Source-linked AI summary
AI-based Fog and Edge Computing: A Systematic Review, Taxonomy and Future Directions
Sundas Iftikhar, Sukhpal Singh Gill, Chenghao Song, Minxian Xu, Mohammad Sadegh Aslanpour, Adel N. Toosi, Junhui Du, Huaming Wu, Shreya Ghosh, Deepraj Chowdhury, Muhammed Golec, Mohit Kumar, Ahmed M. Abdelmoniem, Felix Cuadrado, Blesson Varghese, Omer Rana, Schahram Dustdar, Steve Uhlig
TL;DR
Resource management in fog and edge computing must handle limited and heterogeneous resources, changing workloads, and unpredictable environments. The paper conducts a systematic literature review of AI/ML-based approaches, develops a taxonomy, compares existing techniques, and identifies open challenges and future directions. It reports that AI-based methods have been applied across many management tasks and can support accurate decisions, metric prediction, and improved service quality, while precision–resource trade-offs and security constraints remain open concerns.
Problem
Fog and edge resource management is difficult because resources, workloads, and environments are heterogeneous, dynamic, limited, and unpredictable.
Method
The paper conducts a CRD-guideline Systematic Literature Review that evaluates studies, develops a taxonomy, compares approaches, analyzes outcomes, and identifies research directions.
Results
AI-based methods span resource estimation, discovery, matching, offloading, load balancing, orchestration, placement, and consolidation, with reported benefits for decision accuracy, time overhead, QoS, utilization, throughput, SLA violations, energy consumption, and fault tolerance.
Takeaways & Limitations
The taxonomy supports selecting AI-based methods for resource management across fog and edge environments under dynamic conditions.
Takeaways & Limitations
AI resource managers face an unresolved precision–resource trade-off between lightweight inference and more resource-intensive, precise models.
Abstract
from arXiv · showhide
Resource management in computing is a very challenging problem that involves making sequential decisions. Resource limitations, resource heterogeneity, dynamic and diverse nature of workload, and the unpredictability of fog/edge computing environments have made resource management even more challenging to be considered in the fog landscape. Recently Artificial Intelligence (AI) and Machine Learning (ML) based solutions are adopted to solve this problem. AI/ML methods with the capability to make sequential decisions like reinforcement learning seem most promising for these type of problems. But these algorithms come with their own challenges such as high variance, explainability, and online training. The continuously changing fog/edge environment dynamics require solutions that learn online, adopting changing computing environment. In this paper, we used standard review methodology to conduct this Systematic Literature Review (SLR) to analyze the role of AI/ML algorithms and the challenges in the applicability of these algorithms for resource management in fog/edge computing environments. Further, various machine learning, deep learning and reinforcement learning techniques for edge AI management have been discussed. Furthermore, we have presented the background and current status of AI/ML-based Fog/Edge Computing. Moreover, a taxonomy of AI/ML-based resource management techniques for fog/edge computing has been proposed and compared the existing techniques based on the proposed taxonomy. Finally, open challenges and promising future research directions have been identified and discussed in the area of AI/ML-based fog/edge computing.
1. Introduction
Fog and edge computing bring processing closer to IoT devices to address cloud latency and QoS challenges, while AI/ML methods support resource management under dynamic workloads. This review systematically examines these approaches, proposes a taxonomy, compares existing techniques, and identifies future directions.
- Motivation: Fog and edge computing address these challenges by moving computing resources closer to end devices and sensors.This strategy can partially reduce cloud load and speed communications.
- Motivation: Cloud-centered processing faces latency, bandwidth, and QoS challenges as IoT devices generate increasingly large volumes of data.The paper cites projected IoT data generation of 79.4 zettabytes in 2025 and notes that traditional cloud infrastructure is not designed for this scale.
- Resource management challenge: Dynamic and diverse workloads make static heuristic resource management inadequate for fog and edge environments.Traditional threshold-based approaches are configured offline for particular workload scenarios and cannot scale applications in changing environments.
- AI/ML approaches: AI/ML techniques are investigated for modeling metrics, predicting conditions, and improving resource provisioning and orchestration decisions.The review discusses applications including resource efficiency, load balancing, energy efficiency, and SLA assurance.
- Review scope: The study uses a systematic literature review to examine AI/ML resource management, synthesize existing work, and identify research gaps.It follows CRD guidelines and reviews studies addressing the background, current status, techniques, and applicability challenges of AI/ML in fog and edge computing.
- Contributions: The paper proposes a taxonomy, compares existing studies across identified categories, and discusses open issues and future directions for Edge AI.The contributions include reviewing AI/ML approaches, comparing studies using taxonomy-related parameters, and identifying promising research directions.
2. Review Methodology
The review uses a structured Systematic Literature Review process to identify, evaluate, classify, and analyze research on AI/ML-based resource management in fog/edge computing.
- Review design: The study follows an SLR methodology based on CRD guidelines, including evaluation, taxonomy creation, analysis, comparison, findings, and future directions.The process is divided among seven explicitly stated review steps.
- Review design: Research questions define the evaluation rules and guide searches across multiple electronic and supplementary sources.The sources include major digital libraries and reference-list investigations.
- Search strategy: The search used constructed strings combining fog, edge, cloud, AI/ML, resource-management tasks, challenges, metrics, simulators, workloads, algorithms, and methods.The search strings were formulated from the research questions and detailed in Table 3.
- Selection and quality assessment: More than 320 papers entered initial screening, after which exclusions removed brief, non-peer-reviewed, non-indexed, non-English, survey, and low-quality works.The stated exclusion criteria also covered book chapters and publications lacking rigorous peer review.
- Selection and quality assessment: The final review included 135 articles selected from prestigious journals and conferences after screening and quality evaluation.Quality assessment considered objectivity, internal consistency, and bias using CRD recommendations.
- Data extraction: All 135 publications were read for data extraction, while random samples and team discussions were used to cross-check collection accuracy.The extraction methodology and forms are documented in the appendices.
3. AI/ML For Fog/Edge Computing: Background and Current Status
Fog/edge computing brings resources closer to users to support latency-sensitive IoT applications, while AI/ML techniques address diverse workloads and resource-management challenges across application domains.
- Fog and edge computing bring computational resources closer to end users to support latency-sensitive applications such as healthcare, transportation, agriculture, and smart homes.
- Resource provisioning selects, deploys, and manages software and hardware resources at runtime, balancing overprovisioning costs against underprovisioning risks such as SLA violations.
- Runtime offloading must select suitable cloud or fog resources for computation-intensive workloads despite resource constraints and complex fog/edge network architectures.
- Scheduling seeks optimal heterogeneous fog/edge resources for upcoming workloads, but its objective-function formulation is NP-Complete and difficult to solve exactly.
- Load balancing distributes workloads across virtualized fog nodes to reduce response time, network delay, waiting time, and scalability problems.
- Existing cloud resource-management solutions do not directly transfer to fog/edge environments because their network conditions, infrastructure distribution, and node capabilities differ.
3.5. Resource Management in Fog/Edge computing
Fog/edge resource management relocates computation toward end devices to reduce latency and energy use, requiring coordinated discovery, offloading, placement, allocation, and orchestration across constrained heterogeneous nodes.
- Edge computing uses resource-limited, heterogeneous devices, while fog combines edge and cloud capabilities closer to users than centralized cloud infrastructure.
- Resource management is closely related to task offloading, which transfers computation-intensive tasks to resource-rich platforms for latency-sensitive applications.
- Resource discovery identifies available resources, their locations, and their usable duration through centralized or distributed mechanisms.
- Offloading decisions must account for heterogeneous node loads because concentrating requests on resource-rich nodes can waste capacity and increase process waiting times.
- Resource orchestration coordinates allocation across nodes by migrating offloaded task data to improve service capability and system-wide load balance.
- Application placement moves compute-intensive services, modules, applications, or models to edge or fog nodes for execution and storage.
- Server consolidation through VM migration can save system energy at the cost of increasing latency for individual tasks.
3.8. Performance Metrics
Fog/edge architectures are evaluated across layered infrastructure using monitoring metrics, analysis methods, planning measures, workloads, benchmarks, and simulators tailored to different computing layers.
- Fog/edge architectures are commonly separated into IoT, edge, fog, and cloud layers with progressively greater computing capabilities.
- Resource utilisation, throughput, resource load, and latency monitor how effectively the architecture processes incoming workloads.
- Machine learning and deep learning can analyze monitoring data, including large time-series datasets, to extract meaningful performance information.
- Planning metrics include decision number, contradictory decisions, completion ratio, and cache hit ratio for evaluating optimization decisions.
- Simulators test architectures before deployment, with tools differing according to IoT, edge, fog, or cloud layers.
- PureEdgeSim models distributed, dynamic, heterogeneous edge-to-cloud infrastructures, while iFogSim supports fog-cloud application deployment and resource-management testing.
- DeFog benchmarks five computation-intensive AI applications spanning object classification, speech recognition, mobile gaming, edge gateways, and face detection.
4. AI Based Techniques For Resource Management in Fog/Edge Computing
The review organizes AI-based resource-management techniques into machine learning, deep learning, and reinforcement learning approaches for prediction, scheduling, allocation, provisioning, and offloading in fog/edge systems.
- The paper presents a taxonomy of AI-based techniques for resource management in fog/edge computing and summarizes their applications across management problems.
- Machine learning supports modeling, classification, prediction, and forecasting for workload and resource-usage analysis in edge/fog environments.
- Deep learning models such as CNNs and LSTMs are used for spatial-feature extraction and prediction of fluctuating workload time series.
- Hybrid LSTM-DNN and reinforcement-learning methods support mobility prediction and resource allocation for vehicular fog computing.
- Reinforcement learning is widely used for resource sharing and allocation decisions involving heterogeneous, capacity-constrained CPU, GPU, and storage resources.
- Deep reinforcement learning methods address task offloading, scheduling, resource allocation, and provisioning using algorithms including soft actor-critic, DQN, policy gradients, A3C, and Q-learning.
- Adaptive task scheduling is motivated by incomplete information about task arrivals and can target execution delay, energy consumption, load balancing, and other KPIs.
5. Taxonomy
The proposed taxonomy organizes AI-based fog and edge computing frameworks across infrastructure, objectives, platforms, resource-management mechanisms, evaluation metrics, AI methods, and target applications.
- The taxonomy covers infrastructure, objectives, platform, resource management, metrics, AI-based methods, and target application.
- Infrastructure: Frameworks span single-cloud, multi-cloud, hybrid-cloud, and community-cloud infrastructure models.
- Objectives: Dominant objectives include resource efficiency, energy consumption, cost efficiency, SLA assurance, and QoS.
- Resource Management: Resource-management techniques include resource matching, task offloading, load balancing, application placement, orchestration, discovery, and estimation.
- Metrics: Common evaluation metrics include resource utilization, throughput, latency, energy consumption, SLA violations, and fault tolerance.
- Target Applications: AI-based approaches have been applied broadly across fog and edge environments, including IoT, smart-city, healthcare, and transportation scenarios.
6. Result Outcomes
The review systematically characterized the fog/edge literature using an iterative selection process and categorized studies by resource, QoS, and real-world factors. Its 2015–2022 taxonomy emphasizes recent research, especially publications from 2022.
- 320 articles were collected, with 135 shortlisted after the iterative selection process.
- The taxonomy references studies from 2015 to 2022, with the majority of referred papers published in 2022.
- The review methodology was inspired by Kitchenham’s systematic literature review procedure and drew on ACM, IEEE, Springer, Scopus, and other sources.
- The taxonomy covers resource management, QoS parameters, and real-world factors including IoT, healthcare, security, and privacy.
- The resource aspect comprised 61 papers, while QoS parameters comprised 49 papers.
7. Open Issues and Future Directions
The paper identifies unresolved challenges in deploying AI-based resource management across heterogeneous, constrained, distributed, and security-sensitive fog/edge environments. Future work must balance AI capability with resource, precision, adaptation, and security requirements.
- Heterogeneous languages, hardware architectures, and processing units complicate resource-management decisions across fog and edge devices.
- Edge sustainability requirements conflict with the resource-hungry nature of AI models, creating a benefit-versus-resource-consumption trade-off.
- Environmental Sustainability: Future work must determine when to use lightweight or precise AI models while satisfying IoT application precision requirements.
- Security versus Efficiency: AI-based managers require application observations for training and inference, making it challenging to satisfy security requirements without degrading them.
- Coexisting AI applications and AI managers raise the challenge of adapting applications to consume different amounts of CPU or accelerator resources.
- Distributed edge and fog networks require substantial shared data for AI inference and training across scattered devices.
8. Summary and Conclusions
The review synthesizes how AI and ML address resource management in fog and edge environments, compares approaches through a taxonomy, and identifies future research needs. It finds broad application coverage and rapid progress, while highlighting systematic-review and fine-grained-control gaps.
- The study advances related surveys through comprehensive comparisons of fog and edge computing approaches.
- AI-based methods address resource estimation, discovery, matching, offloading, load balancing, orchestration, placement, and consolidation.
- Advanced methods including DNN, Q-learning, DQN, and reinforcement learning optimize utilization, throughput, SLA violations, energy consumption, and fault tolerance.
- The review combines fog and edge computing with an AI-based optimization framework covering the whole resource-management process.
- Microservice and serverless approaches are identified as promising directions for fine-grained resource control.
Appendix A: A quality assessment forms
Appendix A describes the quality-assessment questions used in the review, distinguishing preliminary examination from evaluation-specific questions.
- Preliminary examination: Table 6 presents the questions used during the preliminary examination.
- The appendix organizes quality assessment around preliminary examination and evaluation questions.
- Specific questions: Table 7 presents the specific questions used during evaluations.
Appendix B: Data items extracted from all articles
Appendix B identifies Table 8 as the repository of data items extracted from all reviewed articles.
- Table 8 shows the data items extracted from all articles.
- The extraction covers articles included in the review.
- Table 8 is titled “Data items extracted from all articles.”
Appendix C: Journals and Conferences for publishing articles about AI/ML in Fog/Edge Computing.
Appendix C lists prominent journals and conferences publishing research on AI/ML-based resource management in fog/edge computing.
- Table 9 lists journals and conferences publishing articles about AI/ML in fog/edge computing.
- The publication venues include journals, conferences, workshops, and other listed venue categories.
- Table 9 defines N as papers reporting AI/ML-based resource-management techniques as their primary research focus.
Appendix D. List of Acronyms
Appendix D provides a list of acronyms used in the paper.
- Table 10 presents the paper’s list of acronyms.
- The acronym list is included as a dedicated appendix resource.
- Appendix D identifies Table 10 as the location of the acronym reference.
Authors Biography
The authors are researchers and academics working across cloud, fog, edge, IoT, AI, deep learning, networking, security, and resource management.
- The contributors hold academic and research positions at universities and research institutions across the United Kingdom, Australia, China, India, Pakistan, Germany, and the United States.
- Their research spans cloud, fog, and edge computing, including resource management, orchestration, scheduling, optimization, and energy efficiency.
- The biographies report publication, reviewing, editorial, conference, award, patent, and professional-membership activities among the contributors.
- The group also covers Internet of Things, wireless and mobile networks, software-defined networking, serverless computing, autonomous systems, and security and privacy.
- Several authors specialize in machine learning, deep learning, AI, trajectory data mining, or model-based evaluation for computing and network systems.