PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Augmented Edge and Fog Computing: Trends and Challenges

Shreshth Tuli, Fatemeh Mirhakimi, Samodha Pallewatta, Syed Zawad, Giuliano Casale, Bahman Javadi, Feng Yan, Rajkumar Buyya, Nicholas R. Jennings

arXiv:2208.00761v2cs.DC

TL;DR

Distributed computing settings create resource-management challenges for AI models, including scalability and generalizability. This survey reviews AI methods used in Fog resource management and finds overlap across decision domains; scalability remains a challenge as Fog nodes or workloads increase.

  • Problem

    AI models can have poor scalability, and their generalizability depends on stability, robustness, and adaptability.

    Source: §6.2 · p. 21

  • Method

    The survey reviews AI-based Fog resource-management techniques, classifying methods as classical, machine learning, reinforcement learning, or deep learning.

    Source: §7 · p. 24

  • Results

    The reviewed methods show significant overlap across decision domains in the AI models they use.

    Source: §7 · p. 24

  • Takeaways & Limitations

    The overlap across decision domains suggests that some design decisions matter more than others.

    Source: §7 · p. 24

    AI models may scale poorly as the number of Fog nodes or workloads increases.

    Source: §6.2 · p. 21

Abstract

from arXiv · show

In recent years, the landscape of computing paradigms has witnessed a gradual yet remarkable shift from monolithic computing to distributed and decentralized paradigms such as Internet of Things (IoT), Edge, Fog, Cloud, and Serverless. The frontiers of these computing technologies have been boosted by shift from manually encoded algorithms to Artificial Intelligence (AI)-driven autonomous systems for optimum and reliable management of distributed computing resources. Prior work focuses on improving existing systems using AI across a wide range of domains, such as efficient resource provisioning, application deployment, task placement, and service management. This survey reviews the evolution of data-driven AI-augmented technologies and their impact on computing systems. We demystify new techniques and draw key insights in Edge, Fog and Cloud resource management-related uses of AI methods and also look at how AI can innovate traditional applications for enhanced Quality of Service (QoS) in the presence of a continuum of resources. We present the latest trends and impact areas such as optimizing AI models that are deployed on or for computing systems. We layout a roadmap for future research directions in areas such as resource management for QoS optimization and service reliability. Finally, we discuss blue-sky ideas and envision this work as an anchor point for future research on AI-driven computing systems.

1. Introduction

Distributed computing spans IoT, Edge, Fog and Cloud resources, while AI increasingly supports automated resource-management decisions. This survey reviews AI augmentation across the Fog continuum, with QoS and resource management as its focus.

1. Introduction

  • IoT connects resource-limited Edge devices with Cloud resources, while Fog devices link the layers into a continuum spanning Edge to Cloud.The paper frames AI as a way to automate dynamic design decisions in resource management.Source: §1 · p. 1

1.1. Motivation of Research in AI-Based Augmentation

  • AI on Fog augments worker-layer applications, whereas AI for Fog uses broker-layer models for placement, scheduling and fault-remediation decisions.AI on Fog includes applications such as traffic surveillance; AI for Fog targets resource-management services.Source: §1.1 · pp. 1–2

1.2. Our Contributions

  • The survey examines deployment of AI applications on Fog, AI-based resource-management decisions, or both, organized around deployment, scheduling and maintenance.It classifies reviewed approaches by their characteristics and methodologies and identifies future directions.Source: §1.2 · p. 3

1.3. Related Surveys

  • Unlike earlier surveys with narrower coverage, this work connects AI on Fog and AI for Fog across the continuum and synthesizes root-cause issues and future directions.Earlier surveys often focused on particular application areas or resource-management tasks rather than their overlap.Source: §1.3 · pp. 3–4

1.4. Article Structure

  • The review proceeds from computing paradigms and service architectures to AI methods, trend analysis, limitations, future directions and conclusions.The paper identifies deployment, scheduling and maintenance as the scopes for its AI-method review.Source: §1.4 · p. 4

2. Background

The Fog continuum combines resources and service models across Edge and Cloud, giving resource managers a range of controls and trade-offs to optimize. The relevant QoS priorities vary by application and industry.

2.1. Related Computing Paradigms

  • Cloud resources offer scalable, on-demand capacity but can be multiple network hops from IoT devices; Edge computing moves resources closer to data for lower response times.Edge nodes are resource-limited and can face bandwidth contention as IoT device counts grow.Source: §2.1 · pp. 4–5

2.2. Shift to Fog Continuum

  • Fog continuum combines Edge, Fog and Cloud resources to manage the resource-latency trade-off across nearby constrained nodes and more distant resource-abundant Cloud nodes.It encompasses Edge-only, Cloud-only and hybrid deployments.Source: §2.2 · p. 5

2.3. Services

  • IaaS exposes infrastructure controls, PaaS exposes platform-level controls, and SaaS gives AI resource managers limited control over application configurations.MLaaS provides out-of-the-box support for data preprocessing, model training and inference.Source: §2.3 · pp. 5–6

2.4. Optimization Parameters

  • Fog QoS optimization spans response time, cost, energy, reliability and AI workload accuracy, with trade-offs shaped by node limits and application requirements.Energy use includes both infrastructure consumption and the inference costs of resource-heavy AI models; distributed nodes also face hardware, software, network and resource-overflow failures.Source: §2.4 · p. 6

2.5. Synergy with Industrial IoT / Industry 5.0 Applications

  • Industry use cases assign different importance to QoS metrics; healthcare prioritizes low response time, accuracy and reliability, while construction emphasizes cost.The survey considers energy as an indicator of carbon footprint and reports reliability as the most common important metric across the discussed applications.Source: §2.5 · pp. 6–7

3. AI Integration in Systems

The section surveys tools and platforms for AI-augmented Fog research, benchmarks used to evaluate Fog systems, and practical constraints on training AI models at the Edge.

3. AI Integration in Systems

  • The section frames AI–Fog integration as an interface between Fog systems and AI methods, with approaches developed at simulation and deployment levels.It also notes that credible research increasingly needs emulators that reproduce industrial deployment scenarios on physical platforms.Source: §3 · p. 7, §3.1.2 · p. 8

3.1.1. Simulated Platforms

  • Fog simulators enable scaled testing, but their approximations can produce noisy results or deviate from physical observations.Examples include iFogSim, IoTSim-Edge, PureEdgeSim, SimEdgeIntel, Deep FogSim, RelIoT, YAFS, and SimFaaS, with different modeling capabilities and constraints.Source: §3.1.1 · pp. 7–8

3.1.2. Physical Platforms

  • Physical test platforms include Edge-capable virtualization and Kubernetes-based systems, while FogBus, EiF, and COSCO support integration and AI-assisted orchestration.OpenStack supports multiple virtualization technologies; KubeEdge and OpenEdge provide Edge–Cloud communication and workload resource assignment, while COSCO couples AI resource management with simulation.Source: §3.1.2 · p. 8

3.2. AI Benchmarks for Fog Systems

  • Fog benchmarks cover heterogeneous real-time workloads, computer-vision applications, and other AI tasks, while workload traces provide datasets for simulation-based testing.DeFog includes six workloads; other suites include AIoTBench, EdgeAIBench, AIBench, EdgeBench, and IoTBench, and trace datasets include Bitbrain, Azure, Google Cluster, and SMD.Source: §3.2 · pp. 8–9

3.3. AI Modelling and Engineering

  • Fog model training is resource-intensive, can take longer on constrained Edge devices than Cloud nodes, competes with other applications, and is difficult when device availability is outside engineers’ control.Federated learning shares parameter updates rather than local data, but iterative synchronization can create network contention and longer waits; centralized learning requires a high-end system but avoids that bandwidth use.Source: §3.3 · p. 9

4. State of the Art Methods

The paper reviews provisioning, scheduling, and maintenance methods for optimizing QoS in Fog systems.

4. State of the Art Methods

  • The review covers provisioning, scheduling, and maintenance as the three resource-management aspects.These aspects are considered in relation to simulator and framework constraints and Fog-system optimization parameters.Source: §4 · p. 9
  • The methods target QoS metrics including response time, cost, energy, accuracy, and resiliency.Source: §4 · p. 9
  • The review examines state-of-the-art methods for each resource-management aspect in Fog systems.Source: §4 · p. 9

4.1. Interface Between AI and Fog

The section organizes AI-augmented Fog resource management around the data models use and the decisions they make for deployment, scheduling, and maintenance.

4.1. Interface Between AI and Fog

  • Fog resource-management inputs include workload utilization, host capacities, network parameters, gateway bandwidths, user and node locations, latency, and mobility.Source: §4.1 · p. 9
  • Deployment covers placing AI models on constrained Fog nodes and allocating or deallocating Fog devices for workloads.Source: §4.1 · pp. 9–10
  • Scheduling covers task placement and migration, including workflow scheduling when incoming tasks have precedence constraints.Source: §4.1 · p. 10
  • Maintenance includes real-time fault or anomaly detection and proactive or reactive recovery, while Fog resource limits can constrain redundancy and increase response times and SLA violations.The review also notes that most data-driven methods reach local optima, and methods designed to avoid them do not guarantee global optima.Source: §4.1 · p. 10

4.2. AI Augmented Deployment

The section surveys methods for deploying AI models on Fog infrastructure and provisioning Fog resources, including their stated tradeoffs and limitations.

4.2.1. AI Augmented DNN Deployment

  • Pruning reduces model memory and compute requirements, while gradient compression, quantization, and periodic aggregation target communication overhead in federated learning.Low-precision choices may require retraining, and multiple precision models increase training time.Source: §4.2.1 · pp. 10–11
  • DNN deployment methods use model compression, layer splitting, scheduling, and adaptive choices to meet application or latency constraints on resource-limited Edge devices.Gillis uses reinforcement learning to select compression or layer splitting and partitions models across serverless functions, but cannot jointly optimize latency and SLA.Source: §4.2.1 · pp. 10, 12
  • Semantic model splitting can parallelize inference and reduce latency, but generally lowers accuracy and requires retraining rather than reusing standard pretrained models.Data splitting instead partitions input batches across network instances, reducing response time at the cost of greater network overhead.Source: §4.2.1 · p. 12
  • TinyML approaches target complex models on constrained embedded devices through techniques including hyper-dimensional computing, memory swapping, and attention condensation.Source: §4.2.1 · p. 12

4.2.2. AI Augmented Resource Provisioning

  • Resource provisioning is commonly framed as predicting dynamic workload demand and then optimizing provisioning decisions for cost, utilization, or QoS.Bootstrapping Cloud VMs or Edge nodes takes time, and non-stationary workload utilization requires dynamic adaptation.Source: §4.2.2 · p. 12
  • Provisioning methods include regression and DNN demand forecasting, optimization search, hybrid prediction-and-optimization, and reactive reinforcement learning.The section reports that continuous relaxation can adversely affect Decision-NN performance, Semi-Direct has limited scalability with workload size, and Narya faces exposure bias.Source: §4.2.2 · pp. 12–13
  • Policy-gradient methods outperform traditional reinforcement-learning Q-learning and Monte Carlo-based approaches, according to the cited state-of-the-art methods.Source: §4.2.2 · p. 13

4.3. AI Augmented Scheduling

AI-augmented scheduling addresses QoS tradeoffs for independently scheduled tasks and precedence-constrained workflows. The survey covers optimization, surrogate, stochastic, reinforcement-learning, and simulation-coupled approaches, alongside their limitations.

4.3.1. AI Augmented Scheduling of Bag-Of-Tasks

  • Bag-of-tasks scheduling must trade off conflicting QoS measures such as response time, cost, and energy.In this workload model, tasks can be scheduled independently.Source: §4.3.1 · p. 13
  • POND combines UCB exploration with MaxWeight decisions, but MaxWeight policies can incur instability, high delays, and poor adaptation to volatile workloads.POND tracks constraint violations with a virtual queue at each Fog node.Source: §4.3.1 · p. 13
  • Surrogate schedulers estimate QoS to guide search or gradient updates, while black-box optimization can reduce interpretability and continuous approximations can yield sub-optimal decisions.Gradient-free search can converge more slowly and scale less well; non-local jumps can raise migration overheads and response times.Source: §4.3.1 · pp. 13–14
  • Stochastic surrogates model QoS distributions or mean and variance for robust decisions, while Gaussian Processes may be inaccurate in heterogeneous Fog environments.Bayesian Neural Networks and Natural Parameter Networks are cited as more sophisticated alternatives.Source: §4.3.1 · p. 14
  • Reinforcement-learning schedulers explore or exploit state-action choices, but slow adaptation can increase wait times, response times, and SLA violations.Examples include deep Q-learning, policy gradients, actor-critic agents, and multi-agent methods.Source: §4.3.1 · p. 15
  • Coupled simulation can provide additional QoS estimates to address exposure bias from stale execution traces and limited training-data diversity.GOBI* interleaves decision optimization with co-simulation [28]; EDSS uses a co-simulator to assess ML-generated scheduling choices.Source: §4.3.1 · p. 15

4.3.2. AI Augmented Scheduling of Workflows

  • Workflow schedulers must respect DAG precedence constraints; meta-heuristics include PSO, list scheduling, ACO, and multi-objective methods such as AMO.List scheduling performs poorly with non-preemptable jobs and heterogeneous requirements or machines.Source: §4.3.2 · p. 15
  • Workflow surrogate optimization trades better QoS estimates for migration overhead, while MCDS uses co-simulation to generate long-term estimates and facilitate stable learning.Other workflow approaches include game-theoretic scheduling for heterogeneous devices and RL methods that exclude precedence-violating actions.Source: §4.3.2 · pp. 15–16

4.4. AI Augmented Maintenance

AI-augmented maintenance covers resilience, fault detection and prediction, and fault remediation in Fog systems. The survey describes approaches that balance reliability needs against resource use and model limitations.

4.4. AI Augmented Maintenance

  • Resilience concerns include contention, service availability, and security and privacy, with mitigation choices carrying resource or efficiency tradeoffs.Hot-standby replication can support service continuity but adds energy and cost through redundant computation.Source: §4.4 · p. 16

4.4.1. AI Augmented Fault-Detection and Prediction

  • Fault-detection methods often reconstruct or forecast time-series data and use prediction deviations as anomaly scores.LSTM-NDT uses moving-average error thresholds, while Omnianomaly uses reconstruction probabilities and an adjusted POT threshold; Omnianomaly improves on NDT at the expense of high training times.Source: §4.4.1 · p. 16
  • Detection models use varied architectures, but recurrent models can be costly and poorly scalable, and some approaches struggle with insufficient data or long-term dependencies.Graph-attention methods model feature and temporal correlations; newer low-overhead architectures face limitations in capturing long-term dependencies.Source: §4.4.1 · pp. 16–18

4.4.2. AI Augmented Fault Remediation

  • Fault-aware scheduling and remediation use prediction, RL, surrogates, and evolutionary methods to allocate or migrate tasks before failures.The approaches include LBOS, ELBS, PCFT, CSAVM, DDQP, and PreGAN.Source: §4.4.2 · p. 18
  • Load balancing can proactively reduce fault risk without requiring additional compute infrastructure after failures.Methods use monitoring, migration, dynamic thresholds, and prediction-based strategies to distribute workload across Fog devices.Source: §4.4.2 · p. 18
  • Auto-scaling and forecasting methods adapt Fog resources to workload changes, while failure scores can guide checkpointing, migration, or replication.Other approaches model VM availability from dependent failure probabilities or use fuzzy recurrent QoS surrogates for VM auto-scaling [286].Source: §4.4.2 · pp. 18–19
  • Straggler-aware methods predict slow tasks or server failures and use replication decisions that balance failure probability against redundancy cost.JFP models temporal failure dependencies with a Dynamic Bayesian Network; START predicts straggler occurrence.Source: §4.4.2 · p. 19

5. Classification of State-Of-The-Art

The survey classifies Fog resource-management methods by AI approach and by deployment, scheduling, and maintenance scope. Its categories overlap across scopes, with classical methods, neural optimization and approximation, and reinforcement learning among the organizing groups.

5. Classification of State-Of-The-Art

  • Classical AI includes search, regression, and meta-heuristics; these methods and neural design recur across deployment, scheduling, and maintenance.Examples include risk-based optimization, regression-based prediction, and PSO across provisioning, workflow scheduling, and load balancing.Source: §5 · p. 19
  • Reinforcement-learning approaches span tabular methods, DQNs, and policy-gradient learning, with neural approximation used to scale beyond large physical-system state-action spaces.DQN and policy-gradient methods appear across placement, provisioning, scheduling, and fault-tolerant scheduling.Source: §5 · p. 19
  • Table 7 classifies state-of-the-art techniques by AI method, supporting comparison of method categories across resource-management scopes.The survey describes overlap among deployment, scheduling, and maintenance approaches.Source: §5 · p. 19, Table 7 · p. 20
  • Neural optimization and approximation use DNNs as QoS surrogates or decision-makers across provisioning, scheduling, and fault-tolerant management.The survey lists FCNs, CNNs, LSTMs, RNNs, GNNs, and GANs across these tasks [180, 182, 183, 173, 174, 175, 176, 28, 236, 259, 260, 258, 116, 14, 268, 269, 270, 41, 283, 286, 287, 288, 289].Source: §5 · p. 20

6. Trends, Challenges and Future Directions

The section surveys emerging trends in AI-augmented Fog systems, identifies limitations and deployment challenges, and outlines application areas and research directions.

6.1. Trends

  • Deep surrogate models approximate QoS metrics with DNNs, while co-simulated digital twins generate diverse traces, support A/B tests, and provide new data for adaptation.The digital twins address exposure bias, data saturation, and non-stationary workloads.Source: §6.1 · pp. 20–21
  • Transformers process sequences in parallel, while geometric models represent system structure; trends also include resource-efficient and unsupervised models and scalable AI methods.Resource-efficient models target constrained, decentralized deployments; unsupervised models avoid reliance on labeled data, and DQNs use fewer parameters than tabular RL for Q estimates.Source: §6.1 · p. 21, §6.2 · p. 21

6.2. Limitations

  • Fog AI limitations include scalability and generalizability; federated learning can support training when centralized data collection is impractical, but must address Edge/Fog heterogeneity.The section also identifies reliability problems from missing, sensitive, noisy, or malicious data, adversarial examples, and meta-heuristic convergence to local optima.Source: §6.2 · pp. 21–22
  • Adaptability is constrained by distribution shifts, biased or outdated data, and resource limits; co-simulator fine-tuning can incur high overhead, while AI-model lifecycle work limits agility.Transfer learning and co-simulation are described as approaches to adaptation.Source: §6.2 · p. 22

6.3. Emerging Challenges

  • Emerging challenges include deploying complex models on legacy infrastructure, tuning or designing models for new settings, and explaining resource-management decisions despite accuracy and scalability trade-offs.Interpretability is especially relevant to sensitive sectors such as healthcare and autonomous vehicles.Source: §6.3 · p. 22

6.4. Application Areas

  • Application directions span healthcare data quality and security, AI-managed 5G/6G networks, predictive supply-chain management, and distributed, reliable AI for smart cities.The section also notes low-latency communications as support for technologies including robot-assisted surgery.Source: §6.4 · pp. 22–23

6.5. Methods

  • Proposed methods include self-supervised and model-driven RL, analog AI, decentralized management, and AI-driven simulations for data generation and decision support.Analog AI may reduce memory, compute, and energy demands but can lose precision; decentralized management distributes load and avoids a single point of failure.Source: §6.5 · pp. 23–24
  • The section also proposes co-designing decisions across deployment, scheduling, and maintenance, and exploring foundation models for Fog tasks and generalization to unseen settings.Co-simulators are approximations of real Fog systems, and the proposed richer simulators are a future direction rather than an established result.Source: §6.5 · p. 24

7. Conclusions

The survey reviews AI augmentation of Fog systems across deployment, scheduling, and maintenance, distinguishing AI on Fog from AI for Fog. It classifies methods into four broad categories and argues for holistic research that draws on historical knowledge and considers computing advances together.

7. Conclusions

  • The review covers AI-based Fog resource management for AI deployment, workload scheduling, and system maintenance, spanning AI on Fog and AI for Fog.It classifies approaches as classical methods, machine learning, reinforcement learning, and deep learning.Source: §7 · p. 24
  • Overlap among AI methods across decision domains points to the importance of some design choices and hints at gaps in current research.Source: §7 · pp. 24–25
  • The authors call for research that distills historical knowledge across AI decision-making methods and considers computing advances together to develop holistic AI-based methods.They connect this approach to emerging application areas, future technologies, and next-generation users.Source: §7 · p. 25

Related papers

Loading 2208.00761v2…