Source-linked AI summary
SAREF-based Ontology for Distributed AI Workflows across the Edge-Fog-Cloud Continuum
Viorica Rozina Chifu, Tudor Cioara, Vasile Ofrim, Liana Toderean, Ionut Anghel, Laura Daniele, Cornelis Bouter
TL;DR
Existing semantic models provide limited support for distributed AI workflows across heterogeneous edge-fog-cloud environments, where AI processes and resources use incompatible representations. This paper proposes DAI-ECC, a SAREF-compliant extension of SAREF4SYST that unifies workflow and infrastructure semantics. Competency questions were successfully answered, and proof-of-concept evaluations reported 90-100% deployment success with decision times below 80 ms.
Problem
Existing semantic models provide limited support for representing distributed AI workflows, lifecycle features, and heterogeneous execution environments interoperably.
Method
DAI-ECC extends SAREF4SYST with concepts for AI pipelines, executable jobs, computational resources, deployment constraints, and communication relationships.
Results
All competency questions were successfully answered using SPARQL querying and semantic reasoning in distributed AI workflow and infrastructure evaluations.
Takeaways & Limitations
DAI-ECC provides a semantic foundation for interoperability, automated reasoning, and resource-aware orchestration of distributed AI workflows.
Takeaways & Limitations
The ontology requires ongoing maintenance to incorporate concepts for evolving paradigms such as federated learning, generative AI, and edge intelligence.
Abstract
from arXiv · showhide
Nowadays semantic models provide limited support for representing distributed AI workflows and their execution across heterogeneous edge, fog, and cloud environments. Therefore, AI processes and resources are often described using incompatible semantic representations, affecting the interoperability, orchestration, and reuse. To address these challenges, this paper proposes a SAREF-compliant ontology for representing distributed AI workflows across the edge-fog-cloud continuum. We extend the SAREF4SYST ontology with concepts for modeling AI pipelines, executable AI jobs, computational resources, deployment constraints, and communication relationships, providing a unified semantic model of both AI workflows and heterogeneous computing infrastructures. The ontology enables semantic interoperability, automated reasoning, and resource-aware orchestration of distributed AI applications while remaining fully aligned with the ETSI SAREF ecosystem. The ontology is evaluated using proof-of-concept smart grid energy services orchestration scenarios and validated using competency questions showing its ability to support AI workflow deployment, execution reasoning, and workload adaptation across heterogeneous edge, fog, and cloud environments. All competency questions were successfully validated using SPARQL querying and semantic reasoning. Experimental results demonstrate deployment success rates of 90-100% with average orchestration decision times below 80 ms across heterogeneous edge-fog-cloud environments, highlighting its effectiveness on ensuring semantic interoperability for distributed AI orchestration.
1. Introduction
Existing semantic models do not adequately represent distributed AI workflows, their lifecycle, or heterogeneous execution environments. DAI-ECC addresses this gap with a SAREF-compliant ontology integrating AI workflows and computational infrastructures.
- Existing ontologies provide limited interoperable support for AI processes and their computational, distribution, and lifecycle features.
- Heterogeneous metadata schemas for computing nodes, AI processes, data assets, and communication flows limit unified interpretation and automated orchestration.
- AI workflows require semantic representation of training, validation, deployment, and inference stages across edge, fog, and cloud environments.
- DAI-ECC extends SAREF4SYST with concepts for AI pipelines, executable jobs, computational resources, deployment constraints, and communication relationships.
- The ontology unifies AI workflow and infrastructure descriptions to support semantic interoperability and machine-interpretable resource descriptions.
- A federated learning smart-grid scenario demonstrates DAI-ECC’s applicability within an orchestration engine for managing AI processes.
2. Related Work
Related ontologies address distributed infrastructure, IoT, and AI artefacts separately, but do not integrate workflow semantics with computational requirements and deployment across the edge-fog-cloud continuum. DAI-ECC extends SAREF to provide that integrated model.
- Cloud, IoT, and edge ontologies standardize infrastructure resources, services, devices, and contextual information, but primarily address non-AI interoperability needs.
- AI-focused semantic models represent tasks, datasets, models, experiments, workflows, and evaluation results for knowledge sharing and reproducibility.
- Existing ontologies do not jointly represent AI workflows, computational requirements, lifecycle, communication patterns, and deployment across edge, fog, and cloud environments.
- DAI-ECC extends the SAREF ecosystem with a unified model supporting interoperability, automated reasoning, and resource-aware orchestration.
3. Ontology development methodology
DAI-ECC was developed through an adapted four-phase LOT methodology, deriving ontology requirements from smart-grid scenarios and implementing them as a validated SAREF4SYST extension. Maintenance keeps the ontology adaptable as distributed AI technologies evolve.
- The adapted LOT methodology comprises ontology requirements specification, implementation, publication, and maintenance.
- Ontology Requirements Specification: Smart-grid use cases identified entities, relationships, exchanged information, and reasoning capabilities needed for orchestration decisions.
- Ontology Requirements Specification: The requirements cover heterogeneous nodes and capabilities, AI workflows and tasks, execution requirements, deployment constraints, data exchanges, and resource relationships.
- Ontology Implementation: DAI-ECC was implemented as an ETSI SAREF4SYST extension using conceptual modeling, formalization, refinement, and validation activities.
- Ontology Implementation: Reasoning capabilities support federated pipeline execution, resource-aware deployment, and workload migration across heterogeneous resources.
- Ontology Maintenance: Maintenance incorporates user feedback, additional scenarios, and periodic consistency and interoperability validation as distributed AI technologies develop.
4. DAI-ECC ontology
DAI-ECC extends SAREF4SYST with concepts for distributed infrastructures, communication, AI workflows, jobs, and data objects. Its properties represent resource capabilities, connectivity, execution constraints, pipeline ordering, and data flow.
- Ontology structure: DAI-ECC reuses SAREF4SYST concepts and adds infrastructure and AI-workflow concepts for modeling distributed edge-fog-cloud systems.Figure 2 distinguishes reused SAREF4SYST concepts, infrastructure extensions, and AI workflow-specific concepts by color.
- Distributed infrastructures: ComputeNode specializes s4syst:System and represents an execution entity classified as Edge, Fog, or Cloud.Node descriptions include processing capacity, memory, storage, accelerator features, and location for matching AI-job requirements to capabilities.
- Distributed infrastructures: NetworkLink, NetworkInterface, and DataConnectionPoint model communication links, node access points, and data access points within distributed AI pipelines.NetworkLink profiles include bandwidth, throughput, latency, jitter, and packet loss rate.
- AI processes: AIPipeline represents a workflow of AIJob instances, whose constraints include latency, location, CPU cores, memory, bandwidth, accelerators, and accelerator memory.AIJob subclasses cover preprocessing, training, inference, evaluation, and aggregation, with validation and testing as evaluation specializations.
- AI processes: DataObject represents datasets, trained models, model updates, and intermediate results consumed, produced, or exchanged by AI jobs.DataObjectType distinguishes Dataset, Model, and ModelUpdate, while accessPolicy represents data-use restrictions.
- Relations and reasoning: Object properties connect jobs, pipelines, compute nodes, accelerators, and data resources while representing execution, pipeline order, and data access.hasStep, isStepOf, and hasNextStep encode pipeline composition and execution sequence; runsAIJob and runsOn encode job placement.
5. Evaluation results
DAI-ECC was evaluated through smart-grid workflow orchestration scenarios, randomized infrastructures, deployment experiments, and competency-question validation. The evaluation showed resource-aware placement, migration handling, failure detection, successful querying and reasoning, and decision times below 80 ms.
- Evaluation setup: The HEDGE-IoT orchestrator used DAI-ECC to manage AI-driven workflows across heterogeneous edge–fog–cloud infrastructures.The evaluation addressed workflow deployment, resource allocation, workload adaptation, and migration.
- Evaluation setup: Over 100 random infrastructures were generated across small, medium, and large size classes with independently assigned node types and hardware profiles.The size classes contained 8–16, 17–30, and 31–40 nodes, respectively; node types were assigned with 60% edge, 30% fog, and 10% cloud probabilities.
- Evaluation setup: The orchestrator instantiated workflows with 5, 10, 15, 20, and 25 data sources and matched job requirements against node capabilities for resource-aware placement.Workflow stages included CPU, memory, storage, and GPU requirements, while nodes were described by computational profiles.
- Deployment scenarios: Deployments produced three outcomes: success without migration, success requiring migration, or failure from insufficient resources.The scenarios distinguish direct placement, migration-assisted deployment, and infeasible deployment.
- Competency-question validation: All evaluated competency questions were successfully answered through SPARQL querying and semantic reasoning over workflows and computational infrastructures.The queries identified deployable nodes and jobs that could not be deployed, supporting resource-aware deployment analysis.
6. Conclusions
The paper presents DAI-ECC as a SAREF-compliant semantic foundation for distributed AI workflows across heterogeneous edge–fog–cloud infrastructures. Validation and a federated-learning energy-services case study support its querying, reasoning, and orchestration applicability.
- Contribution: DAI-ECC extends SAREF4SYST with concepts for AI pipelines, executable jobs, computational resources, deployment constraints, and communication relationships.It provides a unified semantic model connecting AI workflow semantics with heterogeneous computing infrastructures.
- Validation: The ontology was validated for logical consistency, inferencing capabilities, and competency-question performance using standard reasoning tools.All competency questions were successfully answered through SPARQL querying and semantic reasoning.
- Applicability: A HEDGE-IoT case study demonstrated the model's applicability to federated-learning energy services across heterogeneous edge–fog–cloud infrastructures.The conclusion positions compatibility with SAREF as a basis for broader applications and future standardization efforts.