A Systematic Literature Review on Load-Balancing Techniques in Fog Computing: Architectures, Strategies, and Emerging Trends

Abstract

Fog computing has emerged as a promising paradigm to extend cloud services toward the edge of the network, enabling low-latency processing and real-time responsiveness for Internet of Things (IoT) applications. However, the distributed, heterogeneous, and resource-constrained nature of fog environments introduces significant challenges in balancing workloads efficiently. This study presents a systematic literature review (SLR) of 113 peer-reviewed articles published between 2020 and 2024, aiming to provide a comprehensive overview of load-balancing strategies in fog computing. This review categorizes fog computing architectures, load-balancing algorithms, scheduling and offloading techniques, fault-tolerance mechanisms, security models, and evaluation metrics. The analysis reveals that three-layer (IoT–Fog–Cloud) architectures remain predominant, with dynamic clustering and virtualization commonly employed to enhance adaptability. Heuristic and hybrid load-balancing approaches are most widely adopted due to their scalability and flexibility. Evaluation frequently centers on latency, energy consumption, and resource utilization, while simulation is primarily conducted using tools such as iFogSim and YAFS. Despite considerable progress, key challenges persist, including workload diversity, security enforcement, and real-time decision-making under dynamic conditions. Emerging trends highlight the growing use of artificial intelligence, software-defined networking, and blockchain to support intelligent, secure, and autonomous load balancing. This review synthesizes current research directions, identifies critical gaps, and offers recommendations for designing efficient and resilient fog-based load-balancing systems.

1. Introduction

The exponential growth of Internet of Things (IoT) devices has driven demand for computational paradigms that support low-latency, real-time processing. Fog computing has emerged as a pivotal solution that bridges the gap between cloud data centers and IoT endpoints by decentralizing computing resources and bringing them closer to the data source. Unlike traditional cloud computing, which relies on centralized infrastructures, fog computing operates through geographically distributed nodes—including gateways, routers, and edge servers—which are capable of handling data locally. This proximity significantly reduces latency, improves responsiveness, and supports context-aware applications across various domains such as smart healthcare, autonomous vehicles, and industrial automation.

However, the decentralized and heterogeneous nature of fog environments introduce several technical challenges. Among them, load balancing stands out as a critical issue. Load balancing in fog computing involves the efficient distribution of tasks across available fog nodes to optimize resource utilization, reduce response time, and ensure service continuity. Unlike centralized cloud environments, fog systems must operate under constrained resources and dynamic network conditions, where nodes may frequently join or leave, workloads fluctuate, and computational capacities vary widely.

Additionally, fog systems must address the Quality of Service (QoS) demands of diverse IoT applications. For example, real-time applications such as video surveillance or telemedicine require prompt data processing, which makes intelligent and adaptive load-balancing mechanisms essential. Other applications may demand privacy-preserving features and fault-tolerant architectures to protect sensitive data and maintain operational reliability. These challenges highlight the need for robust, flexible, and context-aware load-balancing strategies in fog computing.

In light of these complexities, this study conducts a systematic literature review (SLR) to explore the state of the art in load balancing for fog computing environments. Specifically, the objectives of this review are as follows:

• To analyze and categorize fog computing architectures and their implications for load balancing.
• To classify load-balancing algorithms and compare their strengths and weaknesses.
• To identify the most frequently used performance metrics and evaluation tools.
• To highlight emerging trends, ongoing challenges, and research gaps.

This review synthesizes findings from 113 peer-reviewed studies published between 2020 and 2024. By doing so, it aims to provide a comprehensive understanding of current practices and future directions for load balancing in fog computing, with an emphasis on improving QoS, scalability, and energy efficiency in IoT-driven environments.

2. Related Works

Numerous systematic reviews and surveys have been conducted to explore load balancing in fog computing. These studies have examined various algorithmic approaches, performance metrics, and challenges. However, most of them are limited either by their scope, methodological rigor, or timeliness. This section outlines key contributions from the existing literature while highlighting their limitations and the rationale for this review.

Kashani and Mahdipour [1] conducted a systematic study categorizing load-balancing algorithms into approximate, exact, fundamental, and hybrid types. Their analysis, based on 49 papers from 2013 to 2021, emphasized response time and energy consumption as core evaluation metrics. However, the review was limited by a relatively narrow time window and a strong focus on simulation-based research, with minimal discussion on real-world applicability or emerging AI-driven methods.

Kaur and Aron [2] presented a systematic review of static and dynamic load-balancing techniques in fog computing, focusing on improving energy efficiency and resource utilization. While their taxonomy and proposed architecture were valuable, the study primarily analyzed early-stage research and did not sufficiently address modern advancements such as intelligent or hybrid algorithms.

Sadashiv [3] provided a comprehensive survey comparing various load-balancing strategies based on their impact on latency, scalability, and energy usage. Although they introduced useful classifications and performance trade-offs, the study relied heavily on theoretical models and case studies without addressing the lack of practical implementations or validation across diverse environments.

Shakeel and Alam [4] reviewed 60 heuristic, meta-heuristic, and hybrid algorithms used in both cloud and fog contexts, offering an extensive framework and classification scheme. Their analysis covered QoS indicators such as response time and resource utilization but lacked a detailed focus on fog-specific architectures and their operational constraints. Moreover, their inclusion of cloud computing diluted the specificity required for fog-centric insights.

Ebneyousef and Shirmarz [5] provided a taxonomy of load-balancing algorithms and highlighted the importance of QoS in fog computing. They underscored the need for adaptive and scalable algorithms but did not delve into algorithmic trends such as the use of reinforcement learning or blockchain integration. Additionally, their dataset was restricted to 50 articles, potentially omitting emerging contributions.

While these prior studies offer valuable insights, they exhibit several limitations, including outdated datasets, insufficient performance evaluations, limited architectural analysis, and a lack of focus on recent innovations such as AI integration, privacy-preserving strategies, and real-world deployments. Furthermore, most reviews did not categorize workload types or evaluation tools in depth. In contrast, this study presents a broad and up-to-date systematic literature review of 113 peer-reviewed articles (2020–2024). It provides the following:

• A detailed classification of fog computing architectures and their scalability, security, and application domains.
• A comprehensive taxonomy of load-balancing algorithms, including heuristic, meta-heuristic, ML/RL, and hybrid strategies.
• An in-depth evaluation of performance metrics, workload types, and assessment tools.
• A synthesis of current challenges, emerging trends, and research opportunities, including AI-driven, blockchain-based, and privacy-aware approaches.

By addressing these dimensions, this review contributes to a more complete and actionable understanding of load balancing in fog computing environments.

3. Methodology

This study follows a systematic literature review (SLR) methodology grounded in the guidelines proposed by Keele [6], with a focus on transparency, repeatability, and rigor. The methodology is structured across several stages: the formulation of research questions, the design of the search strategy, a study selection based on inclusion and exclusion criteria, quality assessment, data extraction, and synthesis.

3.1. Research Questions

The primary objective of this SLR is to analyze the state of the art in load balancing within fog computing. To achieve this, seven research questions (RQs) were formulated:

• RQ1: What are the various architectures of fog computing, and how do they differ in terms of functionality, scalability, and application?
• RQ2: What types of load-balancing strategies or algorithms are applied in fog computing environments?
• RQ2.1: What are the advantages and disadvantages of these strategies?
• RQ3: What performance metrics are most commonly used to evaluate load-balancing algorithms in fog computing?
• RQ4: What workload types are frequently used to evaluate load balancing in fog computing (e.g., static vs. dynamic)?
• RQ5: What evaluation tools are commonly employed for assessing load-balancing algorithms in fog computing?

• RQ6: What methods are used to assess load-balancing effectiveness in fog computing environments?
• RQ7: What are the key challenges, emerging trends, and unresolved issues related to load balancing in fog computing?

3.2. Study Selection Criteria

To ensure a high-quality and relevant dataset, predefined inclusion and exclusion criteria were applied (

Table 1. Inclusion and exclusion criteria.

Criteria	Inclusion	Exclusion
Literature	Peer-reviewed journal articles (Q1/Q2)	Non-indexed journals, retracted papers, conference proceedings
Publication Date	2020–2024	Prior to 2020
Language	English	Non-English
Focus	Fog computing with emphasis on load balancing	Studies on edge computing or unrelated topics

). Only journal articles from Q1 and Q2 indexed sources, written in English, and published between 2020 and 2024 were considered. Conference papers, book chapters, and articles irrelevant to fog computing or load balancing were excluded.

3.3. Search Strategy Design

A systematic search was conducted in October 2024 across three major scientific databases: Scopus, Web of Science (WoS), and IEEE Xplore. Keywords and Boolean operators were defined based on the research questions and refined iteratively. Search string example: (“load balance*” OR “load distribution” OR “load scheduling” OR “workload distribution”) AND (“fog computing”) AND (“algorithm” OR “method” OR “strategy” OR “approach”).

This process yielded 322 articles: Scopus (80), IEEE (64), and WoS (178). After removing duplicates and applying eligibility screening, 113 studies were selected for final inclusion. Figure 1 illustrates the steps of selection for the final set of 113 studies, which were used for data extraction and synthesis.

3.4. Quality Assessment

To ensure the reliability of the selected studies, a quality assessment framework was applied using five criteria (

Table 2. Quality assessment criteria.

Criterion ID	Description
C1	Are the research objectives clearly defined?
C2	Are the methods well described and appropriate?
C3	Is the experimental design justifiable?
C4	Are the performance results measured and reported in detail?
C5	Are limitations acknowledged and conclusions supported by results?

Three reviewers independently evaluated each paper, with disagreements resolved via consensus or majority vote.

). Each study was scored using a binary scale: Yes (1 point) and No (0 points). Only studies with a score of ≥3 were retained for synthesis.

3.5. Data Extraction

A structured data extraction form was developed to systematically capture relevant information from each selected study. This included the following:

• Architecture type, layers, and node configurations (RQ1);
• Load-balancing strategy, classification, and brief description (RQ2);
• Advantages and disadvantages of algorithms (RQ2.1);
• Performance metrics and evaluation tools (RQ3–RQ5);
• Workload types (RQ4);
• Assessment methods (RQ6);
• Challenges, trends, and open issues (RQ7).

Each paper was independently reviewed by two authors to ensure accuracy and consistency.

3.6. Data Synthesis

Extracted data were synthesized using quantitative tabulation and qualitative analysis and organized around each research question. Microsoft Excel was used to compile the results, create visualizations, and identify thematic patterns. Where appropriate, the results were categorized by algorithm type, metric usage frequency, architectural design, or application domain.

3.7. Threats to Validity

This review acknowledges several potential threats to validity:

• Selection bias: To mitigate this, we used multiple databases and rigorous inclusion/exclusion criteria.
• Publication bias: Only Q1/Q2 journal articles were selected, which may exclude the relevant gray literature.
• Reviewer bias: Dual review and majority voting were employed to reduce subjectivity during selection and extraction.
• Tool limitations: Some studies lacked transparency in tool usage, making cross-comparison more difficult.

To further reduce bias, we employed backward and forward snowballing techniques to identify any missing but relevant studies.

4. Results and Discussion

This section presents the findings of the systematic literature review (SLR), organized according to the research questions outlined in Section 3. Each subsection synthesizes data across the 113 selected studies and provides both quantitative and qualitative insights into the current landscape of load balancing in fog computing.

4.1. RQ1: Fog Computing Architectures

The analysis reveals that three-layer architectures (IoT–Fog–Cloud) dominate current fog computing designs, referenced in over half the studies (e.g., [7,8,9,10,11,12]). These architectures facilitate efficient task distribution by enabling low-latency processing at the fog layer and long-term storage at the cloud layer. Dynamic clustering is a recurrent feature, enhancing scalability and fault tolerance. Hierarchical architecture appears in studies such as [13,14], introducing additional fog layers to enable regional aggregation and decision-making. These designs support geographically distributed deployments, making them suitable for applications such as intelligent transportation systems. Flat (decentralized) models [15] are leveraged in highly dynamic environments where node independence is critical, such as mobile ad hoc networks. Cloud–Fog–Edge (CFE) architectures, described in [16,17], offer more modular task assignment through fine-grained control over edge and fog layers.

While often discussed in relation to edge computing, fog computing represents a broader paradigm. Edge computing typically refers to computation that takes place directly on or near the data-generating devices (e.g., sensors, gateways). In contrast, fog computing extends this model by introducing an intermediate layer of processing nodes-known as fog nodes- which may include local servers, routers, or base stations. These nodes provide storage, computation, and networking services closer to the edge but still one step above the data source. This hierarchical structure supports context-aware processing, improved scalability, and more granular resource management. In this review, we specifically focus on fog computing, while occasionally referencing edge computing where it overlaps or is part of a fog-based architecture.

Microservices-based and peer-to-peer architectures [18,19] introduce flexibility and autonomous decision-making, particularly for real-time analytics and smart home scenarios. Security, task scheduling efficiency, and fault tolerance were commonly addressed across all architectural types. Notably, studies integrating blockchain [20] and software-defined networking (SDN) Xia, et al. [21] proposed architectures with advanced trust and resource control mechanisms.

Figure 2, Figure 3 and Figure 4 are visualizations of the results in

Table A1. Summary of fog computing architectures across research studies.

Study	Number of Layers	Architecture Description	Number of Nodes	Number of Clusters
[10]	3	Cloud layer, fog layer, terminal nodes layer	50–300	6–14
[23]	Multiple	Multiple layers, including cloud, fog nodes, and end devices.	NA	NA
[7]	3	IoT devices, fog nodes, cloud/central server	50	NA
[22]	3	End-device layer, fog layer, cloud layer	2–30 fog nodes	NA
[15]	Flat	The architecture is non-hierarchical, which suggests a flat structure rather than multiple layers	NA	NA
[13]	3	Local fog layer, adjacent fog nodes, cloud layer	4–16	NA
[11]	3	IoT devices (edge), fog layer, cloud layer	NA	NA
[8]	3	Device, fog, control layers	100–600 IoT, 40 fog nodes, 3 cloud nodes	Clusters for task management
[35]	3	IoT devices, fog nodes, cloud	75 fog nodes	5 clusters
[12]	3	Cloud computing layer, fog computing layer, user devices	11 nodes	NA
[9]	3	IoT layer, fog layer, cloud layer	12 IoT devices, 4 access points, 4 gateways, 6 fog nodes, 1 cloud data center	4 clusters
[25]	3	Devices layer, fog layer, cloud layer	236–578 vehicles, fog nodes, mobile and dedicated	Dynamically created clusters
[54]	3	Edge devices, fog nodes, cloud infrastructure	22 IoT nodes, 3 fog nodes, 1 cloud node	NA
[52]	3	IoT devices, fog computing, cloud services	5–25 fog nodes	NA
[57]	Multiple	Distributed layers, including sensors, fog nodes, and cloud layers	NA	NA
[59]	3	IoT devices, fog computing resources, cloud data centers	50 IoT devices, 4 fog providers (50 VMs each)	4 clusters
[53]	3	User devices, fog nodes, cloud systems	5, 15, 30, 50 VRs	NA
[66]	Multiple	Multi-layer architecture integrating edge, fog, and cloud environments	10–25 IoT devices, 10 fog nodes	NA
[31]	4	Production equipment, fog computing, cloud computing, user layers	10 fog nodes	NA
[67]	3	IoT devices, fog layer, central cloud layer	3 fog servers	NA
[68]	Variable	Varies per deployment scenario	NA	NA
[69]	3	NA	5 Fog nodes	4 cluster
[75]	Variable	various architectures, including sequential, tree-based, and DAG-based	NA	NA
[74]	7	Multiple fog nodes and a centralized cloud environment	1073 fog nodes distributed across 7 layers	Puddles based on geographic proximity
[29]	3	Infrastructure tier, fog tier, global management tier	1000 OBUs, 50 RSUs	RSUs grouped by LSDNC
[45]	2	Fog layer, cloud layer	NA	NA
[16]	3	Dew layer, fog layer, cloud layer	NA	NA
[61]	2	Fog layer, cloud layer	5 cloud nodes, 3 fog nodes	NA
[70]	Multiple	Multi-layered fog computing paradigm extending from IoT sensors to cloud	Varies, heterogeneous fog nodes	Clusters of IoT devices
[71]	3	Cloud layer, fog layer, IoT layer	100–500 fog nodes	NA
[99]	3	IoT devices, fog layer, cloud services	Multiple fog nodes and cloud nodes	NA
[96]	3	End devices, cog nodes, cloud layer	16–64 fog nodes	NA
[63]	3	End-user devices, fog layer, cloud services	40–100 VMs, 50 physical machines	NA
[30]	4	Acquisition layer, image layer, computing layer, robot layer	5 fog nodes	NA
[72]	3	IoT layer, fog layer, cloud layer	10–25 fog nodes	NA
[39]	3	IoT layer, fog layer, cloud layer	1 master node, multiple SBC devices	Homogeneous or heterogeneous clusters
[32]	4	IoT gateways, fog broker, fog cluster, applications layer	5 fog nodes	1 cluster
[52]	4	Cloud layer, proxy server layer, gateway layer, edge device layer	Varies based on simulation	NA
[34]	3	Edge layer, fog layer, cloud layer	500–2000 fog nodes	NA
[42]	3	Cloud layer, fog layer, end-user layer	6 fog nodes	NA
[14]	Multiple	A multi-layered approach involving edge, fog, and cloud layers	2 Raspberry Pis	Single Kubernetes cluster
[21]	4	Cloud computing layer, SDN control layer, fog computing layer, user layer	10 fog nodes, 2 cloud nodes	SDN control allows for dynamic clustering
[40]	3	Edge layer, fog layer, cloud layer	29 fog nodes	Dynamic clusters
[80]	3	Edge layer, fog layer, cloud layer	4–5 fog nodes	Dynamic clusters
[76]	3	The sensing layer, fog layer, cloud layer	Multiple fog nodes	Dynamic clusters
[89]	3	Cloud layer, fog layer, end-user layer	4 fog nodes, 1 cloud node	NA
[77]	3	End-user layer, fog layer, cloud layer	2 fog nodes, 1 cloud node	NA
[36]	3	Lightweight nodes layer, access points layer, dedicated computing servers’ layer	400 lightweight nodes, 10 access points	Dynamic logical groupings
[62]	3	Edge/IoT layer, fog layer, cloud layer	50 fog nodes, 1 cloud node	Dynamic groupings
[37]	3	IoT device layer, fog layer, cloud layer	20 fog nodes, 40 volunteer devices, 1 cloud data center	Logical grouping based on proximity
[100]	2	IoT layer, fog layer	10 fog nodes	NA
[78]	2	Vehicular layer, RSU layer	6 RSUs, 1000–3000 vehicles	Dynamically created clusters
[101]	3	Edge layer, fog layer, cloud layer	100 edge devices, 20 fog devices, 5 cloud servers	NA
[38]	3	IoT layer, fog layer, edge server layer	1 edge server, multiple fog nodes	Dynamic clusters based on workload
[24]	2	Fog layer, cloud layer	3 fog servers, 1 cloud server	Single fog cluster
[26]	3	IoT layer, fog layer, cloud layer	Dynamic, based on parked vehicles	Dynamically formed clusters
[81]	3	Vehicular fog layer, fog server layer, cloud layer	Vehicular fog nodes, RSUs, cloud nodes	Dynamically created clusters
[20]	4	Perception layer, blockchain layer, SDN and fog layer, cloud layer	Distributed RSUs, 500 vehicles	Dynamically created clusters
[82]	3	Vehicular layer, fog layer, cloud layer	UAVs and RSUs, dynamic	Dynamic swarms
[46]	2	Fog layer, IoT device layer	4 fog nodes, 50 IoT devices	4 clusters
[64]	3	Fog layer, edge layer, cloud layer	5 UEs per edge server, 5 edge servers, 1 cloud server	5 clusters
[55]	3	IoT device layer, fog layer, cloud layer	2 fog servers, 1 cloud server	2 clusters
[18]	3	IoT layer, fog layer, cloud layer	Multiple fog nodes, dynamic environment	Dynamic clustering
[56]	2	Fog layer, cloud layer	15 fog nodes, varying IoT devices	Dynamic clusters
[90]	3	IoT layer, fog layer, cloud layer	Varies based on demand	Dynamic clustering
[83]	2	Fog layer, cloud layer	Distributed fog nodes, cloud nodes	Dynamic clusters
[91]	3	Microgrid layer, fog layer, cloud layer	3 microgrids, fog nodes	3 clusters
[27]	3	IoT layer, fog layer, cloud layer	9 edge servers, dynamic IoT devices	Dynamic clusters
[84]	3	Edge layer, fog layer, cloud layer	Multiple appliances, smart sensors	Dynamic clusters based on HEMS
[17]	4	IoT layer, edge layer, fog layer, cloud layer	Multiple edge, fog, and cloud nodes	Dynamic clusters
[85]	3	Containerized workload layer, fog layer, cloud layer	6 nodes, Kubernetes	Dynamic clusters
[28]	3	Edge, fog, cloud	100 fog nodes, 10–20 IoT devices	Task-based clusters
[86]	3	IoT devices layer, fog landscape layer, cloud layer	Multiple fog cells	Fog colonies as clusters
[97]	2	Fog layer, cloud layer	NA	NA
[98]	3	Sensor layer, fog layer, cloud layer	Multiple fog nodes	Dynamic clusters
[19]	4	Sensor layer, fog devices, proxy servers, cloud data centers	30 microdata centers	Hierarchical MDCs
[102]	3	IoT layer, fog layer, cloud layer	NA	NA
[87]	3	IoT layer, fog layer, cloud layer	50 fog nodes	NA
[88]	3	IoT layer, fog layer, cloud layer	NA	NA
[60]	3	IoMT layer, fog layer, cloud layer	100–500 fog devices	NA
[103]	4	Fog devices, cloud data center	5 devices at each tier	NA
[65]	Multiple	NA	10–70 heterogeneous fog nodes	3 clusters
[92]	Multiple	Multiple layers, including fog nodes and cloud	500 fog nodes, 200 fog–cloud interfaces	NA
[104]	4	Edge layer, Base station layer, fog layer, cloud layer	20 heterogeneous virtual machines in the fog layer	NA
[41]	3	Edge, fog, cloud	100 fog nodes, 3 data centers	NA
[50]	3	IoT layer, fog layer, cloud layer	5–20 fog nodes in experiments	NA
[105]	3	EDC, Fog nodes, cloud data centers	Distributed fog nodes between data sources and the cloud	NA
[33]	4	Perception layer, fog layer, cloud layer, communication layer	Dynamic based on vehicles and lanes	NA
[106]	3	End users, fog nodes, cloud layer	100 fog nodes, centralized cloud servers	NA
[47]	3	IoT layer, fog layer, cloud layer	Varies with the application; fog servers categorized into overloaded, balanced, and underloaded	NA
[107]	3	IoT devices layer, fog layer, cloud layer	20 fog nodes, 6 cloud nodes	NA
[108]	3	IoT layer, fog layer, cloud layer	Dynamic, fog nodes per region, cloud nodes	NA
[93]	3	IoT device layer, fog layer, cloud layer	1–5 fog nodes as gateways	10 clusters
[51]	3	End-user layer, fog layer, cloud layer	2 to 200 fog nodes, dynamically grouped into clusters	10 clusters (20 nodes each)
[48]	3	IoT layer, fog layer, cloud layer	200 fog servers, dynamic categorization	NA
[109]	4	WGL, FCL, CCL, RAL	Dynamic, based on the number of IoT devices, fog nodes, and cloud servers	NA
[94]	3	Infrastructure layer, fog layer, cloud layer	10 fog nodes, 1 cloud node	NA
[44]	3	IoT layer, fog layer, cloud layer	15 IoT devices, 8 fog nodes, 30–180 VMs, 1 cloud data center	NA
[95]	3	IoT layer, fog layer, cloud layer	5–20 fog nodes	NA
[110]	3	IoT layer, fog layer, cloud layer	20 fog nodes, 30 cloud nodes	NA
[111]	2	Fog layer, consumer layer	5 fog nodes in residential areas	NA
[112]	3	IoT layer, fog layer, cloud layer	Varies dynamically based on IoT devices, fog nodes, and cloud servers	NA
[113]	3	IoT tier, fog tier, data-center tier	100 fog nodes, 10–20 IoT devices, 3 cloud data centers	NA
[49]	3	IoT layer, fog layer, cloud layer	30 fog nodes, multiple cloud nodes	NA
[2]	3	IoT layer, fog layer, cloud layer	5–20 fog nodes with VMs, 1 cloud node	NA
[58]	3	IoT layer, fog layer, cloud layer	Multiple fog nodes (small servers, routers, gateways)	NA
[43]	3	IoT devices, fog nodes, cloud data centers	15 fog nodes	NA
[114]	2	Load balancer at the first stage and virtual machines/web servers at the second stage	Single load balancer, multiple virtual machines/web servers	NA
[115]	3	Mobile devices, fog computing layer, cloud layer	10 fog nodes, 2 cloud nodes	NA
[116]	3	Edge, fog, cloud	1 edge node per consumer, 1 fog node per load point, centralized cloud data center	NA
[117]	3	IoT devices, fog layer, cloud layer	Mobile IoT devices, fog gateways, brokers, and processing servers	NA

. It can be seen in Figure 2 that most architectures have three layers, aligning with the classic IoT–Fog–Cloud hierarchy (e.g., [7,11,22,23]). This highlights the consistency in design philosophy across studies, reflecting the layered abstraction that supports scalability and modularity in fog computing environments. Meanwhile, the scatter plot (Figure 3) shows the relationship between the number of architectural layers and the number of nodes. There is no strong linear correlation between the number of layers and the number of nodes, suggesting that architectures with more layers do not necessarily imply more nodes, and the node count is more likely influenced by the deployment scope or application type (e.g., smart cities vs. industrial systems). The countplot (Figure 4) shows the distribution of cluster types. It can be seen that dynamic clustering dominates, highlighting the adaptive nature of fog computing to workload, mobility, and real-time requirements [18,24,25,26,27]. However, some studies use task-based clustering [28] or form fixed clusters, often based on geography or role (e.g., RSUs in vehicular networks [29], fog colonies [30]).

The insights from Table A1 and the related figures (Figure 2, Figure 3 and Figure 4) regarding layering trends: (a) Three-layer architectures are the de facto standard in fog computing research. (b) Studies with more than three layers [30,31,32,33] often involve complex domains like robotics, vehicular networks, or blockchain, which necessitate more nuanced abstractions. In terms of node scale variability, node counts vary widely from as few as 2 [22] to over 2000 [34], showing a huge span in deployment scales. This reflects fog computing’s flexibility, applicable to both local edge setups and large distributed infrastructures. Finally, cluster formation is either static (e.g., predefined groups in [9,35]) or dynamic based on proximity [36,37], workload [2,38], or node types (e.g., master–worker models in [39]). Figure 5 presents a 3D plot showing the relationships among layers, nodes, and clustering. It is noticed that there is no tight clustering or linear plane, which suggests that the data are widely dispersed. This reveals that architectures with few layers can still support many nodes, and systems with many nodes may use simple or complex clustering, but no single formula dominates.

4.2. RQ2: Load-Balancing Strategies and Algorithms

The reviewed literature identifies nine major categories of load-balancing strategies used in fog computing, each with distinct strengths and limitations. Fundamental strategies, such as round robin and random assignment, offer low complexity but lack adaptability in dynamic environments. Exact optimization techniques like linear programming deliver optimal results but are computationally intensive and unsuitable for real-time fog scenarios. Heuristic approaches are the most widely adopted due to their flexibility and ease of implementation, although they may yield suboptimal solutions and often require system-specific tuning.

Meta-heuristic algorithms, including Particle Swarm Optimization and Genetic Algorithms, excel in handling multi-objective optimization but are resource-heavy and often demand extensive parameter calibration. Machine learning (ML) and reinforcement learning (RL) methods are gaining popularity for their ability to predict and adapt to workload patterns in real time. However, they face significant challenges in fog environments. These include high training latency, resource demands during inference, limited interpretability (especially with deep models), and difficulties in deploying or updating models on low-power, distributed nodes. Additionally, data scarcity and privacy constraints in fog scenarios can hinder effective model training.

Fuzzy logic-based strategies are well suited for systems dealing with uncertain or imprecise inputs, although they rely on manually defined rules and do not scale well without adaptive mechanisms. Game theory models enable decentralized resource negotiation but often involve high computational costs and depend on accurate system modeling. Probabilistic methods provide lightweight solutions for stochastic systems but struggle with dynamic changes. Lastly, hybrid approaches; combinations of heuristics, ML, meta-heuristics, and fuzzy logic; are increasingly used to balance accuracy, adaptability, and resource efficiency, though they come with increased implementation complexity. Overall, while heuristic methods dominate current practice, ML/RL and hybrid models are emerging as promising solutions for adaptive and intelligent load balancing in fog computing.

Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10 and

Table A2. The advantages and disadvantages of load-balancing categories.

Category	Advantages	Disadvantages	Representative Papers
Fundamental Strategies	Simple and predictable behavior in static environments.	-Lack of adaptability to dynamic workloads.	[57,88]
	Low computational overhead, ideal for systems with minimal complexity.	-Inefficient resource utilization in heterogeneous systems.
Exact Optimization	Guarantees optimal solutions for latency and resource utilization.	-High computational demands.	[19,62]
	Ideal for theoretical modeling and small-scale systems.	-Poor scalability, unsuitable for real-time and large-scale fog scenarios.
Heuristic Appoaches	Simple and flexible with moderate computational efficiency.	-Often yield suboptimal solutions.	[20,21,74]
	Adaptable to system constraints, suitable for dynamic environments.	-Highly dependent on parameter tuning and system-specific configurations.
		-Limited exploration capabilities.
Meta-Heuristic Strategies	Excel in exploring complex solution spaces and multi-objective optimization.	-Computationally intensive.	[2,49,95,110,111,112]
	Well suited for dynamic environments, balancing global exploration and local convergence.	-Require extensive tuning of fitness functions.
		-Risk of convergence to local optima without proper hybridization or enhancements.
ML/RL	Offer intelligent automation, adapting to changing workloads in real time.	-Require large training datasets.	[7,33,41]
	Optimize across multiple objectives, suitable for dynamic and complex fog systems.	-High implementation complexity.
		-Vulnerable to overfitting and poor performance in underrepresented scenarios.
Fuzzy Logic	Handle uncertainty and imprecision effectively.	-Dependence on expert-defined rules limits scalability.	[18,44,45]
	Low resource consumption and easily adaptable to diverse scenarios with clear fuzzy rules.	-Struggle in highly dynamic environments.
Game Theory	Ensures fairness and efficiency in distributed and competitive environments.	-Computationally complex to achieve equilibrium.	[38,46]
	Well suited for resource allocation in multi-agent systems.	-Dependent on accurate real-time data, limiting scalability in dynamic systems.
Probabilistic/Statistical	Simple and efficient for tasks involving statistical variability.	-Oversimplify system dynamics, limiting applicability in deterministic or complex scenarios.	[64,69]
	Effective in handling uncertainty within fog environments.	-Limited adaptability can lead to suboptimal performance.
Hybrid Approaches	Combine strengths of multiple methods, such as meta-heuristics, ML/RL, or exact methods.	-High complexity in implementation and parameter tuning.	[2,10,20,26,31]
	Effective for multi-objective optimization in dynamic and complex environments.	-Often requires specialized hardware for deployment.

complement this discussion by illustrating the distribution, co-occurrence, and performance trade-offs among the reviewed strategies. Figure 6 and Figure 7 present the results related to

Table A3. Categorization of load-balancing strategies in fog computing.

Category	Description	Representative Papers
Fundamental Strategies	Basic approaches with predictable behavior are suitable for simple static environments.	[57,58,88]
Exact Optimization	Optimization techniques guarantee the theoretical best outcomes, ideal for small-scale scenarios.	[19,62,111,115]
Heuristic Approaches	Flexible and efficient algorithms suited for dynamic environments with moderate complexity.	[8,14,20,21,23,29,32,34,40,42,45,52,54,61,63,68,69,70,71,74,76,77,78,89,96,99,100,101]
Meta-Heuristic Strategies	Advanced strategies exploring complex solution spaces, balancing exploration and exploitation.	[2,22,35,47,49,106]
ML/RL	Machine learning and reinforcement learning models for real-time adaptive optimization.	[17,18,33,41,55]
Fuzzy Logic	Techniques for managing uncertainty using fuzzy rules, applicable to imprecise data scenarios.	[8,11,18,22,45]
Game Theory	Game theory-based methods ensure fairness and efficiency in distributed systems.	[38,46]
Probabilistic/Statistical	Probabilistic models handling variability are effective in uncertain fog environments.	[64,69]
Hybrid Approaches	Integrated methods combining strengths of multiple approaches for multi-objective optimization.	[2,10,20,26,31]

 Figure 6 shows the number of studies per strategy. It is noticed that the top category is heuristic approaches, which dominate the landscape with 80+ studies. Meta-heuristic strategies and hybrid approaches show a strong presence, indicating growing interest in adaptive and multi-objective methods. Additionally, machine/deep/reinforcement learning methods (ML/RL) are emerging but still trail behind heuristics. Figure 7 presents the strategy proportions. Heuristics account for more than 50% of all categorized strategies. Smaller yet important segments are exact optimization (~5%) and fuzzy logic/game theory (niche but valuable in uncertain or decentralized environments), and ML/RL appears in ~10% of studies, showing fast growth, especially in autonomous or real-time decision-making.

In summary, heuristic approaches are practical, scalable, and ideal for real-time load balancing in volatile environments and cited in studies like [24,27,29,40,41]. Meta-heuristic strategies are suitable for complex, multi-variable problems requiring a balance between exploration and exploitation and used in [9,12,22,28,36]. ML/RL strategies are increasingly adopted for systems needing adaptive optimization. Studies like [7,33,42,43] highlight learning-based load prediction and decision-making. Fuzzy logic is helpful in systems with imprecise inputs or needing human-like reasoning, which is seen in [8,44,45]. Game theory ensures fairness and efficiency in multi-agent fog environments, which is applied in [38,46] for decentralized negotiation and decision-making. Finally, hybrid approaches combine the strengths of other strategies for robust, versatile load balancing and are employed in [10,26,28,47].

Regarding the distribution of studies by the number of categories (Figure 8), 49 studies used a single strategy, showing strong domain-specific alignment. Yet over 30 studies appear in two to four categories, revealing a move toward multi-strategy fusion in fog environments. This shows that real-world fog computing challenges often require more than one theoretical approach, especially in dynamic or uncertain conditions. The heatmap (Figure 9) reveals how often different categories are used together in the same studies. Heuristic approaches co-occur with almost every other strategy: strongly with meta-heuristic (20+ overlaps), frequently with ML/RL, hybrid, and even fuzzy logic, hybrid approaches commonly overlap with meta-heuristic, heuristic, ML/RL, and fuzzy logic, and ML/RL + meta-heuristics is a powerful emerging combo, often used in adaptive scheduling, predictive balancing, and real-time orchestration [47,48,49].

The Venn diagram (focused on heuristic, meta-heuristic, and hybrid approaches) in Figure 10 reveals a substantial intersection where studies (like [47,50,51]) implement all three approaches, indicating high system complexity and a large, shared region between heuristic and meta-heuristic, showing a common pairing for performance tuning under constraints.

4.3. RQ2.2: Advantages and Disadvantages in Load-Balancing Strategies

The strengths and limitations of each category were analyzed across all studies: (a) Heuristic methods offer low complexity and high adaptability but often converge on suboptimal solutions. (b) Meta-heuristics provide robust optimization but require intensive parameter tuning. (c) ML/RL algorithms enable intelligent decision-making but demand large training datasets and are prone to overfitting. (d) Exact optimization ensures optimality but lacks scalability. (e) Hybrid approaches integrate benefits across techniques but often introduce implementation complexity.

Based on Table A3, Figure 11, Figure 12 and Figure 13 present some statistics. The distribution of scheduling strategies is presented in Figure 11 and Figure 12. The most prevalent is hybrid scheduling, which appears in 72 studies, making it the dominant strategy, followed by latency-aware and deadline-aware strategies, appearing in over 55 studies each. The least represented are security-aware scheduling and cost-aware scheduling; while still significant, these are covered in fewer studies (~40 each), possibly due to their specialized nature. Meanwhile, the heatmap of the co-occurrence of scheduling strategies (Figure 13) shows the high overlap zones: hybrid scheduling co-occurs with nearly every other category, acting as a meta-layer combining various constraints, and strong links also exist: latency-aware ↔ deadline-aware, energy-aware ↔ cost-aware, and context-aware ↔ resource-aware. This suggests that real-world fog applications frequently balance multiple metrics: energy, delay, cost, security, and more.

4.4. RQ3: Performance Metrics

Performance evaluation across studies primarily involved the following metric categories: (a) Latency and response time: most frequently reported (e.g., [8,9,15]). (b) Energy consumption: particularly relevant in mobile and battery-constrained environments [10,24]. (c) Resource utilization: CPU, memory, and bandwidth usage [11,52]. (d) Task completion time and throughput: common in smart city and real-time applications [7,53]. (e) Fault tolerance and deadline adherence: evaluated via metrics like task failure rate and missed deadlines [31,54].

Several studies also reported composite metrics (

Table A4. Categorization of performance metrics in fog computing.

Primary Category	Subcategory	Percentage
1. Latency and Time Metrics	Latency (ms)	16%
	Task Completion Time (sec)
	Response Time
	Communication Delay
	Waiting Time in Queue
	Turnaround Time (TAT)
	Service Delay Time
	Temporal Delay
	Processing Time
	Mean Service Time (MST)
2. Resource Utilization Metrics	Resource Utilization (CPU%)	7%
	Memory Usage
	Number of Computing Resources
	Network Utilization
	Number of Used Devices
	Bandwidth Utilization
3. Energy Efficiency Metrics	Energy Consumption (W)	4%
	Brown Energy Consumption
4. Reliability and Fault Metrics	Fault Tolerance (Yes/No)	9%
	Failure Rate
	Access Level Violations
	Deadline Violations
	Number of Deadlines Missed
5. Cost Metrics	Communication Cost	7%
	Execution Cost
	Service Cost
	Cost of Data Transmission
6. Load-Balancing and Distribution Metrics	Load-Balancing Level (LBL)	10%
	Workload Distribution
	Queue Length
	Task Delivery
	Workflows
7. Network and Scalability Metrics	Network Lifetime	8%
	Network Bandwidth
	Jitter
	Congestion
	Scalability
8. Prediction and Accuracy Metrics	Prediction Interval Coverage (PIC)	5%
	Prediction Accuracy (AAPE)
	Prediction Efficiency
	Accuracy
	Accuracy of IRS Classification
9. Quality of Service (QoS) Metrics	QoS Satisfaction Rate	18%
	Blocking Probability (bp)
	Success Ratio (SR)
	Fairness Index
	Throughput (Tasks/sec)
	Efficiency
	Scalability
	Prediction Interval Coverage (PIC)
	Prediction Accuracy (AAPE)
	Average Processing Time (APT)
	Success Ratio (SR)
	Blocking Probability (bp)
	Task Delivery
10. Security Metrics	Encryption Time	2%
	Decryption Time
	Handover Served Ratio (HSR)
	Onboard Unit Served Ratio (OSR)
11. Specialized Metrics	Solution Convergence	14%
	Queue Management
	Loss Rate
	Consensus Time
	Social Welfare
	Offloading Success
	Task Rejection Rate
	Node Selection Success Rates

) such as QoS satisfaction rates, prediction accuracy, and success ratios. Figure 14 presents the offloading strategy distribution. The most used strategies are hybrid offloading and dynamic offloading, which lead with 65+ studies each, reflecting the complexity and adaptiveness of modern fog systems. The other popular strategies are partial offloading and latency/energy-aware strategies, which are also widespread, showing the need to balance performance and efficiency. In terms of the specialized focus areas, application-specific offloading and context-aware offloading are common in targeted domains like AR/VR, healthcare, or location-based services.

The heatmap (Figure 15) shows the strongest co-occurrences: hybrid ↔ dynamic ↔ resource-aware. These three strategies frequently co-occur, as fog environments require real-time, flexible offloading based on the device state and network load. Similarly, energy-aware ↔ latency-aware ↔ context-aware is a common trio, especially in mobile and sensor-driven applications, where energy and delay are key, and context drives adaptiveness.

4.5. RQ4: Workload Types

Workloads were classified as follows: (a) Dynamic (95.6% of studies): real-time IoT data with unpredictable volumes [10,15,55]. (b) Static: benchmarking scenarios with fixed workloads [23,56]. (c) Hybrid: combining predictable and real-time tasks [57,58]. (d) Data intensive: special category for large-scale analytics and multimedia processing [13]. (e) Dynamic workloads dominate the evaluation landscape due to the volatile nature of fog environments.

Figure 16 and Figure 17 present the studies listed in

Table A5. Distribution of workload types in reviewed fog computing studies.

Workload Type	Paper IDs	Description	Frequency	Percentage
Dynamic	[7,8,9,10,15,20,21,22,31,35,74,118]	Workload is dynamically generated based on real-time conditions, reflecting the variability and unpredictability of IoT systems.	108	95.60%
Static	[23,56]	Fixed workloads without runtime changes, often used for benchmarking or theoretical analysis.	2	1.80%
Dynamic and Data Intensive	[13]	Combines dynamic nature with high data volume or computational intensity.	1	0.90%
Static/Dynamic	[57,58]	Incorporates both static and dynamic workloads to evaluate hybrid or flexible algorithms.	2	1.80%

. The distribution of fault tolerance strategies (Figure 16) shows that (dominant techniques) task migration, checkpointing, and replication are the most adopted methods, each appearing in 65+ studies. These strategies focus on recovery, continuity, and proactive protection, critical in dynamic fog environments. The supportive techniques (failure detection, self-healing, and resilient scheduling) occur in 50+ studies, showcasing a trend toward intelligence and autonomy in fault handling. Finally, comprehensive resilience; hybrid fault tolerance; appears in most studies, combining multiple techniques for layered protection.

In regard to the heatmap (Figure 17), the top co-occurring pairs are as follows: checkpointing ↔ replication, task migration ↔ resilient scheduling, and self-healing ↔ failure detection. This clustering suggests that fog systems rely heavily on failover and recovery coordination, especially when real-time services are at risk.

4.6. RQ5: Evaluation Tools

Evaluation environments ranged from simulators to real-world testbeds: (a) iFogSim (most common): used in 24 studies for modeling latency, energy, and network utilization [59,60]. (b) CloudSim and PFogSim: applied to hybrid fog–cloud scenarios [43]. (c) OMNeT++ and YAFS: chosen for detailed network simulations [61,62]. (d) Custom simulators: built using Python, Java, Docker, or MATLAB to simulate unique conditions [23,36]. Real-world testbeds: Implemented using Raspberry Pi and microcontroller-based setups in [63]. Some studies integrated multiple tools to bridge network and computer modeling (e.g., [58]).

Results presented in

Table A6. Mapping of evaluation tools for studies in fog computing.

Study	Tool Name
[11,23,24,69,99]	Custom Simulator
[7]	COSCO Simulator
[15,54]	Discrete-event Simulator (YAFS)
[40,71]	Python (SimPy)
[32]	Python (SciPy)
[83,91,102]	Python (Custom)
[65]	Java (Custom)
[57,66]	C (Custom)
[8]	IoTSim-Osmosis
[22,31,35]	MATLAB
[79,95,111,114,117]	iFogSim
[68,85]	Kubernetes and Istio
[37,58]	OMNeT++
[74]	PFogSim
[27]	EdgeCloudSim
[20,81]	SUMO Simulator
[90,115]	Mininet-WiFi
[43,49,58,113]	CloudSim
[14]	Truffle Suite and Ganache
[72,116]	Docker
[28,84]	Real-World Testbeds
[82]	AnyLogic
[2,87]	FogSim
[30]	PySpark
[39]	CloudAnalyst
[93]	Apache Karaf

reveal that the most common approaches are intrusion detection and data encryption, highlighting a dual focus on preventing and protecting against attacks. Furthermore, authentication, privacy preservation, and trust management follow closely, highlighting the importance of data integrity and user legitimacy. The emerging trends are blockchain-based security and lightweight security, which are gaining traction due to their applicability in decentralized and resource-constrained environments. The top co-occurrences are intrusion detection ↔ authentication, privacy ↔ data encryption, and trust management ↔ lightweight and blockchain. These patterns confirm that security is rarely isolated; instead, strategies are tightly coupled to address multiple threat vectors simultaneously.

4.7. RQ6: Assessment Methods

Studies utilized various assessment approaches: (a) simulation-based evaluation: dominant method due to its reproducibility. (b) Prototype implementation: used in research targeting specific use cases like smart cities or industrial IoT [28,55]. (c) Theoretical modeling: used for algorithm development and proof of concept [62,64]. Few studies incorporated empirical measurements from real-world deployments, indicating a gap in practical validation.

Table 3. Distribution of studies across current challenges in fog computing.

Challenge	Paper(s)
Resource Heterogeneity	[8,9,10,12,13,15,22,25,31,35,54,57,59,66,67,68,69]
Dynamic Workloads	[8,9,10,12,13,15,16,22,29,30,34,39,42,63,70,71,72,73]
Latency Sensitivity	[8,10,12,13,15,22,52,54]
Energy Management and Efficiency	[8,10,13,15,20,21,22,25,30,34,35,36,37,38,39,53,54,55,57,59,66,69,70,71,74,75,76,77,78]
Security and Privacy	[8,10,14,15,19,21,35,38,40,41,45,46,54,55,56,57,59,62,63,64,65,66,69,70,71,72,74,76,79,80,81,82,83,84,85,86,87,88]
Scalability	[8,12,13,14,15,20,22,27,35,36,37,39,41,44,48,52,55,63,66,77,80,88,89,90,91,92,93,94,95]
Mobility and Environmental Changes	[32,39,63,72,96]

shows that iFogSim is by far the most widely used tool, referenced in 90+ studies (

Table A7. Classification and definition of tools used for evaluating load-balancing algorithms in fog computing.

Tool Name	Classification	Definition
iFogSim	Widely Recognized Simulator	A simulator designed for fog and IoT environments, providing capabilities for modeling latency, energy consumption, and network parameters.
CloudSim	Widely Recognized Simulator	A simulation platform for modeling cloud computing environments, extended for fog computing.
PFogSim	Widely Recognized Simulator	An extension of CloudSim designed for large-scale, heterogeneous fog environments.
FogSim	Widely Recognized Simulator	A modular simulator tailored for fog and cloud computing environments.
YAFS (Yet Another Fog Simulator)	Widely Recognized Simulator	A simulator focusing on fog-specific features and performance metrics.
OMNeT++	Widely Recognized Simulator	A discrete-event network simulation framework for task distribution and communication modeling.
MATLAB	Widely Recognized Simulator	A computational environment supporting algorithm implementation and simulation, often used for fuzzy logic and deep learning.
Mininet/Mininet-WiFi	Network and IoT-Specific Tool	A network emulator for testing SDN and IoT architectures.
SUMO	Network and IoT-Specific Tool	A traffic simulation tool used for vehicular mobility and task-offloading evaluations.
IoTSim-Osmosis	Network and IoT-Specific Tool	A Java-based simulator built on the CloudSim framework for IoT and fog computing.
Truffle Suite	Blockchain and SDN Tool	A development framework for blockchain applications, used for writing and testing smart contracts.
Ganache	Blockchain and SDN Tool	A personal Ethereum blockchain simulator for testing blockchain-based implementations.
Docker	Custom Simulator Framework	A containerization platform used for deploying custom simulation environments.
SimPy	Python-Based Framework	A process-based discrete-event simulation framework in Python.
SciPy	Python-Based Framework	A Python library for optimization and simulation in custom environments.
Custom Simulators	Custom Simulator Framework	Tailored simulation environments designed for specific research scenarios.
Real-World Testbeds	Real-World Testbed	Physical setups using devices like Raspberry Pi for practical evaluations.
AnyLogic	Widely Recognized Simulator	A multi-method simulation modeling tool for evaluating complex systems.
EdgeCloudSim	Widely Recognized Simulator	A simulation environment for edge–fog–cloud computing scenarios.

). Its popularity stems from native support for IoT, fog, and mobility to strong community adoption and extensions (e.g., iFogSim2, iFogSim-Mobility). YAFS and EdgeCloudSim follow, but with significantly fewer studies. They offer more flexibility (YAFS) or more focused simulation on edge environments (EdgeCloudSim). CloudSim and its variants still appear in fog research, mainly for legacy support or hybrid cloud–fog architectures. GreenCloud, EmuFog, and PureEdgeSim are more specialized and used in niche studies, often focusing on energy, network topology, or edge-centric architectures. Finally, custom simulators are found in a handful of studies (e.g., [58,65]), typically to support novel frameworks or architectures not yet supported in public tools.

While iFogSim remains the most widely adopted simulator in fog computing research—cited in over 90 studies—its dominance warrants critical examination. One of its primary limitations lies in its simplified modeling of network dynamics. iFogSim typically assumes stable and deterministic network conditions, which do not accurately reflect the variability and unpredictability of real-world fog and IoT networks. Moreover, its workload generation models are often static or predefined, limiting its applicability for evaluating systems that require dynamic or event-driven workloads. These constraints may lead to over-optimistic performance estimates and reduce the generalizability of findings derived from iFogSim-based simulations.

In contrast, tools like YAFS (Yet Another Fog Simulator) offer greater flexibility in modeling stochastic behavior and network latency. YAFS supports dynamic node behavior and complex event-driven simulations, making it suitable for environments with mobility or failure-prone devices. OMNeT++, though more complex to configure, provides high-fidelity network simulation capabilities and is ideal for evaluating fog systems with detailed communication protocols and cross-layer interactions. Researchers focused on the accurate modeling of network variability, message delays, or routing protocols may find YAFS or OMNeT++ more appropriate than iFogSim.

Ultimately, the selection of a simulation tool should be aligned with the research objectives. While iFogSim offers ease of use and strong community support for application-level load balancing, alternatives like YAFS and OMNeT++ provide richer abstractions for network-level modeling and dynamic behaviors. A more diversified toolset can enable a deeper and more realistic understanding of fog computing performance under different operational conditions.

Despite their methodological convenience, simulation-based evaluations present a narrow view of fog computing performance under real-world conditions. The limited adoption of physical testbeds in the current research stems from several practical barriers. First, fog environments often involve heterogeneous and distributed devices, making testbed design complex and resource intensive. Deploying and maintaining such infrastructure requires not only significant financial investment but also expertise in embedded systems, networking, and security. Second, real-time testing introduces variability due to hardware failures, network delays, and environmental noise factors that are difficult to control and reproduce consistently. Additionally, there is often limited access to large-scale open-source fog testbeds, which restricts community-wide experimentation.

To address these challenges, future research could leverage affordable prototyping platforms such as Raspberry Pi clusters, Arduino boards, or Jetson Nano kits to build lightweight test environments. Cloud-integrated fog simulators (e.g., iFogSim2 with IoT hardware extensions) and federated edge–cloud frameworks can also be explored to simulate partial real-world behavior. Moreover, partnerships between academia and industry can play a vital role in providing access to realistic test environments, operational data, and deployment feedback. These approaches will help bridge the current gap between theoretical evaluations and operational viability, thus enhancing the credibility and applicability of fog-based load-balancing strategies.

4.8. RQ7: Challenges, Trends, and Future Directions

Fog computing continues to face several critical challenges that hinder its broader adoption and practical deployment. One of the primary challenges involves managing the heterogeneous capabilities of fog nodes, which can vary significantly in terms of computational power, memory, and energy efficiency. This heterogeneity complicates the process of load balancing and requires adaptable solutions that can operate effectively across diverse environments. Another pressing challenge is achieving scalability, particularly in high-density IoT environments where the sheer volume of data and the number of connected devices demand efficient, real-time processing with minimal delays.

Ensuring energy efficiency and reducing latency simultaneously remains a difficult task, especially given the resource-constrained nature of many fog nodes. As real-time responsiveness is critical in applications like healthcare and autonomous systems, achieving this balance is an ongoing concern. Security and privacy also remain at the forefront of research efforts, as decentralized fog architectures inherently expose data to a broader range of potential threats. Protecting sensitive information while maintaining performance necessitates lightweight yet effective security frameworks.

In terms of emerging trends, several technologies are shaping the next generation of fog computing systems. The adoption of artificial intelligence (AI) and machine learning (ML) is enabling adaptive load-balancing strategies that can learn and predict optimal resource allocation patterns in dynamic environments. Reinforcement learning (RL) is particularly being explored for its potential to autonomously optimize resource usage over time. Additionally, the integration of blockchain technology is gaining attention for its ability to support secure, tamper-proof data handling and decentralized trust management, which are vital in untrusted and distributed settings. Software-defined networking (SDN) is another key trend, offering centralized control over fog resources while maintaining the decentralized nature of fog infrastructure. This can facilitate the efficient orchestration and dynamic reconfiguration of resources in response to changing workloads.

Looking ahead, future research is expected to prioritize the development of real-time, energy-aware hybrid algorithms that can effectively balance multiple competing objectives such as latency, energy consumption, cost, and security. Furthermore, there is a strong need to expand real-world testbed deployments to complement simulation-based evaluations. This would provide more accurate assessments of system performance in practical conditions, accounting for hardware limitations, network variability, and user behavior. Another promising direction is cross-layer optimization, which involves coordinating decisions across application, network, and infrastructure layers to achieve holistic and context-aware scheduling.

Lastly, the lack of standardized evaluation frameworks presents a major obstacle to progress in the field. Without consistent benchmarking methodologies, it remains difficult to compare results across studies or replicate findings. Developing unified, transparent, and comprehensive evaluation frameworks is essential for ensuring reproducibility and enabling the meaningful comparisons of different load-balancing strategies.

While security and privacy have been acknowledged as critical challenges in fog computing, there is growing emphasis on implementing efficient yet robust mechanisms tailored for resource-constrained environments. Among the most widely discussed approaches is lightweight cryptography, which offers encryption and authentication using smaller key sizes and faster algorithms. These are particularly suitable for fog nodes with limited processing power and memory. However, they often involve trade-offs in terms of reduced cryptographic strength or susceptibility to emerging attack vectors and must be carefully selected based on threat models and application criticality.

Another emerging approach is the adoption of zero-trust architectures (ZTAs), where no device or user is inherently trusted, even within the network perimeter. ZTA involves continuous verification, micro-segmentation, and strict access controls. While this model enhances security posture against insider threats and lateral attacks, it introduces additional overhead from continuous monitoring and policy enforcement, which can impact system latency and increase energy consumption—especially if not optimized for fog-level operations.

Furthermore, secure data offloading and trusted execution environments (TEEs) are also being explored to protect sensitive data during transmission and processing. These solutions often rely on hardware support, which may not be uniformly available across all fog nodes, posing deployment challenges. Overall, integrating these security mechanisms must be balanced against real-time responsiveness and energy efficiency, which are often critical in fog applications such as healthcare, vehicular systems, or industrial automation.

Future research should focus on security–performance co-optimization, where security mechanisms are integrated in a way that minimally disrupts core performance metrics. Approaches such as adaptive security models, which adjust protection levels based on workload or context, may provide promising directions for aligning safety and efficiency in fog computing deployments.

Table A8. Emerging trends and technologies in fog computing research.

Emerging Trend	Key Technologies/Methodologies	Relevant Studies
AI and Machine Learning	Predictive load balancing, intelligent resource management	[7,33,41]
Edge Computing Integration	Enhancing processing capabilities for IoT	[10,78]
Reinforcement Learning (RL)	Privacy-aware load balancing, adaptive resource allocation	[15,61,66,75]
Intelligent Scheduling	Real-time data processing, decision-making adaptations	[28,42,85,109]
Multi-Layer Scheduling Methods	Complex scheduling for large-scale applications	[13,52,74]
Clustering Integration	Improving load balancing in vehicular networks	[24]
AI-Driven Task Allocation	Real-time prioritization for IoT systems	[11,32,92]
Software-Defined Networking (SDN)	Flexibility in network management	[20,21,77]
Serverless Computing	Function as a Service (FaaS) for IoT applications	[12]
Adaptive Offloading Techniques	Machine learning for resource management in vehicular fog	[25]
Fault Tolerance Mechanisms	Hybrid approaches incorporating RL	[54]
Scalability and Resilience	Managing large-scale systems and dynamic environments	[52,66,101]
Energy Management	Balancing energy efficiency with computational demands	[42,93,95]
Security and Privacy Concerns	Addressing challenges in IoT data	[19,64,76]
Interoperability Between Fog and Cloud	Seamless task distribution across resources	[37,66,106]
Hybrid Optimization Approaches	Combining meta-heuristic algorithms for better performance	[2,44,104]
Game–Theoretic Frameworks	Workload distribution strategies	[38]
Dynamic Resource Allocation	Adapting systems to real-time changes in workload	[34,80]
Blockchain Integration	Secure data handling and load balancing	[14,33,76]
Integration of Renewable Energy Sources	Energy-efficient fog devices	[70,108]
Community-Based Placement	Distributing workloads effectively across nodes	[94]
Hybrid Algorithms	Combining multiple methodologies for enhanced performance	[43,50]
Dynamic Offloading	Real-time task-offloading strategies	[24,80,89,119]
Fuzzy Logic Integration	Intelligent scheduling strategies	[117]
Nature-Inspired Algorithms	Resource optimization strategies	[105,114]

-related visualizations are presented in Figure 18 and Figure 19. The metrics of trend distribution are presented in Figure 18. The most frequent trend is latency/delay and energy consumption, which dominate and appear in over 80 studies each. These are central due to fog computing’s real-time and power-sensitive nature. Resource utilization, cost efficiency, throughput, and load balancing follow closely, confirming multi-faceted practices in the near future. Finally, security/privacy and fault tolerance, and custom metrics are used in more specialized domains such as healthcare or vehicular networks. The co-occurrence rends (heatmap) (Figure 19) show frequent pairings: latency + energy, utilization + cost, security + privacy, and latency + QoS. These clusters reflect how trends are strategically bundled: performance with efficiency, protection with compliance, and responsiveness with service satisfaction.

5. Implications

The architecture choice should match the use case: (a) simple two-layer or three-layer designs are suitable for general IoT applications. (b) Advanced domains (e.g., autonomous vehicles and health systems) benefit from multi-layer or hybrid models. With regard to dynamic Clustering Is Essential: real-time data flow, mobility, and heterogeneous resources make static clustering insufficient, and the integration of SDN (as in [21]) or orchestration platforms (like Kubernetes in [14,85]) can enable intelligent cluster management. Regarding that scalability and interoperability are ongoing challenges, the diversity in node and cluster configurations reveals a lack of standardization, which might hinder interoperability across platforms.

The design and architecture can start simply and scale smartly: in the beginning, utilize heuristics in pilot deployments; then, transition to ML/meta-heuristics as the complexity increases. For autonomous or smart environments, the priority is for ML/RL or hybrid strategies to handle dynamic and uncertain workloads. Meanwhile, while exact optimization is ideal in theory, it is often impractical for large-scale or real-time fog systems.

Regarding strategies, there is no one-size-fits-all strategy; load balancing in fog computing requires blending multiple paradigms, especially for heterogeneous, large-scale, or mobile scenarios. It could be said that heuristics are the glue; i.e., heuristics serve as the core logic layer, often enhanced by meta-heuristics, learning, or fuzzy reasoning, depending on system complexity. Finally, in terms of future trends, such as intelligent hybridization, studies like [47,50,51] showcase layered or hierarchical load-balancing models, combining rule-based, statistical, and AI-driven components.

Based on the above observations,

Table 4. Strategy recommendations.

Category	Best With	Example Use Case
Heuristic	Real-time, moderate complexity	Smart traffic coordination
Meta-Heuristic	Complex optimization problems	Vehicle routing and resource migration
ML/RL	Adaptive/predictive control	Energy-aware scheduling in fog systems
Fuzzy Logic	Input uncertainty or vagueness	Health monitoring with fuzzy sensor data
Hybrid Approaches	Multi-objective, large-scale, or mobile systems	Smart cities with dynamic resource allocation

presents a tailored recommendation matrix for selecting load-balancing strategies in fog computing environments. Each strategy is mapped to scenarios where it performs optimally, along with real-world examples to guide implementation. Heuristic approaches are best suited for systems requiring quick, real-time decisions under moderate complexity, such as smart traffic control. Meta-heuristic strategies are ideal for solving complex optimization problems with constraints, commonly found in vehicle routing or task-offloading scenarios. ML/RL strategies enable adaptive and predictive load balancing, especially useful in energy-aware fog environments. Fuzzy logic excels in handling vague or imprecise input data, as seen in health monitoring systems using uncertain sensor readings. Hybrid approaches combine multiple methods to address diverse objectives; like latency, cost, and mobility; making them highly effective in large-scale or heterogeneous systems, such as smart cities. This strategic mapping offers a practical reference for researchers and practitioners to align system requirements with the most effective load-balancing methods.

In terms of scheduling, hybrid scheduling is essential. Complex fog environments demand multi-objective optimization (e.g., [2,28,97]). Especially vital in smart city, healthcare, and mobility-driven use cases. Latency and deadlines are core concerns, and time-sensitivity is a major driver, particularly in IoT and edge applications [27,38]. Energy and cost often travel together as fog nodes, which are resource constrained, so studies often address these factors together [41,76]. Finally, security is gaining attention, appearing in 40+ studies, which shows a rising concern for secure offloading, authentication, and data handling [32,47,98].

Table 5. Scheduling strategy recommendation matrix.

Category	Best When	Example Use Case
Resource-aware	System has limited fog resources like CPU or bandwidth	Workload balancing in industrial IoT systems
Deadline-aware	Tasks must be completed within strict time constraints	Real-time video analytics or emergency response
Energy-aware	Reducing power usage is a high priority	Battery-constrained sensor networks or mobile fog nodes
Latency-aware	Application requires real-time responsiveness	Autonomous vehicles, remote surgery, AR/VR streaming
Security-aware	Sensitive data or user privacy are involved	Healthcare data processing or financial transactions
Cost-aware	Cloud offloading or communication costs must be minimized	Fog–cloud collaboration in smart cities
Context-aware	Environmental or user context must adapt scheduling decisions	Location-based task delegation or user mobility prediction
Hybrid scheduling	Multiple conflicting objectives must be balanced	Smart cities managing energy, latency, and cost together

presents practical mapping between scheduling strategy types and their optimal application contexts in fog computing environments. Each strategy is associated with scenarios where it is most effective, along with representative real-world use cases. This matrix serves as a design guideline for selecting scheduling algorithms that align with the specific operational goals and constraints of fog-based systems.

Regarding the offloading in Fog is contextual and dynamic; the popularity of dynamic and context-aware offloading confirms that systems must continuously adapt to environmental and device conditions. Similarly, hybrid techniques dominate real deployments. Real-world fog deployments often need to balance latency, energy, and computation load, which explains the dominance of hybrid strategies. On the other hand, while full offloading is useful for constrained devices, partial offloading is more flexible and efficient in bandwidth-limited or real-time cases. Eventually, cross-layer decision-making is common; co-occurrence across 4+ categories shows that fog-based offloading decisions are not isolated but rather integrated with scheduling, resource management, and security.

Table 6. Offloading strategy recommendations.

	Best When	Example Use Case
Full Offloading	End device is severely resource constrained, or idle task tolerance is high	Wearable health sensors offloading all processing to fog/cloud
Partial Offloading	Partial processing is possible on device to save bandwidth or preserve privacy	Edge preprocessing for smart surveillance before cloud analytics
Dynamic Offloading	Network or device status changes rapidly; decisions must be adaptive	Mobile users in vehicular fog systems or smart retail scenarios
Application-Specific Offloading	Offloading depends on application logic or requirements	Augmented reality or IoT applications with tailored offloading models
Resource-Aware Offloading	Fog/cloud resources are heterogeneous, and availability varies	Load balancing in multi-tier fog infrastructure
Latency-Aware Offloading	Real-time response is critical	Tactile internet or remote healthcare diagnostics
Energy-Aware Offloading	Energy savings are more important than performance	Battery-powered IoT devices or sensor networks
Context-Aware Offloading	Offloading needs to adapt to user behavior or environmental data	Location-based task assignment in smart cities
Hybrid Offloading	Multiple objectives like latency, energy, and context must be optimized together	Smart transportation systems with high mobility and QoS constraints

presents practical guidance for selecting offloading strategies in fog computing based on system characteristics and application needs. Each strategy is aligned with specific operational contexts and real-world use cases. This matrix supports developers and researchers in choosing the most suitable offloading mechanisms to align with diverse operational goals and application scenarios in fog-based systems.

Regarding fault management, tolerance in fog is not optional. The consistent use of migration, replication, and recovery strategies confirms that faults are expected and must be managed proactively. Furthermore, multi-layer protection is commonplace. The overlap of 5–7 strategies in most studies indicates that single-method fault tolerance is insufficient for real-world fog deployments. On the other hand, predictive and self-healing trends are rising. The increasing use of AI-based detection and self-healing shows a shift toward autonomous fog systems that can adapt to failures without human intervention. Finally, hybrid approaches are now standard practice. Systems that combine detection, redundancy, and recovery strategies are more resilient, especially in mission-critical applications like healthcare or industrial automation.

Then, it could be said that security is multi-dimensional. No single technique suffices. Systems must incorporate both proactive (e.g., intrusion detection) and preventive (e.g., access control and encryption) layers. Fog security must be lightweight. The rise in lightweight and blockchain-based security solutions reflects a growing demand for scalable yet efficient protection. Finally, trust and privacy are critical; with sensitive data traversing fog nodes, trust management and privacy-preserving methods are essential, especially in healthcare and financial systems.

6. Conclusions and Future Work

6.1. Conclusions

This systematic literature review offers a comprehensive examination of load-balancing strategies within the context of fog computing. By analyzing 113 peer-reviewed articles published between 2020 and 2024, this study identifies prevailing architectural models, algorithmic advancements, evaluation metrics, and emerging trends that shape current research in the field.

This review reveals that three-layer architectures (IoT–Fog–Cloud) are the most widely adopted, as they provide an effective balance between localized responsiveness and centralized control. More complex hierarchical and multi-layer configurations further enhance scalability and are especially well suited for data-intensive and latency-sensitive applications. In terms of load-balancing strategies, dynamic approaches; particularly heuristic, meta-heuristic, and hybrid algorithms; dominate the landscape due to their adaptability in heterogeneous and fluctuating environments. Increasingly, machine learning and reinforcement learning techniques are being leveraged to support intelligent, predictive load balancing.

Performance evaluation in the reviewed studies frequently focuses on latency, response time, energy consumption, and resource utilization. Over 95% of the studies simulate dynamic workloads, underlining the importance of context-aware and real-time distribution strategies. iFogSim emerges as the most commonly used simulation tool, followed by CloudSim and several custom-built environments. Nevertheless, a notable limitation persists: the absence of real-world deployments in many studies restricts the generalizability and practical applicability of their results.

Despite substantial progress, fog computing continues to face critical challenges, particularly in scalability, security, and energy efficiency. Promising research directions include the integration of artificial intelligence for adaptive scheduling, blockchain for decentralized trust management, and software-defined networking (SDN) for enhanced control and orchestration. These technologies have significant potential to address the unique constraints of privacy-sensitive and resource-limited environments. Ultimately, this review provides researchers and practitioners with a structured roadmap of existing methodologies, while highlighting unresolved issues and areas that demand further exploration.

6.2. Future Work

Despite substantial progress in fog computing and load-balancing strategies, several critical challenges and research gaps remain unaddressed. One of the primary limitations is an overreliance on simulation-based evaluations. While simulations offer a controlled environment for preliminary testing, they often fail to capture the complexities of real-world deployments. Future research should therefore focus on the development of real-world testbeds to validate algorithmic performance under operational conditions, providing insights into practical constraints such as hardware limitations, network variability, and user dynamics.

Another important area of focus is energy-aware and green computing. As fog nodes are typically deployed near end-users and may operate with constrained power resources, there is a pressing need for energy-efficient strategies. These strategies must optimize system performance without compromising device longevity or reliability. Integrating energy consumption as a key metric in load-balancing decision-making is essential for sustainable fog computing environments.

Moreover, the dynamic and heterogeneous nature of fog environments necessitate the development of context-aware load-balancing mechanisms. Future approaches should not only respond to workload fluctuations but also incorporate contextual information such as user mobility, geographical distribution, and application-specific Quality of Service (QoS) requirements. Contextual adaptation will enhance the responsiveness and efficiency of resource management in highly dynamic IoT scenarios.

Security and privacy also remain critical concerns. Many IoT applications process sensitive or personal data, making it imperative that future load-balancing solutions integrate security features at their core. Approaches that incorporate encryption, trust management frameworks, and blockchain technologies into load-balancing protocols could offer enhanced data protection while maintaining performance.

In addition, most existing algorithms focus on optimizing a single objective, such as latency or energy consumption. However, real-world applications often require a balance across multiple criteria, including performance, cost, energy efficiency, and security. Future research should explore multi-objective optimization techniques that can effectively manage these competing priorities within a unified framework.

Finally, the lack of standardized evaluation frameworks hinders the ability to conduct fair and reproducible comparisons across different studies. The development of a unified benchmarking framework is essential to promote transparency, improve comparability, and accelerate progress in the field.

By addressing these challenges, future research can pave the way for the development of resilient, scalable, and intelligent fog computing systems capable of supporting the diverse and evolving needs of next-generation IoT applications.

While this review provides a qualitative synthesis of architectural and algorithmic strategies in fog computing, future research should move toward quantitative meta-analysis to enhance comparability and rigor. A statistical aggregation of key performance metrics; such as latency reduction, energy savings, and task completion times; would enable a more objective evaluation of the competing strategies. Such a synthesis could be derived from benchmarking results across tools like iFogSim, YAFS, and OMNeT++ using standardized workloads.

Moreover, there is a critical need for deployment-focused case studies that go beyond simulation. Industrial prototypes and smart city implementations; such as those in healthcare monitoring or vehicular networks; can offer vital insights into real-world constraints, including hardware limitations, environmental volatility, and user behavior. These studies are particularly important for evaluating interoperability and scalability, as fog systems often span heterogeneous platforms, protocols, and administrative domains.

To unify these efforts, we advocate for the development of an open benchmarking framework that defines standardized evaluation scenarios, metrics, datasets, and reporting formats. Such a framework would greatly improve transparency, reproducibility, and cross-study comparability- key pillars for advancing the technical maturity of fog computing research and facilitating real-world adoption.