PaperlayerView exact revision on arXiv

Source-linked AI summary

Scheduling in Cloud,Fog, and Edge Computing: A Comparative Analysis and Decision Framework for Technique Selection

Rana Ghazali, Douglas G. Down

arXiv:2610.13830v1cs.DC

TL;DR

Scheduling across cloud, fog, and edge is challenging, while existing reviews provide limited guidance for choosing techniques across environments and scheduling granularities. This review compares technique families across requirements and develops a decision framework; its synthesis describes where each family is appropriate and the trade-offs involved.

  • Problem

    Existing reviews provide limited guidance for selecting scheduling techniques across cloud, fog, and edge environments and task, job, and workflow granularities.

    Source: §1 · pp. 1–2

  • Method

    The review comparatively analyzes scheduling studies across technique families, environments, granularities, objectives, constraints, and deployment requirements to develop a selection framework.

    Source: §1 · p. 2

  • Results

    No technique family is universally superior: heuristics suit stable, low-overhead decisions, metaheuristics complex multi-objective problems, learning-based methods dynamic settings, and hybrids combine capabilities with added complexity.

    Source: §8 · p. 56, §9 · p. 57

  • Takeaways & Limitations

    Technique selection should match workload, environment, objectives, resource constraints, system dynamics, and computational budget rather than assume one family fits all problems.

    Source: §1 · p. 2

    Evaluation relies substantially on simulation, and limited real-world testbeds and operational workloads restrict conclusions about deployment robustness, workload drift, inference overhead, and practical scalability.

    Source: §6.3.3 · p. 41

Abstract

from arXiv · show

The increasing heterogeneity and dynamism of cloud, fog, and edge computing make scheduling a challenging problem involving diverse resource capabilities, workload characteristics, latency and energy constraints, and competing optimization objectives. Although numerous scheduling methods have been proposed, existing reviews often provide fragmented coverage of computing environments, scheduling granularities, or technique families, offering limited guidance on selecting an appropriate approach for specific requirements. This review presents a comparative analysis of task, job, and workflow scheduling across the cloud, fog, and edge paradigms, covering heuristic, metaheuristic, learning-based, and hybrid techniques. Studies published between 2020 and 2026 are analyzed according to scheduling granularity, computing environment, optimization objectives, constraints, performance metrics, and evaluation tools. The analysis highlights trade-offs among computational efficiency, optimization capability, adaptability, and implementation complexity across technique families. Heuristic methods offer low computational overhead for relatively stable scheduling conditions, whereas metaheuristic methods are suited to complex and multi-objective optimization problems. Learning-based methods provide greater adaptability to dynamic and uncertain environments but introduce data, training, inference, and convergence requirements. Hybrid methods combine complementary capabilities but generally increase computational and implementation complexity. Based on these findings, a decision framework is developed to guide technique selection according to workload characteristics, environmental conditions, optimization objectives, resource constraints, and computational requirements.

Keywords: Cloud computing; Fog computing; Edge computing; Task scheduling; Workflow scheduling; Job

The keywords identify the paper’s focus on cloud, fog, and edge scheduling, including task, job, and workflow concerns.

Keywords

  • The review focuses on cloud, fog, and edge computing, scheduling, resource allocation, and scheduling optimization.Its technique keywords include heuristic, metaheuristic, learning-based, and hybrid approaches.Source: Keywords: Cloud computing; Fog… · p. 1

1. Introduction

Scheduling across cloud, fog, and edge environments must account for diverse resources, workloads, constraints, objectives, and computational granularity. This review addresses fragmented prior coverage with a comparative analysis and a requirement-oriented framework for selecting among technique families.

1. Introduction

  • Cloud, fog, and edge scheduling spans heterogeneous resources and distinct task, job, and workflow granularities, with dependencies and communication costs affecting technique suitability.An approach suited to independent tasks may be less appropriate for dependency-constrained workflows or large-scale jobs.Source: §1 · p. 1
  • Prior reviews provide limited guidance on matching scheduling requirements to techniques and often separate computing paradigms, granularities, or technique families.The review analyzes 28 surveys published between 2020 and 2026 to characterize this landscape.Source: §1 · p. 2
  • The review analyzes 183 primary studies across cloud, fog, and edge using scheduling granularity, environment, objectives, constraints, evaluation, and deployment requirements.It develops a comparative, requirement-oriented synthesis rather than an algorithm-by-algorithm catalogue.Source: §1 · p. 2
  • The decision framework relates scheduling entity, environment, workload, objectives, resource constraints, system dynamics, and computational budget to technique selection.It does not assume any technique family is universally superior.Source: §1 · p. 2
  • The review jointly compares task, job, and workflow scheduling across heuristic, metaheuristic, learning-based, and hybrid techniques.It examines how scheduling granularity affects technique requirements and selection.Source: §1 · p. 2
  • The analysis synthesizes research trends, dominant objectives, evaluation practices, trade-offs, research gaps, and future directions across the four technique families.It highlights challenges involving dynamic workloads, multi-objective optimization, realistic evaluation, scalability, reliability, QoS, and practical deployment.Source: §1 · pp. 2–3

2. Research methodology

The review uses a systematic comparative methodology to relate scheduling requirements and computing environments to the suitability of technique families. It identifies and selects recent studies, classifies them across common dimensions, and synthesizes the findings into selection guidance.

2. Research methodology

  • The methodology classifies task, job, and workflow studies into four technique families and compares their environments, requirements, and workload characteristics.Its three stages are literature identification and selection, comparative classification and synthesis, and a technique-selection framework.Source: §2 · p. 3

2.1. Literature Search and Study Selection

  • The search covered major scientific databases using terms for computing environments, scheduling entities, and technique families, with backward reference searching to find additional studies.The search was limited to 2020–2026 publications.Source: §2.1 · p. 3
  • Eligible primary studies addressed cloud, fog, edge, or integrated environments; scheduling or related resource allocation; and an evaluated technique from one of the four families.Studies also needed sufficient methodological information and simulation-based or experimental evaluation.Source: §2.1 · pp. 3–4
  • The final corpus comprised 28 representative surveys and 183 primary studies: 12 heuristic, 105 metaheuristic, 22 learning-based, and 44 hybrid approaches.Studies outside the period or scope, duplicates, insufficiently described or unevaluated work, and studies not assignable to a technique family were excluded.Source: §2.1 · p. 4

3. Comparison with existing surveys

Existing surveys vary in their coverage of computing environments, scheduling granularity, and technique families, leaving limited guidance on how requirements relate to technique suitability. This review compares surveys and scheduling paradigms through a decision-oriented perspective spanning the cloud–fog–edge continuum.

3. Comparison with existing surveys

  • Existing surveys often focus on one paradigm or scheduling granularity, while relationships between requirements and technique suitability are less consistently analyzed.The review extends beyond task scheduling and load balancing to jointly consider task, job, and workflow scheduling.Source: §3 · pp. 4–5
  • The review compares four technique families across cloud, fog, and edge, linking workload, objectives, constraints, and computational overhead to suitability rather than naming a universal winner.Table 2 compares surveys by computing environment, granularity, technique coverage, analytical dimensions, statistical synthesis, and selection guidance.Source: §3 · p. 5, Table 2 · p. 5
  • Cloud, fog, edge, and integrated environments differ in resource capacity, centralization, network conditions, and dynamics, shaping scheduling requirements across the continuum.The review frames technique choice as a joint consideration of environment, workload, objectives, constraints, and available computational resources.Source: §3 · p. 6

5. Scheduling Requirements and Optimization Objectives

Scheduling requirements span performance, economic, energy, QoS, and user-centric objectives, alongside system and application constraints. Technique selection should reflect their combined demands and the target environment.

5. Scheduling Requirements and Optimization Objectives

  • Objectives include makespan, execution and response time, throughput, utilization, cost, energy, carbon emissions, reliability, availability, SLA satisfaction, and deadline compliance.Latency-sensitive fog and edge applications may prioritize response time and communication delay over global computational efficiency.Source: §5 · p. 7
  • Resource capacity, dependencies, deadlines, budgets, data locality, bandwidth, mobility, fault tolerance, and workload variability can determine feasibility before optimization quality is considered.Multiple objectives and constraints make scheduling a constrained decision problem with competing requirements.Source: §5 · p. 7
  • The appropriate technique depends on the combined environment, workload, objectives, constraints, and computational resources, rather than on the computing paradigm alone.Cloud environments generally offer abundant resources and centralized information; fog and edge impose stronger latency, energy, capacity, and responsiveness constraints.Source: §5 · pp. 6–7

6. Taxonomy of Scheduling techniques

The review classifies scheduling approaches by their underlying decision mechanisms, enabling comparison across different computing paradigms and scheduling scenarios.

6. Taxonomy of Scheduling techniques

  • The taxonomy groups scheduling techniques into heuristic, metaheuristic, learning-based, and hybrid families.Figure 1 presents this classification.Source: §6 · p. 7, Figure 1 · p. 8
  • The classification emphasizes decision mechanisms rather than individual algorithms, providing a common basis for comparing techniques across paradigms and scenarios.Source: §6 · p. 7
  • Task-priority and load-balancing methods appear among the named categories in the scheduling taxonomy.Source: §6 · pp. 7–8

6.1 Heuristic techniques

Heuristic scheduling offers rapid, relatively low-overhead decisions, but its suitability depends on the information available, the workload, and the environment. The review compares three heuristic families and provides selection guidance that weighs simplicity against information needs, adaptability, and optimization scope.

6.1 Heuristic techniques

  • Heuristics use problem-specific rules and simplified decisions for rapid, low-overhead scheduling, but may perform poorly in dynamic, heterogeneous, or multi-objective settings.Source: §6.1 · p. 8

6.1.1 Classification of Heuristic Scheduling Techniques

  • Heuristic techniques are classified by decision information into task-priority, resource-aware, and load-balancing families.Task-priority methods order tasks using characteristics such as arrival time, execution time, priority, or deadline.Source: §6.1.1 · p. 8, Figure 2 · p. 9

6.1.1.1Task priority-based scheduling

6.1.1.2 Resource-aware scheduling

  • Resource-aware heuristics select using execution or completion estimates; their effectiveness depends on estimate accuracy and current resource information, with monitoring and runtime estimation as trade-offs.MONWS adds task categories and makespan, cost, and utilization objectives; cost-aware Max-Min [36] and BCSV extend related criteria.Source: §6.1.1.2 · p. 10

6.1.1.3 Load-balancing-based scheduling

  • Round Robin and Opportunistic Load Balancing distribute workloads with low scheduling overhead, but quantum choice affects Round Robin and application-specific objectives may receive limited attention.Adaptive Round-Robin variants adjust the quantum to workload characteristics, adding scheduling calculations; such load balancing may fit dynamic fog and edge settings.Source: §6.1.1.3 · p. 10

6.1.2 Cross-technique comparison of heuristic scheduling families across computing paradigms

  • Heuristic variants add cost, energy, QoS, or load balancing to basic rules, improving applicability to complex settings while requiring more computation and accurate runtime information.Task-priority, resource-aware, and load-balancing families emphasize, respectively, ordering, resource capabilities and estimates, and workload distribution.Source: §6.1.2 · pp. 10–12, Table 5 · p. 11

6.1.3 Statistical analysis of heuristic-based strategies

  • 79% of heuristic studies address cloud computing and 21% fog; task scheduling accounts for 63%, workflow scheduling 31%, and job scheduling 6%.Single-objective studies account for 16%, bi-objective 37%, and multi-objective 47%; simulations account for 63%, while 37% report real-world implementations.Source: Figure 3 · p. 12, §6.1.3 · p. 12

6.1.4 Decision framework for selecting heuristic techniques

  • Resource-aware methods suit heterogeneous cloud and workflow scheduling; priority approaches suit deadline-sensitive fog and real-time settings; load balancing suits scalable distributed and edge environments.No heuristic family fits all conditions; combining priority, resource, and load-balancing information broadens decision inputs but increases computational and implementation requirements.Source: §6.1.4 · pp. 12–13, Table 6 · p. 13

6.2 Metaheuristic Techniques

This section compares metaheuristic and related optimization approaches for task and workflow scheduling across cloud, fog, edge, hybrid, and distributed environments, including their search mechanisms and varied optimization objectives. It emphasizes selecting methods according to workload and infrastructure conditions, objectives, and computational constraints, while considering trade-offs in search coverage, convergence, scalability, and tuning costs.

6.2 Metaheuristic Techniques

  • Metaheuristics are problem-independent, non-deterministic methods that explore complex search spaces for near-optimal solutions within reasonable computational time.The review distinguishes nature-inspired from non-nature-inspired methods.Source: §6.2 · p. 13, §6.2.1 · p. 14

6.2.1.1 Nature-inspired techniques

  • Nature-inspired methods address multidimensional, multimodal optimization and are grouped into evolutionary, swarm, physics- or chemistry-based, human-based, ecological or biological, and emerging techniques.The review classifies these methods before examining representative algorithms and their strengths and limitations.Source: §6.2.1.1 · p. 14

6.2.1.1.1 Evolutionary-based methods

  • Evolutionary algorithms suit complex, multi-objective scheduling, but encoding, objective formulation, parameter control, and computational overhead shape their effectiveness.The review distinguishes genetic, differential evolution, co-evolutionary, and memetic families by their selection-relevant differences.Source: §6.2.1.1.1 · pp. 14–16
  • Genetic algorithms support diverse scheduling objectives, while their variants target scalability, load balancing, deadlines, latency, and dynamic conditions; chromosome design should reflect the environment.Differential evolution offers adaptive global search for multi-objective problems, but discrete scheduling requires suitable encoding and operators.Source: §6.2.1.1.1 · pp. 14–15
  • Co-evolution decomposes coupled scheduling decisions into interacting subproblems, while memetic methods add local search or domain knowledge; both can improve fit at added complexity or computational cost.These approaches are most relevant when explicit decomposition or problem-specific refinement benefits complex scheduling.Source: §6.2.1.1.1 · pp. 15–16
  • Foraging algorithms offer distinct trade-offs: ACO handles discrete mappings and dependencies, PSO is fast and lightweight, while CSA and BOA need adaptive or problem-specific measures against stagnation.ACO can incur overhead and parameter sensitivity; PSO often needs adaptation for discrete, many-objective, or constrained workflows.Source: §6.2.1.1.2 · pp. 17–19
  • Hunting-inspired methods balance exploration and exploitation through distinct search structures, but diversity loss, premature convergence, parameter sensitivity, and computational cost require attention.Selection should match the task: CSO supports local refinement and global exploration, GWO uses hierarchical guidance, WOA adapts exploration, and HHO alternates search modes.Source: §6.2.1.1.2 · pp. 19–21
  • Cuckoo Search variants target distinct needs: MOCS offers a lightweight multi-objective baseline, IMOCS supports Pareto latency–energy trade-offs, and MoDCSA addresses discrete, reliability-aware IoT–fog scheduling.CS uses Lévy-flight exploration and abandonment of inferior solutions; Pareto archives and diversity preservation can help limit premature convergence in multi-objective problems.Source: §6.2.1.1.2 · p. 23

6.2.1.1.3 Physics-based methods

  • Physics-based methods trade search capability against complexity: GSA variants address multi-objective search or load distribution, MVO variants target QoS or convergence, and SA variants support priority, adaptive convergence, or time-aware offloading.Their added mechanisms can increase computational or parameter demands, particularly for large-scale or resource-constrained edge scheduling.Source: §6.2.1.1.3 · pp. 24–25

6.2.1.2.1 Search-based methods

  • Search-based methods suit large combinatorial problems when iterative exploration is feasible, with GRASP favoring randomized exploration, ILS repeated local improvement, and VNS systematic neighborhood adaptation.Adaptive variants can help with faster convergence or uncertainty, while search designs should reflect dependencies, heterogeneous resources, latency, energy, cost, or mobility.Source: §6.2.1.2.1 · pp. 25–26

6.2.1.2.2 Mathematical-based methods

  • Mathematical methods precisely represent objectives and constraints and can provide optimal solutions on tractable instances, but their scalability limits favor small or moderate, offline, or benchmark uses.ILP supports exact or relaxed formulations, MILP handles discrete and continuous decisions, and MINLP represents nonlinear relationships; relaxation, decomposition, or hybridization can improve practicality at scale.Source: §6.2.1.2.2 · pp. 26–27

6.2.2 Cross-technique comparison of metaheuristic-based scheduling families across computing paradigms

  • Metaheuristics are not interchangeable: Table 7 and the review show differing search mechanisms, multi-objective capabilities, convergence behavior, computational needs, and suitability across workloads and computing environments.The appropriate method depends on workload, objective, and infrastructure conditions rather than on a universally superior technique.Source: §6.2.2 · p. 27, Table 7 · p. 27
  • Edge-focused applications include latency-sensitive scheduling, IoT edge task clustering, and time-constrained applications in heterogeneous edge computing.The comparison also reports suitability for IoT and mobile crowdsensing applications in fog-cloud computing.Source: §6.2.2 · pp. 28–29
  • Cloud workflow use cases include multi-objective, security-aware, deadline-budget-constrained, scientific, and large-scale workflow scheduling.Reported applications include complex dependencies and workflow-based cloud scheduling problems.Source: §6.2.2 · pp. 28–29
  • Distributed workflow applications include parallel IoT workflows across fog and cloud servers and scientific workflows in green DVFS-enabled edge-cloud computing.A separate reported use case targets large-scale, real-world workflows with complex dependencies.Source: §6.2.2 · pp. 28–29
  • Reported optimization criteria span execution time, makespan, cost, energy consumption, load balancing, resource utilization, throughput, and reliability.Some cases explicitly combine multiple objectives, while others focus on energy, latency, deadlines, or budget constraints.Source: §6.2.2 · pp. 28–29
  • Cloud task applications include multi-objective, QoS-based, large-scale, and energy-efficient scheduling, including heterogeneous cloud settings.Other reported cases address energy-aware parallel tasks in multi-cloud environments and distributing independent tasks across heterogeneous nodes.Source: §6.2.2 · pp. 28–29
  • The cited entries identify ECSA, LS-CSO, DBOA, and FBOA; described applications include computationally intensive heterogeneous edge environments and scientific workflow scheduling, including DVFS-based scheduling.Source: §6.2.2 · pp. 29–30
  • The cited cat-swarm methods include MCSOFLB [94], associated with makespan, throughput, cost, resource utilization, fault-tolerant scheduling, and load balancing, alongside MDCSO [92] and PCSO [93].MDCSO [92] uses task size and processing capacity; PCSO [93] prioritizes tasks based on unit electricity cost at data-center locations.Source: §6.2.2 · p. 30
  • Grey-wolf variants target multi-objective task scheduling, including fault-tolerant, load-balanced multi-cloud scheduling with resource utilization, makespan, and throughput objectives [95, 96].MoDGWA [98] is described as cost-aware and targets makespan, reliability, and cost.Source: §6.2.2 · p. 30
  • Cloud task-scheduling variants include IWC, which uses adaptive population sizing, and GCWOA, which combines opposition initialization with adaptive mobility.Their listed objectives include cost, execution time, resource utilization, completion time, load balancing, and resource utilization.Source: §6.2.2 · pp. 30–31
  • MCWOA targets cloud scheduling that meets strict QoS requirements while pursuing effectiveness across makespan, energy consumption, and delay.Its described mechanisms include Sobol initialization, task-oriented modifications, adaptive coefficients, and Pareto exploitation.Source: §6.2.2 · p. 31
  • HHO variants address cloud task and workflow scheduling, including SLA-aware workflows and energy-efficient scheduling with job dependencies.The listed variants include SLA-aware HHO, Multi-Tactic HHO, and Enhanced HHO.Source: §6.2.2 · p. 31
  • Firefly-based methods span trust-aware scheduling on heterogeneous resources and workflow scheduling, with objectives including makespan, availability, throughput, utilization, and energy.Examples include trust-aware firefly optimization (TAFFA), Modified Firefly Optimization Algorithm [117], and Adaptive Firefly Algorithm [118].Source: §6.2.2 · p. 31
  • Sunflower and cuckoo-search variants target energy-aware, multi-objective scheduling for cloud, fog, and heterogeneous mobile-edge tasks.The listed objectives include makespan, cost, energy consumption, throughput, and delay.Source: §6.2.2 · p. 32
  • Monarch and gravitational-search methods address conflicting QoS objectives in large-scale cloud and heterogeneous edge–fog–cloud scheduling.The listed objectives include makespan, reliability, load balancing, execution time, completion time, and cost.Source: §6.2.2 · p. 32
  • Multi-Verse Optimization variants are applied to static and QoS-aware cloud scheduling and trust-based workflow scheduling in heterogeneous cloud environments.The listed concerns include makespan, resource utilization, reliability, and throughput.Source: §6.2.2 · p. 32
  • Simulated-annealing variants support priority-aware scheduling in dynamic fog, latency-sensitive edge scheduling, and energy-aware task scheduling across edge and cloud.PANTAS uses parallel simulated-annealing instances and time-arrival scheduling for joint offloading.Source: §6.2.2 · p. 32
  • Reported applications include deadline-aware IoT scheduling in virtualized fog-cloud, cloud resource allocation, scientific workflows, and precedence-constrained tasks on heterogeneous systems.The reviewed use cases also include resource-constrained computing and budget-constrained cloud workflows.Source: §6.2.2 · pp. 32–33
  • Other applications target energy-efficient, fault-aware cloud tasks, uncertainty-tolerant cloud-manufacturing workflows, static small-scale fog-cloud scheduling, and location-aware cloud-fog jobs.The reviewed scope also includes real-time fog applications and deadline-constrained, security-aware workflows in hybrid clouds.Source: §6.2.2 · p. 33
  • Adaptive metaheuristics can respond to changing workloads or resources, but add design and parameter-tuning requirements; exact optimization scales poorly, placing metaheuristics between exact methods and lightweight heuristics.For large cloud problems, parallel or hybrid search can use abundant computation; edge scheduling may require fewer candidates or iterations, potentially sacrificing solution quality.Source: §6.2.2 · p. 34

6.2.3 Statistical analysis of metaheuristic-based strategies

Metaheuristic scheduling research is concentrated in cloud settings and bio-inspired methods, while optimization metrics and evaluation tools show a strong emphasis on makespan and simulation. The section also identifies constraints that shape fog and edge use.

6.2.3 Statistical analysis of metaheuristic-based strategies

  • Cloud represents 60% of metaheuristic scheduling research, compared with fog at 28% and edge at 12%.Fog approaches increasingly use adaptive, migration-aware, and multi-objective mechanisms; constrained resources, mobility, and communication overhead challenge edge adoption.Source: §6.2.3 · p. 34, Figure 6 · p. 35
  • Bio-inspired methods account for 60% versus 21% for evolutionary methods; ACO (17%), PSO (15%), HHO (15%), and WOA (13%) lead within bio-inspired approaches.Their prevalence reflects adaptability and methodological maturity, not universal superiority.Source: §6.2.3 · p. 34, Figure 6 · p. 35
  • Makespan is the most frequently optimized metric (23.7%), followed by cost and energy consumption (15.5% each) and execution time (12.4%).The findings suggest a shift toward multi-objective scheduling that balances temporal, economic, and energy-efficiency requirements.Source: §6.2.3 · p. 34
  • CloudSim is the predominant evaluation environment (49%), followed by MATLAB (17%).Reliance on controlled simulation may not fully capture real-world workload variability, contention, communication delays, and infrastructure dynamics.Source: §6.2.3 · p. 35

6.2.4 Guideline for selecting metaheuristic-based techniques

Metaheuristic selection is framed as a requirement-driven decision across infrastructure, scheduling structure, objectives, and workload dynamics. The framework matches algorithm capabilities to application constraints rather than assuming a universally superior or most popular method.

6.2.4 Guideline for selecting metaheuristic-based techniques

  • Selection should jointly consider computing infrastructure, scheduling structure, optimization objectives, and workload dynamics.The decision framework organizes choices around application requirements and contexts.Source: §6.2.4 · p. 35, Table 8 · p. 36
  • Cloud settings prioritize scalability and multi-objective allocation, while fog and edge settings emphasize heterogeneity, energy efficiency, latency, QoS, and computational overhead.Adaptive or lightweight variants are favored where latency, energy, heterogeneity, and resource limitations matter.Source: §6.2.4 · p. 35
  • Algorithm choice should reflect workload and constraints: cost- or energy-aware methods suit economic or sustainability priorities, while deadline- and QoS-sensitive tasks need explicit constraint handling.For workflows, GA and local-search approaches are identified as suitable for preserving task dependencies.Source: §6.2.4 · p. 35

6.3.1 Classification of learning-based techniques

Learning-based scheduling spans clustering, predictive deep learning, and reinforcement learning, with different mechanisms and resource demands. The approaches trade simplicity against predictive capability, adaptation, and training or deployment complexity.

6.3.1 Classification of learning-based techniques

  • Learning-based methods are classified as unsupervised learning, deep learning, and reinforcement learning, with RL divided into value-based, policy-based, and actor–critic methods.Clustering organizes tasks or resources, deep learning extracts features and predicts, and RL learns scheduling policies through environmental interaction.Source: §6.3.1 · p. 37

6.3.1.1 Unsupervised learning techniques

  • K-Means-based scheduling groups tasks or resources using features such as task length and resource demands, progressing from simple grouping to resource-aware, context-aware, and multi-objective clustering.Basic clustering is lightweight but depends on feature selection and cluster configuration; fuzzy or adaptive extensions add flexibility and complexity.Source: §6.3.1.1 · p. 37

6.3.1.2 Deep Learning Techniques

  • CNNs extract task–resource features, while RNN and LSTM models capture temporal patterns; these approaches can reduce reliance on manually designed rules but need more data and computation than clustering.Deep learning is most appropriate when nonlinear or temporal patterns can be learned from sufficiently representative historical data; greater predictive capability brings higher training and deployment requirements.Source: §6.3.1.2 · pp. 37–38

6.3.1.3 Reinforcement Learning Techniques

  • Reinforcement learning treats scheduling as sequential decisions in which an agent observes a state, selects an action, and receives a reward tied to scheduling performance.It can learn through interaction in dynamic environments, but training time, state–action dimensionality, reward design, convergence stability, and computational overhead remain limitations.Source: §6.3.1.3 · p. 38
  • Value-based RL estimates state–action rewards: Q-Learning suits smaller discrete spaces, while DQN extends value estimation to larger state spaces.Reviewed variants address sample efficiency, value stability, partial observability, search-space reduction, and multi-objective scheduling, at the cost of added complexity and training requirements.Source: §6.3.1.3 · p. 38
  • Policy-based methods directly optimize scheduling policies for complex or continuous decision spaces, while actor–critic methods pair an action-selecting actor with a value-estimating critic.PPO constrains policy updates; actor–critic extensions add uncertainty handling, decentralized learning, trust, reliability, and workflow-specific representations, with greater computational and architectural costs.Source: §6.3.1.3 · pp. 38–39

6.3.2 Cross-technique comparison of learning-based scheduling families across computing paradigms

Learning-based scheduling families differ in predictive capability, adaptability, and computational demands. Their suitability depends on the workload, available data and resources, objectives, and deployment constraints.

6.3.2 Cross-technique comparison of learning-based scheduling families across computing paradigms

  • Unsupervised learning offers lightweight task or resource grouping, deep learning learns nonlinear representations and temporal patterns, and RL supports dynamic decisions through environmental interaction.Their respective constraints include limited adaptation after workload changes, representative-data and computation needs, and RL training, reward-design, and convergence demands.Source: §6.3.2 · p. 39
  • Within RL, Q-Learning suits smaller state–action spaces, DQN scales value estimation to larger spaces, policy methods suit complex decisions, and actor–critic combines policy and value learning.PPO offers more controlled policy updates than REINFORCE but adds optimization overhead; advanced RL variants also raise architectural and training complexity.Source: §6.3.2 · p. 39
  • No increase in algorithmic sophistication guarantees practical suitability; family choice depends on workload dynamics, state and action dimensionality, data availability, objectives, computational resources, and deployment constraints.The comparison frames technique selection around these requirements rather than sophistication alone.Source: §6.3.2 · p. 39

6.3.3 Statistical analysis of learning-based strategies

The reviewed learning-based studies emphasize task scheduling, increasingly use adaptive RL/DRL, and rely substantially on simulation; workflow scheduling and real-world validation receive less attention.

6.3.3 Statistical analysis of learning-based strategies

  • Fog computing accounts for 36% of reviewed studies, while cloud and edge computing each account for 32%.These percentages describe the reviewed sample, not the entire learning-based scheduling literature.Source: §6.3.3 · p. 41
  • The reviewed methods progress from unsupervised grouping and deep-learning prediction toward RL/DRL for sequential adaptation to changing system states.This progression reflects increasing scheduling-environment complexity, not universal preference for one learning paradigm.Source: §6.3.3 · p. 41
  • Makespan, execution time, energy consumption, and resource utilization dominate objectives, while reliability, fault tolerance, security, scalability, cost, and QoS appear less often.The review identifies broader multi-objective formulations as an opportunity for future studies.Source: §6.3.3 · p. 41
  • Evaluation relies substantially on simulation, including CloudSim for cloud studies and Python implementations for RL and deep learning.Fog- and edge-oriented studies also use iFogSim, SimPy, SUMO, OMNeT++, and Fog-Bus2; limited testbed and operational-workload use constrains conclusions about deployment robustness, workload drift, inference overhead, and practical scalability.Source: §6.3.3 · p. 41
  • The review’s overall findings motivate workflow-aware scheduling, real-world validation, adaptation to workload changes, and deployment-oriented performance measures.These priorities follow the observed task emphasis, simulation reliance, and increasing use of adaptive RL/DRL.Source: §6.3.3 · p. 41
  • 86% of reviewed learning-based studies address task scheduling, compared with 14% for workflow scheduling.Figure 8 summarizes these distributions.Source: §6.3.3 · p. 41, Figure 8 · p. 42

6.3.4 Guideline for selecting learning-based techniques

The proposed framework selects learning-based scheduling techniques according to problem requirements rather than algorithmic popularity. It maps learning paradigms to scheduling conditions while accounting for adaptability, complexity, training demands, and optimization scope.

6.3.4 Guideline for selecting learning-based techniques

  • The framework relates learning paradigms to scheduling requirements and application conditions, without ranking algorithms universally.Selection considers scheduling entity, environment, workload, objectives, resource constraints, system dynamics, and computational budget.Source: §6.3.4 · p. 42, Table 11 · p. 42
  • Clustering suits workload classification, task grouping, or resource organization, particularly in relatively stable environments; deep learning suits prediction or representation learning from historical data.Deep learning is considered when complex nonlinear relationships must be captured.Source: §6.3.4 · p. 42
  • Value-based reinforcement learning, including Q-Learning and DQN variants, fits discrete scheduling decisions requiring adaptation to changing system states.Source: §6.3.4 · p. 42
  • Policy-based methods suit continuous or large action spaces, PPO supports stable policy updates, and actor–critic methods address complex problems with large state spaces or multiple objectives.Actor–critic methods combine policy optimization with value estimation.Source: §6.3.4 · p. 42
  • Technique selection must balance adaptability, computational complexity, state/action-space requirements, training demands, and optimization scope; combining learning mechanisms may help highly dynamic or multi-objective problems but adds complexity.Source: §6.3.4 · p. 42

6.4 Hybrid techniques

This section examines hybrid scheduling techniques that combine complementary algorithms to balance search capability, adaptability, and performance across cloud, fog, and edge environments. It considers their suitability in relation to workload and environmental constraints, optimization objectives, and computational and implementation costs.

6.4.1 Classification of hybrid-based techniques

  • Hybrid scheduling is classified by six pairings: heuristic–heuristic, heuristic–metaheuristic, heuristic–learning, metaheuristic–metaheuristic, metaheuristic–learning, and learning–learning.The review compares representative studies in each category and analyzes their advantages and limitations.Source: §6.4.1 · p. 43

6.4.1.1 Heuristic + Heuristic

  • Heuristic–heuristic methods combine lightweight rules: HMMRR emphasizes fairness and responsiveness, while HAMM targets mixed-size workload balancing; predefined rules limit adaptability in dynamic environments.These hybrids suit fast decisions, simple implementation, and low computational cost when global optimization is less important.Source: §6.4.1.1 · p. 43

6.4.1.2 Heuristic + Metaheuristic

  • Heuristic–metaheuristic methods use heuristics for initialization or solution construction and metaheuristics for broader exploration, improving optimization quality but raising costs that can constrain edge use.Cloud examples include MaxGen and HEPGA, while fog examples include PSO-RR and GAMMR.Source: §6.4.1.2 · p. 44

6.4.1.3 Heuristic + Learning- based

  • Heuristic–learning hybrids combine scheduling rules with data-driven adaptation, as in DQ-HEFT and FODAS, but require training data, model execution, memory, and convergence.They are most appropriate when changing workload and resource conditions justify learning, particularly in fog and edge environments.Source: §6.4.1.3 · p. 44

6.4.1.4 Metaheuristic + Metaheuristic

  • Metaheuristic–metaheuristic hybrids target complex, multi-objective problems across cloud, fog, and edge, but their computational and implementation complexity can limit latency-sensitive or resource-constrained edge use.Examples combine evolutionary, swarm, and other search methods to improve exploration, convergence, diversity, and objectives such as makespan, energy, cost, and resource utilization.Source: §6.4.1.4 · pp. 44–45

6.4.1.5 Metaheuristic + Learning-based

  • Metaheuristic–learning hybrids pair global optimization with prediction or adaptive decisions, supporting changing workloads in cloud, fog, and edge, but adding model and computational overhead.Examples include GA-DQN, EDLB, and REACT; these methods suit settings where environmental dynamics justify the added learning complexity.Source: §6.4.1.5 · p. 45

6.4.1.6 Learning-based + Learning-based

  • Learning–learning hybrids combine temporal or supervised modeling with reinforcement or online learning; DRL-LSTM and A3C-R2N2 [212] can adapt to non-stationary workloads but need data and resources.They are most appropriate when workload dynamics are substantial and resources can support training and inference.Source: §6.4.1.6 · p. 46

6.4.2 Cross-technique comparison of hybrid-based scheduling families across computing paradigms

  • Table 12 compares hybrid algorithms by scheduling level, performance metrics, user constraints, application domains, and simulation tools across computing environments.The section describes the comparison as covering task-, job-, and workflow-level scheduling.Source: §6.4.2 · p. 46, Table 12 · p. 46
  • Heuristic–heuristic hybrids minimize complexity and support interpretable allocation but suit stable, simple conditions; metaheuristic–metaheuristic hybrids address large, conflicting searches at higher computational and parameter-tuning cost.The latter are less suitable for severely resource-constrained edge devices or applications requiring extremely low decision latency.Source: §6.4.2 · p. 49
  • Metaheuristic–learning hybrids use workload and resource information to adapt in dynamic environments, while learning–learning hybrids offer continuous adaptation with higher data, computation, coordination, and stability requirements.Learning effectiveness depends on training-data availability, representativeness, and quality; interpretability and generalization to unseen workloads remain concerns.Source: §6.4.2 · pp. 49–50

6.4.3 Statistical analysis of hybrid-based strategies

  • Cloud accounts for 54% of hybrid studies, fog 34%, and edge 12%, reflecting cloud resource capacity and the deployment constraints of computationally expensive methods at the edge.The comparatively smaller edge share does not imply lower research relevance; it reflects the tension between the potential benefits of adaptive, multi-objective scheduling and edge resource limits.Source: §6.4.3 · p. 50
  • Task scheduling represents 68% of reviewed studies, workflow scheduling 27%, and job scheduling 5%, with job coordination comparatively underexplored.Workflow scheduling is less prevalent because task dependencies, communication overhead, and heterogeneous execution paths add modeling and optimization complexity.Source: §6.4.3 · p. 50
  • Metaheuristic + Metaheuristic accounts for 48% of hybrid combinations and Metaheuristic + Learning 20%, while less frequent learning-rich combinations carry greater implementation and validation burdens.Metaheuristic pairs balance global exploration with local exploitation or convergence; learning components add prediction or adaptation to combinatorial search. Lower adoption of Learning + Learning does not establish inferior performance.Source: §6.4.3 · pp. 50–51
  • Makespan (22%), energy consumption (18%), and cost (15%) are the most frequently evaluated metrics, reflecting trade-offs between completion time and resource or operational efficiency.The passage frames hybridization as useful for navigating these conflicting objectives rather than optimizing each metric independently.Source: §6.4.3 · p. 51
  • CloudSim (43%) and Python (18%) are the most represented evaluation environments, reflecting their accessibility and flexibility rather than universal superiority.CloudSim supports controlled comparisons with consistent infrastructure assumptions, while Python enables flexible prototyping, particularly for learning-based components.Source: §6.4.3 · p. 51

6.4.4 Guideline for selecting hybrid-based techniques

  • Hybrid selection should match technique complexity and adaptability to application requirements, workload dynamics, resource constraints, scheduling granularity, and computing environment.The framework treats solution quality and algorithmic complexity as a trade-off, not a basis for declaring one hybrid universally superior.Source: §6.4.4 · pp. 51–52
  • Cloud, fog, edge, and integrated environments require different hybrid choices, balancing optimization quality and adaptability against latency, resource limits, and computational overhead.Cloud task scheduling may use HMMRR or HAMM for low complexity, GA–PSO or GA–GWO for global optimization, and DQ-HEFT or GA-DQN for dynamic workflows. Lightweight clustering with Round Robin suits constrained fog nodes; edge settings prioritize responsiveness, while REACT targets dynamic edge–cloud conditions. For variable integrated edge–cloud workloads, A3C-R2N2 combines asynchronous actor–critic learning and recurrent modeling, but data, training, inference, and stability constraints can favor simpler methods.Source: §6.4.4 · p. 52
  • Heuristic–heuristic hybrids suit low-overhead decisions, metaheuristic–metaheuristic hybrids suit complex multi-objective searches, and learning-enhanced hybrids suit dynamic or uncertain workloads.As methods progress toward learning–learning hybrids for stochastic, non-stationary settings, data needs, computational overhead, convergence and stability challenges, and reduced interpretability also increase.Source: §6.4.4 · pp. 52–53

7. Simulation Platforms for Scheduling Evaluation: Comparative Analysis, Research Trends, and Open Challenges

Simulation platforms shape which scheduling behaviors can be evaluated and how valid, reproducible, and generalizable the results are. The review compares platform specializations, traces a shift toward multi-layer and hybrid simulation, and identifies persistent comparability and realism challenges.

Comparative analysis of simulation platforms

  • Cloud-centric simulators dominate cloud scheduling evaluation, but centralized assumptions limit their representation of fog and edge network dynamics, mobility, and latency.Fog- and edge-oriented tools model hierarchical cloud–fog–edge architectures, while network-oriented tools focus on communication, mobility, and vehicular edge scenarios.Source: §7 · p. 54

Research trends

  • Simulation platforms have shifted from cloud resource management toward multi-layer modeling of latency, mobility, distributed execution, and adaptive scheduling.This evolution tracks research moving from centralized-cloud objectives toward IoT, cyber-physical, latency-sensitive, and AI-driven applications.Source: §7 · p. 54

Comparative analysis of simulation platforms

  • Simulation choice should follow the target paradigm, scheduling objectives, and metrics because no platform supports every evaluation requirement.CloudSim-based frameworks support resource allocation, workflow scheduling, and energy-aware optimization; fog simulators suit latency-sensitive and IoT applications, while network simulators support communication-aware studies.Source: §7 · p. 54, Table 14 · p. 55, Table 15 · p. 55

Research trends

  • Hybrid simulation environments are increasingly used to model computational, networking, and mobility characteristics together.Examples include CloudSim with OMNeT++ and iFogSim with SUMO.Source: §7 · p. 54

Open challenges

  • Inconsistent platform assumptions hinder fair comparisons, while simplified workload, network, and mobility models contribute to a simulation-to-reality gap.Customized extensions, proprietary workloads, undocumented settings, and missing standardized benchmarks also make independent validation difficult.Source: §7 · pp. 54–55

8. Discussion

The discussion frames technique selection as a match between scheduling requirements and environment rather than an algorithm ranking. It summarizes family-specific trade-offs and outlines research needs in adaptation, scalability, multi-objective design, validation, and resource-aware intelligence.

Technique selection

  • Heuristics suit stable, latency-sensitive workloads; metaheuristics suit complex multi-objective search; learning-based and hybrid methods address adaptation or competing requirements at added cost.Learning-based methods need training data and computational resources; hybrid methods generally increase implementation complexity and computational overhead.Source: §8 · pp. 55–56
  • The decision framework matches technique families to optimization objectives, workload dynamics, computational constraints, adaptability needs, and the target environment.Cloud can accommodate computationally intensive techniques more readily, while fog and edge place greater emphasis on latency, energy efficiency, and computational feasibility.Source: §8 · p. 56, Table 16 · p. 56

Future research directions

  • Future schedulers should adapt to workload drift and changing resources without complete retraining, while evaluating adaptation speed, data needs, generalization, and overhead.Scalability assessments should also measure execution time, memory, communication overhead, convergence, and scheduling quality as system size grows.Source: §8 · pp. 56–57
  • Multi-objective, constraint-aware hybrid schedulers should justify component roles and quantify optimization gains against training or inference overhead.For resource-constrained edge settings, lightweight intelligence is needed to balance decision quality, latency, memory, energy, and adaptability.Source: §8 · pp. 56–57
  • Real-world testbeds, realistic workloads, standardized benchmarks, and reproducible protocols are needed to strengthen practical validity and fair comparison.The proposed directions also include reliability, security, interpretability, privacy, fairness, and decentralized approaches.Source: §8 · pp. 56–57

9. Conclusion

The review synthesizes scheduling across cloud, fog, and edge, concluding that no technique family is universally superior. It offers a requirement-oriented framework for choosing techniques and identifies unresolved validation, scalability, and multi-objective challenges.

Conclusion

  • The review finds no universally superior technique family and supports selection by scheduling requirements and environment rather than reported performance rankings.Heuristics favor low overhead and stable workloads; metaheuristics suit complex multi-objective problems; learning-based and hybrid methods offer adaptability or combined capabilities with added requirements and complexity.Source: §9 · p. 57
  • The decision framework translates comparative evidence into guidance based on scheduling granularity, workload dynamics, environment, objectives, constraints, adaptability, and computational budget.It is intended to help researchers and practitioners identify suitable technique families for specific scheduling scenarios.Source: §9 · p. 57
  • The review identifies limited real-world validation, dynamic-workload scalability, conflicting objectives, and learning-based or hybrid complexity as unresolved challenges.Future work should emphasize standardized benchmarking, realistic validation, adaptive resource-aware scheduling, and integrated cloud–fog–edge frameworks.Source: §9 · p. 58
Loading 2610.13830v1…