Distributed intelligence on the Edge-to-Cloud Continuum: A systematic literature review

Motivation

• Data processing and ML are shifting from centralized approaches toward multi-tier infrastructures that place resources closer to data generation.Edge or Fog processing can support prompt insight extraction and ultra-low latency while balancing the benefits of centralization and decentralization.Source: §1 · p. 1

Deployment challenges

• Edge-to-Cloud deployments must reconcile interoperability, mobility, latency, network efficiency, privacy, and hardware-resource constraints across highly distributed infrastructures.The review also calls for extensive, reproducible evaluation of workflow combinations across accuracy, training time, network overhead, energy consumption, and processing latency.Source: §1 · pp. 1–2

Review scope

• The review surveys state-of-the-art ML and Data Analytics libraries and frameworks, learning paradigms, and systems and testbeds for experimental evaluation.Its coverage includes simulation, emulation, and deployment systems, as well as large-scale testbeds for complex Edge-to-Cloud workflows.Source: §1 · p. 2

Reproducibility

• The review assesses experiment reproducibility through artifact access, experimental-setup definition, and access to results.It frames reproducibility as necessary for others to verify scientific claims.Source: §1 · p. 2

Open challenges

• The review identifies open challenges in holistic performance understanding, workflow optimization, efficient AI deployment on heterogeneous infrastructure, and reproducible experiment analysis.These directions address the deployment and evaluation of learning-based analytics across the Computing Continuum.Source: §1 · p. 2

Scope and motivation

• Earlier Computing Continuum surveys focused on domains such as resource management, security, architectures, robotics, and blockchain, while this review covers ML and Data Analytics across Edge, Cloud, and hybrid environments.The authors describe their review as the first systematic exploration of applying ML, Data Analytics, and their combination to distributed intelligence across these environments.Source: §2 · pp. 2–3

Machine Learning across the continuum

• Prior ML surveys discuss continuum challenges and trends, including scalability, deployment, hardware heterogeneity, efficient edge ML, offloading, and communication-efficient distributed training.These reviews also cover future prospects and resource-limited-device training techniques.Source: §2 · p. 2

Deep learning and IoT

• Deep-learning reviews cover edge and hybrid edge-cloud applications, systems and packages, deployment architectures, and IoT data and application challenges.Topics include resource management, energy, migration, privacy, data availability, preprocessing, and online provisioning.Source: §2 · p. 3

Data Analytics across the continuum

• Data Analytics reviews cover IoT Big Data processing and learning frameworks, infrastructure for smart-city analytics, and platforms for real-time Edge-to-Cloud pipelines.The cited reviews also identify challenges such as interoperability, privacy, and characterization of smart-city applications.Source: §2 · p. 3

Distinctive contributions

• This review distinguishes itself by analyzing experimental evaluations and validation, including testbeds, simulation and deployment systems, frameworks, metrics, datasets, hardware, and reproducibility.It also discusses open challenges and research opportunities identified through the review.Source: §2 · p. 3

Methodology

• The review methodology is based on prior systematic-review approaches.Source: §3 · p. 3
• The methodology divides the review into three processes: planning, conducting, and reporting.Source: §3 · p. 3
• The paper describes the activities corresponding to each of the three review processes.Source: §3 · p. 3

3.1. Planning the Review

• Complex workflows combining computing, analytics, and learning often require hybrid execution from IoT devices to Cloud/HPC systems, making the continuum difficult to understand holistically.Source: §3.1 · p. 3

3.1.1. Identify the Need for the Review

• The review aims to provide a taxonomy of libraries, frameworks, and learning paradigms for intelligent Edge-to-Cloud workflows, and to identify systems, testbeds, and research opportunities.Source: §3.1.1 · pp. 3–4

3.1.2. Define the Research Questions

• The review asks about state-of-the-art ML and Data Analytics methods, their combination, experimental-research solutions and reproducibility, and open challenges.Source: §3.1.2 · p. 4

3.1.3. Define the Search String

• The search string combines IoT, edge or fog, big data, stream processing, and learning or intelligence terms.Source: §3.1.3 · p. 4

3.1.4. Define the Sources of Research

• The review searched ScienceDirect, ACM, IEEE Xplore, Springer Link, and selected Usenix venues.Source: §3.1.4 · p. 4

3.1.5. Define Inclusion and Exclusion Criteria

• Eligible articles had to address at least one research question and evaluate or propose systems, frameworks, or architectures for intelligence on Edge, Cloud, or Edge-to-Cloud environments.The search scope was journal and conference articles, magazines, and book chapters published from January 2016 through August 2021.Source: §3.1.5 · p. 4

3.1.6. Define the Data Extraction Procedure

• The data-extraction form records bibliographic details, contributions, cited learning or analytics frameworks, experimental approach, and open challenges or future work.Source: §3.1.6 · p. 4

3.2. Conducting the Review

• The search returned 1,159 articles across ScienceDirect, ACM, IEEE Xplore, Springer Link, and Usenix.The database counts were 242, 206, 325, 290, and 96, respectively.Source: §3.2 · p. 4

3.2.1. Identify the Primary Studies

• Screening abstracts and conclusions removed out-of-scope articles, leaving 69 papers for quality evaluation and information extraction.Source: §3.2.1 · p. 4

3.2.2. Evaluate the Quality of the Studies

• Quality evaluation checked whether selected articles addressed techniques or approaches enabling intelligent analytics on the Edge-to-Cloud Continuum.Source: §3.2.2 · p. 4

3.2.3. Extract the Relevant Information

• For each of the 69 selected articles, the reviewers read the full paper and recorded information using the previously defined form.Source: §3.2.3 · p. 4

3.3. Reporting the Review

• The review issues two report types: general statistics about the studies and answers to the research questions.Source: §3.3 · p. 4

3.3.1. Present an Overview of the Studies

• Global statistics are aligned with the research questions, while the extracted studies are grouped, taxonomized, and summarized to answer them.Source: §3.3.1 · p. 4, §3.3.2 · pp. 4–5

3.3.2. Present the Answers to the Research Questions

• Figure 2 shows selected-article percentages by publication year and scientific database; no article published in 2016 survived screening, and Table 1 groups studies by area and computing paradigm.Source: Figure 2 · p. 5, §3.3.2 · p. 5, Table 1 · p. 5

4. Machine Learning Methods on the Edge-to-Cloud Continuum

• The review maps learning methods and experimental resources, including frameworks, tasks, metrics, hardware, models, and datasets, and compares inference packages across heterogeneous Edge devices.Source: §4 · p. 5, §4.1 · p. 6

4.1. Inference on the Edge

• Mistify reduces DNN porting time by more than 10 times, while Bonsai outperforms named baselines and Edgent targets low-latency collaborative inference.Other reported approaches include embedded learning that reduces data transfer and offloading that may improve performance on resource-constrained devices.Source: §4.1 · pp. 5–6
• Edge inference studies report benefits from minimizing data transfer or model size, partitioning or offloading models, and parallelizing inference across resource-constrained devices.Source: §4.1 · p. 6

4.2. Distributed Training on the Edge

• The review also describes lightweight Edge methods for real-time human identification and tracking, with reported decent accuracy at resource consumption affordable by Edge devices.Source: §4.2 · pp. 6–7
• On Edge devices, collaborative training strategies are reported to reduce training time without sacrificing accuracy, while lightweight models can lower resource use while preserving accuracy.CLONE trains locally on private data and asynchronously aggregates model parameters; an iterative collaborative-processing system coordinates workers through a central controller.Source: §4.2 · pp. 6–7

4.3. Distributed training across the Edge-to-Cloud Continuum

• Synthetic Gradients yielded comparable performance to back propagation for a four-layer model, but VGG16 showed accuracy degradation and slower convergence; DLion reached target accuracy faster than Gaia and Ako.Source: §4.3 · pp. 7–8
• Collaborative Edge-to-Cloud training is reported to reduce training time without significant accuracy loss, data sent to the Cloud, and end-to-end processing latency.Source: §4.3 · p. 8

4.4. Main takeaways

• No single library or framework fits all continuum needs; the review offers solutions and artifacts to help researchers and engineers identify options relevant to their projects.Edge ML is constrained by accelerator memory and storage, unreliable low-bandwidth links, and battery limits.Source: §4.4 · p. 8

5.1. Data Processing on the Edge

• EdgeWise uses congestion-aware scheduling and a fixed-size worker pool, achieving up to 3 times Storm’s throughput while keeping latency low; related Apache NiFi systems distribute processing across Edge devices, [37],.ECStream offloads atomic tasks to peer Edge devices so time-constrained applications can run entirely at the Edge.Source: §5.1 · p. 12
• Edge analytics frameworks support local stream processing and collaborative query pipelines, with reported benefits in recognition performance, low delay and network use, and scalability as Edge nodes are added,,.The Edge Intelligence framework reports improved performance without degrading recognition accuracy; Fed4Edge evaluations report improved scalability with more Raspberry Pi devices.Source: §5.1 · p. 12
• SCEDA uses reinforcement learning to schedule around Edge-node connectivity and concept drift, seeking timely analytics while limiting accuracy loss and achieving accuracy comparable to core analytics [9].The review summarizes these approaches as using Edge capacity for distributed processing of data-intensive, latency-sensitive applications, including full processing at the Edge.Source: §5.1 · p. 12

5.2. Big Data Processing across the Edge-to-Cloud

• An Edge cluster of 6 nodes performs up to 5-6 times faster than Cloud deployments, while hierarchical data fusion enables decisions 20x and 90x faster at lower levels than centralized Cloud analysis.The first comparison uses an intelligent-surveillance evaluation; the second compares distributed fusion with sending all low-level sensor readings to a central Cloud.Source: §5.2 · p. 12
• Research describes a collaborative Edge-Cloud framework for live analytics in wireless IoT networks, a framework to deploy and optimize Big Data Analytics applications across the Edge-to-Cloud continuum, and Cameo’s latency-targeted event prioritization.Experiments show Cameo reduces query latency in single- and multi-tenant settings.Source: §5.2 · pp. 12–13
• Collaborative Edge-to-Cloud analytics primarily improves performance by minimizing application processing latency compared with Cloud-only approaches.Source: §5.2 · p. 13

5.3. Main takeaways

• The review organizes methods into two categories: data processing on the Edge and Edge-to-Cloud Big Data processing.Source: §5.3 · p. 13
• Recent work emphasizes collaborative Edge-Cloud processing for live analytics rather than novel processing frameworks designed to run only on the Edge.Source: §5.3 · p. 13

6.1. Combining Data Analytics and Learning on the Cloud

• Cloud analytics combines distributed tools such as Mahout, Spark MLlib, and FlinkML with frameworks including TensorFlow and PyTorch, while no single tool suits every problem,.Research directions include parallel and distributed ML algorithms and addressing the lack of online ML algorithms for real-time streaming data.Source: §6.1 · p. 13
• Spark SVM on Mesos outperformed MapReduce SVM, Storm SVM, RBFNN, and DBN in classification accuracy, processing speed, and convergence rate.The evaluations varied training-set sizes.Source: §6.1 · p. 15
• For 10K-element streams at 15–100 frames per second, the Edge-Cloudlet-MultiResource system took 49% less time and saved 99% bandwidth than a centralized Cloud-only approach.Source: §6.1 · p. 15

6.2. Combining Data Analytics and Learning on the Edge-to-Cloud Continuum

• Stream to Cloud and Edge (S2CE) is proposed for mining Big Data streams across hybrid Cloud and Edge infrastructure, with distributed processing, data fusion, and preprocessing.Source: §6.2 · p. 15
• Online Learning or Sequential Learning accounts for 21%, while Lifelong Learning or Continual Learning accounts for 5% of reviewed papers, [133].Online models update existing knowledge as data arrive rather than rebuilding the model, and stream-based learning suits IoT analysis and edge deployment.Source: §6.2 · p. 16
• DF-DDNN reduces latency by up to 33% versus the traditional IoT-Cloud model, while 5G I-IoT outperforms 4G-IoT and 5G-IoT in channel utilization.DF-DDNN integrates data flow and distributed deep learning in IoT-Edge environments; 5G I-IoT combines Big Data mining, Deep Learning, and Reinforcement Learning across Cloud, Fog, and Edge.Source: §6.2 · p. 19
• Other approaches combine batch learning, online learning, and stream mining for public-transport delay prediction, validated with real delay data streams.The review also groups studies into Cloud-based analytics and machine learning, and analytics combined with machine learning across the Edge-to-Cloud Continuum.Source: §6.2 · p. 19, §6.3 · p. 19
• Cloud-based CRBM traffic modeling can suffer severe accuracy impacts when connectivity outages interrupt communication between Fog and Cloud nodes.The proposed architecture combines a Fog-to-Cloud data distribution layer with online learning to learn and predict traffic telemetry.Source: §6.2 · p. 19

6.3. Main takeaways

• Recent efforts deploy Federated, Transfer, Multi-task, Reinforcement, and Online Learning on distributed Edge devices to improve predictive analytics, communication, connectivity, and latency.These efforts seek fast and accurate predictions, optimized communication channels, fewer connectivity issues, and latency low enough to meet Quality of Service requirements.Source: §6.3 · p. 19

7. Experimental Research and Reproducibility

• Simulation, emulation, and deployment systems support experimental research, while the review also assesses selected studies’ testbed size and reproducibility support.Simulation enables large-scale behavior analysis and configuration tuning; emulation offers more realistic results than simulation at lower cost than real deployments.Source: §7 · p. 19, Table 9 · p. 20, §7.1.1 · p. 22

7.1.1. Simulation Systems

• Cloud, Fog, and Edge simulators model infrastructure and workloads, with Fog systems also supporting mobility, federation, placement, and resource-management evaluation.CloudSim models data centers, virtual machines, and provisioning policies; EdgeCloudSim and IoTSim-Edge model computing and networking resources or heterogeneous device behavior.Source: §7.1.1 · pp. 22–23

7.1.2. Emulation Systems

• Fogbed, EmuFog, and RADICAL-DREAMER emulate resource provisioning, application workloads, or deployment configurations for Fog and Edge-to-Cloud settings.EmuFog supports topology design, Fog-node placement, and Docker applications over an emulated network; RADICAL-DREAMER models tasks, workloads, and distributed resources.Source: §7.1.2 · p. 23

7.1.3. Deployment Systems

• E2Clab supports reproducible experiments with real-world workloads, mapping application parts to physical testbeds, varying and scaling scenarios, and monitoring results.Kubernetes manages containerized applications across hosts, while KubeEdge extends orchestration and device management to Edge hosts.Source: §7.1.3 · p. 23

7.2. Large-Scale Experimental Testbeds for Edge-to-Cloud Experiments

• Testbed initiatives combine resources for Edge-to-Cloud experiments, including Grid5000 with FIT IoT-Lab and Chameleon with CHI@Edge.FIT IoT-Lab has more than 1.5K nodes and supports energy and network monitoring; CHI@Edge provides access to real IoT/Edge devices.Source: §7.2 · pp. 23–24

7.3. Support to Experimental Reproducibility

• The review evaluates reproducibility support through access to artifacts, description of the experimental setup, and access to results.Artifact access includes items such as datasets, code, applications, and configuration files; results access includes logs and runtime metrics.Source: §7.3 · p. 24
• 68% of studies do not provide experimental artifacts, 95% do not provide results, and 70% evaluate on small-scale testbeds.By contrast, 76% describe the experimental setup in detail; 20% use medium-scale and 10% large-scale testbeds.Source: §7.3 · p. 24, Figure 6 · p. 24, Figure 7 · p. 24

7.4. Main takeaways

• The findings call for rigorous experimental methods and deployment systems that guide large-scale validation and manage monitoring, results, and experiment provenance.The review identifies EnOSlib and E2Clab as existing efforts, while noting further advances are still needed.Source: §7.4 · pp. 24–25

Frameworks and libraries

• The most common AI frameworks are TensorFlow and PyTorch in the Cloud and MXNet and Caffe2 at the Edge; Cloud analytics commonly uses Spark, Flink, and Kafka, while Edge uses Nifi.The review identifies very few open-source Data Analytics frameworks designed for the Edge.Source: §8 · p. 25

Learning paradigms

• Distributed ML/DL, Online Learning, Reinforcement Learning, Transfer/Multi-task Learning, and Federated Learning are the most cited paradigms, but few studies explore their trade-offs at scale or jointly.Source: §8 · p. 25

Hardware heterogeneity

• Studies insufficiently analyze Edge hardware heterogeneity, relying mainly on Raspberry Pi devices or emulations of resource-limited devices.The review recommends including GPU- and TPU-enabled devices alongside CPUs in future evaluations.Source: §8 · p. 26

Testbed scale

• Most reviewed evaluations use small-scale testbeds, and the authors recommend larger-scale validation of distributed ML, DL, and data-stream approaches.The review describes small-scale setups as fewer than six machines or devices on average.Source: §8 · p. 26

Reproducibility

• Most studies describe their experimental setup but do not share artifacts or results publicly, limiting support for reproducibility.The review recommends rigorous experimental methodologies and notes the use of reproducibility badges by some conferences and journals.Source: §8 · p. 26

9.1. Understanding Performance of Application Workflows on the Edge-to-Cloud Continuum

• Understanding end-to-end performance requires identifying infrastructure bottlenecks and how system, network, and hardware choices affect latency and energy use.Deployment configuration is a multi-objective, NP-hard optimization problem with a vast search space; proposed directions include parallel evaluation and surrogate models for optimization and sensitivity analysis.Source: §9.1 · p. 26, §9.2 · p. 27
• Modeling, simulation, and emulation struggle to represent Edge-device heterogeneity and volatility, so experimental evaluation remains the main route to accurate performance insights.The review also notes the difficulty of reproducing representative application behavior across the many configurations of hybrid Edge-Cloud infrastructure.Source: §9.1 · p. 26

9.3. Enabling Intelligence on the Highly Heterogeneous Edge-to-Cloud Continuum

• Potential directions include incremental model updates for streaming data, knowledge transfer, distributed stream processing, and model compression or search for resource-limited deployment.The review also identifies distributed Edge training, Neural Architecture Search, Hyperparameter Search, and Knowledge Distillation as open challenges.Source: §9.3 · pp. 27–28
• The review reports no studies exploring these AI challenges at scale on hybrid Edge-to-Cloud infrastructures, despite the need to evaluate learning-paradigm trade-offs and workflow performance.Relevant measures include memory and energy use, model accuracy, training time, network overhead, and processing latency.Source: §9.3 · pp. 27–28

9.4. Supporting Reproducible Analysis of Complex Edge-to-Cloud Workflows

• Reproducible evaluation requires defined experimental methods, access to artifacts and results, and provenance capture across heterogeneous resources and networks.Provenance systems must limit capture overhead on constrained Edge devices and account for network overhead when transmitting provenance data.Source: §9.4 · p. 28

• The review covers methods for ML and deep-learning inference, centralized and distributed training, Edge data processing, and Big Data stream analytics across the Edge-to-Cloud Continuum.Source: §10 · p. 28
• It reviews collaborative learning paradigms and systems, then surveys simulation, emulation, deployment systems, testbeds, and their support for experiment reproducibility.Source: §10 · p. 28
• Open challenges include holistic performance understanding and optimization, efficient AI-workflow deployment, and reproducible large-scale analysis of Edge-to-Cloud workflows.Source: §10 · p. 28