The rapid growth of data-intensive applications across the edge-cloud continuum presents a dual challenge: immense energy consumption and complex data governance. This paper presents the GLACIATION project, which tackles these issues through a platform for energy-eficient and privacy-preserving data operations. We showcase an integrated approach that combines a Distributed Knowledge Graph (DKG) with AI-driven orchestration to automate and optimize data management. Key innovations include the Green Index, a real-time metric for sustainable energy-aware workload shifting; a bio-inspired scheduling engine using Ant Colony Optimization; and a lightweight, zero-shot Large Language Model (LLM) interface that enables semantic exploration of the DKG via natural language. The efectiveness of this framework is demonstrated through its validation in real-world industrial and public sector pilot deployments.
The proliferation of IoT and data-intensive applications is driving a shift towards the edge-cloud computing continuum, a paradigm that introduces significant challenges related to data governance, latency, and energy consumption. As data generation at the edge explodes, the need for intelligent, automated, and eficient data operations becomes critical. Traditional cloud-centric models are often ill-equipped to handle the complex trade-ofs between processing data locally to reduce latency and moving it to centralized clouds for powerful analytics, all while adhering to strict privacy regulations and minimizing the environmental footprint.
This paper outlines the key innovations of the GLACIATION platform. We begin by describing the project’s architecture and the core semantic metadata model (GLC-MRM) that enables interoperability. We then detail the platform’s AI-driven optimization capabilities, including the use of Ant Colony Optimization for data discovery, a bi-level scheduling approach for managing trade-ofs, and a lightweight Large Language Model (LLM) interface for democratizing data access. Finally, we present the validation of our approach through real-world industrial pilots, with a focus on the Green Index for sustainable, energy-aware orchestration in the energy sector. GLACIATION (Green responsibLe privACy preservIng dAta operaTIONS) is a Horizon Europe project designed to address the significant energy consumption and carbon emissions resulting from the rapid growth of big data analytics across the edge-to-cloud continuum. The project’s central ambition is to create a platform for energy-eficient, privacy-preserving data operations. The core of this initiative is the development of a novel Distributed Knowledge Graph (DKG) that spans the entire edge-core-cloud (S. Gupta)
CEUR
ceur-ws.org architecture. By leveraging AI-enforced minimal data movement and optimizing the physical location of analytics and data storage, GLACIATION aims to achieve substantial reductions in power consumption.
The technical approach revolves around a modular, microservices-based platform that supports the DKG. The conceptual architecture, illustrated in Figure 1, is comprised of key components including a Metadata service to manage data annotations, a Trade-of service to balance latency and resource use, and a Prediction service for forecasting data popularity and workload needs. A vital element is the project’s metadata framework, which provides the tools to embed privacy and trust requirements directly into data operations. The eficacy and generality of the GLACIATION platform will be validated through four demanding, real-world use cases in the public-service, manufacturing, enterprise, and energy sectors, led by partners MEF/Sogei, Dell, SAP, and IPTO, respectively1.
GLACIATION introduces the GLACIATION Metadata Reference Model (GLC-MRM) [
The GLC-MRM is designed for extensibility through a use-case-driven specialization methodology, where the core ontology is adapted with modular vocabularies to describe specific assets, services, and contexts. This includes detailed specializations for Kubernetes resources (nodes, pods, workloads), IoT devices, and crucial energy metadata such as the Green Index. To ensure broad interoperability, the model is formally mapped to established external ontologies, including the DCAT vocabulary for describing datasets and the ETSI SAREF ontology for modeling devices and energy-related measurements. 1The source code and use cases of the GLACIATION platform are available at https://github.com/glaciation-heu This semantic framework, supported by technologies like SHACL for data validation and JSON-LD for lightweight data exchange, enables both declarative SPARQL queries and automated policy enforcement, forming the cornerstone of the platform’s orchestration capabilities.
The GLACIATION optimization strategy is centered on a Distributed Knowledge Graph (DKG), which
serves as the foundation for a novel metadata fabric codenamed IceStream. The approach taken is
inspired by the Agri-Gaia project [
This rich semantic representation facilitates powerful optimization mechanisms through standardized protocols. The architecture supports declarative, decision-making queries via the SPARQL query language, allowing the orchestration engine to retrieve precise metadata for various microservices. For instance, the Trade-of Service queries the DKG to acquire metadata on latency, resource usage, energy, privacy, and security to inform workload placement decisions. Furthermore, the system leverages the Shapes Constraint Language (SHACL) to validate RDF data against a set of predefined conditions and policies, ensuring data quality, consistency, and compliance with domain-specific constraints. This combination of semantic querying and constraint validation provides a robust framework for implementing and enforcing contextual policies, enabling dynamic, automated, and intelligent optimization across the platform.
GLACIATION employs a bio-inspired approach for distributed data discovery and movement, utilizing
a variant of Ant Colony Optimization (ACO) [
The scheduling and orchestration logic is framed as a Bi-level Multi-Objective Optimization Problem
(BMO-TSLB), addressed by an Improved Multi-Objective Ant Colony Optimization (IMOACO)
algorithm [
To democratize access to the Distributed Knowledge Graph (DKG), we introduce a framework for
question answering that leverages Large Language Models (LLMs) in a zero-shot setting [
Our preliminary evaluations demonstrate the viability and efectiveness of this approach. For the task of generating SPARQL queries from natural language questions, using models such as Llama2-7B, Llama3-8B, and Mistral-7B, we observed that increasing the number of examples (n-shot prompting, n ≥ 1) generally leads to significant improvements in BERT scores (Precision, Recall, and F1) for pretrained LLMs compared to a 0-shot setting. Notably, a fine-tuned Mistral-7B model (Mistral-7B FT), even when ifne-tuned on a diferent dataset, achieved the best overall F1 score of 0.9671 with 5-shot prompting. This fine-tuned model also significantly improved performance in a 0-shot setting compared to its nonifne-tuned counterpart, indicating its adaptability to new datasets for the SPARQL2Q task. The query enhancer component within our framework has been particularly efective in significantly improving the quality of generated queries. For example, the average Acc@10 score (percentage of correct answers out of 10 runs) increased from 0.22 (without enhancer) to 0.57 (with enhancer) across a set of 30 questions. For specific complex questions, the Acc@10 improved from zero to 0.9 and 1.0 respectively with the enhancer.
The GLACIATION framework is validated across three diverse, real-world pilot deployments, demonstrating its practical applicability in industrial and public sector settings. These use cases serve to test and measure the efectiveness of the platform’s core capabilities, including energy-aware workload distribution, privacy-preserving data orchestration, and semantic policy enforcement.
The pilots showcase the integrated functionality of the GLACIATION approach. At the Independent
Power Transmission Operator (IPTO) in Greece, the system utilizes the Green Index derived from
real-time SCADA data to perform energy-aware workload distribution for grid anomaly detection tasks
across three interconnected data centers [
In a smart manufacturing scenario in Ireland, the platform manages privacy-preserving orchestration
using data from robotic and IoT sensors, enforcing data sovereignty and access control policies to govern
data movement and computation. A third pilot with MEF/SOGEI in Italy focuses on optimizing data
movement and energy consumption in a public administration context. Across these deployments, the
Open Policy Agent (OPA) framework is utilized to implement semantic policy control, with real-time
dashboards providing visibility into performance and compliance [
The calculation of the Green Index begins with the net power balance ( ) at a substation, indicating whether the site is a net exporter or importer of energy. This local value is then combined with the national ratio of renewable energy sources ( ) to produce a RES-adjusted power balance, . To ensure comparability, this value is normalized using a signed percentile rank over a one-week historical window and scaled to a final range of 0 to 1. The resulting index is strongly correlated with national 2 emissions, where a higher index value (approaching 1) signifies greater availability of green energy and signals optimal conditions for executing workloads.
This paper presented the GLACIATION framework, a platform for energy-optimized and semanticallyaware data management. We demonstrated how its integrated technologies—a Distributed Knowledge Graph (DKG) based on the GLC-MRM ontology, an AI-driven scheduling engine using Ant Colony Optimization, and the Green Index—create an intelligent orchestration platform.
The successful validation highlights the potential of semantic technologies to solve complex optimization problems in distributed environments. The DKG, paired with a lightweight LLM interface, is a significant step toward making sophisticated data infrastructures more autonomous, eficient, and accessible. Future work will focus on scaling the AI models, optimizing data movement strategies for heterogeneous networks, and improving human-in-the-loop support for domain experts.
This work is funded by the European Union’s Horizon Europe research and innovation programme under grant agreement No 101070141 (GLACIATION).
The author(s) have not employed any Generative AI tools.