Edge devices powered by renewable energy sources, such as solar panels, face the challenge of operating under uncertain and intermittent energy availability. This paper proposes a Reinforcement Learning (RL) scheduler for heavy tasks, e.g. image classification running on solar-powered devices with an accumulator.The scheduler maximizes the total number of processed images by deciding whether to process an image immediately or store it for later processing, all while remaining energy-aware to prevent device shutdown. Crucially, the method operates without any prior knowledge of future solar energy production, making it suitable for devices lacking an internet connection for forecasting. The RL method processes 99.8% of the images processed by the optimal solution obtained through Integer Linear Programming (ILP), missing on average only 224 seconds of captured frames per day. This demonstrates that RL is a viable alternative to optimal solvers even in forecast-free scenarios

Efficient resource management in Kubernetes is crucial for optimizing performance and cost in cloud computing environments. Traditional autoscaling methods react to workload changes but often fail to predict them, leading to underutilization of resources or performance degradation. This paper introduces a model-free Reinforcement Learning (RL) based autoscaler that leverages a deep Q-Network (DQN) alongside a Long Short-Term Memory (LSTM) network for workload forecasting. By predicting future request rates, our autoscaler proactively adjusts the number of container replicas, ensuring compliance to Service Level Objectives (SLOs) while minimizing resource usage. Experimental results demonstrate that our approach outperforms standard Kubernetes auto-scaling strategies, achieving a resource consumption reduction of up to 10% and improving performance levels up to 30% compared to the default auto-scaling algorithm.

DNNs are widely used for complex tasks like image and signal processing, and they are in increasing demand for implementation on Internet of Things (IoT) devices. For these devices, optimizing DNN models is a necessary task. Generally, standard optimization approaches require specialists to manually fine-tune hyper-parameters to find a good trade-off between efficiency and accuracy. In this paper, we propose OptDNN, a software that employs innovative and automatic approaches to determine optimal hyper-parameters for pruning, clustering, and quantization. The models optimized by OptDNN have a smaller memory footprint, faster inference time, and a similar accuracy to the original models.

By spreading out computing workloads over multiple levels, the Edge-to-Cloud continuum paradigm improves the performance of applications that are sensitive to latency. However, real-time scheduling is difficult on the Edge computing layer since it consists of a variety of nodes with varying uptime. In this paper, we address this issue by proposing an online and adaptive scheduling algorithm based on a continuous learning Reinforcement Learning. Our algorithm determines each work request individually, optimizing scheduling policies to meet real-time application requirements while taking environmental energy and battery limits into account. We validate the efficacy of our approach in dynamically assigning tasks, particularly in scenarios where Edge nodes exhibit variable speeds and unpredictable failures, while efficiently managing energy resources and battery constraints through extensive simulations and comparisons with static scheduling strategies.

DNNs, commonly employed for complex tasks such as image and language processing, are increasingly sought for deployment on Internet of Things (IoT) devices. These devices operate with constrained resources, including limited computational power, memory, slower processors, and restricted energy requirements. Consequently, optimizing DNN models becomes crucial to minimize memory usage and computational time. However, traditional optimization methods require skilled professionals to manually fine-tune hyperparameters, striking a balance between efficiency and accuracy. This paper introduces an innovative solution for identifying optimal hyperparameters, focusing on the application of pruning, clusterization, and quantization. Initial empirical analyses were conducted to understand the relationships between model size, accuracy, pruning rate, and the number of clusters. Building upon these findings, we developed a framework that proposes two algorithms: one for discovering optimal pruning and the second for determining the optimal number of clusters. Through the adoption of efficient algorithms and the best quantization configuration, our tool integrates an optimization procedure that successfully reduces model size and inference time. The optimized models generated exhibit results comparable to, and in some cases surpass, those of more complex state-of-the-art approaches. The framework successfully optimized ResNet50, reducing the model size by 6.35x with a speedup of 2.91x, while only sacrificing 0.87

Fog Computing is a widely adopted paradigm that allows distributing the computation in a geographic area. This makes it possible to implement time-critical applications and opens the study to a series of solutions that permit smartly organizing the traffic among a set of fog nodes, which constitute the core of the Fog Computing paradigm. As a typical smart city setting is subject to a continuous change in traffic conditions, it is necessary to design algorithms that can manage all the computing resources by properly distributing the traffic among the nodes in an adaptive way. In this paper, we propose a cooperative and decentralized algorithm based on Reinforcement Learning that is able to perform online scheduling decisions among fog nodes. This can be seen as an improvement over the power-of-two random choices paradigm used as a baseline. By showing results from our delay-based simulator and then from our framework “P2PFaaS” installed on 12 Raspberry Pis, we show how our approach maximizes the rate of the tasks executed within the deadline, outperforming the power-of-two random choices both in a fixed load condition and with traffic extracted from a real smart city scenario.

The knowledge of environmental depth is essential in multiple robotics and computer vision tasks for both terrestrial and underwater scenarios. Moreover, the hardware on which this technology runs, generally IoT and embedded devices, are limited in terms of power consumption, and therefore, models with a low-energy footprint are required to be designed. Recent works aim at enabling depth perception using single RGB images on deep architectures, such as convolutional neural networks and vision transformers, which are generally unsuitable for real-time inferences on low-power embedded hardware. Moreover, such architectures are trained to estimate depth maps mainly on terrestrial scenarios due to the scarcity of underwater depth data. Purposely, we present two lightweight architectures based on optimized MobileNetV3 encoders and a specifically designed decoder to achieve fast inferences and accurate estimations over embedded devices, a feasibility study to predict depth maps over underwater scenarios, and an energy assessment to understand which is the effective energy consumption during the inference. Precisely, we propose the MobileNetV3S75 configuration to infer on the 32-bit ARM CPU and the MobileNetV3LMin for the 8-bit Edge TPU hardware. In underwater settings, the proposed design achieves comparable estimations with fast inference performances compared to state-of-the-art methods. Moreover, we statistically proved that the architecture of the models has an impact on the energy footprint in terms of Watts required by the device during the inference. Then, the proposed architectures would be considered to be a promising approach for real-time monocular depth estimation by offering the best trade-off between inference performances, estimation error and energy consumption, with the aim of improving the environment perception for underwater drones, lightweight robots and Internet of things.

In Edge and Fog Computing environments, it is usual to design and test distributed algorithms that implement scheduling and load balancing solutions.The operation paradigm that usually fits the context requires the users to make calls to the closer node for executing a task, and since the service must be distributed among a set of nodes, the serverless paradigm with the FaaS (Function-as-a-Service) is the most promising strategy to use.In light of these preconditions, we designed and implemented a framework called P2PFaaS.The framework, built upon Docker containers, allows the implementation of fully decentralised scheduling or load balancing algorithms among a set of nodes.By relying on three basic services, such as the scheduling service, the discovery service, and the learner service, the framework allows the implementation of any kind of scheduling solution, even if based on Reinforcement Learning.Finally, the framework provides a ready-to-go solution that can be installed and has been tested both on x86 servers and ARM-based edge nodes (like, for example, the Raspberry Pi).

We consider the case of a set of energy harvesting edge nodes, equipped with photovoltaic panels that implement some kind of monitoring service. To ensure that the service operates in an optimal way, nodes have sometimes offload some of their data to other nodes. We show that this kind of task offloading (migration) can improve service performance by avoiding temporary interruptions and prolonging the overall service lifetime. We present a centralized algorithm based on Linear Programming optimization problem solution and a distributed implementation.

When dealing with distributed applications in Edge or Fog computing environments, the service latency that the user experiences at a given node can be considered an indicator of how much the node itself is loaded with respect to the others. Indeed, only considering the average CPU time or the RAM utilisation, for example, does not give a clear depiction of the load situation because these parameters are application- and hardware-agnostic. They do not give any information about how the application is performing from the user's perspective, and they cannot be used for a QoS-oriented load balancing. In this article, we propose a load balancing algorithm that is focused on the service latency with the objective of levelling it across all the nodes in a fully decentralised manner. In this way, no user will experience a worse QoS than the other. By providing a differential model of the system and an adaptive heuristic to find the solution to the problem in real settings, we show both in simulation and in a real-world deployment, based on a cluster of Raspberry Pi boards, that our approach is able to level the service latency among a set of heterogeneous nodes organised in different topologies.

Virtual reality (VR) is a technology that has the potential to innovate whole sectors as well as the way we interact with digital environments. AR and VR devices are a perfect way to gain knowledge, information, and to practice operation plans during, before, and after a disaster took place, exploiting their capabilities. VR headsets like the Meta Quest 2 allow to perform rendering of applications locally or exploit the offloading to a Server in the cloud, and, in a disaster management scenario, the energy consumption of the device is a fundamental knowledge in order to make rational choices on which of the two types of rendering to perform. In this paper, we investigate the energy efficiency of the Meta Quest 2 in both of the two rendering approaches performing tests based on a benchmark application developed in Unreal Engine. From the results of our experiments, we found that remote rendering, performed via AirLink, allowed us to obtain higher FPS and overall quality, as well as keep the device GPU and CPU usage at lower values than with the local rendering paradigm. However, from the energy efficiency perspective, test results showed that besides the lighter CPU and GPU work using the remote rendering approach, the energy consumption, in the overall execution, exceeds the one using the local rendering paradigm, because of the network communication with the server.

Green-powered edge computing architectures allow bringing computation in areas that are not reached by the power grids. More often, in applications for Precision Agriculture and Smart Cities, we could have a set of nodes that are coupled with an accumulator which is, during the day, re-charged by the energy harvested by small solar panels. With the latest advances in technology, the edge node is generally assimilated to be a low-power Single Board Computer (SBC), and it is able to carry out even relatively demanding tasks. For example, it can run deep learning models to images or video sequences captured in loco by cameras. However, due to the differences in terms of power consumption and weather conditions, each node experiences a different lifespan, some nodes may even shut down prematurely, causing the interruption of the portion of the deployed service. In this paper, we propose three decentralized algorithms that solve the problem by making the nodes cooperatively balance the traffic in order to level and maximize their lifespan. By comparing the approaches in two different experiments by using a cluster of Raspberry Pi 4 we show that our solutions allow to increase the lifespan of the service of 10% on average wrt the case in which no algorithm is applied.

Fog and Edge Computing are two paradigms specifically suitable for real-time and time-critical applications, which are usually distributed among a set of nodes that constitutes the core idea of both Fog and Edge Computing. Since nodes are heterogeneous and subject to different traffic patterns, distributed scheduling algorithms are in charge of making each request meet the specified deadline. In this paper, we exploit the approach of Reinforcement Learning based decision-making for designing a cooperative and decentralized task online scheduling approach which is composed of two RL-based decisions. One for selecting the node to which to offload the traffic and one for accepting or not the incoming offloading request. The experiments that we conducted on a cluster of Raspberry Pi 4 show that introducing a second RL decision increases the rate of tasks executed within the deadline of 4% as it introduces more flexibility during the decision-making process, consequently enabling better scheduling decisions.

In this article, we consider a load balancing protocol based on the power of randomchoices that is adapted to a fog deploy in which several independent fog nodes equipped with a set of servers orVMare serving the same geographical area. The protocol is based on a simple but effective mechanism based on a threshold T. When a fog node receives a unit of computation or a job, it immediately executes the job if the number of its occupied servers is lower than T, otherwise the node executes a randomized algorithm by probing F other fog nodes in the area, and delegates the execution of the job to the least loaded one, provided the workload is lower than the probing node. Through a mathematical analysiswe show that probing just one node (F = 1) when there are less than two free VMs provides the same performance of the well known power-of-two random choices centralized algorithm, but at a much lower delay and control overhead costs. Also, simulations are used to address the node heterogeneity and, with a real testbed, we offer results that prove the effective benefit of the proposed solution in practical applications.

The distributed nature of edge computing infrastructures requires a significant effort to avoid overload conditions due to uneven distribution of incoming load from sensors placed over a wide area. While optimisation algorithms operating offline can address this issue in the medium to long term, sudden and unexpected traffic surges require an online approach where load balancing actions are taken at a smaller time scale. However, when the service time of a single request becomes comparable with the latency needed to take and actuate load balancing decisions, the design of online approaches becomes particularly challenging. This paper focuses on the class of online algorithms for load balancing based on resource sharing among random nodes. While this randomisation principle is a straightforward and effective way to share resources and achieve load balance, it fails to work properly when the interval between decision making and decision actuating times (called schedule lag ) becomes comparable with the time required to execute a job, a condition not rare in edge computing systems, and provokes stale (out-of-date) information to be involved in scheduling decisions. Our analysis combines (1) a theoretical model that evaluates how stale information reduces the effectiveness of the balancing mechanism and describes the correlation between the system state at decision making and decision actuating times; (2) a simulation approach to study a wide range of algorithm parameters and possible usage scenarios. The results of our analysis provides the designers of distributed edge systems with useful hints to decide, based on the scenario, which load balancing protocol is the most suitable.

When deploying a distributed application in the Fog or Edge computing environments, the average service latency among all the involved nodes can be an indicator of how much a node is loaded with respect to the other. Indeed, only considering the average CPU time, or the RAM utilisation, for example, does not give a clear depiction of the load situation because these parameters are application- and hardware-agnostic. They do not give any information about how the application is performing from the user perspective and they cannot be used for a QoS-oriented load balancing of the system. Moreover, due to the displacement of the nodes and the heterogeneity of the computing devices the necessity of a load balancing algorithm is clear. In this paper, we propose a load balancing approach that is focused on the service latency with the objective to level it across all the nodes in a fully decentralized manner, in this way no user will experience a worse QoS than the other. By providing a differential model of the system and an adaptive heuristic to find the solution to the problem, we show both in simulation and in a real-world deployment that our approach is able to level the service latency among a set of heterogeneous nodes organized in different topologies.

Fog Computing is today a wide used paradigm that allows to distribute the computation in a geographic area. This not only makes possible to implement time-critical applications but opens the study to a series of solutions which permit to smartly organise the traffic among a set of Fog nodes, which constitute the core of the Fog Computing paradigm. A typical smart city setting is subject to a continuous change of traffic conditions, a node that was saturated can become almost completely unloaded and this creates the need of designing an algorithm which allows to meet the strict deadlines of the tasks but at the same time it can choose the best scheduling policy according to the current load situation that can vary at any time. In this paper, we use a Reinforcement Learning approach to design such an algorithm starting from the power-of-random choice paradigm, used as a baseline. By showing results from our delay-based simulator, we demonstrate how such distributed reinforcement learning approach is able to maximise the rate of the tasks executed within the deadline in a way that is equal to every node, both in a fixed load condition and in a real geographic scenario.

We consider the case of a group of communicating edge nodes only supplied with photovoltaic (PV) panels. This configuration makes the deploy green if all the data is elaborated locally, but a minimal (if nothing) communication with a remote server. This model has several characteristics. Since the amount of storable energy is limited by the battery capacity, the solar energy at an edge node with fully charged battery cannot be further accumulated, which is an indirect source of inefficiency. On the other hand, the energy consumed by edge nodes is tied to the activity of the nodes. Roughly an edge node in an idle state consumes less than one third of the power consumed when elaborating data, while the energy cost to move data among the edge nodes is usually less than the one required to elaborate them (which is the case in many applications). As a consequence an edge node which is running out of green energy can conveniently offload its computations to energy richer nodes in idle states. This paper reports some preliminary results of this form of green cooperation.

Virtual Reality (VR) is now a well-established technology which offers realistic and immersive virtual worlds to the user usually by means of Head-Mounted Displays (HMDs). Actually, these devices can also be backed by a cloud server which can host the game server or even directly render the virtual world, as in the well-known cloud gaming paradigm. However, due to the drastically low latencies that technology requires, it is more convenient, when possible, to use servers that are as close to the users as possible. As a consequence, implementing the game or the render server in the Fog Computing layer is a concrete possibility. A research investigation has been carried out by using the Meta Quest 2 device, which are the QoE trade-offs, in terms of graphic quality and network performance, both in the case in which the HMD performs the 3d rendering locally by using a sample game written with Unreal Engine and in the case in which the 3d rendering is done in the Fog by means of the nVidia CloudXR framework and Oculus Air Link. The results of the proposed experiments reveal that the remote rendering offers a stable frame rate against a higher quality image. Instead, local rendering sets the best possible graphics quality against the optimal frame rate. Additionally, the utilization of remote rendering to perform video compression in the case of decreasing the available bandwidth to adjust graphics quality and FPS has also been analyzed. The same does not hold for Motion-to-Photon latency, which increases with distance, reducing the general QoE.