinct HI approaches and evaluate their utility for image classification.

Edge computing has revolutionized the world of mobile and wireless networks world thanks to its flexible, secure, and performing characteristics. Lately, we have witnessed the increasing use of it to make more performing the deployment of machine learning (ML) techniques such as federated learning (FL). FL was debuted to improve communication efficiency compared to conventional distributed machine learning (ML). The original FL assumes a central aggregation server to aggregate locally optimized parameters and might bring reliability and latency issues. In this paper, we conduct an in-depth study of strategies to replace this central server by a flying master that is dynamically selected based on the current participants and/or available resources at every FL round of optimization. Specifically, we compare different metrics to select this flying master and assess consensus algorithms to perform the selection. Our results demonstrate a significant reduction of runtime using our flying master FL framework compared to the original FL from measurements results conducted in our EdgeAI testbed and over real 5G networks using an operational edge testbed.

In recent years, the research community started to extensively study how edge computing can enhance the provisioning of a seamless and performing Machine Learning (ML) experience. Boosting the performance of ML inference at the edge became a driving factor especially for enabling those use-cases in which proximity to the data sources, near real-time requirements, and need of a reduced network latency represent a determining factor. The growing demand of edge-based ML services has been also boosted by an increasing market release of small-form factor inference accelerators devices that feature, however, heterogeneous and not fully interoperable software and hardware characteristics. A key aspect that has not yet been fully investigated is how to discover and efficiently optimize the provision of ML inference services in distributed edge systems featuring heterogeneous edge inference accelerators - not neglecting also that the limited devices computation capabilities may imply the need of orchestrating the inference execution provisioning among the different system's devices. The main goal of this demo is to showcase how ML inference services can be agnostically discovered, provisioned, and orchestrated in a cluster of heterogeneous and distributed edge nodes.

The conventional federated learning (FedL) architecture distributes machine learning (ML) across worker devices by having them train local models that are periodically aggregated by a server. FedL ignores two important characteristics of contemporary wireless networks, however: (i) the network may contain heterogeneous communication/computation resources, while (ii) there may be significant overlaps in devices' local data distributions. In this work, we develop a novel optimization methodology that jointly accounts for these factors via intelligent device sampling complemented by device-to-device (D2D) offloading. Our optimization aims to select the best combination of sampled nodes and data offloading configuration to maximize FedL training accuracy subject to realistic constraints on the network topology and device capabilities. Theoretical analysis of the D2D offloading subproblem leads to new FedL convergence bounds and an efficient sequential convex optimizer. Using this result, we develop a sampling methodology based on graph convolutional networks (GCNs) which learns the relationship between network attributes, sampled nodes, and resulting offloading that maximizes FedL accuracy. Through evaluation on real-world datasets and network measurements from our IoT testbed, we find that our methodology while sampling less than 5% of all devices outperforms conventional FedL substantially both in terms of trained model accuracy and required resource utilization.

The more stringent requirements of many applications services - especially in terms of latency and bandwidth - is leading to the migration from data center based architecture towards a mobile-edge computing paradigm. This emerging network architecture aims to increase the overall infrastructure efficiency by delivering low-latency and bandwidth-efficient services. In this context, a not fully investigated aspect is represented by the possibility of placing edge-computing tasks on low-power nodes. This paper seeks to provide insights for future deployments, by conducting an empirical study on the performance evaluation of a Single-Board Computer and a Small-Form Factor Computer when acting as a mobile-edge computing server. Our results aim to determine the potentiality and the performance bounds for these devices, during the execution of applications characterized by different performance requirements.

Modern vehicles are equipped with several interconnected sensors on board for monitoring and diagnosis purposes; their availability is a main driver for the development of novel applications in the smart vehicle domain. In this paper, we propose a Docker container-based platform as solution for implementing customized smart car applications. Through a proof-of-concept prototype-developed on a Raspberry Pi3 board-we show that a container-based virtualization approach is not only viable but also effective and flexible in the management of several parallel processes running on On Board Unit. More specifically, the platform can take priority-based decisions by handling multiple inputs, e.g., data from the CANbus based on the OBD II codes, video from the on-board webcam, and so on. Results are promising for the development of future in-vehicle virtualized platforms.

The constant and rapid evolution of hardware and software components, along with the proliferation of newer standards and technologies, makes the creation of IoT environments devoted to test purposes extremely complicated. In this work, we present the design of an IoT emulation platform (ELIoT), which aim on making the deployment of IoT test environments easier and more flexible. ELIoT provides support for well-known IoT protocols - e.g., Constrained Application Protocol (CoAP) and Lightweight M2M (LWM2M) -, facilitating the study of interactions between different IoT entities. In order to make ELIoT easily portable and highly customizable, we exploit the lightweight and flexible characteristics of emerging virtualization technologies, such as Docker containers, for emulating both simple and complex IoT devices. A comprehensive performance evaluation aims to assess the feasibility of our implementation, by testing ELIoT scalability properties and other relevant performance metrics such as communication latency and throughput.

The constant and fast increase in the number of heterogeneous Internet of Things (IoT) devices that populate everyday life environments brings new challenges to the full exploitation of the computation, memory, sensing, and actuation resources associated to them. In this context, device virtualization solutions and platforms may definitely play a key role in enabling the desired tradeoff between flexibility and performance. This paper focuses on lightweight virtualization technologies for IoT devices, suitably thought to effectively deploy new integrated applications and to create a novel distributed and virtualized ecosystem. Two different frameworks for container-based IoT service provisioning are compared, the one based on a direct interaction between two cooperating devices and the other based on the presence of a manager supervising the operations between cooperating devices forming a cluster. In the latter case, accounting for the growing impetus to move intelligence toward the edge of the network, management features are implemented at the network access point to provide short latency responses. We also introduce the outcomes of a thorough performance evaluation campaign conducted via a real IoT testbed. The measurements, performed by accounting for the constraints of typical IoT nodes, shed light on the actual feasibility of container-based IoT frameworks.

Virtualization of operating systems provides a common way to run different services in the cloud. Recently, the lightweight virtualization technologies claim to offer superior performance. In this paper, we present a detailed performance comparison of traditional hypervisor based virtualization and new lightweight solutions. In our measurements, we use several benchmarks tools in order to understand the strengths, weaknesses, and anomalies introduced by these different platforms in terms of processing, storage, memory and network. Our results show that containers achieve generally better performance when compared with traditional virtual machines and other recent solutions. Albeit containers offer clearly more dense deployment of virtual machines, the performance difference with other technologies is in many cases relatively small.