Scientific Publications of the University of Toulouse II Le Mirail
Not a member yet
92205 research outputs found
Sort by
: Préface et traduction de Julie Vatain-Corfdir
No full textN° ISBN : 978-2-8107-1343-1On ne compte plus les réécritures du mythe d’Orphée, mais qu’en est-il du point de vue d’Eurydice ? Jadis noyée sous les accents de la lyre, sa voix se fait pleinement entendre dans cette réécriture tant poétique que féministe. Le destin tragique de la jeune femme se joue dans une esthétique épurée, qui réinvente la tradition antique, et où la beauté visuelle de la scène n’a d’égale que son étrangeté. Par son traitement intime du mythe et sa dramaturgie qui fait vivre et revivre la perte, Eurydice offre aussi l’une des grandes méditations de la scène contemporaine sur le deuil
Receptiveness of the wine industry to fungus-resistant grape varieties in the south of France
No full textInternational audienceThe adoption of fungus-resistant grape varieties (FRGs) represents a promising pathway for steering viticulture towards more sustainable production methods by reducing the use of phytosanitary inputs. At the time of writing, the dissemination of these varieties remains limited, partly due to constraints within the wine industry (i.e., cost of planting and cultivar limitations associated with Protected Designation of Origin).This study was conducted with commercial wines made from two types of grape: Vitis vinifera and FRGs. A panel of 96 participants from the wine industry in the Occitanie region (south of France) conducted sensory evaluations. The panel performed the evaluations both blind and having been informed about type of grape in a combination of short CATA (Check-All-That-Apply) sessions followed by questions exploring their interest in FRGs and expected plantations in the coming years.The results of the sensory analyses underlined the absence of any notable difference in liking or in the sensory profiles of the wines, whether tasted blind or not. Indeed, disclosure of the type of grape used to make the wines did not alter participants’ perceptions or evaluations. Furthermore, the analysis of questionnaire data revealed a typology of three adopter profiles: i) “sceptics”: older professionals from private wineries who were generally unfavourable towards the adoption of FRGs, ii) “receptives”: cooperative members who showed measured support for innovation, and iii) “observers”: young, non-decision-making individuals with heterogeneous opinions. This industry panel predicted that there could be 25 % of vineyard areas planted with FRGs in Occitanie within the coming 30 years.This study offers insights into the future adoption of this new plant material within the wine industry
Towards Multi-Policy Hierarchical Scheduling in Linux for Containerized Space Applications
No full textInternational audienceThis paper examines the challenges of running containerized flight software on Linux for space systems, focusing on how to achieve both predictability and flexibility in scheduling. While containers provide lightweight isolation, they complicate the temporal determinism required by safety-critical spacecraft. Hierarchical scheduling and support for multiple policies are increasingly necessary to accommodate heterogeneous onboard workloads, from hard real-time control loops to best-effort processing.To address these challenges, a methodology is introduced for designing, analyzing, and implementing hierarchical multi-policy scheduling on Linux. The approach is structured around a meta-model capturing the key concepts of hierarchical reservation, a formal analysis based on the Compositional Scheduling Framework (CSF) for sizing container budgets, and a kernel-level execution model built on Linux task groups. Building on these principles, the paper presents the Task Group Bandwidth Server (TGBS), a kernel mechanism that enforces per-container CPU reservations using deadline-server abstractions and full virtual runqueues. Unlike the Hierarchical Constant Bandwidth Server (HCBS), which is limited to real-time (RT) workloads, TGBS applies uniformly to fair-share (FAIR), RT, and deadline-based (DL) scheduling classes.The experimental evaluation assesses functional behavior, real-time latency, and the execution of workloads sized analytically. Functional tests confirm correct enforcement of hierarchical reservations and priority ordering across scheduling classes. Latency measurements show that the overhead introduced by virtual runqueues remains moderate and comparable to HCBS. Finally, schedulability experiments demonstrate that reservations dimensioned with CSF are respected for real-time tasks in a mixed-policy environment, while highlighting the need for extended analytical models to fully capture FAIR scheduling.Overall, the results indicate that hierarchical reservations provide a practical path toward predictable and isolated execution of mixed-policy workloads on Linux. The work establishes a foundation that links analytical reservation sizing with kernel mechanisms and identifies future directions, including multicore support and improved multi-policy schedulability analysis
book Review: Alice Munro’s Bestiary: A Book of Human and Non-Human Animals, Héliane Ventura (2024)
No full textInternational audienceReview of: Alice Munro’s Bestiary: A Book of Human and Non-Human Animals , Héliane Ventura (2024) Newcastle Upon Tyne: Cambridge Scholars Publishing, 237 pp., ISBN 978-1-03640-870-1, h/bk, GPB 64.9
Apprentissage de Champs de Propriétés Matériaux à partir d’Images et de PINNS
No full textNational audienceLes progrès en imagerie et en corrélation d’images numériques (CIN) permettent aujourd’hui d’accéderà des champs de déplacements très riches, ouvrant notamment la voie à l’identification de champs depropriétés matériaux. Toutefois, ce type d’identification correspond à un problème inverse en grandedimension, qui peut mettre en difficulté les méthodes classiques, souvent trop couteuses dans ce cas.Dans ce contexte, nous développons une approche basée sur les réseaux de neurones informéspar la physique (PINN), capable d’identifier des champs de propriétés matériaux à partir de donnéesexpérimentales CIN ou directement à partir des images. Contrairement aux approches classiquesde machine learning, ces réseaux ne nécessitent pas un entraînement lourd basé sur de nombreuxscénarios en apprentissage supervisé lors d’une phase offline avant de pouvoir être exploités effica-cement en phase online. En effet, les paramètres physiques inconnus peuvent être intégrés commeparamètres du réseau de neurones, permettant ainsi une identification du modèle physique « à lavolée » au cours de l’entraînement. Numériquement, cette approche reste quasiment insensible à ladimensionnalité des paramètres physiques à inférer, grâce à la structure particulière des réseauxqui facilite l’évaluation des gradients via la différentiation automatique.La méthode développée repose sur une formulation mixte dans laquelle les champs de déplacementet de contrainte sont représentés par deux réseaux de neurones distincts. Plusieurs innovationssont introduites pour rendre l’approche performante en mécanique expérimentale : en particulier,formulation d’équilibres mécaniques globaux afin d’exploiter les forces de réaction mesurées, utilisationde Fourier Features pour capturer les variations locales des solutions, et stratégie d’initialisation etd’optimisation alternée pour permettre la convergence dans des espaces de grande dimension.L’approche est d’abord validée sur des données synthétiques (déplacements et images), où ellepermet de reconstruire avec précision des distributions complexes du module de Young, tant enélasticité linéaire qu’en hyperélasticité, tout en réduisant fortement l’effet du bruit. Elle est ensuiteappliquée à des données expérimentales CIN, et permet alors de reconstruire des champs de modulede Young révélant l’endommagement. La méthodologie s’est révélée non seulement précise, maiségalement efficace, ne nécessitant que des ressources informatiques standar
Forest-to-grassland conversion reduces ground beetles (Coleoptera: Carabidae) diversity in tropical Andean montane habitats
No full textInternational audienceThe deforestation of natural habitats due to agricultural expansion and human settlements is a major driver of global biodiversity loss. This study assessed the impact of natural habitat conversion into grasslands on ground beetle (Coleoptera: Carabidae) assemblages along two elevational gradients (~ 1900–3000 m asl) in the southwestern tropical Andes of Colombia. Sampling was performed using pitfall traps, litter sifting, and nocturnal hand collection at six localities, each comprising a natural forest area and a deforested grassland area. Environmental variables were recorded to characterize land use. We found that these conditions differed markedly between forests and grasslands, with forests exhibiting more stable temperature and humidity, and greater leaf litter depth. Ground beetle richness and abundance were consistently higher in forests, while grasslands were dominated by a few generalist species. Assemblages in forests showed high species turnover, including a large proportion of flightless, microendemic species, whereas grassland communities were homogenized across cordilleras and elevations. Implications for insect conservation Our study highlights the detrimental effects of deforestation on carabid diversity and assemblage structure. The observed loss of diversity implies a high risk of extinction for flightless, endemic forest specialists restricted to small areas. These findings emphasize the importance of preserving natural forest areas at different elevations and on different slopes, highlighting the need for taxonomic studies to inform targeted conservation strategies in tropical montane ecosystems
K-Means and Gaussian Mixture Models on Lie Groups: Application to Geometrical Clustering
No full textIn this article, we derive and implement two new clustering algorithms dedicated to Lie groups, adapted from the well known K-Means algorithm and Gaussian mixture models. More precisely, the K-Means algorithm is reformalized by taking into account the fact that observations belong to a Lie group (LG) with an appropriate intrinsic metric and Gaussian mixture model are defined for data living in LGs. The consistency and the performance of the resulting are numerically validated for data lying on the LG SE(2), by comparison with state-of-the-art methods for synthetic data and pseudo-real data generated using an ultra-sound sensor model
"Enjeux clinique de la motricité et de l'acte, introduction de la journée"
No full textInternational audienc
ETP4HPC SRA 6 White Paper - Energy Efficiency and Sustainability. ETP4HPC
No full textThis is a white paper released as part of the ETP4HPC’s Strategic Research Agenda 6. Energy efficiency and sustainability (EE&S) have become central challenges for modern High-Performance Computing (HPC) and AI infrastructures. While energy savings are often viewed as the main issue, sustainability also includes ecological, economic, and societal dimensions. A full life-cycle perspective—from planning and procurement to operation and decommissioning—is necessary to avoid rebound effects such as the Jevons paradox[1]. HPC and AI systems increasingly interact with their environment. Their multi-megawatt, highly variable power demand influences grid stability and requires better forecasting and closer coordination with energy providers. Future operations must incorporate energy- and carbon-aware scheduling, while heat reuse concepts and standardized sustainability metrics (PUE, WUE, ERE) become integral to system design and evaluation. Growing hardware heterogeneity (CPUs, GPUs, accelerators, and emerging neuromorphic or quantum devices) offers efficiency potential that is still underutilized due to insufficient software optimization. Progress requires stronger hardware–software co-design, greater support for developers in exploiting advanced architectural features, and systematic use of monitoring data to identify inefficient applications. European semiconductor initiatives further enable alignment between hardware and computational requirements. Software, algorithms, and workflows are equally important. Energy-efficient algorithms, data reduction techniques (compression, filtering, deduplication), improved I/O strategies, and optimized workflows can significantly reduce resource usage. Strengthening reproducibility and robust experiment packaging across diverse systems helps prevent wasted compute time and supports sustainable software lifecycles. Digital twins (DTs) are emerging as key tools for lifecycle management. They support planning, operational optimization, predictive analysis, and end-of-life decisions. Effective DTs require harmonized monitoring, standard metrics, and shared methodologies. A European knowledge initiative could accelerate adoption and ensure consistent practices across HPC sites. Raising awareness and empowering stakeholders—developers, users, operators, vendors, and funding bodies—is essential. Transparent monitoring infrastructures and vendor-agnostic sustainability indicators enable informed decision-making. User feedback mechanisms (dashboards, reports, incentives) can encourage more energy-conscious behaviour. Funding agencies, particularly the European Commission, should embed sustainability metrics, full life-cycle assessments [ISO 14040/14044], and continuous reporting requirements into procurement and project evaluations. In the post-exascale era, simple performance scaling through increased energy use is no longer viable. With growing AI workloads and the need to reduce energy demand and CO₂ emissions, coordinated progress across hardware, software, operations, infrastructure, and user behaviour is imperative. Only a combined effort by all stakeholders will enable high-performing HPC and AI systems to meet future needs while aligning with long-term sustainability goals. [1] https://en.wikipedia.org/wiki/Jevons_parado
Ordonnancement en ligne robuste de coflows à partir de prédictions
No full textIn this paper, we introduce an online scheduling framework for minimizing the total weighted average completion time of coflows, in scenarios where both flow sizes and release times are unknown. Instead of relying on exact flow sizes, the framework leverages potentially inaccurate predictions. At each decision point, it selects a subset of coflows by solving a Minimum Unscheduled Weight Problem (MUWP), and then applies a variant of the Sincronia algorithm-adapted to operate directly on the predicted sizes. We prove that the proposed approach achieves provably strong competitive guarantees, and extensive simulations demonstrate that it outperforms baseline methods both on average and in worst-case instances