Over the past decade, key advancements in Artificial Intelligence (AI) and Edge Computing (EC) have led to the development of EdgeAI services to provide intelligent and low latency responses essential for mission-critical applications. However, the expansion of EdgeAI services to the network extremes can face challenges such as load fluctuations causing delays in AI inference and concerns over energy efficiency. This paper proposes “model swapping” where the model employed by the EdgeAI service is swapped on-the-fly with another readily available model so that cost and energy savings are achieved during runtime inference tasks. The ModelSwapper can achieve this by employing a lowcost algorithmic technique that explores meaningful trade-offs between the computational overhead and the model accuracy. By doing so, edge nodes adapt to load fluctuations by substituting complex models with simpler ones, thus meeting desired latency requirements, albeit with potentially higher uncertainty. Our evaluation with two EdgeAI services (object detection, NLU) demonstrates that ModelSwapper can significantly reduce energy usage and inference delays by at least $27 \%$ and $68 \%$ respectively, with only a $1 \%$ reduction in accuracy.

Energy profiling and optimization are expected to be crucial factors impacting the realisation of the Internet of Things (IoT) as more intelligence is deployed at the network extremes to achieve better response times in the proximity of where data are harvested. To improve the performance of streaming analytics jobs, several schedulers have been designed to tackle key challenges in edge computing realms, including resource heterogeneity and highly volatile network links. However, energy-aware scheduling for streaming analytic jobs is at best, not adequately examined. In this article, we introduce PowerStorm, a scheduler for streaming analytic jobs that is designed to explore trade-offs between performance and energy consumption in geodistributed edge computing settings. We implement our scheduler for Apache Storm and show the scheduler’s energy saving capabilities over the Yahoo streaming benchmark with worker nodes featuring heterogeneous power and resource capabilities on both a physical and emulated testbed.

As data analytics and machine learning increasingly permeate our cities, factories, and homes, the computing infrastructure for data-intensive systems becomes more challenging. That is, the vision of pervasive, intelligent, and cyber-physical IoT systems will not be realized with centralized cloud resources alone. Such resources are simply too far away from sensor-equipped devices and users, resulting in high latency, bandwidth bottlenecks, and unnecessary energy consumption. In addition, there are often privacy and security requirements that mandate distributed architectures. As a result, new distributed computing paradigms are emerging that promise to bring computing and storage closer to data sources and users. The emerging distributed computing environments of edge and fog computing provide additional resources within mobile networks, ISP infrastructure, and even LEO satellites. These diverse and dynamic computing environments pose significant challenges to the performance, dependability, and efficiency of data-intensive systems running on such infrastructure. At the same time, it is far less clear how to properly benchmark, evaluate, and test the behavior of systems that span IoT devices, edge nodes, and cloud resources. For example, IoT sensor data stream processing systems can be leveraged to continuously optimize the operation of urban infrastructures (such as public transportation systems, water networks, or medical infrastructures). The behavior of such systems must be thoroughly assessed before they can be deployed to edge and fog infrastructure. In addition, these systems must be evaluated reproducibly under the expected computing environment conditions, including variations of those conditions, given the inherently unsteady nature of IoT environments. In addition, there is growing concern about the energy consumption and greenhouse gas emissions of ICT (and especially distributed ML-based applications), which further warrants close examination of the behavior of new data-intensive applications. Despite significant research and development efforts to improve benchmarking, experimentation tools, and reproducible practices for data-intensive systems spanning from the edge to the cloud, more research is urgently needed. We therefore invited high-quality research papers on this topic for this special issue of Software: Practice and Experience, and we were able to select two out of four submissions for this special issue with the help of our reviewers. The first accepted paper is titled “faas-sim: A Trace-Driven Simulation Framework for Serverless Edge Computing Platforms”.1 It is co-authored by Philipp Raith, Thomas Rausch, Alireza Furutanpey, and Schahram Dustdar. The paper presents the design and implementation of a new simulation framework, “faas-sim,” for modeling and evaluating serverless software architectures spanning the edge-cloud continuum based on a scenario description, a given network topology, and workload traces. The new simulator is demonstrated by using it for performance estimation, resource planning, co-simulation, and scientific evaluation. The authors also evaluate faas-sim's network simulation and resource utilization. Furthermore, they highlight traces that come with faas-sim and provide an overview of published research that has used faas-sim. The second accepted paper is titled “Software-in-the-Loop Simulation for Developing and Testing Carbon-Aware Applications”.2 It is co-authored by Philipp Wiesner, Marvin Steinke, Henrik Nickel, Yazan Kitana, and Odej Kao. As an alternative to relying on purely simulated or purely real testbeds, the paper proposes the use of software-in-the-loop simulation and hybrid testbeds for testing carbon-aware software applications in the context of energy simulations. The paper describes the design and implementation of a prototype, “Vessim,” as well as two experiments demonstrating the capabilities and features of the novel tool. In this way, the paper shows how a message broker can reliably and realistically connect currently running applications under test to real-time simulations, while the energy demand is continuously measured or modeled. We are grateful to the Editor-in-Chief of the Journal, Dr. Rajkumar Buyya, for inviting us to organize this special issue. We are also grateful for the valuable support from the administrative office of the journal. In addition, we are very grateful for the thorough and thoughtful reviews provided by our reviewers. Finally, we appreciate the hard work and trust of the authors who submitted papers to our special issue.

Energy consumption and carbon emissions are expected to be crucial factors for Internet of Things (IoT) applications. Both the scale and the geo-distribution keep increasing, while Artificial Intelligence (AI) further penetrates the “edge” in order to satisfy the need for highly-responsive and intelligent services. To date, several edge/fog emulators are catering for IoT testing by supporting the deployment and execution of AI-driven IoT services in consolidated test environments. These tools enable the configuration of infrastructures so that they closely resemble edge devices and IoT networks. However, energy consumption and carbon emissions estimations during the testing of AI services are still missing from the current state of IoT testing suites. This study highlights important questions that developers of AI-driven IoT services are in need of answers, along with a set of observations and challenges, aiming to help researchers designing IoT testing and benchmarking suites to cater to user needs.

As drone technology penetrates even more application domains, Machine Learning (ML) is becoming a key driver enabling intelligence in the sky. However, ML Practitioners and Drone Application Operators are faced with several challenges when wanting to test ML-driven drone applications early in the design phase. These include the development and configuration of experiment use-cases over a robotics simulator along with the collection and assessment of desired KPIs which can range from ML algorithm accuracy to drone resource utilization and the impact of "intelligence" to the drone’s energy footprint. This demonstration showcases FlockAI, an open and modular by design framework supporting users with the rapid deployment and repeatable testing during the design phase of ML-driven drone applications over the Webots robotics simulator. Through realistic use-cases, the demonstration will show how FlockAI can be used to design drone testbeds with "ready-to-go" drone templates, deploy ML models, configure on-board/remote inference, monitor and export drone resource utilization, network overhead and energy consumption to pinpoint performance inefficiencies and understand if various trade-offs can be exploited.

Micro-Datacenters (DCs) are emerging as key en-ablers for Edge computing and 5G mobile networks by pro-viding processing power closer to IoT devices to extract timely analytic insights. However, the performance evaluation of data stream processing on micro-DCs is a daunting task due to difficulties raised by the time-consuming setup, configuration and heterogeneity of the underlying environment. To address these challenges, we introduce BenchPilot, a modular and highly customizable benchmarking framework for edge micro-DCs. BenchPilot provides a high-level declarative model for describing experiment testbeds and scenarios that automates the bench-marking process on Streaming Distributed Processing Engines (SDPEs). The latter enables users to focus on performance analysis instead of dealing with the complex and time-consuming setup. BenchPilot instantiates the underlying cluster, performs repeatable experimentation, and provides a unified monitoring stack in heterogeneous Micro-DCs. To highlight the usability of BenchPilot, we conduct experiments on two popular streaming engines, namely Apache Storm and Flink. Our experiments compare the engines based on performance, CPU utilization, energy consumption, temperature, and network I/O.

With the proliferation of raw Internet of Things (IoTs) data, Fog Computing is emerging as a computing paradigm for delay-sensitive streaming analytics with operators deploying big data distributed engines on Fog resources [1]. Nevertheless, the current (Cloud-based) distributed analytics solutions are unaware of the unique characteristics of Fog realms. For instance, task placement algorithms consider homogeneous underlying resources without considering the Fog nodes' heterogeneity and the non-uniform network connections, resulting in sub-optimal processing performance. Moreover, data quality can play an important role, where corrupted data, and network uncertainty may lead to less useful results. In turn, energy consumption can critically impact the overall cost and liveness of the underlying processing infrastructure. Specifically, scheduling tasks on nodes with energy-hungry profiles or battery-powered devices may temporarily be beneficial for the performance, but it may increase the overall cost, or/and the battery-powered devices may not be available when needed. A Fog-enabled analytics stack must allow users to optimize Fog-specific indicators or trade-offs among them. For instance, users may sacrifice a portion of the execution performance to minimize energy consumption or vice versa. Except for the performance issues raised by Fog, the state-of-the-art distributed processing engines offer only low-level procedural programming interfaces with operators facing a steep learning curve to master them. So, query abstractions are crucial for minimizing the deployment time, errors, and debugging.

We are now witnessing the extensive deployment of drones in a diverse set of applications with Machine Learning (ML) constituting a key enabler empowering the uptake of drone technology. With the advancements of robotics and edge computing, on-board ML is on the uprise. However, testing ML solutions for drones before release to production is a daunting task for ML practitioners. This usually involves the testing on a robotics emulator to collect various key performance indicators ranging from algorithm correctness to resource utilization. Thus, to thoroughly evaluate performance, a true understanding of the ML algorithm impact on the drones most scarce resource is required. Without a doubt, this is the drones battery, which entails continuously monitoring energy consumption. In this paper we introduce HornEt, a modular framework enabling the customization and composition of various monitorable components to produce realistic energy models that can be used during the testing of ML-driven drone applications. To show the wide applicability of our framework, we introduce a proof-of-concept use-case illustrating the energy profiling of a drone application at different levels of granularity.

Internet-enabled physical devices with “smart” processing capabilities are becoming the tools for understanding the complexity of the global inter-connected world we inhabit. The Internet of Things (IoT) churns tremendous amounts of data flooding from devices scattered across multiple locations to the processing engines of almost all industry sectors. However, as the number of “things” surpasses the population of the technology-enabled world, real-time processing and energy-efficiency are great challenges of the big data era transitioning to IoT. In this article, we introduce a lightweight adaptive monitoring framework suitable for smart IoT devices with limited processing capabilities. Our framework, inexpensively and in place dynamically adjusts the monitoring intensity and the amount of data disseminated through the network based on a low-cost adaptive and probabilistic learning model capable of capturing at runtime the current evolution and variability of the data stream. By accomplishing this, energy consumption and data volume are reduced, allowing IoT devices to preserve battery and ease processing on cloud computing and streaming services. Experiments on real-world data from cloud services, internet security services, wearables and intelligent transportation services, show that our framework achieves a balance between efficiency and accuracy. Specifically, our framework reduces data volume by 74 percent, energy consumption by at least 71 percent, while maintaining accuracy always above 89 percent.

Fog Computing is emerging as the dominating paradigm bridging the compute and connectivity gap between sensing devices and latency-sensitive services. However, experimenting and evaluating IoT services is a daunting task involving the manual configuration and deployment of a mixture of geodistributed physical and virtual infrastructure with different resource and network requirements. This results in sub-optimal, costly and error-prone deployments due to numerous unexpected overheads not initially envisioned in the design phase and underwhelming testing conditions not resembling the end environment. In this paper, we introduce Fogify, an emulator easing the modeling, deployment and large-scale experimentation of fog and edge testbeds. Fogify provides a toolset to: (i) model complex fog topologies comprised of heterogeneous resources, network capabilities and QoS criteria; (ii) deploy the modelled configuration and services using popular containerized descriptions to a cloud or local environment; (iii) experiment, measure and evaluate the deployment by injecting faults and adapting the configuration at runtime to test different “what-if” scenarios that reveal the limitations of a service before introduced to the public. In the evaluation, proof-of-concept IoT services with real-world workloads are introduced to show the wide applicability and benefits of rapid prototyping via Fogify.

For more than the better parts of the last decades, we are witnessing the proliferation of IoT devices, as well as an exponential growth in the volume of data generated outside of datacenters. With the generated data at the extremes of the network and the restricted device-to-cloud bandwidth, data mitigation is becoming the major barrier of cloud-based IoT services [1]. To alleviate these challenges, Fog Computing extends the Cloud's capabilities closer to IoT devices.

With consumers embracing the prevalence of ubiquitously connected smart devices, Edge Computing is emerging as a principal computing paradigm for latency-sensitive and in-proximity services. However, as the plethora of data generated across connected devices continues to vastly increase, the need to query the "edge" and derive in-time analytic insights is more evident than ever. This paper introduces our vision for a rich and declarative query model abstraction particularly tailored for the unique characteristics of Edge Computing and presents a prototype framework that realizes our vision. Towards this, the declarative query model enables users to express high-level and descriptive analytic insights, while our framework compiles, optimizes and executes the query plan decoupled from the programming model of the underlying data processing engine. Afterwards, we showcase a number of potential use-cases which stand to benefit from the realization of query-driven descriptive analytics for edge computing. We conclude by elaborating on the open challenges that still must be addressed to realize our vision and potential research opportunities for the academic community to further advance the current State-of-the-Art.

Fueled by the wide societal interest for decentralized services, blockchain has emerged into a highly desired paradigm extending well beyond financial transactions. The next generation of blockchain services are now storing more and more data on distributed ledgers. Therefore, the need to perform analytic queries over blockchains is more evident than ever. However, despite the wide public interest and the release of several frameworks, efficiently accessing and processing data from blockchains is challenging. This paper introduces Datachain, a lightweight, flexible and interoperable framework deliberately designed to ease the extraction of data hosted on distributed ledgers. Through high-level query abstractions, users connect to underlying blockchains, perform data requests, extract transactions, manage data assets and derive high-level analytic insights. Most importantly, due to the inherent interoperable nature of Datachain, queries and analytic jobs are reusable and can be executed without alterations on different underlying blockchains. To illustrate the wide applicability of Datachain, we present a realistic use-case on top of Hyperledger and BigchainDB.

With the widespread adoption of temporal graphs to study fast evolving interactions in dynamic networks, attention is needed to provide graph metrics in time and at scale. In this paper, we introduce ATMoN, an open-source library developed to computationally offload graph processing engines and ease the communication overhead in dynamic networks over an unprecedented wealth of data. This is achieved, by efficiently adapting, in place and inexpensively, the temporal granularity at which graph metrics are computed based on runtime knowledge captured by a low-cost probabilistic learning model capable of approximating both the metric stream evolution and the volatility of the graph topology. After a thorough evaluation with real-world data from mobile, face-to-face and vehicular networks, results show that ATMoN is able to reduce the compute overhead by at least 76%, data volume by 60% and overall cloud costs by at least 54%, while always maintaining accuracy above 88%.

Automatic resource provisioning is a challenging and complex task. It requires for applications, services and underlying platforms to be continuously monitored at multiple levels and time intervals. The complex nature of this task lays in the ability of the monitoring system to automatically detect runtime configurations in a cloud service due to elasticity action enforcement. Moreover, with the adoption of open cloud standards and library stacks, cloud consumers are now able to migrate their applications or even distribute them across multiple cloud domains. However, current cloud monitoring tools are either bounded to specific cloud platforms or limit their portability to provide elasticity support. In this article, we describe the challenges when monitoring elastically adaptive multi-cloud services. We then introduce a novel automated, modular, multi-layer and portable cloud monitoring framework. Experiments on multiple clouds and real-life applications show that our framework is capable of automatically adapting when elasticity actions are enforced to either the cloud service or to the monitoring topology. Furthermore, it is recoverable from faults introduced in the monitoring configuration with proven scalability and low runtime footprint. Most importantly, our framework is able to reduce network traffic by 41 percent and consequently the monitoring cost, which is both billable and noticeable in large-scale multi-cloud services.

As more knowledge is vastly added to the devices fuelling the Internet of Things (IoT) energy efficiency and real-time data processing are great challenges that must be tackled. In this paper, we introduce ADMin, a low-cost IoT framework that reduces on device energy consumption and the volume of data disseminated across the network. This is achieved by efficiently adapting the rate at which IoT devices disseminate monitoring streams based on run-time knowledge of the stream evolution, variability and seasonal behavior. Rather than transmitting the entire stream, ADMin favors sending updates for its estimation model from which values can be inferred, triggering dissemination only when shifts in the stream evolution are detected. Results on real-life testbeds, show that ADMin is able to reduce energy consumption by at least 83%, data volume by 71%, shift detection delays by 61% while maintaining accuracy above 91% in comparison to other IoT frameworks.

Real-time data processing while the velocity and volume of data generated keep increasing, as well as, energy-efficiency are great challenges of big data streaming which have transitioned to the Internet of Things (IoT) realm. In this paper, we introduce AdaM, a lightweight adaptive monitoring framework for smart battery-powered IoT devices with limited processing capabilities. AdaM, inexpensively and in place dynamically adapts the monitoring intensity and the amount of data disseminated through the network based on the current evolution and variability of the metric stream. Results on real-world testbeds, show that AdaM achieves a balance between efficiency and accuracy. Specifically, AdaM is capable of reducing data volume by 74%, energy consumption by at least 71%, while preserving a greater than 89% accuracy.

To optimize the cost and performance of complex cloud services under dynamic requirements, workflows and diverse cloud offerings, we rely on different elasticity control processes. An elasticity control process, when being enforced, produces effects in different parts of the cloud service. These effects normally evolve in time and depend on workload characteristics, and on the actions within the elasticity control process enforced. Therefore, understanding the effects on the behavior of the cloud service is of utter importance for runtime decision-making process, when controlling cloud service elasticity. In this paper, we present a novel methodology and a framework for estimating and evaluating cloud service elasticity behaviors. To estimate the elasticity behavior, we collect information concerning service structure, deployment, service runtime, control processes, and cloud infrastructure. Based on this information, we utilize clustering techniques to identify cloud service elasticity behavior, in time, and for different parts of the service. Knowledge about such behavior is utilized within a cloud service elasticity controller to substantially improve the selection and execution of elasticity control processes. These elasticity behavior estimations are successfully being used by our elasticity controller, in order to improve runtime decision quality. We evaluate our framework with three real-world cloud services in different application domains. Experiments show that we are able to estimate the behavior in 89.5% of the cases. Moreover, we have observed improvements in our elasticity controller, which takes better control decisions, and does not exhibit control oscillations.

Over the past decade, Cloud Computing has rapidly become a widely accepted paradigm with core concepts such as elasticity, scalability and on demand automatic resource provisioning emerging as next generation Cloud service-must have-properties. Automatic resource provisioning for Cloud applications is not a trivial task, requiring for both the applications and platform, to be constantly monitored, capturing information at various levels and time granularity. In this paper we describe the challenges that occur when monitoring elastically adaptive Cloud applications and to address these issues we present JCatascopia, a fully automated, multi-layer, interoperable Cloud Monitoring System. Experiments on different production Cloud platforms show that JCatascopia is a Monitoring System capable of supporting a fully automated Cloud resource provisioning system with proven interoperability, scalability and low runtime footprint. Most importantly, JCatascopia is able to adapt in a fully automatic manner when elasticity actions are enforced to an application deployment.