Adaptive virtualization for AI-enabled Cloud-edge Continuum

Información del proyecto

CloudSkin

Identificador del acuerdo de subvención: 101092646

DOI

10.3030/101092646

Proyecto cerrado

Fecha de la firma de la CE 17 Noviembre 2022

Fecha de inicio 1 Enero 2023

Fecha de finalización 31 Diciembre 2025

Financiado con arreglo a

• Digital, Industry and Space

Coste total

€ 3 405 322,51

Aportación de la UE

€ 3 405 322,50

Coordinado por

UNIVERSITAT ROVIRA I VIRGILI
Spain

Periodic Reporting for period 2 - CloudSkin (Adaptive virtualization for AI-enabled Cloud-edge Continuum)

Período documentado: 2024-07-01 hasta 2025-12-31

Resumen del contexto y de los objetivos generales del proyecto

Currently, most data processing takes place in centralized cloud data centers, with limited computation at the edge. To reduce dependency on non-EU providers and strengthen European technological sovereignty, the project developed a cognitive Cloud–edge continuum platform capable of exploiting heterogeneous computing and storage resources across distributed infrastructures. Using AI-driven decision-making, the platform dynamically selects the optimal execution location between cloud and edge, adapting to workload and application requirements.

The platform follows a modular three-layer architecture covering orchestration, execution, and infrastructure services.

First, intelligent management is delivered through an AI-enabled “Learning Plane” applying machine learning and predictive analytics to optimize workload placement, resource utilization, and network traffic while enforcing service-level objectives (SLOs). It integrates serverless and container-based systems, enabling automated migration, multi-site deployment, and real-time adaptation validated in AI-driven scenarios such as surgical pipelines and automotive analytics.

Second, a universal execution layer enables stack-identical execution across cloud and edge environments. This is achieved through C-Cells, a WebAssembly-based abstraction combined with Trusted Execution Environments (TEEs) for confidential computing. C-Cells allow unmodified legacy applications, including MPI and OpenMP workloads, to run securely across heterogeneous infrastructures. Execution is extended to storage nodes, enabling sandboxed near-data processing, elastic migration, and fast serverless acceleration.

Third, storage is transformed into an active component of the continuum. Programmable mechanisms support in-transit data transformation, intelligent routing, and automated data tiering across cloud and edge sites, delivering low-latency, adaptive scaling, and fault-tolerant streaming for real-time IoT and video analytics workloads.

The platform was validated in AI-assisted surgery, metabolomics analytics, and connected automotive applications, demonstrating real-time edge processing, reduced cloud dependency, secure execution of legacy stacks without modification, transparent migration across sites, and unified orchestration of serverless, containerized, and WebAssembly workloads.

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

Extensive technical and scientific work was conducted to design, implement, integrate, and validate the Cloud–edge continuum platform across heterogeneous cloud, HPC, and edge infrastructures.

A key achievement was the AI-driven orchestration framework (“Learning Plane”). Machine learning models were developed for workload forecasting, SLO enforcement, and adaptive resource allocation. Integration with orchestration frameworks enabled automated task placement and dynamic migration between cloud and edge sites. The system achieved coordinated multi-site orchestration and real-time adaptation of AI pipelines while reducing communication overhead.

Another major achievement was portable and confidential execution across the continuum. C-Cells, a WebAssembly-based abstraction, support stack-identical execution across infrastructures. Legacy and HPC applications run without modification across cloud and edge resources. Integration with TEEs enables confidential computing with minimal developer effort, achieving near-native performance and reduced startup latency. Elastic execution, including live migration and near-data processing, was validated in operational scenarios.

WebAssembly execution was extended into the storage layer, enabling computation close to data and reducing data movement for I/O-bound workloads. Storage infrastructure was enhanced to support automated data tiering, in-transit transformation, buffering, and intelligent routing across distributed deployments. Predictive auto-scaling and fault-tolerant streaming mechanisms support continuous analytics with low latency.

All components were integrated into a composable three-layer architecture linking orchestration, execution, and storage services. End-to-end interoperability was demonstrated in surgical workflows, metabolomics analytics, and connected automotive scenarios. The platform reached operational maturity across testbeds, delivering AI-driven optimization, confidential execution, near-data processing, and coordinated multi-site orchestration.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

The project delivered an operational Cloud–edge continuum platform integrating AI-driven orchestration, secure portable execution, and programmable storage.

Scientifically, it advanced distributed computing by demonstrating predictive AI-based management of heterogeneous infrastructures. The Learning Plane enables workload-aware placement, SLO-driven allocation, and dynamic migration across cloud and edge nodes. WebAssembly-based abstractions provide universal runtimes supporting elastic migration, near-data execution, and confidential computing, allowing unmodified HPC applications to run securely across the continuum. Programmable streaming and tiered storage services introduce low-latency, fault-tolerant data management with in-transit processing. The project produced 52 peer-reviewed publications, including 7 Q1 journals and 16 CORE A/A* conferences.

Industrial impact includes ultra-low-latency AI workloads, optimized resource utilization, and reduced operational costs. Eighteen exploitable results (TRL 4–8) target healthcare, mobility, and agriculture. Healthcare pilots support secure processing of surgical and metabolomics data. Agricultural use cases improve irrigation efficiency and resource management. Connected mobility benefits from low-latency edge analytics and AI-based assistance. Core components are open-source and aligned with EU reference architectures, supporting SME adoption and commercialization.

Societally, the platform promotes privacy-preserving data processing and sustainable resource management while reinforcing the European cloud–edge ecosystem through modular, interoperable components. Further uptake requires large-scale pilots, continued innovation, productization of mature components, alignment with EU standards, and access to investment.

Cloudskin three-layered architecture