Live video analytics requires multi-stage processing pipeline with low-latency that traditional cloud-centric approaches cannot always provide. While edge computing brings resources closer to users, edge nodes are resource-limited, heterogeneous, and carry dynamic workloads. Current load-balancing strategies often overlook key contextual factors such as communication overhead and compute pressure. Without full "foreseeable" context, systems suffer from latency spikes and high frame loss under bursty workloads or mobile deployments.

Federated learning (FL) enabled creating models that are competitive to centralized machine learning models, without compromising user privacy. Participating FL clients train local models on their data and only share model weights. An FL server aggregates these weights into global weights that are pushed to clients for the next training round. Despite FL research growth, most of this work is conceived in experimental simulated environments that do not reflect its applicability to real-world scenarios. Also, existing open-source FL testbeds/frameworks have drawbacks that prohibit convenient deployment over a large spectrum of heterogeneous clients in realistic environments. These drawbacks include simulations, unrealistic data sets, not supporting heterogeneity, and not having realistic environment control in terms of network and client churn, for example. In this article, we introduce (RealFL) a novel, realistic, open-source, and extendable platform for FL that supports a large scale of heterogeneous clients. It enables a realistic assessment of FL solutions by controlling various environmental parameters, e.g., network, client churn, data distribution, training complexity, and client heterogeneity. Using these parameters, we assess RealFL performance through an extensive evaluation. Preliminary evaluation shows a performance gap of up to 72% in training time and 27% in accuracy between FL-simulated environments and RealFL. Moreover, extensive evaluation reveals that realistic environmental parameters could affect accuracy by up to 52.7%, training time by up to 77.5%, and communication overhead by up to 98%.

Federated Learning is being hailed as a privacy-preserving machine learning alternative, by allowing models to be distributively trained on source devices owning their data. Most FL solutions, and their assessments, however, assume superior environmental reliability, despite the more realistic variances in environmental factors such as device and network capacity, data distribution, and device churn. As such, we argue in this paper, that there is a growing chasm between current FL assessment setups and the evolving FL assessment needs. Motivated by this chasm, we conduct, to the best of our knowledge, the first empirical measurement study of FL performance given realistic environmental factors. Our study quantifies the impact of these environmental factors on FL performance in terms of training time, accuracy, and communication overhead. Our findings have broad implications for the future development of FL including client admission control and scheduling optimizations.

Federated Learning (FL) enabled creating models that are competitive to centralized Machine Learning models while preserving privacy by allowing clients to train data locally. Despite FL research growth, most of the work assessment and existing open-source FL testbeds/frameworks have drawbacks that prohibit convenient deployment over a large spectrum of heterogeneous clients in realistic environments. These drawbacks include simulations, unrealistic datasets, not supporting heterogeneity, and not having a realistic environment control in terms of network and client churn, for example. In this paper, we introduce (RealFL) a novel, realistic, open-source, and extendable platform for FL that supports a large scale of heterogeneous clients. It enables realistic assessment of FL solutions by controlling various environmental parameters; e.g. network, client churn, data distribution, training complexity, and client heterogeneity. Using these parameters, we assess RealFL performance through an extensive evaluation. The results show a performance gap of up to 77.5% in training time and 23.9% in accuracy between FL unrealistic environments and RealFL.

Federated Learning (FL) has enabled Machine Learning (ML) applications to capture a larger spectrum of data by allowing such data to remain on-device, a desirable privacy guarantee in many applications. However, the highly iterative nature of FL optimization algorithms requires low-latency and high-throughput connections to clients. Unfortunately, realistic FL training scenarios include heterogeneous clients that are restricted by computation and communication, thereby slowing down or even failing FL training. In this paper, we propose FedTeams; a trust-based and resource-aware FL system that minimizes training latency, while improving accuracy. To achieve this, we mitigate the risk of straggling and weakly-connected clients by leveraging social trust and allowing these clients to offload their data to more powerful trusted peers that can train on their behalf. In specific, we formulate and solve an optimization problem that leverages the FedTeam’s trust graph and client resource information to optimize the distribution of training and minimize training latency. We evaluate FedTeams in a simulated environment, demonstrating up to a 81.6% decrease in training latency and 11.2% increase in global model accuracy when compared to existing state-of-the-art solutions.

The affordability and deployment-flexibility of Unmanned Air Vehicles (UAVs) have ignited the development of many smart applications, including surveillance, disaster management, and smart farming. Drone's energy consumption is a critical issue and it can be controlled through different factors, depending on the application. One approach is to minimize energy consumption by defining a minimal number of strategic target-coverage locations that the drone needs to traverse and efficiently plan the drone's route through these locations. In this paper, we provide solutions that efficiently allow UAVs to cover multiple targets using their cameras. These solutions identify a minimum set of strategic locations that cover the targets and plan the drone's routes across these locations. We address the problem with the objective of minimizing the total energy consumed by the drone during its mission. We model the problem as mixed-integer programming problem and provide a set of heuristics; with and without target clustering. We evaluate the system using simulations. The results indicate the significance of clustering in minimizing the number of strategic locations and saving the drone's energy. Moreover, flexibility in selecting cluster centers provides further reduction in the strategic locations and energy consumption.

Mobile and IoT devices are becoming increasingly capable computing platforms that are often underutilized. In this paper, we propose RAMOS, a system that leverages the idle compute cycles in a group of heterogeneous mobile and IoT devices that can be clustered to form an edge FemtoCloud. At the heart of this system, we formulate a multi-objective, resource-aware task assignment and scheduling problem. The scheduler runs in two main modes; latency-minimization and energy-efficiency. Under the latency-minimization mode, it strives to maximize the computational throughput of the constructed FemtoCloud while maintaining the energy consumption below an operator specified threshold. Under the energy-efficient mode, it minimizes the total energy consumed in the FemtoCloud while meeting defined tasks deadlines. Due to the NP-Completeness of this scheduling problem, we design a set of heuristics to solve it. We implement a prototype of our system and use it to evaluate its performance and efficiency. Our results demonstrate the system's ability to meet different scheduling objectives while adhering to pre-specified time and energy constraints. Compared to other schedulers, RAMOS achieves 10 to 40 percent completion time improvement under latency minimization mode and up to 30 percent more energy-efficiency under the energy-efficient mode.

Many novel IoT-based applications now require large compute resources, high-privacy, and low-latency. This demand has triggered the rise of fog and edge computing to complement the high-latency and low-privacy cloud. Fog computing provides lower latency by bringing computational servers closer to the user, typically within the city's vicinity. However, due to the high cost of deploying such fog servers at scale, and poor network infrastructures in many countries and areas, edge computing has been introduced. Edge computing argues for leveraging compute resources, typically within a user's immediate environment, on distributed ensembles of devices called FemtoClouds. In this paper, we propose Maestro, a system that aids users by offloading computational jobs from them to multiple FemtoClouds in their immediate vicinity. We propose an integrated architecture for Maestro, which incorporates a new scheduling algorithm that assigns compute tasks to FemtoClouds. We implement a full prototype of Maestro, and evaluate its performance on our experimental testbed, as well as through emulation. Our results show that our system and scheduler outperforms state-of-the-art by up to 55%.