While anomaly detection is essential for cloud computing and predictive maintenance, approaches that bridge continual learning with environmental sustainability remain largely unexplored. In production environments, evolving data distributions cause performance degradation of machine learning models, and naive adaptation may lead to forgetting of past patterns that are likely to reoccur. In this study, we formulate temporal concept-incremental anomaly detection for both univariate and multivariate time series. We introduce time-based and entity-based concept extraction strategies together with Wasserstein-distance diagnostics for inter- and intra-concept distribution drift, and define an evaluation protocol that measures detection quality, knowledge retention, and energy consumption. We further propose Selective Temporal Replay (STR), a replay strategy for autoencoder-based continual anomaly detection that combines diversity-based initialization, novelty-gated admission, retention-aware replacement, and age-weighted preservation of stored sequences. The empirical study comprises more than 5,400 experiments across four datasets, nine autoencoder architectures, and ten continual learning strategies. On Yahoo! A1, STR obtains the highest ROC-AUC in seven of the nine architectures and is the only strategy with positive backward transfer on all four datasets; under the smallest memory budgets, it obtains the highest ROC-AUC among replay methods on Yahoo! A1 and SMD. The sustainability analysis identifies when selective replay justifies its additional energy cost and when regularization-based continual learning strategies are preferable.
The increasing density and heterogeneity of Internet of Things (IoT) deployments impose stringent requirements on latency, energy efficiency, and reliability within edge computing environments. These systems typically operate under highly dynamic, bursty workloads, where rapid fluctuations in queue occupancy and resource utilization can lead to congestion, increased latency, and deadline violations. Conventional task offloading strategies, which rely on static heuristics or data-intensive machine learning models, often fail to achieve real-time adaptability, low computational overhead, and decentralized decision-making simultaneously. This study introduces an event-driven neuromorphic task offloading framework for edge-IoT systems utilizing Spiking Neural Networks (SNNs). The offloading problem is formulated as a multi-objective optimization process that jointly addresses latency, energy consumption, and deadline compliance. An SNN-based controller is integrated into edge nodes to approximate optimal decisions via spike-driven temporal dynamics, enabling adaptive, energy-aware orchestration without centralized optimization or continuous model inference. The proposed framework is evaluated using a hybrid simulation environment that integrates YAFS and Brian2. Theoretical analysis demonstrates bounded decision dynamics and sparse computational complexity. Experimental results indicate that the proposed approach reduces latency by up to 26
Cloud-assisted vehicular computing enables latency-sensitive applications by offloading tasks among vehicles, roadside units (RSUs), and remote clouds, yet it faces stringent challenges due to rapid topology changes, heterogeneous resources, and diverse task requirements. This paper presents a comprehensive study of task offloading and resource scheduling in a vehicle–RSU–cloud environment. We propose a hierarchical computation framework that flexibly supports task execution across local vehicles, neighboring vehicles, RSUs, and cloud resources. To capture realistic operational constraints, we develop joint models incorporating task attributes, security requirements, pricing factors, and heterogeneous computing/communication capabilities. Building on these models, we design an efficient task migration and resource scheduling strategy that improves overall system performance under dynamic network conditions. Extensive simulation-based evaluation and comparative analysis demonstrate that the proposed approach enhances offloading efficiency and robustness, especially in scenarios with multiple scheduling targets, offering a practical basis for scalable vehicular edge–cloud collaboration.
The rapid expansion of the Internet of Vehicles (IoV) has exacerbated challenges in computation offloading. In such highly dynamic environments, the simultaneous optimization of latency and energy consumption is critical. Existing approaches, such as standard Deep Q-Networks (DQN), often struggle with slow convergence and difficulties in achieving an optimal trade-off between metrics when confronted with large state spaces and temporal dependencies induced by high vehicular mobility. This paper proposes LA-DRQN, a novel hierarchical decision-making framework that synergistically integrates Learning Automata (LA) at the strategic level for macro-policy selection with a Deep Recurrent Q-Network (DRQN) at the tactical level for fine-tuning computation offloading ratios. Key contributions include: (1) replacing stochastic mobility models with Cellular Automata (CA)-based simulations to enhance environmental realism; (2) designing a dynamic Roadside Unit (RSU) connectivity mechanism incorporating distance-dependent latency and bandwidth modeling; and (3) developing an instantaneous multi-objective reward function to expedite convergence. Extensive experiments in iFogSim environment over approximately 7,200 episodes demonstrate the superiority of LA-DRQN over standard baselines and state-of-the-art algorithms. Experimental results reveal that the proposed method achieves a 24.7
Disk bandwidth is a critical resource for I/O-intensive applications that must transfer large volumes of data to and from persistent storage. Most multi-tenant infrastructures efficiently allocate CPU and memory resources to concurrent workloads, but typically lack mechanisms for allocating I/O bandwidth. As a result, users often resort to exclusive node reservations to avoid disk contention, which can lead to underutilisation of other node resources if not fully exploited. Another common issue is that users do not know the exact resource requirements of their applications. Even when this is known, applications rarely maintain peak resource usage throughout their entire execution, resulting in wasted resources that could otherwise benefit other users. Today, many users prefer cloud serverless platforms because of their ease of use and flexible billing. However, these platforms have inherent limitations and may not be suitable for workloads with specific requirements. In this paper, we present a serverless scaling mechanism that dynamically adjusts disk I/O bandwidth for containerised applications by scaling their allocation up or down based on real-time usage and configurable weights. In addition, the system incorporates automatic extension management capabilities for virtual disk devices, such as logical volumes. Our approach can be integrated with other serverless scaling mechanisms, such as CPU and memory management, to provide a comprehensive resource scaling solution. The experimental results have shown significant performance improvements, with overall runtime reductions of up to 53
The grid computing infrastructure requires runtime systems that handle three simultaneous tasks of managing data locations, tracking communication costs, and maintaining system operation stability to execute its large-scale iterative and data-parallel workloads. Existing runtime systems fail to achieve proper dynamic execution conditions because they treat data placement scheduling and synchronization as separate elements, which creates unstable operations. The paper introduces a complete runtime control system that achieves optimal results through its base-driven model for making migration scheduling and coordination choices. The proposed model integrates both locality awareness and migration cost with a Lyapunov-inspired stability control component into a unified feedback system. The formal model establishes computation, communication, and synchronization expenses, which combine with queue-based coordination feedback to direct runtime choices through an established process. The framework evaluation occurs through a controlled event-driven simulation environment that replicates the analytical model's runtime control structure. The evaluation uses various workload intensity configurations to assess how the proposed method performs against typical static and heuristic dynamic runtime strategies. The experimental analysis studies three areas of execution efficiency, communication and synchronization overhead, and runtime stability. The results show that coordination stability integration into runtime control systems creates better execution reliability and execution prediction accuracy for grid-based iterative processing systems, while unified feedback-driven runtime management becomes essential for future large-scale distributed analytics platforms.
Dynamic tasks or resources localization in distributed edge computing systems with inaccuracies in node positioning represent a critical research direction. Existing closed-form methods often introduce auxiliary variables to pseudo-linearize the performance models, followed by parameter estimation via weighted least squares (WLS). Although such a procedure is frequently refined in a subsequent step, the inclusion of extra variables tends to amplify estimation errors, causing the results to deviate from the Cramér–Rao lower bound (CRLB) under high-noise scenarios. To address this limitation, we propose a two-phase closed-form algorithm. In the first phase, extra variables are eliminated using an orthogonal projection matrix, and an initial solution is derived via least squares (LS). Since the omitted terms contain relevant information about task location and dynamics, this preliminary estimate remains suboptimal. A second phase is therefore designed to enhance its accuracy. The proposed estimator has potential for real-time implementation, because it avoids iterative convergence and provides deterministic computational steps. Both theoretical analysis and numerical simulations demonstrate that the proposed estimator achieves the CRLB under moderate Gaussian noise conditions. Simulation results further confirm the computational efficiency of the method and its superiority over existing closed-form alternatives.
Serverless architecture enables the execution of intelligent applications without the management of underlying infrastructure. As cloud services continue to evolve, containers are proving to be a full and efficient solution in virtualization technology for serverless environments. However, resource allocation in serverless systems becomes more complex due to the inappropriate scheduling of functions. Moreover, response latency is another crucial aspect of the serverless paradigm which causes a significant impact on the performance. Therefore, to tackle the aforementioned challenges and to balance the trade-off between cost, resource utilization, and response time in serverless clusters, a multi-agent deep reinforcement learning model for function scheduling (MDRL-FP) utilizing the Proximal Policy Optimization (PPO) is proposed. The MDRL-FP model addresses the aforementioned challenges by considering Virtual Machine (VM) overheads, types, and constraints. The MDRL-FP model identifies load patterns and VM categories from historical data traces to generate efficient allocation policies. The results achieved through experiments show that the proposed model achieves a reduction of 2-24
LightCam is a lightweight, privacy-preserving federated learning framework for resource-constrained camera networks in smart city surveillance. It targets three key challenges: strict protection of identity-bearing video and audio, limited computation and bandwidth at edge cameras, and heterogeneous data and system conditions. LightCam decomposes a multimodal model into a frozen backbone and a small task head, so cameras only train and upload the head, greatly reducing local cost and exposure of shallow features. Local data are split into non-sensitive anchor samples and privacy-critical samples; DP-SGD is applied only to the private subset with a scene-aware dynamic privacy budget. A resource-aware aggregation mechanism further weights client updates by data volume, update stability, and device status. Experiments on a multimodal benchmark constructed from CelebA, CIFAR-10, and UrbanSound8K show that LightCam achieves 85.82 ε =2.0 ), while reducing per-round communication by over two orders of magnitude and maintaining stable convergence under heterogeneous and impaired clients.
The growing number of Internet of Things (IoT) devices and interconnected systems has significantly increased computational demands, necessitating efficient task scheduling in Fog-Cloud-IoT (FCIoT) networks. While fog computing offers reduced latency, optimal performance is challenged by energy constraints and dynamic network conditions. This work formulates the FCIoT task scheduling problem and proposes an energy-efficient, deadline-aware method incorporating a dynamic thresholding mechanism. We enhance the Heat Transfer Relation-based Optimization Algorithm (HTOA) to create an Improved HTOA (IHTOA), which adaptively schedules tasks across fog and cloud layers. Simulation results demonstrate that IHTOA substantially improves task scheduling in FCIoT environments. The proposed method significantly outperforms other approaches, demonstrating notable improvements over Extended Classifier System (XCS), Golden Eagle Optimizer (GEO), Non-Dominated Sorting Genetic Algorithm II (NSGA-II), and HTOA in resource allocation efficiency and average response time. Furthermore, it substantially reduces deadline violation occurrences and extends network lifetime compared to the other approaches.
Achieving reliable real-time responsiveness under highly dynamic 6G conditions remains a challenging problem, despite recent progress in edge intelligence and cloud-assisted inference. This challenge is particularly critical in latency-sensitive applications, where even small communication or processing delays can significantly degrade system performance. The issue becomes more pronounced in sensorimotor feedback systems, where delays directly affect control accuracy, system stability, and user experience. Therefore, there is a need for adaptive and learning-driven mechanisms that can maintain ultra-low latency and robust performance under fluctuating network conditions. In this paper, we propose a Violin Real-Time Responsive Task Distributed Collaborative Model (VRRT–DCM), which coordinates inference across sensing, edge, and cloud layers to ensure timely response while supporting long-term analytical processing. Unlike conventional static partitioning approaches, the proposed framework integrates three key components: (i) a lightweight LSTM–CNN hybrid model designed for resource-constrained edge environments, (ii) a divergence-aware dynamic task partitioning strategy that balances latency, energy consumption, and model consistency, and (iii) an asynchronous edge–cloud synchronization mechanism that enables continuous learning without disrupting real-time operations. Furthermore, we formulate a joint optimization problem that unifies inference latency, energy expenditure, and model drift into a single utility-driven decision framework. The proposed approach is evaluated through hardware-in-the-loop experiments and trace-driven simulations under realistic 6G network conditions. Experimental results demonstrate that VRRT–DCM achieves sub-30 ms latency, reduces communication overhead by more than 40
Virtualization is a transformative technology that enables multiple virtual machines to operate concurrently on a single physical host, thereby rationalizing resource utilization in cloud data centers. These data centers consist of multiple servers and consume a significant amount of energy, which requires cloud providers to have mechanisms that optimize the placement of virtual machines in these servers. An optimal placement reduces energy consumption and improves resource utilization rate. In this paper, we introduce a novel algorithm based on a recent bio-inspired metaheuristic called Manta Ray Foraging Optimization (MRFO) to solve the virtual machine placement (VMP) problem. Although MRFO has been successfully applied to various engineering optimization tasks, to the best of our knowledge, it has never been used in the context of VMP. Accordingly, this paper investigates for the first time an enhanced version of the MRFO algorithm adapted for VMP problem in cloud data centers. The proposed algorithm is evaluated with CloudSim toolkit under various performance metrics, including energy consumption, resource utilization, and number of active servers, under several cloud environment heterogeneity levels. Simulation results are compared with several baseline and hybrid algorithms widely adopted in this field. The proposed algorithm achieves an improvement of 11.53
Fog computing extends cloud capabilities to the network edge, enabling low-latency and context-aware service delivery for IoT applications. Deployment of the Fog devices plays a major role in offering the optimal services to the end users. While existing Fog Device Deployment (FDD) strategies focus on maximising connectivity and edge coverage, they often overlook logical edge-to-edge communication and network cohesion. This work extends the existing JAYA based FDD model by deriving a modified fitness formulation that penalises fragmented networks using disjoint edge clusters and includes logical edge-edge communication through shared fog nodes. A detailed comparison of these two fitness functions exhibits that the proposed formulation produces much more cohesive and well-connected Fog device network. A comparative analysis on synthetic topologies demonstrates superior cohesion, reduced fragmentation, and smoother convergence, validating the proposed improvements.
Cloud computing has emerged as the backbone of modern digital infrastructure, offering scalability, flexibility, and cost efficiency through its on-demand and pay-as-you-go models. However, the exponential growth in virtualized workloads has intensified the challenges of optimal resource allocation and energy management in large-scale data centers. Existing methods often rely on static or heuristic bin-packing strategies that fail to dynamically adapt to fluctuating workloads, leading to resource imbalance and increased power consumption. The primary objective of this work is to minimize energy consumption and enhance utilization efficiency by optimizing multidimensional resources, including CPU, RAM, and bandwidth, through dynamic learning. A Quadrant-based Difference (QBD) mechanism is proposed to handle the 3D resources. Further, QBD with Reinforcement Learning (QBDRL) for 3D Bin Packing problem is proposed for energy-efficient Virtual Machine (VM) placement. The proposed QBDRL algorithm integrates a QBD with Q-Learning to continuously adjust VM placement according to workload variations, reducing the number of active hosts while maintaining performance stability. Experimental evaluation on the CloudSim platform using PlanetLab traces demonstrates that QBD and QBDRL achieve 44.12
With the rapid evolution of intelligent transportation systems in next-generation networks, the efficient and accurate object detection has become essential for maintaining the highway safety and reliability. Although the centralized or cloud-based object detection provides huge computing resources, it inevitably suffers from the high latency and bandwidth overhead. On the contrast, the edge-only computing paradigm is with low latency, but has to face challenges such as the small object pixel proportion, low localization accuracy of general models and small crack semantic segmentation. To bring together advantages from both cloud and edge computing, we propose a cloud–edge collaborative two-stage object detection and semantic segmentation method. First, to tackle the weak detection by the small proportion of object pixels, a detection model based on super-resolution feature generation is developed, which processes the dataset by removing low-quality subsets and supplementing the self-made data, and then uses relevant backbone networks with other components to enhance the representation ability of small targets. Second, for the semantic segmentation of small targets, a segmentation model based on the local perception is proposed. In particular, by applying it to Mask2Former with an additional auxiliary convolution layer, the model can well capture local details and low-dimensional semantic information, improving the segmentation precision. Finally, experiments demonstrate that the proposed cloud–edge collaborative two-stage method achieves the higher detection accuracy and richer segmentation details over existing models.
Apache Spark has become a cornerstone of modern big data analytics, increasingly powered by the scalability and cost efficiency of cloud computing. However, the heterogeneity of cluster resources and the diverse resource-performance characteristics of Spark applications make resource allocation and scheduling under dynamic workloads inherently challenging. Existing studies on Spark resource optimization primarily target single-application settings and typically do not explicitly model system-level behavior under dynamically arriving workloads. Meanwhile, research on Spark application scheduling has largely focused on task or executor scheduling, with resource allocation often handled in an ad hoc or heuristic manner. In this paper, we propose a unified framework for Spark application scheduling that integrates fine-grained executor-level resource allocation. Building upon this framework, we present a scheduling algorithm that jointly optimizes cost, application performance, and makespan via deep reinforcement learning–based resource optimization. To support effective RL training, we further construct performance prediction models for Spark applications and design a simulator that captures key cluster characteristics relevant to scheduling. Extensive experiments conducted in both simulated and real environments, using real workload traces, demonstrate that our proposed algorithm achieves a better trade-off among cost, application performance, and makespan compared with baseline strategies.
As the demand for utilizing HPC resources in large-scale scientific computing continues to grow, LHC experiments such as ALICE are actively pursuing the integration of HPC resources into their global grid environments. However, supercomputers have structural and operational characteristics—such as system architecture, security policies, and software deployment methods—that differ significantly from those of general-purpose servers. These differences pose practical limitations to the application of existing standardized Grid integration models to HPC systems. In particular, HPC systems that require interconnection across multiple security domains face even more complex integration challenges. We propose a neutral integration strategy for HPC systems that does not depend on their specific architectures. By using existing constraints in a reversed way, the approach allows flexible integration with minimal service interruptions and operational overhead, even during infrastructure changes like system upgrades or migrations. This method was applied to Nurion, the national supercomputer operating across multiple security domains, and successfully integrated it into the ALICE Grid Site. Over approximately 17 months of operation, the site demonstrated stable job processing and efficient utilization of multicore resources. We also found CPU over-consumption behavior in the ALICE middleware through continuous system monitoring. By presenting a feasible and system-independent strategy for HPC integration, this paper offers a practical reference architecture for building next-generation HPC-based grid computing infrastructures in preparation for the HL-LHC era, where upgraded detector performance will require even greater computational capacity.
The rapid expansion of the Internet of Things (IoT) has significantly increased the number of connected devices, leading to a dramatic rise in diverse and latency-sensitive computational demands. In heterogeneous Fog–Cloud environments, efficient task scheduling is crucial to meeting stringent requirements for energy efficiency and quality of service (QoS). However, this problem is inherently complex due to the multidimensional nature of tasks—characterized by computational length, deadline, and input file size—which strongly influence bandwidth utilization, transmission costs, and energy consumption. In addition, the dynamic availability of Fog and Cloud resources over time introduces further challenges in ensuring workload balance and system efficiency. To address these issues, this paper first formulates the task scheduling problem as a Mixed Integer Non-Linear Programming (MINLP) model that incorporates communication, processing, and queuing delays, aiming to minimize energy consumption while maximizing deadline satisfaction. Building on this formulation, we propose an Improved Priority-aware Genetic Algorithm (IPGA) that introduces three key innovations: (1) a Pareto-based non-dominating sorting strategy that classifies tasks using all three critical attributes—length, deadline, and input size—enabling more accurate and intelligent task categorization; (2) a dynamic resource-aware allocation mechanism that distributes tasks between Fog and Cloud nodes proportionally to their available resources, thereby balancing workloads, reducing makespan, and enhancing QoS; and (3) a guided mutation strategy that selectively reassigns tasks from overloaded nodes to underutilized ones, improving convergence while reducing overall energy consumption. Extensive simulations across diverse scenarios confirm the effectiveness of the proposed approach, demonstrating up to 60
Scheduling tasks in multi-cloud settings is complex since different cloud service providers have different latency and cost limits, which must be balanced. In a multi-cloud setting, ineffective scheduling results in more significant execution costs, more latency, missed deadlines, underutilization of resources, and worse service performance. Existing approaches frequently lack effective mechanisms for balancing cost and delay, are unable to adjust dynamically to changes in workload, and ineffectively utilize resource heterogeneity, which results in less-than-ideal task distribution and higher operating expenses. By applying the concepts of quantum computing, it enhances the Gannet Optimization Algorithm (GOA) and speeds up convergence. To further enhance the global optimum solution, QIGTLC integrates Termite Life Cycle Optimizer (TLCO) with quantum-inspired GOA. The suggested approach first completes activities on the most economical and significant cloud providers while making sure that the latency rate, energy usage, and makespan are minimized. The remaining jobs that cannot be planned within the allotted resources are dynamically split among several cloud platforms according to performance, cost, and availability factors. Lastly, the suggested task scheduling strategy maximizes overall resource efficiency and minimizes execution costs by distributing workloads among several cloud providers. Experiments are carried out on NASA and HPC2N, two real-world workloads, and the outcomes are contrasted with other existing methods. The outcomes show that the suggested strategy performs better in terms of cost, energy usage, latency, deadline violation cost, and makespan, demonstrating its efficacy in multi-cloud settings.
With the rapid development of big data and real-time stream processing technologies, Apache Flink, as one of the most mainstream stream processing frameworks today, is widely used for large-scale data processing and real-time analysis. However, in resource-level heterogeneous cluster environments, the default data partitioning strategy of Flink suffers from load imbalance and low resource utilization. Existing studies primarily focus on load balancing between cluster nodes while neglecting the uneven distribution of loads among parallel task instances. To address these issues, this paper constructs a three-tier load balancing model, which considers the load balancing optimisation of clusters comprehensively from the three levels of task instances, computing nodes and data transmission. Based on this model, load-aware dynamic data partitioning algorithm (LADP) is proposed, which is designed to not only achieve load balancing between cluster nodes and task instances during partitioning but also prioritize low-latency downstream nodes to alleviate load skew among task instances and reduce data transmission latency. Additionally, through load skew repair strategy and dynamic balanced allocation strategy, LADP can adaptively adjust data distribution for overloaded nodes, making stream data distribution more balanced. Experimental results show that compared to the default partitioning strategy, Dr-Stream, DPS, and St-Stream, LADP increases the system’s average throughput by 17.83
