EN FR
EN FR
MIMOVE - 2025

2025Activity reportProject-Team‌​‌MIMOVE

RNSR: 201421139W

Creation of‌​‌ the Project-Team: 2018 February​​ 01

Each year, Inria​​​‌ research teams publish an‌ Activity Report presenting their‌​‌ work and results over​​ the reporting period. These​​​‌ reports follow a common‌ structure, with some optional‌​‌ sections depending on the​​ specific team. They typically​​​‌ begin by outlining the‌ overall objectives and research‌​‌ programme, including the main​​ research themes, goals, and​​​‌ methodological approaches. They also‌ describe the application domains‌​‌ targeted by the team,​​ highlighting the scientific or​​​‌ societal contexts in which‌ their work is situated.‌​‌

The reports then present​​ the highlights of the​​​‌ year, covering major scientific‌ achievements, software developments, or‌​‌ teaching contributions. When relevant,​​ they include sections on​​​‌ software, platforms, and open‌ data, detailing the tools‌​‌ developed and how they​​ are shared. A substantial​​​‌ part is dedicated to‌ new results, where scientific‌​‌ contributions are described in​​ detail, often with subsections​​​‌ specifying participants and associated‌ keywords.

Finally, the Activity‌​‌ Report addresses funding, contracts,​​ partnerships, and collaborations at​​​‌ various levels, from industrial‌ agreements to international cooperations.‌​‌ It also covers dissemination​​ and teaching activities, such​​​‌ as participation in scientific‌ events, outreach, and supervision.‌​‌ The document concludes with​​ a presentation of scientific​​​‌ production, including major publications‌ and those produced during‌​‌ the year.

Keywords

Computer​​ Science and Digital Science​​​‌

  • A1.2.1. Dynamic reconfiguration
  • A1.2.3.‌ Routing
  • A1.2.4. QoS, performance‌​‌ evaluation
  • A1.2.6. Sensor networks​​
  • A1.3.2. Mobile distributed systems​​​‌
  • A1.3.5. Cloud
  • A1.3.6. Fog,‌ Edge
  • A1.5. Complex systems‌​‌
  • A1.5.1. Systems of systems​​
  • A1.5.2. Communicating systems
  • A2.5.​​​‌ Software engineering
  • A2.6.2. Middleware‌
  • A3.2.4. Semantic Web
  • A3.2.5.‌​‌ Ontologies
  • A9.2. Machine learning​​​‌
  • A9.4. Natural language processing​

Other Research Topics and​‌ Application Domains

  • B6.4. Internet​​ of things
  • B8.1. Smart​​​‌ building/home
  • B8.2. Connected city​

1 Team members, visitors,​‌ external collaborators

Research Scientist​​

  • Nikolaos Georgantas [Team​​​‌ leader, INRIA,​ Researcher, HDR]​‌

PhD Students

  • Emile Royer​​ [INRIA]
  • Haidong​​​‌ Zhao [INRIA]​

Technical Staff

  • Shahin Abdoul​‌ Soukour [INRIA,​​ Engineer, until Nov​​​‌ 2025, Pre-Doc until​ Sep 2025 / Post-Doc​‌ after]
  • Siamak Solat​​ [INRIA, Engineer​​​‌, from Feb 2025​, Post-Doc]

Interns​‌ and Apprentices

  • Akash Balamurugan​​ [INRIA, Intern​​​‌, from Apr 2025​ until Sep 2025]​‌
  • Luigi Lizzini [INRIA​​, Intern, from​​​‌ Apr 2025 until Sep​ 2025]

Administrative Assistants​‌

  • Diana Marino Duarte [​​INRIA]
  • Eugenie-Marie Montagne​​​‌ [INRIA]

External​ Collaborators

  • Shahin Abdoul Soukour​‌ [Télécom SudParis,​​ from Dec 2025,​​​‌ Post-Doc Researcher]
  • William​ Aboucaya [Université Paris​‌ Dauphine - PSL,​​ Post-Doc researcher]
  • Maroua​​​‌ Bahri [Sorbonne Université​, Associate Professor]​‌
  • Patient Ntumba Wa Ntumba​​ [CNAM Paris,​​​‌ Post-Doc Researcher]

2​ Overall objectives

MiMove has​‌ historically been positioned as​​ a research team addressing​​​‌ distributed computing systems. Such​ systems comprise components that​‌ span global networking and​​ computing infrastructures, mobile networking​​​‌ environments, powerful hand-held user​ devices, and physical-world sensing​‌ and actuation devices. In​​ particular, the Internet of​​​‌ Things (IoT) has been​ one of our main​‌ focuses. In such rich​​ environments, distributed systems have​​​‌ a number of challenging​ features, such as dynamicity​‌ due to volatile resources​​ and user mobility, heterogeneity​​​‌ due to constituent resources​ developed and run independently,​‌ and context-dependence due to​​ the highly changing characteristics​​​‌ of the execution environment,​ whether technical, physical or​‌ social.

In this context,​​ we have addressed in​​​‌ the past several phases​ in the lifecycle of​‌ distributed systems: system design,​​ deployment and runtime. We​​​‌ have tackled aspects such​ as: system interoperability &​‌ composition, resource allocation &​​ system performance, reliable mobile​​​‌ crowdsensing for environmental monitoring,​ collaborative participatory processes. In​‌ our solutions, we have​​ introduced system models, analyses,​​​‌ algorithms and protocols for​ capturing and managing the​‌ characteristics of the systems​​ under study, as well​​​‌ as designed and developed​ related middleware tools and​‌ architectures.

More recently, while​​ keeping the essential character​​​‌ of the team, we​ have been gradually aligning​‌ our activities to a​​ new orientation. We are​​​‌ focusing our distributed system​ research on distributed machine​‌ learning (ML) systems, including​​ federated ML systems, addressing​​​‌ both the training of​ ML models and their​‌ use for ML inference.​​ We situate distributed ML​​​‌ systems of interest in​ the resource/compute continuum edge-fog-cloud,​‌ combined with the IoT.​​ In this setting, ML​​​‌ systems have to deal​ with the specificities related​‌ to the resource environment,​​ but also to handle​​​‌ continuous IoT data streams.​ The latter requirement makes​‌ us concentrate on online​​ machine learning.

In this​​​‌ context, our central objective​ is to optimize the​‌ essential system trade-off between​​ performance and resource usage,​​ where we take into​​​‌ account the performance of‌ both ML models (accuracy)‌​‌ and ML systems (quality​​ of service). To this​​​‌ aim, we tackle algorithmic‌ aspects of both ML‌​‌ models (e.g., parameter tuning)​​ and ML systems (e.g.,​​​‌ scheduling, communication). On the‌ other hand, acknowledging the‌​‌ powerful effect that ML-based​​ methods can have on​​​‌ distributed systems, our second‌ research objective is to‌​‌ employ ML models for​​ supporting important system tasks​​​‌ such as resource allocation‌ and scheduling in the‌​‌ compute continuum.

3 Research​​ program

As part of​​​‌ MiMove's new research orientation‌ introduced in the previous‌​‌ section, we have been​​ developing the following research​​​‌ activities.

3.1 Efficient scheduling‌ of ML inference on‌​‌ GPUs

We address optimal​​ scheduling of ML inference​​​‌ tasks on local GPU‌ devices. We take into‌​‌ account the priorities of​​ different tasks (urgency levels)​​​‌ as well as the‌ interference in resource usage‌​‌ that results from task​​ processing parallelization. Scheduling is​​​‌ decided based on dynamic‌ interference prediction.

3.2 Automated‌​‌ machine learning (autoML) on​​ data streams

Automated machine​​​‌ learning aims at automating‌ the optimal tuning of‌​‌ ML algorithms. We address​​ the problem of autoML​​​‌ when applied to IoT‌ data streams in a‌​‌ distributed setting. Distribution is​​ applied as a means​​​‌ for managing both the‌ computational complexity of autoML‌​‌ tasks and the specificities​​ of IoT data collection.​​​‌ In the context of‌ distributed machine learning, we‌​‌ are studying the effect​​ of online federated learning​​​‌ on decision trees.

3.3‌ Robust ML under resource‌​‌ volatility

We address the​​ problem of scheduling ML​​​‌ training and inference processes‌ over a large-scale network‌​‌ of heterogeneous, volatile and​​ constrained resources.

3.4 NLP​​​‌ in goal-oriented requirements engineering‌

In Goal-Oriented Requirements Engineering‌​‌ (GORE), application designers capture​​ requirements of the new​​​‌ system-to-be as a goal‌ hierarchy. We leverage domain‌​‌ knowledge graphs (KGs) as​​ sources for inspiring and​​​‌ refining goals. We develop‌ methods and related tools‌​‌ that provide interactive assistance​​ to designers for goal​​​‌ elicitation. To extract relevant‌ knowledge from a KG‌​‌ in an automated way,​​ we use NLP techniques​​​‌ in several innovative ways.‌

3.5 Reinforcement learning for‌​‌ proactive scheduling of data​​ streams

We address optimal​​​‌ placement of data stream‌ operators in the compute‌​‌ continuum. Extending our previous​​ work on heuristic-based optimal​​​‌ scheduling on fog-cloud resources‌ that minimizes overall resource‌​‌ usage cost, we explore​​ online reinforcement learning (RL)​​​‌ for proactive scheduling, where‌ RL is performed on‌​‌ dynamic simulation data.

3.6​​ Federated learning on IoT​​​‌ data

We explore federated‌ learning (FL) on IoT‌​‌ data, where we deal​​ in particular with low-volume,​​​‌ sparse, imbalanced, non-IID real‌ data from device Wi-Fi‌​‌ / Bluetooth connectivity logs.​​ We apply our FL​​​‌ method to predict long-term‌ occupancy in industrial buildings.‌​‌

4 Application domains

Historically,​​ MiMove's research has had​​​‌ a strong focus on‌ the Internet of Things‌​‌ (IoT). This is still​​ currently the case, in​​​‌ particular coupled with the‌ resource/compute continuum. Numerous application‌​‌ domains result from this​​ setting. Connected cities and​​​‌ smart buildings are among‌ those.

5 Latest software‌​‌ developments, platforms, open data​​​‌

5.1 Latest software developments​

5.1.1 KG2GoalModel

  • Name:
    Goal​‌ model construction from Knowledge​​ Graph
  • Keywords:
    Goal-oriented requirements​​​‌ engineering, Knowledge graph, Natural​ language processing
  • Functional Description:​‌
    In Goal-Oriented Requirements Engineering​​ (GORE), application designers capture​​​‌ requirements of the new​ system-to-be as a goal​‌ hierarchy. We leverage domain​​ knowledge graphs (KGs) as​​​‌ sources for inspiring and​ refining goals. We propose​‌ a method and a​​ graphical tool that provide​​​‌ interactive assistance to designers​ for goal elicitation. To​‌ extract relevant knowledge from​​ a KG in an​​​‌ automated way, we use​ NLP techniques (Semantic Similarity,​‌ Natural Language Inference, Sentiment​​ Analysis, Sentence Compression, LLMs)​​​‌ in several innovative ways.​
  • URL:
  • Contact:
    Abdoul​‌ Abdoul Soukour
  • Participants:
    Abdoul​​ Abdoul Soukour, William Aboucaya,​​​‌ Nikolaos Georgantas

5.1.2 ML​ inference serving system

  • Name:​‌
    ML inference serving system​​ (short name will be​​​‌ disclosed after publishing of​ this work)
  • Keywords:
    Machine​‌ learning, Inference serving system,​​ Scheduling
  • Functional Description:
    Machine​​​‌ learning (ML) inference serving​ systems host deep neural​‌ network (DNN) models and​​ aim to efficiently schedule​​​‌ incoming inference requests across​ available GPU resources. However,​‌ limited support for task​​ prioritization and insufficient latency​​​‌ estimation under concurrency may​ restrict their applicability in​‌ many real-world scenarios. We​​ present a serving system​​​‌ designed to enhance deadline​ satisfaction for mixed-priority inference​‌ traffic under high GPU​​ utilization. To improve latency​​​‌ estimation, our serving system​ models potential contention during​‌ data transfer and accounts​​ for contention in kernel​​​‌ execution through an adaptive​ prediction model. By drawing​‌ on these predictions, it​​ performs priority-aware scheduling and​​​‌ thereby provides differentiated handling.​
  • URL:
  • Contact:​‌
    Haidong Zhao
  • Participants:
    Haidong​​ Zhao, Nikolaos Georgantas

5.1.3​​​‌ OSMAC

  • Name:
    Online SMAC​
  • Keywords:
    Machine learning, Automated​‌ machine learning, Online Learning,​​ Data stream, Bayesian optimization​​​‌
  • Functional Description:
    Online SMAC​ is an autoML optimiser​‌ for online machine learning​​ that determines the best​​​‌ machine learning model for​ a given task, and​‌ what hyperparameters to use.​​ It uses Bayesian optimisation​​​‌ to find the best​ model–hyperparameter combination, inspired from​‌ the SMAC method.
  • Release​​ Contributions:
    First release.
  • News​​​‌ of the Year:
    Created​ the first version.
  • URL:​‌
  • Publication:
  • Contact:​​
    Emile Royer
  • Participants:
    Emile​​​‌ Royer, Maroua Bahri, Nikolaos​ Georgantas

5.1.4 OccupFL

  • Name:​‌
    Federated Learning for Occupancy​​ Prediction
  • Keywords:
    Federated learning,​​​‌ Occupancy prediction, Smart building​
  • Functional Description:
    Connected-device logs​‌ are a common proxy​​ for estimating human occupancy​​​‌ in smart buildings, but​ they pose several challenges​‌ for machine-learning practitioners. As​​ part of the BPI​​​‌ France 2030 CP4SC research​ and innovation project, data​‌ were collected over 511​​ days at 15-minute intervals​​​‌ across eight zones. Accurate​ forecasting of device-connectivity counts​‌ often requires training data​​ that are long enough​​​‌ to capture seasonal patterns.​ These data were ill-suited​‌ to federated learning, especially​​ for long-term prediction, due​​​‌ to (i) low volume,​ (ii) extreme sparsity and​‌ class imbalance, and (iii)​​ demonstrably non-IID distributions. We​​​‌ address these challenges with​ a proof-of-concept federated learning​‌ pipeline that includes: (1)​​ statistical tests confirming non-IID​​​‌ distributions, (2) a synthetic​ data generator that preserves​‌ seasonal patterns while filling​​ gaps, (3) a dynamic​​ FedProx-style server for stable​​​‌ aggregation, and (4) focal-MSE‌ loss functions calibrated to‌​‌ each zone’s imbalance ratio.​​
  • URL:
  • Contact:​​​‌
    Siamak Solat
  • Participants:
    Siamak‌ Solat, Nikolaos Georgantas

5.1.5‌​‌ AI-based adaptive scheduler

  • Name:​​
    AI-based adaptive scheduler
  • Keywords:​​​‌
    Internet of things, Data‌ stream, Fog computing, Cloud‌​‌ computing, Machine learning, Reinforcement​​ learning
  • Functional Description:
    AI-based​​​‌ proactive scheduler optimizes DSPA‌ (Data Stream Processing and‌​‌ Analytics) operator placement across​​ the hierarchical Edge-Fog-Cloud architecture.​​​‌ Unlike traditional reactive heuristics,‌ our scheduler leverages Reinforcement‌​‌ Learning (RL), powered by​​ the Proximal Policy Optimization​​​‌ (PPO) algorithm. We implement‌ a state persistence mechanism‌​‌ that enables cumulative online​​ learning across training episodes,​​​‌ which distinguishes our method‌ from conventional supervised learning‌​‌ approaches that rely on​​ pre-collected datasets. Unlike offline​​​‌ training paradigms, our PPO‌ agent learns directly through‌​‌ continuous interaction with the​​ live YAFS simulation environment,​​​‌ adapting its scheduling policies‌ in real-time based on‌​‌ immediate system feedback. This​​ online learning approach allows​​​‌ the agent to make‌ proactive placement decisions by‌​‌ observing comprehensive state representations​​ encompassing service placements, device​​​‌ resource utilization (CPU, RAM),‌ network link conditions (bandwidth,‌​‌ latency), and application-level metrics.​​
  • URL:
  • Contact:
    Akash​​​‌ Balamurugan
  • Participants:
    Akash Balamurugan,‌ Maroua Bahri, Nikolaos Georgantas,‌​‌ Patient Ntumba Wa Ntumba​​

5.1.6 Interoperability Enabler

  • Name:​​​‌
    Interoperability Enabler for Data‌ Spaces
  • Keywords:
    Data spaces,‌​‌ Data marketplace, Interoperability
  • Functional​​ Description:
    The Horizon Europe​​​‌ SEDIMARK project designed and‌ prototyped a secure decentralised‌​‌ and intelligent data and​​ services marketplace that bridges​​​‌ remote data platforms and‌ allows the efficient and‌​‌ privacy-preserving sharing of vast​​ amounts of heterogeneous, high​​​‌ quality, certified data and‌ services supporting the common‌​‌ EU data spaces. Interoperability​​ Enabler was designed to​​​‌ facilitate seamless integration and‌ interaction among various artefacts‌​‌ within the SEDIMARK ecosystem,​​ including data, AI models,​​​‌ and service offerings. Interoperability‌ Enabler comprises the following‌​‌ components: (i) Data Formatter​​ – Convert JSON data​​​‌ (time-series data) into the‌ SEDIMARK internal processing format‌​‌ (pandas DataFrames), (ii) Data​​ Mapper – Convert data​​​‌ from pandas DataFrames into‌ JSON, (iii) Data Extractor‌​‌ – Extract relevant data​​ from a pandas DataFrame,​​​‌ (iv) Metadata Restorer –‌ Restore metadata to a‌​‌ pandas DataFrame, (v) Data​​ Merger – Merge two​​​‌ DataFrames by matching column‌ names.
  • URL:
  • Contact:‌​‌
    Abdoul Abdoul Soukour
  • Participants:​​
    Abdoul Abdoul Soukour, Maroua​​​‌ Bahri, Nikolaos Georgantas

6‌ New results

6.1 Leveraging‌​‌ domain knowledge in software​​ system goal models

Participants:​​​‌ Shahin Abdoul Soukour,‌ Nikolaos Georgantas.

In‌​‌ Software Engineering (SE), system​​ design is an important​​​‌ phase in the software‌ development lifecycle. During this‌​‌ phase, the architecture and​​ internal components of a​​​‌ software system are defined‌ to meet specific requirements.‌​‌ Effective system design ensures​​ that the final product​​​‌ is robust, scalable and‌ aligned with stakeholder needs.‌​‌ One of the major​​ challenges in system design​​​‌ is effectively capturing and‌ structuring domain knowledge to‌​‌ guide the design process.​​ Requirements Engineering (RE) plays​​​‌ a pivotal role in‌ addressing this challenge.RE is‌​‌ the initial and fundamental​​ step in the design​​​‌ process of any information‌ or software system, focusing‌​‌ on establishing, documenting, analyzing,​​​‌ validating, and managing the​ requirements of a software​‌ system. It takes into​​ account all the activities​​​‌ related to eliciting, specifying,​ and validating needs and​‌ constraints of the stakeholders,​​ ensuring that the final​​​‌ product closely corresponds to​ their expectations. Traditionally, domain​‌ knowledge in RE serves​​ as contextual support, helping​​​‌ to clarify and refine​ requirements that stem mainly​‌ from stakeholder input, regulations​​ or business needs. However,​​​‌ in this research, we​ explore an alternative approach​‌ where domain knowledge is​​ not just an auxiliary​​​‌ resource but an active​ source of inspiration for​‌ generating requirements. Goal-Oriented Requirements​​ Engineering (GORE) is a​​​‌ specific approach within RE​ that emphasizes identifying and​‌ modelling the high-level goals​​ of stakeholders. It focuses​​​‌ on understanding why certain​ requirements are needed and​‌ how they contribute to​​ the overall goals of​​​‌ the system. The goal​ model, an essential component​‌ of GORE, describes the​​ system's goals using a​​​‌ hierarchical structure in which​ high-level goals are refined​‌ (or decomposed) into more​​ specific ones. Despite various​​​‌ automation or semi-automation attempts,​ building goal models for​‌ software system design remains​​ time-consuming and quite complex,​​​‌ often requiring significant manual​ effort. To respond to​‌ these challenges, this PhD​​ research focuses on leveraging​​​‌ domain knowledge in the​ form of a Knowledge​‌ Graph (KG). The KG​​ will assist application designers​​​‌ in creating goals that​ are inspired from this​‌ knowledge, thereby facilitating the​​ construction of goal models.​​​‌ By combining with the​ integration of Natural Language​‌ Processing (NLP) techniques, relevant​​ information from the KG​​​‌ can be captured and​ suggested, aiding the application​‌ designer in building a​​ goal model more efficiently​​​‌ for software system design.​ This thesis makes several​‌ key contributions: the design​​ and implementation of methods​​​‌ for exploring KG by​ using NLP techniques for​‌ semi-automatic goal modelling; the​​ development of a technique​​​‌ to make the formulated​ goals more abstract to​‌ facilitate KG exploration and​​ to extract relevant information;​​​‌ and the development of​ a prototype, which is​‌ an interactive graphical tool,​​ that demonstrates and validates​​​‌ the proposed approach.

6.2​ An Interactive Tool for​‌ Goal Model Construction using​​ a Knowledge Graph

Participants:​​​‌ Shahin Abdoul Soukour,​ William Aboucaya, Nikolaos​‌ Georgantas.

The goal​​ model is an essential​​​‌ model in Goal-Oriented Requirements​ Engineering. It is used​‌ to describe the system's​​ goals using a hierarchical​​​‌ structure in which high-level​ goals are refined into​‌ more specific ones. Constructing​​ a goal model for​​​‌ a new application can​ present challenges, demanding considerable​‌ time and effort. Although​​ there have been attempts​​​‌ to automate or semi-automate​ the construction of goal​‌ models, these tasks remain​​ complex and manual. This​​​‌ paper presents an interactive​ graphical tool that leverages​‌ a domain Knowledge Graph​​ (KG) to assist the​​​‌ application designer in creating​ goals derived from this​‌ knowledge, thereby facilitating the​​ creation of goal models.​​​‌ We use semantic similarity​ measurement and Natural Language​‌ Inference (NLI) to effectively​​ extract and align triples​​​‌ from the KG with​ the high-level initial goals​‌ formulated by the application​​ designer. The extracted triples​​ undergo sentiment analysis and​​​‌ Graph-to-Text (G2T) generation to‌ build meaningful subgoals. Nevertheless,‌​‌ processing KGs with Natural​​ Language Processing (NLP) techniques​​​‌ can be a lengthy‌ process. We introduce a‌​‌ restriction based approach to​​ bound the exploration of​​​‌ the KG to the‌ most promising nodes. By‌​‌ tuning KG exploration bounds​​ while using our tool​​​‌ in a case study,‌ we analyze the trade-off‌​‌ between the quality of​​ the resulting goal model​​​‌ and time performance, which‌ is a key factor‌​‌ for an interactive approach.​​ Our paper highlights the​​​‌ relevance of our restriction‌ based approach to information‌​‌ retrieval in KGs to​​ facilitate goal model generation.​​​‌

6.3 OSMAC: A Dynamic‌ SMAC for Data Streams‌​‌

Participants: Émile Royer,​​ Maroua Bahri, Nikolaos​​​‌ Georgantas.

Automated machine‌ learning (autoML) methods often‌​‌ require multiple passes over​​ data and are computationally​​​‌ intensive, rendering them unsuitable‌ for streaming scenarios where‌​‌ data is continuously generated​​ and distributions evolve over​​​‌ time. The few existing‌ autoML solutions for stream‌​‌ learning mainly rely on​​ random search or genetic​​​‌ algorithms, which struggle to‌ maintain high performance in‌​‌ dynamic environments. By contrast,​​ leading methods in batch​​​‌ learning such as the‌ Sequential Model-based Algorithm Configuration‌​‌ (SMAC) leverage modelbased approaches,​​ suggesting opportunities for improvement​​​‌ in stream settings. To‌ address these challenges and‌​‌ meet the requirements of​​ stream scenarios, we introduce​​​‌ OnlineSMAC, a model-based optimizer‌ for data streams. OnlineSMAC‌​‌ combines Bayesian optimization with​​ an extension of the​​​‌ SMAC optimizer to dynamically‌ select optimal processing pipelines‌​‌ and hyperparameters. Our results​​ show that this approach​​​‌ is highly competitive, achieving‌ performance on par with‌​‌ state-of-the-art stream autoML methods.​​ This highlights the promising​​​‌ potential of using Bayesian‌ optimization for data streams.‌​‌

6.4 ML Inference Scheduling​​ with Predictable Latency

Participants:​​​‌ Haidong Zhao, Nikolaos‌ Georgantas.

Machine learning‌​‌ (ML) inference serving systems​​ can schedule requests to​​​‌ improve GPU utilization and‌ to meet service level‌​‌ objectives (SLOs) or deadlines.​​ However, improving GPU utilization​​​‌ may compromise latency-sensitive scheduling,‌ as concurrent tasks contend‌​‌ for GPU resources and​​ thereby introduce interference. Given​​​‌ that interference effects introduce‌ unpredictability in scheduling, neglecting‌​‌ them may compromise SLO​​ or deadline satisfaction. Nevertheless,​​​‌ existing interference prediction approaches‌ remain limited in several‌​‌ respects, which may restrict​​ their usefulness for scheduling.​​​‌ First, they are often‌ coarse-grained, which ignores runtime‌​‌ co-location dynamics and thus​​ restricts their accuracy in​​​‌ interference prediction. Second, they‌ tend to use a‌​‌ static prediction model, which​​ may not effectively cope​​​‌ with different workload characteristics.‌ In this paper, we‌​‌ evaluate the potential limitations​​ of existing interference prediction​​​‌ approaches, finding that coarse-grained‌ methods can lead to‌​‌ noticeable deviations in prediction​​ accuracy and that static​​​‌ models degrade considerably under‌ changing workloads.

7 Partnerships‌​‌ and cooperations

7.1 European​​ initiatives

7.1.1 Horizon Europe​​​‌

SEDIMARK
  • Title:
    SEcure Decentralised‌ Intelligent Data MARKetplace
  • Duration:‌​‌
    2022 - 2025
  • Partner​​ Institutions:
    • INSTITUT NATIONAL DE​​​‌ RECHERCHE EN INFORMATIQUE ET‌ AUTOMATIQUE (INRIA), France
    • WINGS‌​‌ ICT SOLUTIONS TECHNOLOGIES PLIROFORIKIS​​ KAI EPIKOINONION ANONYMI ETAIREIA​​​‌ (WINGS ICT SOLUTIONS AE),‌ Greece
    • UNIVERSITY COLLEGE DUBLIN,‌​‌ NATIONAL UNIVERSITY OF IRELAND,​​​‌ DUBLIN (NUID UCD), Ireland​
    • FORUM VIRIUM HELSINKI OY​‌ (RADIO- JATELEVISIOTEKNIIKAN TUTKIMUS RTT),​​ Finland
    • SIEMENS SRL, Romania​​​‌
    • ATOS SPAIN SA, Spain​
    • AYUNTAMIENTO DE SANTANDER, Spain​‌
    • METLEN ENERGY & METALS​​ AE (METLEN), Greece
    • UNIVERSIDAD​​​‌ DE CANTABRIA (UC), Spain​
    • FONDAZIONE LINKS - LEADING​‌ INNOVATION & KNOWLEDGE FOR​​ SOCIETY (FONDAZIONE LINKS), Italy​​​‌
    • ATOS IT SOLUTIONS AND​ SERVICES IBERIA SL (ATOS​‌ IT), Spain
    • UNIVERSITY OF​​ SURREY (SURREY), United Kingdom​​​‌
    • EGM (EGM SAS), France​

Participants: Shahin Abdoul Soukour​‌, Nikolaos Georgantas.​​

SEDIMARK aims at designing​​​‌ and prototyping a secure,​ decentralised and intelligent data​‌ and services marketplace, based​​ on Distributed Ledger Technology​​​‌ and Artificial Intelligence, which​ bridges remote data platforms​‌ and allows the efficient​​ and privacy-preserving sharing of​​​‌ vast amounts of heterogeneus,​ high quality, certified data​‌ and services supporting the​​ common EU data spaces.​​​‌ SEDIMARK includes a distributed​ registry of resources (data/services)​‌ stored on edge systems,​​ close to where they​​​‌ are generated and where​ the data are cleaned,​‌ labelled, validated and anonymised.​​ Energy efficient AI techniques​​​‌ will be used for​ automated data quality management,​‌ labelling and classification of​​ data as well as​​​‌ for providing (distributed) analytics​ and advanced services on​‌ top of the data.​​ Semantic interoperability based on​​​‌ common ontologies and data​ models will allow the​‌ easy and efficient discovery,​​ sharing and federation of​​​‌ heterogeneous data from multiple​ sources.

7.2 National initiatives​‌

BPI France 2030 CP4SC​​ project
  • Title:
    Cloud Platform​​​‌ For Smart City
  • Duration:​
    2023 - 2025
  • Partner​‌ Institutions:
    • ATOS/Eviden
    • Ericsson
    • INRIA​​
    • INRAE
    • IFPEN
    • Oslandia
    • Vertical​​​‌ M2M

Participants: Siamak Solat​, Nikolaos Georgantas.​‌

The goal of the​​ CP4SC platform is to​​​‌ assist governments in implementing​ ambitious policies towards achieving​‌ carbon neutrality by ingesting​​ data from various sources​​​‌ across multiple verticals, such​ as mobility, energy management,​‌ and earth and environmental​​ observation. By placing data​​​‌ analysis at the core​ of these activities, the​‌ CP4SC project provides a​​ comprehensive, adaptable, and secure​​​‌ technological solution that meets​ the highest requirements of​‌ organizations involved in complex​​ projects, with a particular​​​‌ focus on mobility, 5G​ connectivity, and secure exchanges.​‌

Inria Challenge Cupseli project​​
  • Title:
    Collaborative Unified Platform​​​‌ for a Scalable and​ Efficient Learning Infrastructure
  • Duration:​‌
    2025 - 2029
  • Partner​​ Institutions:
    • Inria teams: ARGO,​​​‌ MIMOVE, COAST, MAGELLAN, STACK,​ WIDE, OCKHAM, COATI, NEO,​‌ TADAAM, TOPAL
    • Hivenet

Participants:​​ Nikolaos Georgantas, Siamak​​​‌ Solat.

Hivenet offers​ a highly original data​‌ storage architecture, in which​​ data is stored in​​​‌ a distributed and secure​ manner on the spare​‌ storage resources of participants,​​ based on a peer-to-peer​​​‌ structure. This structure ensures​ scalability, resilience and voluntary​‌ sharing of data between​​ users. The aim of​​​‌ this challenge between Hivenet​ and Inria is to​‌ push the limits of​​ distributed AI computing. Its​​​‌ goal is to demonstrate​ that even the most​‌ demanding AI and Big​​ Data applications can run​​​‌ efficiently on heterogeneous, distributed,​ and volatile resources –​‌ while maintaining accuracy, ensuring​​ privacy, and reducing environmental​​​‌ impact.

8 Dissemination

8.1​ Promoting scientific activities

8.1.1​‌ Scientific events: selection

Member​​ of the conference program​​ committees
  • Nikolaos Georgantas ,​​​‌ member of the TPC‌ of the following international‌​‌ conferences: ACM SAC'25, IEEE​​ SOSE'25, IEEE SMARTCOMP'25'26, IEEE​​​‌ WETICE’25, CoopIS'25, MODELSWARD'25'26, ENASE'26.‌
  • Siamak Solat , member‌​‌ of the TPC of​​ the following international conference:​​​‌ IEEE BCCA'25.
Reviewer
  • Siamak‌ Solat , reviewer for‌​‌ the following international conference:​​ ACM SAC'26.

8.1.2 Invited​​​‌ talks

  • Haidong Zhao ,‌ "Strait: Perceiving Priority and‌​‌ Interference in ML Inference​​ Serving", Networks & Systems​​​‌ Workshop (Journées non-thématiques GDR‌ RSD), IRIT – Site‌​‌ ENSEEIHT, Toulouse, March 21,​​ 2025.

8.1.3 Scientific expertise​​​‌

  • Nikolaos Georgantas , member‌ of the EDITE Doctoral‌​‌ School's 2025 selection committee​​ “Communications, Networks and Systems”​​​‌ for PhD fellowships.
  • Nikolaos‌ Georgantas , member of‌​‌ the PhD monitoring committee​​ of Himadri Chhaya-Shailesh (Sorbonne​​​‌ Université), Victor Laforet (Sorbonne‌ Université).

8.2 Teaching -‌​‌ Supervision - Juries -​​ Educational and pedagogical outreach​​​‌

8.2.1 Supervision

  • PhD thesis‌ defense: Shahin Abdoul Soukour‌​‌ , "Leveraging domain knowledge​​ in software system goal​​​‌ models", Sorbonne Université, Sep‌ 19, 2025, Nikolaos Georgantas‌​‌ .
  • Master's degree thesis​​ defense: Akash Balamurugan ,​​​‌ "AI-based scheduler for IoT‌ data streams analytics", Cnam‌​‌ Paris, Oct 16, 2025,​​ Maroua Bahri , Nikolaos​​​‌ Georgantas .
  • 1st year‌ Master's internship: Luigi Lizzini‌​‌ , AI-based scheduling for​​ ML inference tasks, Cnam​​​‌ Paris, Apr-Sep 2025, Nikolaos‌ Georgantas .
  • PhDs in‌​‌ progress:
    • Haidong Zhao (from​​ March 2023): "Efficient ML​​​‌ Inference Scheduling", Sorbonne Université,‌ Nikolaos Georgantas .
    • Emile‌​‌ Royer (from November 2024):​​ "Distributed automated machine learning​​​‌ with application on IoT‌ data", Sorbonne Université, Maroua‌​‌ Bahri , Nikolaos Georgantas​​ .

8.2.2 Juries

  • Nikolaos​​​‌ Georgantas , rapporteur for‌ the habilitation thesis of‌​‌ Joyce El Haddad (Université​​ Paris Dauphine - PSL),​​​‌ Dec 2025.

8.2.3 Educational‌ and pedagogical outreach

  • Emile‌​‌ Royer , participation in​​ a round table on​​​‌ "Pursuing a research career",‌ EPF Engineering School, Apr‌​‌ 2025.

8.3 Popularization

8.3.1​​ Participation in Live events​​​‌

  • Emile Royer , presentation‌ of the interactive animation‌​‌ “The illustrator apprentice” aiming​​ at explaining artificial intelligence​​​‌ to children and the‌ general public, Fête de‌​‌ la science, Campus Pierre​​ et Marie Curie, Sorbonne​​​‌ Université, Oct 2025.

9‌ Scientific production

9.1 Major‌​‌ publications

  • 1 articleR.​​Rafael Angarita, B.​​​‌Bruno Lefèvre, S.‌Shohreh Ahvar, E.‌​‌Ehsan Ahvar, N.​​Nikolaos Georgantas and V.​​​‌Valerie Issarny. Universal‌ Social Network Bus: Towards‌​‌ the Federation of Heterogeneous​​ Online Social Network Services​​​‌.ACM Transactions on‌ Internet Technology2019HAL‌​‌DOI
  • 2 articleA.​​Amel Bennaceur and V.​​​‌Valérie Issarny. Automated‌ Synthesis of Mediators to‌​‌ Support Component Interoperability.​​IEEE Transactions on Software​​​‌ Engineering2015, 22‌HAL
  • 3 articleB.‌​‌Benjamin Billet and V.​​Valérie Issarny. Spinel:​​​‌ An Opportunistic Proxy for‌ Connecting Sensors to the‌​‌ Internet of Things.​​ACM Transactions on Internet​​​‌ Technology172March‌ 2017, 1 -‌​‌ 21HALDOI
  • 4​​ inproceedingsG.Gordon Blair​​​‌, A.Amel Bennaceur‌, N.Nikolaos Georgantas‌​‌, P.Paul Grace​​, V.Valérie Issarny​​​‌, V.Vatsala Nundloll‌ and M.Massimo Paolucci‌​‌. The Role of​​​‌ Ontologies in Emergent Middleware:​ Supporting Interoperability in Complex​‌ Distributed Systems.Big​​ Ideas track of ACM/IFIP/USENIX​​​‌ 12th International Middleware Conference​Lisbon, Portugal2011,​‌ URL: http://hal.inria.fr/inria-00629059/en
  • 5 article​​G.Georgios Bouloukakis,​​​‌ N.Nikolaos Georgantas,​ P.Patient Ntumba and​‌ V.Valérie Issarny.​​ Automated synthesis of mediators​​​‌ for middleware-layer protocol interoperability​ in the IoT.​‌Future Generation Computer Systems​​101December 2019,​​​‌ 1271-1294HALDOI
  • 6​ inproceedingsY.Yifan Du​‌, F.Francoise Sailhan​​ and V.Valerie Issarny​​​‌. Let Opportunistic Crowdsensors​ Work Together for Resource-efficient,​‌ Quality-aware Observations.PerCom​​ 2020: IEEE International Conference​​​‌ on Pervasive Computing and​ CommunicationsAustin / Virtual,​‌ United StatesMarch 2020​​HALDOI
  • 7 article​​​‌S.Sara Hachem,​ A.Animesh Pathak and​‌ V.Valérie Issarny.​​ Service-Oriented Middleware for Large-Scale​​​‌ Mobile Participatory Sensing.​Pervasive and Mobile Computing​‌2014, URL: http://hal.inria.fr/hal-00872407​​
  • 8 articleP.Patient​​​‌ Ntumba, N.Nikolaos​ Georgantas and V.Vassilis​‌ Christophides. Adaptive Scheduling​​ of Continuous Operators for​​​‌ IoT Edge analytics.​Future Generation Computer Systems​‌April 2024HALDOI​​

9.2 Publications of the​​​‌ year

International peer-reviewed conferences​

  • 9 inproceedingsS.Shahin​‌ Abdoul-Soukour, W.William​​ Aboucaya and N.Nikolaos​​​‌ Georgantas. An Interactive​ Tool for Goal Model​‌ Construction using a Knowledge​​ Graph.REFSQ 2025​​​‌ - 31st International Working​ Conference on Requirement Engineering:​‌ Foundation for Software Quality​​Barcelona, SpainApril 2025​​​‌, 15HAL
  • 10​ inproceedingsÉ.Émile Royer​‌, M.Maroua Bahri​​ and N.Nikolaos Georgantas​​​‌. OSMAC: A Dynamic​ SMAC for Data Streams​‌.2025 IEEE 37th​​ International Conference on Tools​​​‌ with Artificial Intelligence37th​ International Conference on Tools​‌ with Artificial Intelligence (ICTAI​​ 2025)Athens, GreeceDecember​​​‌ 2025, 73-80HAL​DOI
  • 11 inproceedingsH.​‌Haidong Zhao and N.​​Nikolaos Georgantas. ML​​​‌ Inference Scheduling with Predictable​ Latency.MAIoT '25:​‌ Middleware for Autonomous AIoT​​ Systems in the Computing​​​‌ Continuum in conjunction with​ the 26th ACM/IFIP International​‌ Middleware Conference (Middleware 2025)​​Nashville, United StatesACM​​​‌December 2025, 25-30​HALDOI

Doctoral dissertations​‌ and habilitation theses

Other scientific publications

Software