Virtual Machine (VM) consolidation has become a fundamental technique for improving resource utilization, reducing energy consumption, and lowering operational costs in cloud data centers. Although numerous VM consolidation methods have been proposed, existing review studies often focus on specific techniques or limited aspects of the consolidation process, lacking a comprehensive phase-oriented analysis. To address this gap, this paper presents a systematic review of approximately 80 studies on dynamic VM consolidation. The selected studies are analyzed and compared according to their methodologies, simulation environments, workload types, optimization objectives, evaluation metrics, and key contributions. Furthermore, the reviewed approaches are systematically categorized based on the four major phases of VM consolidation: host overloading detection, host underloading detection, VM selection, and VM placement. For each phase, heuristic, metaheuristic, and other algorithmic approaches are critically analyzed, highlighting their strengths, limitations, and application trends. Based on this analysis, the paper identifies current research gaps and discusses promising future research directions toward more energy-efficient, scalable, and sustainable cloud resource management. The findings provide researchers and practitioners with a structured reference for understanding the current state of VM consolidation and support the development of more effective resource management strategies.
Cloud computing is an efficient paradigm for storing and processing heterogeneous data produced by IoT applications. However, the increasing demand and the importance of timely execution of critical-mission applications, along with some inherent problems of cloud computing, fog computing has been introduced of fog computing. This paradigm aims to complement and extend cloud-like services more closely to end-users, offering near-optimal latency and reasonable real-time responses. Application placement is considered as a serious challenge addressing through supposing deterministic parameters by most researches, which is not realistic in practice. Despite extensive research on uncertainty issues, studies on uncertainty in cloud computing are limited and none have yet applied meta‑heuristic algorithms to fog computing under uncertainty. Furthermore, the role of uncertainty in resource provisioning, service models, and programming models has not been sufficiently explored in scientific research. In this study, an IoT service placement method in fog environments that considers uncertainty has been presented. The objective is to reduce time costs while simultaneously increasing the service acceptance rate and satisfying quality of service requirements, taking into account the uncertainty in the fog environment. In the proposed approach, a meta-heuristic algorithm has been employed along with two methods for handling random variable uncertainty, namely chance-constrained programming and averaging, targeting to solve the multi-objective optimization problem. Simulation results indicate that the proposed chance-constrained programming method reduces the number of lost services by 17
This research presents HPP-FDRL, a hybrid framework that combines fuzzy logic and deep reinforcement learning for prioritizing critical heart patients. The problem is formulated as task scheduling in a fog–cloud environment. In conventional cloud-based healthcare systems, remote cloud processing may be affected by latency and bandwidth limitations, which can be problematic for time-sensitive workloads. HPP-FDRL therefore uses a fog–cloud architecture in which patient-related tasks can be processed closer to the data source. At the patient layer, health data are collected through IoT sensors and sent to fog nodes that represent treatment centers. At the fog layer, a Takagi–Sugeno fuzzy system estimates patient priority from clinical attributes. An Actor–Critic deep reinforcement learning policy then assigns tasks to fog nodes using a multi-objective reward that considers resource availability, energy consumption, deadlines, and queue fairness. Cloud offloading is also available as an action, while the learned policy tends to use local fog allocation when sufficient capacity is available. The framework is evaluated under static and dynamic workload settings and compared with learning-based, hybrid, metaheuristic, and heuristic approaches. The results show that HPP-FDRL achieves competitive or higher makespan performance for heart-patient tasks while maintaining high fault tolerance. Energy consumption reflects a deliberate trade-off favoring timely, priority-aware processing over pure energy minimization. Further experiments examine the reward and prioritization components, statistical significance, scalability, and cloud-offloading behaviour. The experiments show the performance of HPP-FDRL under the evaluated simulation setting. Overall, the study shows that combining clinical fuzzy prioritization with deep reinforcement learning provides a way to allocate heart-patient tasks in fog–cloud systems while accounting for both patient priority and system resources.
The Internet of Vehicles (IoV) creates new transportation systems through its ability to connect vehicles with infrastructure and cloud computing resources which improve road safety and traffic management and self-driving vehicle functions. Existing research shows that IoV enables advanced applications which include traffic prediction and accident prevention but these developments create major privacy risks because vehicles produce large amounts of confidential information. The development of Federated Learning (FL) presents a useful solution which allows vehicles to work together on Machine Learning (ML) model development while keeping their sensitive information safe on their own devices. The use of FL in IoV systems for cyber-attack detection and traffic prediction has advanced but current methods continue to use standard ML models which cannot handle the complex time-related patterns found in vehicle data. Researchers who have applied deep learning methods in their studies have failed to recognize how attention mechanisms can improve model understanding and effectiveness. The research introduces a new framework which combines Attention-Based CNN-LSTM Encoders and Decoders with an FL system to create Adaptive Risk Aggregation (ARA) technology for IoV environments. The framework uses CNNs for extracting features which LSTMs use to model temporal data and attention mechanisms for selecting relevant features that training process needs which the system needs to accomplish. The proposed research conducts method evaluation through CIC IoV-2024 dataset testing which uses Exactness and Value of Precision and probability of detection and F1-measures as performance metrics to measure testing results. The research provides predictive performance results which exceed its primary contributions because the framework enables sensitive data protection through local storage while organizations can train models together across different vehicle types with reduced communication costs.
This study proposes a cooperative multi-agent deep reinforcement learning framework for cost-aware metro–bus transfer coordination under fixed rail schedules. Rather than jointly redesigning rail and bus timetables, the proposed approach focuses on adaptive bus-side operational control to improve transfer reliability while preserving operational feasibility in large-scale multimodal networks. The problem is formulated as a cooperative multi-agent decision process and solved using QMIX under a centralized training and decentralized execution paradigm. A step-wise reward function jointly captures passenger waiting time, in-vehicle travel time, and operational adjustment cost, while performance is evaluated using a novel cost-aware intermodal synchronization efficiency (CISE) metric. The framework is validated using real operational data from Tehran’s public transport system, including the complete operational metro network and ten major BRT corridors representing the primary high-demand metro–bus feeder subsystem under fixed metro timetables. Comparative experiments against the current timetable, an Independent deep Q-network (I-DQN), and proximal policy optimization (PPO) show that, on regular weekdays, the proposed method reduces weighted average passenger travel time from 1428.30 s to 1381.17 s (3.3
