Edge nodes host latency-sensitive applications on fixed and resource-limited hardware. Under moderate load increases, contention for shared CPU, network, and I/O resources introduces kernel-level delays that result in tail-latency degradation. In practice, many edge workloads are black-box services, making application-level instrumentation impractical. Existing resource metrics such as CPU utilization and Linux Pressure Stall Information (PSI) often fail to explain latency behavior, especially when they saturate. We present K-Sense, a non-invasive eBPF-based framework that infers QoS degradation from kernel behavior. K-Sense collects a small set of kernel-level delay metrics and represents the system state as a point in a multi-dimensional feature space. It then computes a covariance-aware friction signal using the Mahalanobis distance to quantify deviation from a calibrated normal operating region. We evaluate K-Sense on a Kubernetes testbed using real microservice workloads from DeathStarBench and a sentimentanalysis application. Results show that friction tracks application P99 latency under changing load and continues to reflect latency variations when CPU utilization and PSI saturate. These signals provide a practical basis for higher-level actions such as admission control, scheduling, and workload migration without requiring application modification.
Solid-state drives (SSDs) are deployed in fog, edge, and cloud environments owing to their high performance and reliability, making their lifespan critical concerns. Zoned Namespace SSDs (ZNS SSDs) address garbage collection (GC) overhead caused by out-of-place updates in conventional SSDs, providing stable and high write performance. RocksDB/ZenFS is a prominent system that leverages ZNS SSDs, where RocksDB, an LSM-tree-based key-value store, utilizes ZNS SSDs as storage through the ZenFS filesystem. To fully exploit RocksDB's write performance, multiple compaction threads are typically utilized. However, the existing zone allocation scheme in ZenFS often leads to high write amplification and frequent zone finish operations in multi-threaded compaction environments, thereby shortening the lifespan of ZNS SSDs. To address this issue, we propose Lifetime Ratio-based Zone Allocation (LRZA), a zone allocation scheme that considers the lifetime ratios of files grouped within a zone. LRZA eliminates the need for zone finish operations during allocation and effectively reduces write amplification caused by garbage collection. Experimental results demonstrate that LRZA reduces the number of used zones and zone reset operations by $\mathbf{1 0} \boldsymbol{\%}$ and $\mathbf{1 4} \boldsymbol{\%}$, respectively, compared to ZenFS. These improvements contribute to extending the operational lifespan of ZNS SSDs while maintaining efficient write performance.
Processing real-time data in the cloud has become the basis for various IoT applications. This processing typically incurs a high pricing cost, due to the use of paid messaging services (e.g., AWS SQS, or GCP Pub/Sub) that send real-time data from IoT devices at the edge to IoT applications in the cloud. Alternatively, costly AI-enabled IoT devices are employed for federated learning at the edge, which lowers the amount of data sent to the cloud. To reduce such costs, we propose an approach that sends real-time data from the IoT device to the cloud with reduced sampling frequency. In the cloud, we use AI to predict the missing data, and send both real and predicted data to the IoT application. This way, the IoT application receives the same amount of data, although less data is sent from the edge to the cloud (using messaging services). In addition, we measure the prediction error in the cloud, and adjust the sampling frequency at the edge, so that the error remains low. Thus, this approach lowers the cost of IoT applications by reducing the use of paid messaging services, and avoiding costly AI-enabled IoT devices, while also ensuring a low prediction error. For the evaluation, we build a prototype, and we experiment with real-world IoT data. Our experiments show promising results, including a reduction in costs by about 70% with a prediction error of around 5%.
We address joint task offloading and resource allocation in multi-user multi-server Mobile Edge Computing (MEC) networks, formulated as a Mixed-Integer Non-Linear Programming (MINLP) problem minimizing weighted delay and energy consumption. We propose a bi-level decomposition framework: an outer loop using an intention-enhanced Proximal Policy Optimization (PPO) agent with multi-head self-attention for binary offloading decisions, and an inner convex optimization loop for optimal task division and CPU frequency allocation. Simulations demonstrate superior convergence stability and reduced system costs over DRL and heuristic baselines.
Integrated sensing, computing, and communication (ISCC) networks combine both mobile edge computing (MEC) and sensing functionalities. Joint uncrewed aerial vehicle (UAV) trajectory planning and resource allocation are essential tasks for UAV-enabled ISCC networks. However, existing works primarily rely on bit-level data transmission, which is sensitive to errors in unstable post-disaster channels and incurs excessive offloading delays and limited bandwidth resources. Motivated by this, we incorporate semantic communication to substitute conventional bit-level transmission for robust and lightweight data offloading, referred to as an integrated sensing, semantic computing, and communication (ISSCC) system. An optimization of joint UAV trajectory planning, resource allocation, and semantic compression decision is established to minimize the average age of information (AoI) for both MEC and sensing functionalities. To tackle this problem, we propose a deep reinforcement learning (DRL)-based framework where the state-of-the-art actor-critic schemes are incorporated and a continuous relaxation and quantization strategy is adopted. Experiments are conducted in a synthetic post-disaster smart city scenario, and the results demonstrate the effectiveness and superiority of the proposed algorithms. In addition, the introduced semantic communication in the ISSCC system helps reduce the AoI by $\mathbf{6 8. 0 8} \boldsymbol{\%}$ compared with the conventional ISCC system.
The proliferation of mobile edge computing (MEC) has transformed our lifestyle by connecting various objects around us and forming large-scale networks composed of heterogeneous devices. However, this highly distributed architecture also introduces severe new challenges: software vulnerabilities in edge environments lead to serious issues in security. Machine learning-based attack detection is considered as an effective approach to addressing these challenges. While machine learning methods show potential in network intrusion detection, they often struggle to accurately identify some novel malicious behavior patterns due to limitations in generalization. To address these limitations, we propose a novel LLM-driven hierarchical defense framework enhancing threat detection and analysis in MEC environments. Our approach directly processes text-based network traffic using advanced large language model frameworks such as Llama 2 and distributed nodes collaborate using federated learning method to automatically extract attack characteristics. To alleviate communication overhead and computational pressure on end devices, we implement an intermediate edge layer to perform local training and aggregation. The proposed method is evaluated using the CIC-IDS2017 dataset and compare with existing intrusion detection systems. Experimental results indicate that our framework achieves comparable detection performance to a centralized baseline while reducing GPU memory usage and communication latency.
Zoned Namespace (ZNS) SSDs expose internal zone structures to improve performance and endurance, making them attractive for data-intensive workloads in cloud, edge, and fog computing environments. However, evaluating zone placement policies remains difficult due to limited runtime visibility into internal resource utilization. We develop femonitor, a lightweight, real-time monitoring tool for FEMU-based ZNS emulators that collects per-channel and per-plane statistics via vendor-specific NVMe admin commands. femonitor enables fine-grained observation of internal SSD behavior, which is particularly useful for evaluating or debugging storage management strategies. Using femonitor, we demonstrate that various zone placement policies allocate internal resources as intended. Our results show that femonitor introduces negligible overhead (below 0.5%), even with a 1ms polling interval, making it suitable for continuous runtime monitoring of ZNS-based storage systems.
The rapid growth of video streaming services and the increasing coexistence of multiple content providers have placed significant demands on network infrastructure. Efficient content caching is crucial to improving user experience and optimizing network resource utilization. Traditional content delivery networks (CDNs) struggle to meet the dynamic requirements of multi-provider environments. In this paper, we address the problem of multi-provider content caching by adopting fog computing to leverage localized storage and processing capabilities. To this extent, we develop CAMP (Cluster-Aware Multi-Provider Caching), a novel cooperative caching framework designed for multi-tier fog architectures. CAMP consists of two phases: resource allocation and content placement. In particular, the first phase clusters users based on location and provider affiliation, hierarchically assigns fog nodes using a fuzzy membership model, and partitions fog storage proportionally among clusters. The second phase employs a cooperative content placement algorithm that dynamically adjusts content redundancy based on popularity to ensure high availability for popular content while minimizing redundancy for infrequently accessed data. Through extensive simulation using real-world datasets, CAMP demonstrates up to 71% reduction in latency and substantial improvement in fog hit ratios, outperforming state-of-the-art caching strategies.
In Hierarchical Federated Learning (HFL), bandwidth budgets and delay constraints can prevent the timely synchronization of the semantic atlases (summaries) required for clients to learn unseen classes. Conventional synchronization policies typically optimize for geometric freshness. However, staleness becomes most harmful when it changes discrete curriculum gating decisions. We find that reducing feature space drift alone is an unreliable proxy for knowledge void recovery under constraints, because learning depends on decision consistency. Our Decision-Critical Atlas Synchronization (DCAS) policy prioritizes updates that are likely to alter these decisions and accelerates recovery from knowledge voids compared to a geometry-centric baseline.
Deploying Transformer based detectors on mobile edge devices is often limited not by FLOPs, but by deployability: (i) graph fragmentation, where unsupported operators trigger CPU fallbacks and expensive CPU↔NPU synchronization, and (ii) quantization instability, where attention driven activation outliers cause INT8 compilation or runtime failures on fixed-point NPUs. Focusing on RF-DETR-Nano on a Snapdragon 8 Gen 3 NPU, we propose an Order-Aware Compression pipeline that composes Structured Pruning (P), Knowledge Distillation (K), and detector aware Post-Training Quantization (Q) in the specific order $\mathbf{P} \rightarrow \mathbf{K} \rightarrow \mathbf{Q}$. Our key insight is that distilling after pruning adapts the student to the pruned structure and reshapes activation statistics, enabling stable INT8 deployment where naive PTQ crashes (e.g., numerical overflow). On COCO at 384×384, our method converts a fragmented FP32 deployment (61.5 ms p95, CPU fallback) into an NPU-only INT8 model $(\mathcal{F}=0 \%)$ with 5.2 ms p95 latency while maintaining 47.8 mAP, yielding an $11.8 \times$ speedup over FP32 hybrid execution and a 1.5× speedup over native FP16.
Transformer-based models such as BERT have achieved remarkable success across a wide range of natural language processing tasks. However, deploying them on resource-constrained devices, including System-on-Chips (SoCs), remains challenging due to limited memory and computational capacity of these devices. Weight sharing has proved to be an effective approach for significantly reducing model size, yet designing an efficient layer-wise weight sharing framework that achieves high compression while maintaining model performance and low latency is not a trivial task. In this paper, we systematically analyze the impact of key weight sharing hyperparameters on model performance, model size, and latency, and propose a Genetic Algorithm-based optimization framework to automatically identify the optimal configuration for each layer—balancing performance, compression ratio, and inference speed. Additionally, through targeted modifications and efficient on-the-fly decompression, we demonstrate that the proposed weight sharing framework introduces no additional latency overhead and can even reduce inference time. Compared to uniform compression, our approach makes transformer blocks with substantially lower memory requirements and reduced latency while maintaining competitive performance across evaluation metrics.
Model parallelism, including pipeline and tensor parallelism, has been widely adopted to distribute computation across devices and enable large language models (LLMs) to run on resource-constrained edge clusters. However, the prefill and decoding phases of generative LLM inference exhibit distinct computational characteristics and compute-communication tradeoffs, making a single parallelism strategy suboptimal for both. Existing phase-aware techniques primarily target large GPU clusters optimized for high-throughput, batched inference, whereas edge environments typically process small batches or individual requests and prioritize end-to-end inference latency. To address this gap, we present LLMEdger, a latency-optimized, phase-aware parallelism scheduler for generative LLM inference on edge. LLMEdger dynamically selects pipeline and tensor parallelism configurations for the prefill and decoding phases, adapting to input/output lengths and heterogeneous device memory to minimize end-to-end latency. We implement and evaluate LLMEdger on CPU-based edge clusters using the 11B-parameter FLAN-T5-XXL model, which does not fit on a single edge device. Our results show that LLMEdger reduces inference latency by up to 14.95% compared to non-phase-aware parallelization and achieves up to $\text{2. 1 2} \times$ speedup over single-node execution.
The placement of Kubernetes control-plane nodes is critical to ensuring cluster reliability, scalability, and performance, and therefore represents a significant deployment challenge in heterogeneous, multi-region environments. Existing initialisation procedures typically select control-plane hosts arbitrarily, without considering node resource capacity or network topology, often leading to suboptimal cluster performance and reduced resilience. Given Kubernetes's status as the de facto standard for container orchestration, there is a need to rigorously evaluate how control-plane node placement influences the overall performance of the cluster operating across multiple regions. This paper advances this goal by introducing an intelligent methodology for selecting control-plane node placement across dynamically selected Cloud-Edge resources spanning multiple regions, as part of an automated orchestration system. More specifically, we propose a reinforcement learning framework based on neural contextual bandits that observes operational performance and learns optimal control-plane placement policies from infrastructure characteristics. Experimental evaluation across several geographically distributed regions and multiple cluster configurations demonstrates substantial performance improvements over several baseline approaches.
For deep learning inference on edge devices, hardware configurations achieving the same throughput can differ by 2× in power consumption, yet operators often struggle to find the efficient ones without exhaustive profiling. Existing approaches often rely on inefficient static presets or require expensive offline profiling that must be repeated for each new model or device. To address this problem, we present CORAL, an online optimization method that discovers near-optimal configurations without offline profiling. CORAL leverages distance covariance to statistically capture the non-linear dependencies between hardware settings, e.g., DVFS and concurrency levels, and performance metrics. Unlike prior work, we explicitly formulate the challenge as a throughput-power co-optimization problem to satisfy power budgets and throughput targets simultaneously. We evaluate CORAL on two NVIDIA Jetson devices across three object detection models ranging from lightweight to heavyweight. In single-target scenarios, CORAL achieves 96
Unmanned Aerial Vehicles (UAVs) have emerged as a transformative technology for real-time road traffic monitoring, offering enhanced efficiency and responsiveness to modern traffic management systems. However, the resource limitations of UAVs and the dynamic nature of traffic densities present significant challenges for continuous operation. To address these constraints, this study proposes a vehicle-density-aware adaptive offloading mechanism that dynamically alternates between local processing and task offloading to fog nodes, based on real-time traffic conditions. The mechanism operates in three distinct modes: Low-CPU Mode for low vehicle density, Full Offloading Mode for moderate density, and Local Processing Mode for high-density scenarios. Preliminary results reveal that the proposed VD-aware adaptive offloading mechanism effectively balances performance, resource efficiency, and communication costs. It maintains competitive accuracy, optimizes throughput, and dynamically manages CPU utilization and communication overhead. These findings highlight the adaptability and efficiency of the proposed mechanism, making it an ideal solution for UAV-based road traffic monitoring in dynamic and resource-constrained environments.
Containerization has transformed application deployment across diverse environments in the edge-cloud computing continuum, providing lightweight, portable solutions that integrate seamlessly across various cloud-native environments. However, default scheduling often leads to resource underutilization due to overestimated requests. This results in inaccurate demand prediction and static allocation strategies. This paper enhances a bottom-up, self-organizing approach with an NNbased peer selection mechanism, enabling intelligent resource allocation while maintaining decentralization. The proposed NNbased approach enables a data-driven selection strategy that significantly improves resource utilization and system performance; a 21% improvement in resource utilization is reported in a high-traffic scenario compared to the predecessor framework. The findings contribute to the broader research community by demonstrating the potential of combining agent-based modeling with machine learning techniques, paving the way for more adaptive and efficient edge orchestration systems.
This paper presents a novel data management approach for efficient data search and movement in a Distributed Knowledge Graph (DKG) within the Edge-Fog-Cloud continuum. The aim is to minimize data traffic and query time by relocating data closer to where it is needed. We propose a variation of the Ant Colony Optimization algorithm, using pheromone abstraction to identify optimal paths between network nodes. Once established, the algorithm moves data along these paths, evaluating the necessity of each relocation. We evaluate our approach through computer simulations, using query hit rates as the primary metric. We compare hit rates before and after data movement to assess algorithm efficiency. Our findings identify parameters that enhance search efficiency, achieving high hit rates while conserving resources. Future work will involve incorporating realistic network properties and exploring practical implementations.
The increased usage of Internet of Things devices at the network edge and the proliferation of microservice-based applications create new orchestration challenges in Edge computing. These include detecting overutilized resources and scaling out overloaded microservices in response to surging requests. This work presents ADApt, an extension of the ADA-PIPE tool developed in the DataCloud project, by monitoring Edge devices, detecting the utilization-based anomalies of processor or memory, investigating the scalability in microservices, and adapting the application executions. To reduce the overutilization bottleneck, we first explore monitored devices executing microservices over various time slots, detecting overutilization-based processing events, and scoring them. Thereafter, based on the memory requirements, ADApt predicts the processing requirements of the microservices and estimates the number of replicas running on the overutilized devices. The prediction results show that the gradient boosting regression-based replica prediction reduces the MAE, MAPE, and RMSE compared to others. Moreover, ADApt can estimate the number of replicas close to the actual data and reduce the CPU utilization of the device by 14%-28%.
This paper presents a real-time detection, tracking, counting, and distance estimation framework deployed on the Boston Dynamics Spot robot, equipped with RGB and thermal cameras. Leveraging edge computing devices such as the NVIDIA Jetson Nano, the system autonomously processes data in dynamic terrains with minimal human intervention. A custom-trained YOLOv8 model, fine-tuned on a unique dataset tailored for military applications, is integrated with the StrongSORT algorithm for object tracking. Additionally, a novel geometric calculation methodology enables precise angle estimation and spatial mapping, enhancing situational awareness. The framework's capabilities include live visualization of detected objects, area-based counting, and mapping within an operational environment using ROS-based tools. Field demonstrations conducted at the Army Research Lab validate the system's effectiveness in processing complex thermal and RGB data in real-time. The proposed solution offers significant potential for enhancing autonomous robotic deployments in mission-critical applications.
Federated Learning (FL) has emerged as a powerful approach for decentralized model training, yet its deployment in large-scale Internet of Things (IoT) environments faces significant challenges. These challenges include fluctuating bandwidth, frequent node failures, and the resource constraints. Such issues are particularly amplified in multilayer Edge-Fog-Cloud infrastructures. Traditional single-layer FL frameworks often fail to address these issues, leading to disrupted training and poor scalability. To tackle these challenges, we propose FTFormer, a novel fault-tolerant, Transformer-based layer offloading framework designed for multilayer federated split learning. FTFormer leverages a Transformer-based policy network to capture the complex interdependencies among Edge, Fog, and Cloud nodes, including bandwidth variability, compute power, and failure probabilities. Combined with an online Proximal Policy Optimization (PPO) algorithm, the framework dynamically adapts offloading decisions in real time, ensuring efficient task allocation under dynamic conditions. Additionally, FTFormer integrates fault-tolerance mechanisms that enable task re-routing and backup deployment to mitigate the impact of node failures and overloads, maintaining smooth training progress. Using a large-scale event-driven simulator capable of modeling thousands of Edge devices and hundreds of Fog/Cloud nodes, we validate FTFormer's performance. Experimental results show that FTFormer significantly improves training speed, fault resilience, and scalability, outperforming state-of-the-art techniques under high-load and failure-prone scenarios. This work highlights FTFormer as a robust solution for deploying resilient FL in real-world IoT systems.
