The high heterogeneity of resource types in multi-cloud computing environments and the complexity and diversity of billing models make the workflow scheduling problem more complex and challenging when scientific workflows need to be completed within strict deadline constraints. Although existing research has progressed in reducing the completion time of workflow scheduling in multi-cloud environments, many difficulties still exist in minimizing scheduling costs while satisfying deadlines. Therefore, a random forest enhanced particle swarm optimization algorithm (RFPSO) is proposed in this study. The RFPSO algorithm implements intelligent initialization of resource allocation through a random forest model, which improves the efficiency of finding optimal solutions. Moreover, it designs a reflective boundary constraint mechanism and a hierarchical task allocation mechanism based on critical path. This design ensures that critical tasks can prioritize access to higher-performance computing resources, which effectively guarantees that tasks within the workflow are scheduled and completed by their deadlines. In addition, the quality of the optimal solution is improved through a local neighborhood search mechanism. Experimental results on scientific workflow datasets such as Epigenomics and Montage show that, compared with existing state-of-the-art methods, RFPSO reduces execution costs by an average of 57.31%.
With the widespread use of Deep Neural Networks (DNNs) in latency-sensitive applications, efficient DNN inference in edge networks has become increasingly important. Cloud-based inference suffers from communication latency, while on-device inference is constrained by the computation and energy resources of User Devices (UDs). Edge networks with multiple UDs and Edge Servers (ESs) provide a collaborative inference paradigm. However, concurrent DNN tasks in such networks still require joint scheduling and offloading, since inter-layer dependencies, intermediate data transmission, and device queue states jointly affect inference performance. In this work, we propose a Distributed Offloading and Graph-Driven Decision (DOGD) framework for parallel DNN inference optimization in edge networks. DOGD models DNN inference tasks as Directed Acyclic Graphs (DAGs) to capture inter-layer dependencies. A Graph Convolutional Network (GCN) extracts dependency-aware subtask features, which are combined with device load states and used by a Deep Reinforcement Learning (DRL) module to jointly determine the execution order and execution location of ready subtasks. By scheduling ready subtasks across different devices, DOGD exploits cross-task parallelism while adapting to runtime device loads. Extensive experiments under different DNN topologies, task scales, and ES settings show that DOGD consistently outperforms the compared methods. Compared with local execution, DOGD achieves an average inference latency reduction of 64.38% when the number of ESs varies from 2 to 6. Under the Latency-Energy Efficiency (LEE) metric, DOGD improves LEE by 46.29% over PPO for the graph-structured GoogleNet model, demonstrating the benefits of structural feature extraction and joint execution-order and execution-location optimization.
Efficient Vehicle Patrol Scheduling (VPS) is essential for improving urban safety, ensuring security, and optimizing operational performance. Traditional scheduling methods often struggle to balance multiple objectives and constraints, such as limited vehicles, mandatory rest periods, and strict revisit cadences under dynamic traffic conditions. To address these challenges, this study proposes a dual methodological approach: first, a formalized optimization model is introduced to obtain exact solutions for small-scale instances; second, two novel scalable heuristics are developed, Adaptive Hill-Climbing-Based Patrol Scheduling (AHBPS) and Genetic-Based Dynamic Vehicle Patrol Scheduling (GDVPS), designed to handle large-scale and dynamic urban networks. Extensive simulations using real-world urban maps demonstrate that GDVPS consistently outperforms AHBPS in both solution quality and scalability, achieving up to 80% coverage in networks with 1000 locations, while maintaining real-time feasibility. The results confirm that GDVPS provides a robust, dynamic, and scalable scheduling solution, making it a promising candidate for deployment in modern urban safety operations.
Flash Friendly File System (F2FS) is a log-structured file system (LFS) optimized for Flash memory characteristics and is widely deployed on mobile devices, embedded systems, and some Linux platforms that use NAND Flash. File fragmentation and free space fragmentation both affect the performance of F2FS. In this work, we investigate the performance impact of fragmentation through energy consumption characterization. Our measurements indicate that energy consumption increases with the number of file and free space fragments, based on experiments evaluating F2FS while serving I/O requests across diverse workload scenarios. While considerable efforts have been devoted to mitigating file fragmentation, comparatively less attention has been paid to understanding and optimizing free space fragmentation, which predominantly arises from the distribution of invalid blocks. We observe that reclaiming invalid blocks via background garbage collection (GC) incurs over 100 mJ of energy per invocation, yet yields only marginal reductions in free space fragmentation. This motivates us to improve GC effectiveness in reducing free space fragmentation.We propose the free space fragmentation-aware (FragGC) and file system performance-aware GC (PerfGC) scheme. We seek to both reduce GCs and enhance the efficiency of each GC operation. We reassess the definition of a free space fragment through empirical analysis and introduce the free space fragmentation factor as a lightweight metric to quantify the degree of free space fragmentation at the segment level. FragGC optimizes victim segment selection and valid block migration based on this metric. PerfGC adjusts GC frequency according to the impact of free space fragmentation on file system performance. Experimental results on a real platform demonstrate that FragGC and PerfGC reduce GC count compared to traditional F2FS and its latest GC optimization, ATGC. FragGC reduces the time to replay traces by 24.4% to 40.9% for large-scale applications.
Vehicular Edge Computing (VEC) brings cloud resources to the Road Side Units (RSUs) and the vehicles to support computation-intensive services in the Internet of Vehicles. The key challenge is allocating and scheduling dependent tasks across mobile vehicles and edge servers under tight time and energy budgets. We propose a trajectory-aware task mapping framework to coordinate vehicles and RSUs for task processing via Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communications. The applications generated by the vehicles can be modeled as Directed Acyclic Graphs (DAGs), and we jointly optimize task-to-device mapping subject to time and energy constraints. First, we design a Space-Time Long Short-Term Memory (ST-LSTM) model to predict short-term vehicle trajectories and contact durations, thereby enabling the selection of suitable vehicles for V2V task migration. Then, we use a task-priority sorting algorithm to schedule the tasks in the DAG so that all dependent tasks are executed in sequence. Finally, we develop MAIAC, a multi-agent actor-critic method that runs in a distributed manner and uses convolutional and pooling layers to learn spatial and temporal features for practical training. Experimental results show that ST-LSTM reduces trajectory error by more than 10% compared with a standard LSTM model. MAIAC converges faster, achieving lower task latency of up to 54% and reducing energy consumption by up to 50% compared with the state-of-the-art methods. The results demonstrate that combining mobility prediction with multi-agent reinforcement learning enables efficient and reliable task mapping in the VEC scenario.
In many application scenarios, the efficiency of linear operations is sensitive to the storage format of sparse matrix, and the sparse storage formats that support efficient operation of different linear operations often vary. Therefore, in order to effectively improve the performance of linear operations, efficient conversion of sparse matrix storage formats is particularly important. However, there are multiple conversion paths between the two determined storage formats. How to choose the high efficient conversion path is the focus of this paper. Therefore, this paper designs a path optimization strategy for sparse matrix storage format conversion, constructs a sparse matrix storage format conversion model, and realizes the path optimization of efficient storage format conversion based on the Dijkstra algorithm. On this basis, efficient conversion is achieved for five typical sparse storage formats: COO, CSR, ELL, DIA and BCSR. Finally, the experimental results based on two different architectures of CPU and NVIDIA GPU show that some direct conversions between sparse matrix storage formats are not the optimal conversion path, and higher conversion efficiency can be achieved by intermediate format conversion. Moreover, the experimental results demonstrate that the speedup between the proposed method under optimal conditions with the direct conversion in the format conversion can reach 119.071× on CPU and reach 60.534× on NVIDIA GPU. In the worst case, its performance is the same as that of direct conversion. This satisfies the dynamic and changeable requirements of different sparse matrix storage formats in application scenarios to a certain extent. It also effectively improves the performance of sparse matrix storage format conversion.
Spatiotemporal data modeling plays a critical role in a wide range of real-world applications, including traffic forecasting, climate analysis, and urban sensing. While recent advances in transformers and graph neural networks have improved the ability to capture complex spatial and temporal dependencies, existing models often lack semantic understanding and contextual awareness, limiting the accuracy of spatiotemporal feature representation. In this work, we propose LLM4ST, a novel multimodal framework that leveraging large language models (LLMs) to enhance spatiotemporal representation learning. LLM4ST unifies three complementary modalities: (1) tokenized raw spatiotemporal data, (2) natural language descriptions that encode domain knowledge or contextual information, and (3) latent embeddings from spatiotemporal encoding backbones. These heterogeneous inputs are integrated via a transformer-based cross-modal encoder, enabling joint reasoning over numerical patterns and semantic cues. By incorporating descriptive language and token information as an auxiliary modality, LLM4ST introduces semantic knowledge priors, thereby enhancing the model’s generalization. Extensive experiments on real-world benchmarks demonstrate that LLM4ST outperforms strong baselines across traffic prediction. Moreover, ablation studies validate the complementary nature of language and data-driven features. Our work highlights the potential of LLMs as general-purpose priors for spatiotemporal understanding, bridging the gap between symbolic semantics and spatiotemporal representations. The codes are at https://github.com/CodingMaker2/LLM4ST.
Indoor environments are increasingly exposed to natural hazards, posing significant challenges for ensuring both preparedness and effective emergency response. However, existing monitoring solutions typically address either routine environmental control or post-event analysis, lacking mechanisms for dynamic reconfiguration across operational conditions.This paper presents a distributed IoT system for indoor disaster preparedness and response, grounded in an operational framework that provides a unified rationale for analyzing and managing indoor living environments. The proposed system integrates environmental and structural sensing, edge-level event detection, and cloud coordination services to enable continuous monitoring under normal conditions and coordinated reconfiguration during emergencies. A dual-regime operational model distinguishes preparedness and response phases, enabling state transitions triggered by critical events. The system is evaluated through a six-month real-world deployment in an educational environment, which provides qualitative and quantitative evidence of its behavior under operational conditions, and through controlled laboratory tests to assess its dual-regime behavior. The results show that the proposed system supports operational awareness and the reconfiguration of coordinated behavior in indoor environments.
Unmanned Aerial Vehicles (UAVs) are increasingly used for road traffic monitoring due to their mobility and wide-area coverage. However, their limited onboard resources make real-time video analysis challenging under dynamic traffic conditions. To overcome this, computational task offloading to nearby fog nodes is often employed. The main challenge lies in deciding when to process locally or offload, as both traffic and computational load vary continuously. Existing heuristic-based approaches are lightweight but rely on fixed thresholds, leading to unstable switching and degraded performance under fluctuating conditions. Meanwhile, Deep Reinforcement Learning (DRL)–based methods can adaptively optimize offloading but require extensive training and high computational costs, limiting their practicality on UAVs. To address this challenge, we propose Dynamic Vehicle Density-aware Offloading (DVDOffload), an adaptive task offloading technique designed to maximize performance and resource efficiency by adapting the offloading decision to road traffic conditions. The proposed method uses vehicle density as the primary workload indicator and dynamically adjusts offloading thresholds using an Exponential Moving Average (EMA) to ensure adaptive and stable decisions. Experimental results in multiple realistic traffic scenarios show that DVDOffload achieves higher accuracy, faster processing, and lower resource consumption compared to several baseline heuristic and DRL-based approaches in the evaluated UAV–fog traffic monitoring system.
While Mobile Edge Intelligence (MEI) provides crucial low-latency Artificial Intelligence (AI) services for Internet of Things (IoT) and 5G/6G networks, individual edge servers struggle to host large scale AI models due to limited storage and compute capabilities. To overcome this Quality of Service (QoS) bottleneck, this paper investigates joint resource optimization in collaborative MEI networks. We aim to minimize the weighted sum of latency and terminal side energy consumption costs for all users across the network, while satisfying AI model availability and caching constraints. Specifically, we first consider a wired backhaul edge layer scenario based on graph topology. By integrating heterogeneous user distributions, time varying wireless channels, and model storage overheads, the coupled optimization problem of dynamic model caching, collaborative task offloading, and resource allocation is formulated as a Decentralized Partially Observable Stochastic Game (Dec-POSG). This provides a systematic decision making framework for multi agent coordination in distributed dynamic environments. Subsequently, to mitigate the decision bias caused by the limited local observations of edge nodes, we propose a topological feature extraction algorithm based on GraphSAGE. By utilizing neighborhood sampling and information aggregation mechanisms, the algorithm effectively captures the spatial correlation features and potential load states among edge servers, achieving deep perception of complex network topology information. Furthermore, we design and implement a Graph Neural Network enhanced Multi Agent Proximal Policy Optimization (GNN-MAPPO) algorithm. Through an end to end distributed learning strategy, the framework demonstrates the potential for near real time and fine grained dynamic management of caching locations, offloading routes, and physical resources, while maintaining low computational overhead. Finally, simulation experiments conducted across metropolitan-scale scenarios validate the proposed framework. The results show that GNN-MAPPO achieves a peak cache hit rate of 0.92 under a 50% model storage ratio. Compared to the state-of-the-art GA-MARL algorithm, GNN-MAPPO reduces the average task completion latency by 12.8% and the total weighted system cost by 15.1% under high-concurrency workloads of 105 requests per second. Furthermore, our learned heuristic achieves a near-optimal cost within 3.4% of the absolute global optimum computed by an exact mathematical solver in small-scale benchmarks. The entire control-plane decision-making process is executed within 8 ms, which consumes less than 8% of the 100 ms scheduling slot duration, leaving more than 92% of the slot fully available for data plane operations. These quantitative results confirm that GNN-MAPPO successfully balances high decision accuracy and low operational latency in large-scale collaborative mobile edge intelligence systems.
Vehicular monitoring systems have been increasingly adopted to support public safety operations in urban environments. Among available technologies, Automatic License Plate Recognition (ALPR) has attracted attention due to its affordability, scalability, and ease of deployment. However, most existing ALPR applications remain predominantly reactive, relying on static watchlists, deterministic rules, or direct database matching, which may limit their ability to support anticipatory decision-making in dynamic public safety scenarios. To better understand this domain, this article presents a Systematic Literature Review (SLR) on the integration of AI-driven predictive modeling into license plate recognition systems for vehicular monitoring in public safety. The review was conducted across six scientific databases and analyzed different variables, highlighting and comparing critical characteristics in recent literature. The findings show that reactive and rule-based approaches still dominate the field, while predictive solutions remain limited and fragmented, with AI-driven methods gaining relevance to support risk assessment and contextual alert generation. Beyond consolidating existing literature, this review clarifies the transition from reactive plate-recognition applications toward predictive and alert-oriented decision-support systems. Moreover, a technical and governance-oriented taxonomy is proposed to organize the main concepts and guide future research efforts, potentially bringing relevant contributions to the area.
