2025Activity reportProject-Team​‌KERDATA

Our objective:​‌ Enable the Data Continuum.​​

Our research project aims​​​‌ to address some open​ challenges at each of​‌ the aforementioned three levels,​​ while considering the two​​​‌ aforementioned transverse concerns. We​ specifically focus on data-related​‌ challenges posed by the​​ requirements (storage, processing, analytics)​​​‌ of complex workflows executed​ on the Edge-Cloud-HPC continuum​‌ and propose innovative algorithms​​ and software architecture solutions​​​‌ towards a Data Continuum.​

Data-Centric​​​‌ Workflow Composition in the​ Computing Continuum.

The scientific​‌ community has reached consensus​​ that common interfaces for​​​‌ data management in the​ continuum are necessary 70​‌. Unified data abstractions​​ can enable the interoperability​​​‌ of data storage and​ processing across the continuum​‌ and facilitate data analytics​​ at all levels 53​​​‌, alleviating the disconnect​ between application- and storage-oriented​‌ approaches to interoperability. However,​​ no unified data modeling​​​‌ approaches exist for structuring​ and representing data on​‌ a logical level across​​ the computing continuum.

The​​​‌ first steps in this​ research direction involve establishing​‌ what are the essential​​ attributes needed to represent​​​‌ data in the different​ programming models coexisting in​‌ the continuum (e.g., ML​​ models, simulation data, annotations​​​‌ resulting from analysis). We​ will systematically categorize these​‌ attributes to deliver data​​ abstractions and models that​​​‌ can be specialized for​ different tasks. In addition,​‌ we will investigate how​​ to embed metadata in​​​‌ these abstractions so that​ future work can explore​‌ new ways of describing,​​ processing, and tracking data​​​‌ at the workflow level,​ which aligns with the​‌ topics of workflow instrumentation​​ and reproducibility. On a​​​‌ longer term, we will​ also study how these​‌ data models could support​​ data volume and transfer​​​‌ reduction as a step​ towards resource sustainability of​‌ applications in the continuum.​​

Enabling Reproducibility and Trustworthiness​​​‌ in Complex Workflows Across​ the Computing Continuum.

Current​‌ approaches to support workflow​​ reproducibility are based on​​​‌ workflow modelling 104 or​ simulation 26, 112​‌. These approaches raise​​ some important challenges in​​​‌ terms of specification, modelling,​ and validation to support​‌ reproducibility in the Computing​​ Continuum. For example, it​​​‌ is increasingly difficult to​ model the heterogeneity and​‌ volatility of Edge devices​​ or to assess the​​​‌ impact of the inherent​ complexity of hybrid Edge-Cloud​‌ deployments on performance. With​​ the rise of AI​​​‌ workflows, the issue of​ reproducibility is aggravated by​‌ the limitations to our​​ ability to reason about​​​‌ the decision-making process of​ many machine learning models​‌ that act as a​​ black box 33,​​​‌ and the lack of​ comprehensive specifications to the​‌ data that needs to​​ be collected to establish​​​‌ a provenance chain (i.e.,​ a record trail of​‌ the overall state of​​ the application and its​​​‌ intermediate results) that builds​ trust on the workflow's​‌ results 46.

We​​ aim to tackle these​​​‌ challenges through a rigorous​ methodology supporting the automatic​‌ deployment, the complete analysis​​ cycle and the optimization​​ of applications on the​​​‌ Computing Continuum. We started‌ to implement this methodology‌​‌ in the E2Clab 102​​ software tool for workflow​​​‌ lifecycle management across the‌ Continuum. For the short‌​‌ term, in the framework​​ of the STEEL project​​​‌ of the PEPR Cloud,‌ we plan to investigate‌​‌ how to further enrich​​ both the methodology and​​​‌ the tool, in order‌ to support the next‌​‌ generation of scientific testbeds​​ (e.g., SLICES-FR) and​​​‌ the non-trivial reproducibility of‌ ML and DL workflows.‌​‌ This is a particularly​​ challenging direction due to​​​‌ the increased degree of‌ randomness (i.e., in terms‌​‌ of initial parameters and​​ hyperparameters settings) incurred by​​​‌ such applications. We will‌ expand E2Clab to capture‌​‌ provenance metadata during the​​ execution of AI workflows,​​​‌ which includes a detailed‌ record of data sources,‌​‌ processing steps, model configurations,​​ and computational resources utilized.​​​‌ This provenance metadata can‌ be leveraged not only‌​‌ to ensure transparency and​​ traceability throughout the AI​​​‌ lifecycle, but also to‌ conduct resource, energy and‌​‌ performance optimizations. For the​​ longer term we will​​​‌ work towards the definition‌ of ontologies and taxonomies‌​‌ for AI workflow provenance​​ data to build a​​​‌ theoretical foundation for developing‌ provenance data management systems‌​‌ tailored for the different​​ stakeholders involved in AI​​​‌ applications.

Efficient Parallel Continual‌ Learning.

Some scenarios of‌​‌ DL training involve the​​ need to assimilate new​​​‌ training data arriving continuously.‌ This kind of incremental‌​‌ training suffers from catastrophic​​ forgetting (i.e., new patterns​​​‌ are reinforced at the‌ expense of previously acquired‌​‌ knowledge). Training from scratch​​ each time new training​​​‌ data becomes available would‌ result in extremely long‌​‌ training times and massive​​ data accumulation. Rehearsal-based continual​​​‌ learning mixes samples from‌ previous training tasks with‌​‌ samples from new training​​ tasks to alleviate catastrophic​​​‌ forgetting, but research to‌ date has not addressed‌​‌ performance and scalability of​​ these methods 39,​​​‌ 101, 54,‌ 59.

We propose‌​‌ asynchronous data management techniques​​ that enable the design​​​‌ and implementation of a‌ scalable distributed rehearsal buffer‌​‌ abstraction, which is instrumental​​ in enabling continual learning​​​‌ to take advantage of‌ data-parallel techniques. So far,‌​‌ this solution was validated​​ for class-incremental classification problems.​​​‌ The approach could however‌ be easily applied to‌​‌ generative models (in which​​ case we can simply​​​‌ use one class to‌ store all representatives). This‌​‌ is a short-term research​​ direction we intend to​​​‌ explore in the context‌ of our new research‌​‌ project. For the longer​​ term, we plan to​​​‌ further explore other parallelization‌ approaches (model-based and hybrid)‌​‌ to address the challenges​​ posed by evolving datasets​​​‌ in DL models.

Scalable,‌ Secure and Resource-Efficient Federated‌​‌ Learning.

FL aims to​​ achieve an accuracy close​​​‌ to the one achieved‌ by centralized models but‌​‌ in a scalable and​​ resource-efficient manner. Simultaneously, FL​​​‌ is subject to security‌ threats coming from the‌​‌ edge of the network​​ since malicious peers may​​​‌ attempt to manipulate the‌ learning process, compromise the‌​‌ privacy of other peers,​​ or disrupt the training​​​‌ altogether. Clustered FL –grouping‌ clients with similar data‌​‌ distributions and by training​​​‌ personalized models in each​ identified cluster– is a​‌ mechanism to support low​​ resource utilization, but existing​​​‌ approaches 72, 51​ mainly focus on the​‌ achieved accuracy of the​​ clustering mechanisms, overlooking system​​​‌ and infrastructure resource constraints​ like energy consumption. At​‌ the same time, current​​ threat mitigation approaches 47​​​‌, 116, 36​, 84 rely on​‌ robust aggregation, anomaly detection​​ and generative models for​​​‌ defending against poisoning attacks.​ Yet, they either have​‌ limited defensive capabilities due​​ to their underlying design​​​‌ or are impractical to​ use as they rely​‌ on constraining building blocks.​​

For the short term,​​​‌ we plan to explore​ approaches to scalable, secure​‌ and resource-efficient FL considering​​ for the first time​​​‌ the device heterogeneity, the​ training accuracy and the​‌ robustness against malicious activity​​ simultaneously. A first direction​​​‌ consists in devising resource-constrained​ clustering algorithms, specifically tailored​‌ for FL executed at​​ the edge. The goal​​​‌ is to enable transparent​ adaptation to the execution​‌ environment (e.g., node volatility,​​ malicious attacks, network congestion)​​​‌ by automatically tuning the​ FL parameters in order​‌ to improve user-defined performance​​ metrics (e.g., energy efficiency,​​​‌ execution time, accuracy). Generative​ model based approaches have​‌ been gaining increasing interest,​​ and are shown to​​​‌ be more resilient against​ a wider range of​‌ attacks. In this context,​​ we will continue to​​​‌ extend FedGuard 60,​ our novel FL framework​‌ that utilizes the generative​​ capabilities of Conditional Variational​​​‌ AutoEncoders (CVAE) to effectively​ defend against poisoning attacks​‌ with tuneable overhead in​​ communication and computation. We​​​‌ plan to enhance the​ robustness of this approach​‌ through new aggregation operators​​ and under different levels​​​‌ of dataset imbalance, including​ highly imbalanced datasets with​‌ very few samples per​​ client. For the longer​​​‌ term, a more challenging​ direction that we plan​‌ to explore is how​​ FedGuard and other strategies​​​‌ perform in a setup​ where clients get access​‌ to a stream of​​ incoming data (i.e., dynamic​​​‌ datasets).

Workflow I/O Behavior Analysis​​ Methods for Sustainability.

Understanding​​​‌ how workflows and applications‌ use staging areas on‌​‌ the Computing Continuum is​​ decisive for improving scheduling​​​‌ algorithms and deploying an‌ optimized I/O software stack‌​‌ 63. This requires​​ first characterizing these applications​​​‌ and workflows from an‌ I/O point of view,‌​‌ i.e. determine through performance​​ evaluation and empirical study​​​‌ a relatively high-level set‌ of characteristics that describes‌​‌ the data access pattern​​ 100, 97.​​​‌ Data collection and analysis‌ will also leverage semi-supervised‌​‌ clustering methods and federated​​ learning techniques from Axis​​​‌ 1. The result of‌ this characterization can then‌​‌ be used to feed​​ job and I/O scheduling​​​‌ algorithms and improve data‌ movement efficiency. The preliminary‌​‌ step in this characterization​​ is the collection of​​​‌ execution traces, from which‌ detailed studies can be‌​‌ carried out 55.​​ In the field of​​​‌ high-performance computing, Darshan109‌ is the reference tool‌​‌ for I/O monitoring.

We​​ will extend our existing​​​‌ work on PyDarshan 88‌, 87, a‌​‌ Python library for querying​​ Darshan log records, and​​​‌ develop new tools and‌ abstractions applicable throughout the‌​‌ computing continuum that generalize​​ "decision support services" that​​​‌ allow the augmentation of‌ workflow execution plans with‌​‌ I/O and energy behavior​​ information which support our​​​‌ resource management and system‌ architecture research. For example,‌​‌ by identifying the most​​ energy-intense operations, candidates for​​​‌ hardware acceleration can be‌ determined. This short-term direction‌​‌ is the first step​​ to provide support for​​​‌ the design of future‌ strategies for I/O scheduling‌​‌ considering performance/energy trade-offs, that​​ we further plan to​​​‌ investigate for the longer‌ term. This line of‌​‌ research also supports Axis​​ 3 on resource management.​​​‌

Abstraction for HPC/Cloud Storage‌ Convergence.

On HPC and‌​‌ Cloud infrastructures, while the​​ number of processing units​​​‌ has grown to meet‌ the computing power requirements‌​‌ of large-scale applications, the​​ I/O capacity as well​​​‌ as the I/O bandwidth‌ per core have drastically‌​‌ decreased. Thus, data management​​ and analytics becoming the​​​‌ critical bottleneck on large-scale‌ systems, vendors have overcome‌​‌ this problem by deploying​​ new tiers of intermediate​​​‌ storage between the applications‌ and the global shared‌​‌ storage system, usually along​​ with a dedicated (and​​​‌ sometimes proprietary) software layer.‌ These new levels of‌​‌ storage hierarchy feature various​​ capacities, characteristics and performance​​​‌ one has to be‌ aware of to fully‌​‌ utilize them 86.​​​‌ This is especially true​ in the context of​‌ complex hybrid workflows such​​ as in-situ analysis, visualization​​​‌ or code-coupling: the unawareness​ of those underlying tiers​‌ is a serious loss​​ of performance 71.​​​‌ An approach focusing on​ storage convergence across HPC​‌ and Cloud infrastructure is​​ decisive to glue this​​​‌ deep hierarchy and to​ make the most of​‌ these new technologies on​​ one side and ensure​​​‌ effective data sharing between​ components running across the​‌ computing continuum 34,​​ 43.

Identifying a​​​‌ good storage abstraction that​ is accurate enough to​‌ properly describe the wide​​ variety of devices and​​​‌ sufficiently general to be​ portable on various systems​‌ is crucial. In that​​ context, we will work​​​‌ on the development of​ a two-stage abstraction layer​‌ above local (system) and​​ remote (distant platform) storage​​​‌ resources. To do so,​ we will follow a​‌ co-design approach, whereby the​​ HPC, Cloud, and Edge-computing​​​‌ architectures would all benefit​ from an infrastructure-wide level​‌ of abstraction. This work​​ will be a continuation​​​‌ of the research undertaken​ on data aggregation 115​‌, 114, for​​ which an abstraction of​​​‌ network topology and memory​ and storage levels was​‌ necessary for the algorithm's​​ portability, and will also​​​‌ build on existing work​ in the community on​‌ resource abstraction 52,​​ 113. In the​​​‌ longer term, we plan​ to extend this library,​‌ which focuses on physical​​ resources, with a logical​​​‌ layer. More concretely, we​ will build on top​‌ of that library a​​ tier-to-tier data transfer layer​​​‌ enabling compatibility between several​ storage paradigms (block, file,​‌ object).

Exascale In-Situ Analytics.​​

Without a major change​​​‌ in practices, the increased​ computing capacity of the​‌ next generation of computers​​ will lead to an​​​‌ explosion in the volume​ of data produced by​‌ numerical simulations. Managing this​​ data, from production to​​​‌ analysis, is a major​ challenge. While it is​‌ not conceivable to do​​ without a storage system,​​​‌ many experiments are aimed​ at reducing its use.​‌ Thus emerged approaches leveraging​​ in-situ processing, in-transit​​​‌ processing, staging nodes​, helper cores 45​‌, 66. All​​ of these approaches aim​​​‌ to replace the usual​ write-read process by a​‌ means to perform analysis​​ at the same time​​​‌ as simulation, a​ capability of particular interest​‌ to physicists. This need​​ has led to the​​​‌ first implementations of in-situ​ or in-transit analysis systems​‌ in simulation codes, and​​ to the creation of​​​‌ specific middleware for asynchronous,​ scalable post-Exascale systems, such​‌ as Damaris CITATION NOT​​ FOUND: dorier:hal-00715252, 67​​​‌, 49.

Developed​ by the KerData team​‌ since 2011, Damaris is​​ the team's flagship software​​​‌ for Exascale HPC. Damaris​ 4 proposes a middleware-level​‌ approach to scalable asynchronous​​ I/O management and real-time​​​‌ in situ processing of​ data from large-scale MPI-based​‌ HPC simulations. It leverages​​ the idea of dedicated​​​‌ cores for such tasks​ performed asynchronously within multicore​‌ nodes. Initial feedback from​​ application users clearly shows​​​‌ the need to design​ a system that can​‌ dynamically trigger the activation​​ of new analyses during​​ the simulation run. The​​​‌ timing can be decided‌ either by the simulation‌​‌ code or by an​​ analysis. To maintain high​​​‌ performance results, it is‌ also essential to appropriately‌​‌ leverage the possibility to​​ place analysis tasks on​​​‌ GPUs. These are challenges‌ we plan to address‌​‌ by extending our previous​​ work based on the​​​‌ Damaris approach, in the‌ context of the Exa-DoST‌​‌ project of the NumPEx​​ PEPR (2023-2030), to support​​​‌ the needs of Exascale‌ workloads. In particular, two‌​‌ applications are targeted: SKA​​ 65 in collaboration with​​​‌ the CNRS, Observatoire de‌ Paris and Observatoire de‌​‌ la Côte d'Azur and​​ Gysela 74, in​​​‌ collaboration with the CEA.‌ For the longer term,‌​‌ we expect additional application​​ requirements to emerge during​​​‌ the execution of the‌ NumPEx PEPR program (in‌​‌ particular, in collaboration with​​ the NumPEx Exa-DI project​​​‌ 5, which has‌ set up a dedicated‌​‌ process to support new​​ applications, not identified yet).​​​‌ We plan to contribute‌ to the support of‌​‌ such applications that could​​ exhibit new patterns with​​​‌ respect to in situ‌ analysis.

Sustainable Interoperability Across‌​‌ the Computing Continuum.

New​​ endeavors towards interoperability in​​​‌ the continuum are addressing‌ the need for common‌​‌ data spaces through federated​​ data infrastructures in the​​​‌ cloud (e.g., Gaia-X 50‌, European Open Science‌​‌ Cloud (EOSC) 37,​​ German National Research Data​​​‌ Infrastructure (NFDI) 75)‌ and converged research infrastructures‌​‌ for leadership-class supercomputing and​​ cloud resources (e.g., FENIX​​​‌ 6, EuroHPC 107‌, PRACE-RI 85,‌​‌ European Grid Initiative (EGI)​​ 80). For the​​​‌ long term, key players‌ in the public and‌​‌ private sectors are making​​ strong investments in long-term​​​‌ strategic decisions about how‌ the computing continuum will‌​‌ develop. Specifically, quantum computing​​ is receiving massive support,​​​‌ and one of the‌ mandates of the EuroHPC‌​‌ JU is to acquire​​ and deploy quantum technologies​​​‌ in HPC environments once‌ they reach sufficient maturity‌​‌ 7. Furthermore, other​​ new technologies like neuromorphic​​​‌ accelerators 56 will increase‌ the heterogeneity in future‌​‌ HPC systems. These non-conventional​​ architectures can also be​​​‌ found in highly energy-efficient‌ IoT devices 95,‌​‌ 108, fast scientific​​ instruments for large-scale science​​​‌ 110, and new‌ approaches for fast and‌​‌ efficient artificial intelligence in​​ cloud and HPC environments​​​‌ 99, 89,‌ 96. Current works‌​‌ on the integration of​​ emerging technologies into existing​​​‌ computing ecosystems focus on‌ the interoperability and performance‌​‌ of algorithms without considering​​ data-oriented optimizations, and workflow-specific​​​‌ challenges such as task-resource‌ mapping, automation, and provenance‌​‌ are rarely explored 64​​, 98.

Overall,​​​‌ the successful interoperability between‌ the existing and projected‌​‌ platforms in the Computing​​ Continuum will depend on​​​‌ middleware able to interoperate‌ and execute in hybrid‌​‌ scenarios. In the short​​ term, we plan to​​​‌ investigate the design a‌ data exchange layer that‌​‌ will be the core​​ of a workflow composition​​​‌ approach that connects with‌ established data staging and‌​‌ transport layers, alleviating the​​ disconnect between raw data​​​‌ management and knowledge-based workflow‌ management in the continuum‌​‌ for better resource balancing,​​​‌ economy, provenance, and data​ reduction. In addition, the​‌ applications themselves could use​​ this information hub to​​​‌ monitor and record the​ progress of their individual​‌ components in smart ways,​​ enriching existing approaches for​​​‌ in situ analysis and​ workflow reproducibility. To ensure​‌ the sustainability of workflow​​ software solutions in an​​​‌ evolving and hyper-heterogeneous landscape,​ on a longer term​‌ we will study the​​ data and access patterns​​​‌ in hybrid workflows involving​ the interaction with emerging​‌ hardware technologies. The final,​​ long-term goal is to​​​‌ contribute to the development​ of an interoperable software​‌ stack by designing data​​ models that address the​​​‌ challenges in encoding, arrangement,​ locality, and mapping to​‌ high-level data abstractions.

Provisioning Storage Resources​​​‌ on Large-Scale Infrastructures.

While​ for years HPC systems​‌ were the predominant means​​ of meeting the requirements​​​‌ expressed by large-scale scientific​ workflows, today some components​‌ have moved away from​​ supercomputers to extend across​​​‌ the Computing Continuum. This​ migration has been mainly​‌ motivated by the need​​ of specialized data processing​​​‌ such as data filtering​ at the Edge or​‌ data analysis on Cloud​​ infrastructures. From an I/O​​​‌ and storage perspective, this​ means having to deal​‌ with very different paradigms:​​ infrastructures where direct access​​​‌ to resources is extremely​ limited due to a​‌ very high level of​​ abstraction, on-premise supercomputers offering​​​‌ a low-level approach requiring​ tight user control, or​‌ highly-constrained devices limited in​​ terms of access and​​​‌ reconfiguration. One way to​ address that is to​‌ converge the infrastructures composing​​ the Computing Continuum by​​​‌ exploring ways to provision​ storage resources distributed across​‌ hybrid HPC/Cloud/Edge systems to​​ complex scientific workflows combining​​​‌ data production, simulation and​ data analysis. However, this​‌ implies low-level access to​​ systems that are sometimes​​​‌ difficult to reach, or​ to resources in production.​‌ Simulation is one way​​ of exploring storage provisioning​​ 58, 57,​​​‌ 82.

In this‌ context, we will continue‌​‌ our work on the​​ simulation of storage systems​​​‌ implemented within the StorAlloc‌ 91, 92 simulator.‌​‌ This work has enabled​​ us to demonstrate the​​​‌ correct sizing and partitioning‌ of intermediate storage resources‌​‌ and to work on​​ the modeling of storage​​​‌ systems, including the way‌ in which they distribute‌​‌ data over the available​​ storage spaces. In KerData,​​​‌ we will be exploring‌ new methods for I/O-aware‌​‌ scheduling of jobs on​​ hybrid infrastructures 78.​​​‌ Based on post-mortem studies‌ of storage systems, we‌​‌ will also work on​​ characterizing workflows running on​​​‌ the Computing Continuum 106‌ in order to refine‌​‌ job scheduling decisions. In​​ the longer term, our​​​‌ ambition is to propose‌ a calibrated and validated‌​‌ simulator of storage systems​​ distributed across the Computing​​​‌ Continuum. This simulator will‌ enable us to predict‌​‌ the I/O performance and​​ energy cost of complex​​​‌ workflows leveraging Edge, Cloud‌ and HPC resources. To‌​‌ achieve this, we will​​ rely on state-of-the-art simulation​​​‌ frameworks such as WRENCH‌ 57 and SimGrid 58‌​‌.

Storage Disaggregation and​​ Computational Storage.

A key​​​‌ challenge for many large-scale‌ applications is the mismatch‌​‌ between compute power, the​​ dimension of caches and​​​‌ buffers, and the available‌ I/O bandwidth from the‌​‌ network down to the​​ chip level. Important contributing​​​‌ factors to this situation‌ include economies of scale‌​‌ catering to markets with​​ different needs, a prohibitively​​​‌ expensive development process and‌ lack of manufacturing capacity‌​‌ to consider more customized​​ solutions. However, as computing,​​​‌ memory, storage, and network‌ hardware are becoming increasingly‌​‌ modular and re-configurable, it​​ is possible to consider​​​‌ system and storage architectures‌ that would have been‌​‌ prohibitively expensive before 29​​, 38. The​​​‌ enabling technologies for some‌ of these developments are‌​‌ the emerging industry standards​​ such as the Compute​​​‌ Express Link (CXL) 73‌, 83 for more‌​‌ flexible integration of accelerators​​ and disaggregated storage, and​​​‌ the P4 programming language‌ used in re-configurable networking‌​‌ 48.

We will​​ identify the most energy-intensive​​​‌ routines used on the‌ I/O path, both from‌​‌ the service perspective and​​ from the domain perspective,​​​‌ and curate a portable‌ reference library of key‌​‌ algorithms catering to both​​ software and hardware acceleration.​​​‌ We will leverage, for‌ example, open container standards‌​‌ and instruction set architectures​​ such as RISC-V that​​​‌ can be applied both‌ in data centers and‌​‌ resource-constrained edge contexts 117​​. High-level technologies such​​​‌ as containers facilitate the‌ development of algorithmic improvements‌​‌ but often also introduce​​ runtime overhead, while low-level​​​‌ hardware acceleration allows to‌ reduce the energy consumption‌​‌ of computations to a​​ minimum. Unfortunately the hardware​​​‌ acceleration of all desired‌ functionality is not possible‌​‌ because of the need​​ to retain flexibility and​​​‌ limitations due to cost,‌ manufacturing, and physical constraints.‌​‌ Instead a careful selection​​ of routines that should​​​‌ be hardware accelerated needs‌ to be performed for‌​‌ which the priorities shift​​ from application to application.​​​‌ We will start by‌ exploring domain and service-specific‌​‌ (e.g., compression, erasure coding,​​​‌ encryption) approaches first and​ then identify generalized abstractions​‌ for common functionality useful​​ across domains. Ultimately, this​​​‌ research enables building reusable,​ abstract, and fine-grained building​‌ blocks that allow the​​ construction of frugal computational​​​‌ storage architectures including the​ subset of functionality optimized​‌ for a particular application​​ or workflow.

Frugal Data​​​‌ Storage Architectures to Support​ Post-Exascale Workflows.

Post-exascale workflows​‌ such as digital twins​​ and machine learning require​​​‌ fast access to increasing​ amounts of data in​‌ long-term archives which poses​​ challenging using existing storage​​​‌ technologies. Especially, long-term storage​ is latency and bandwidth​‌ constrained while high-performance storage​​ systems tend to be​​​‌ cost, energy, and capacity​ constrained. A major obstacle​‌ for better utilization of​​ existing technologies lies in​​​‌ the requirements of legacy​ applications, but due to​‌ applications and workflows transitioning​​ to new programming models​​​‌ it becomes possible to​ consider new storage system​‌ architectures. This creates an​​ opportunity to research frugal​​​‌ data storage architectures that​ integrate computational storage allowing​‌ to avoid wait times​​ and stress on contended​​​‌ resources such as the​ network and storage subsystems​‌ while also increasing energy​​ efficiency through hardware acceleration.​​​‌

High I/O performance and​ energy-efficient storage designs require​‌ taking a domain- or​​ application-specific approaches and an​​​‌ extension of computational storage​ to long-term data archives​‌ 73, 83,​​ 69. By applying​​​‌ the methods for holistic​ workflow I/O behavior analysis​‌ already discussed in Axis​​ 2 to discover bottlenecks,​​​‌ it becomes possible to​ identify service- and application-specific​‌ I/O bottlenecks and apply​​ I/O acceleration building blocks.​​​‌ Using these identified building​ blocks we will develop​‌ software libraries that allow​​ their remote execution and/or​​​‌ hardware acceleration close to​ the data storage location.​‌ A holistic effort is​​ necessary to combine advances​​​‌ and consolidation in data​ and workflow management –​‌ as promoted by the​​ FAIR principles (Findable, Accessible,​​​‌ Interoperable, Reusable) 118 –​ with emerging open technologies​‌ for computation and storage​​ 29, 38.​​​‌ To this end, we​ will research low-level software​‌ and hardware support for​​ metadata queries well as​​​‌ aggregations on top of​ self-describing data formats needed​‌ by both the application​​ workflow and data management​​​‌ communities. In particular, we​ will investigate the integration​‌ of emerging storage technologies​​ that allow for highly​​​‌ parallel access as found​ in NAND- and NVRAM-based​‌ systems as manufacturing costs​​ go down, as well​​​‌ as DNA-based storage systems​ when array-based synthesis methods​‌ become commercially available 81​​, 94.

HPC​​​‌ Resource Management Faced with​ the Environmental Crisis.

It​‌ is essential to consider​​ the evolution of HPC​​​‌ in the face of​ the climate crisis, and​‌ its impact on our​​ research topics. As in​​​‌ other field, we have​ to consider what a​‌ "lower-tech" version of HPC​​ would be, how to​​​‌ make it usable. This​ is the target of​‌ this axis. The current​​ trend in HPC has​​​‌ been to outbid each​ other for new supercomputers,​‌ renewing them every 6-7​​ years to make them​​​‌ ever more powerful. However,​ this policy seems hardly​‌ sustainable. Regular shortages of​​ components 42, 77​​, the origin (and​​​‌ uniqueness) of the sources‌ of certain materials, coupled‌​‌ with the geopolitical context​​ 61 alone make this​​​‌ growth policy challenging. We‌ will start by trying‌​‌ to evaluate the need​​ for scale from a​​​‌ social perspective: what is‌ the relation between scale‌​‌ and social advance. While​​ the scientific community traditionally​​​‌ relies on various metrics‌ to assess the performance‌​‌ of HPC systems —such​​ as the Top500 list​​​‌ (based on HPL performance),‌ HPCG, Graph500, IO500— these‌​‌ metrics do not capture​​ how HPC contributes to​​​‌ social progress.

Then, in‌ front of the lack‌​‌ of resources, we expect​​ manufacturers to need to​​​‌ take these risks into‌ account in their future‌​‌ machines. American HPC laboratories​​ already have constraints for​​​‌ the 2050 horizon, such‌ as zero-emission procurement. So,‌​‌ the first trend we​​ can expect is an​​​‌ extension of HPC machine‌ lifetimes. This could be‌​‌ followed by a move​​ towards refurbished machines, i.e.​​​‌ machines that use components‌ from other machines. These‌​‌ changes, and the introduction​​ of second-hand hardware, should​​​‌ open up several challenges‌ for system managers. Until‌​‌ now, the number of​​ faults has grown linearly​​​‌ with the number of‌ resources 35. HPC‌​‌ fault tolerance mechanism assume​​ that the Mean Time​​​‌ Between Failure is large‌ in front of system‌​‌ characteristic time (such as​​ the time to checkpoint​​​‌ data). With second-hand material,‌ the number of fault‌​‌ may increase at a​​ much higher rate, while​​​‌ machine performance would not‌ improve since we are‌​‌ not updating the machines.​​ This would render obsolete​​​‌ existing fault-tolerance mechanisms. To‌ this end we will‌​‌ explore new fault-tolerance mechanisms​​ that could be applicable.​​​‌ Heterogeneity linked to resource‌ unavailability and the increased‌​‌ computational complexity motivates the​​ need for a precise​​​‌ description of available resources.‌ In this context, we‌​‌ will explore alternatives for​​ an efficient design of​​​‌ resource management systems to‌ optimize the use of‌​‌ these resources. In addition,​​ non-fatal faults may be​​​‌ invisible, typically slowdowns/varying performance‌ due to wear and‌​‌ tear. We will investigate​​ how one can detect​​​‌ and manage resources that‌ slow down the calculations‌​‌ performed on it. Of​​ course, one of the​​​‌ challenge of this axis‌ will be to work‌​‌ on defining metrics to​​ evaluate the benefits of​​​‌ various solutions. Indeed, not‌ only is it a‌​‌ multi-dimensional problem (TCO analysis),​​ but it should also​​​‌ consider what has been‌ long known on optimization‌​‌ and Jevons paradox 28​​.

Social impact.​​

When considering sufficiency in​​​‌ HPC, we need to‌ question the use of‌​‌ the resources and if​​ we can reduce them.​​​‌ This is the main‌ challenge of the project‌​‌ on result-scalability 11:​​ we aim at proposing​​​‌ ways to correctly resize‌ HPC computations by focusing‌​‌ on an evaluation of​​ the output rather than​​​‌ by considering input-based scaling‌ models.

Environmental impact.

Part‌​‌ of our research focuses​​ on extending the lifespan​​​‌ of HPC machines, in‌ the hope that it‌​‌ could reduce the environmental​​ impact of the field.​​​‌ We have set-up a‌ working group with different‌​‌ teams at Inria Rennes​​ (PACAP, TARAN) to study​​​‌ the challenges extending the‌ life of supercomputers would‌​‌ raise.

7.1.1​​ Damaris

• Keywords:
  Visualization, I/O,​​​‌ HPC, Exascale, High performance‌ computing
• Scientific Description:

  Damaris‌​‌ is a middleware for​​ I/O and data management​​​‌ targeting large-scale, MPI-based HPC‌ simulations. It initially proposed‌​‌ to dedicate cores for​​ asynchronous I/O in multicore​​​‌ nodes of recent HPC‌ platforms, with an emphasis‌​‌ on ease of integration​​ in existing simulations, efficient​​​‌ resource usage (with the‌ use of shared memory)‌​‌ and simplicity of extension​​ through plug-ins.

  Over the​​​‌ years, Damaris has evolved‌ into a more elaborate‌​‌ system, providing the possibility​​ to use dedicated cores​​​‌ or dedicated nodes to‌ in situ data processing‌​‌ and visualization. It proposes​​ a seamless connection to​​​‌ the VisIt visualization framework‌ to enable in situ‌​‌ visualization with minimum impact​​ on run time. Damaris​​​‌ provides an extremely simple‌ API and can be‌​‌ easily integrated into the​​ existing large-scale simulations.

  Damaris​​​‌ was at the core‌ of the PhD thesis‌​‌ of Matthieu Dorier, who​​ received an Accessit to​​​‌ the Gilles Kahn Ph.D.‌ Thesis Award of the‌​‌ SIF and the Academy​​ of Science in 2015.​​​‌ Developed in the framework‌ of our collaboration with‌​‌ the JLESC – Joint​​ Laboratory for Extreme-Scale Computing,​​​‌ Damaris was the first‌ software resulted from this‌​‌ joint lab validated in​​ 2011 for integration to​​​‌ the Blue Waters supercomputer‌ project. It scaled up‌​‌ to 16,000 cores on​​ Oak Ridge’s leadership supercomputer​​​‌ Titan (first in the‌ Top500 supercomputer list in‌​‌ 2013) before being validated​​ on other top supercomputers.​​​‌ Active development is currently‌ continuing within the KerData‌​‌ team at Inria, where​​ it is at the​​​‌ center of several collaborations‌ with industry as well‌​‌ as with national and​​​‌ international academic partners.

  Damaris​ has been selected to​‌ be one of the​​ key software pieces of​​​‌ software for the NumPEx​ PEPR project, which aims​‌ to provide the software​​ infrastructure for the future​​​‌ Exascale machine to be​ hosted in France in​‌ 2025 (Alice Recoque, Jules​​ Vernes project). The capabilities​​​‌ within Damaris will further​ studied in collaboration with​‌ CEA within the NumPEx​​ exploratory PEPR project.

• Functional​​​‌ Description:
  Damaris is a​ middleware for data management​‌ and in-situ visualization targeting​​ large-scale HPC simulations. Damaris​​​‌ enables: - In-situ data​ analysis by using selected​‌ dedicated cores/nodes of the​​ simulation platform. - Asynchronous​​​‌ and fast data transfer​ from HPC simulations to​‌ Damaris. - Semantic-aware dataset​​ processing through Damaris plug-ins,​​​‌ - Writing aggregated data​ (by hdf5 format) or​‌ visualizing them either by​​ VisIt or ParaView. -​​​‌ Dask analytics supports.
• Release​ Contributions:
  v1.12.1 of Damaris​‌ provides basis for an​​ overhaul of the the​​​‌ Plugin layer: adding event​ triggers on specific hooks,​‌ reorganizing event functioning, and​​ enabling/adding data dependency for​​​‌ events. It includes also​ the (missing) implementions of​‌ the management of Parameter​​ for the string and​​​‌ label types, and the​ handling of some typos​‌ and bugs.
• News of​​ the Year:
  In 2025,​​​‌ an extendable Scheduling layer​ has been added (yet​‌ to be released): to​​ reduce communication costs. Also,​​​‌ to enable dynamic analysis​ handling capability in Damaris,​‌ two main activities have​​ been carry out (yet​​​‌ to be released). The​ development of a dynamic​‌ expression module, and an​​ overhaul of the the​​​‌ Plugin layer (Harmonization of​ plugin definition, with possibility​‌ to pass specific data​​ to each plugin, dynamic​​​‌ event creation, triggers (with​ condition), event/data dependency, data​‌ availability across iteration with​​ ‘sliding window’). Furthermore, in​​​‌ the context of NumPEx​ PEPR project, we enhanced​‌ the Damaris / PDI​​ interoperability. We continued the​​​‌ development of the Damaris​ plugin for PDI (to​‌ be release soon), and​​ started working on the​​​‌ PDI plugin in Damaris.​ With this, the simulation​‌ instrumented with PDI (https://pdi.dev/main/),​​ can use Damaris to​​​‌ perform asynchronous data analysis​ using dedicated resources, and​‌ the ones instrumented with​​ Damaris could have full​​​‌ access to PDI ecosystem.​ In addition, Damaris is​‌ now part of the​​ NumPEx Software Catalog (https://numpex-pc5.gitlabpages.inria.fr/tutorials/projects/catalog/index.html).​​​‌
• URL:
  https://­project.­inria.­fr/­damaris/
• Contact:
  Gabriel​ Antoniu
• Participant:
  8 anonymous​‌ participants
• Partner:
  ENS Rennes​​

7.1.2 E2Clab

• Name:
  Edge-to-Cloud​​​‌ lab
• Keywords:
  Distributed systems,​ Cloud, Reproducibility, Experimentation, Computing​‌ Continuum, Evaluation, Large scale,​​ Provenance
• Scientific Description:

  E2Clab​​​‌ is a framework that​ implements a rigorous methodology​‌ that provides guidelines to​​ move from real-life application​​​‌ workflows to representative settings​ of the physical infrastructure​‌ underlying this application in​​ order to accurately reproduce​​​‌ its relevant behaviors and​ therefore understand and optimize​‌ end-to-end performance.

  E2Clab allows​​ a rigorous analysis of​​​‌ possible application configurations in​ a controlled testbed environment​‌ to understand their behavior​​ and related performance trade-offs.​​​‌ E2Clab can be generalized​ to other applications in​‌ the Edge-to-Cloud Continuum. E2Clab​​ is currently used by​​​‌ the Pl@ntNet team to​ understand and optimize the​‌ performance of the application.​​ It is also used​​ by our partners from​​​‌ Instituto Politécnico Nacional for‌ automatic experiment deployments in‌​‌ the context of the​​ SmartFastData associate team.

  In​​​‌ an effort to enhance‌ the reproducibility capabilities of‌​‌ E2Clab, we extended it​​ to enable efficient provenance​​​‌ date capture across the‌ Edge-to-Cloud Continuum. Specifically, we‌​‌ leverage simplified data models,​​ data compression and grouping,​​​‌ and lightweight transmission protocols‌ to reduce overheads for‌​‌ collecting such data on​​ the IoT/Edge. This integration​​​‌ makes E2Clab a promising‌ platform for the performance‌​‌ optimization of applications through​​ reproducible experiments.

• Functional Description:​​​‌
  E2Clab is a framework‌ that implements a rigorous‌​‌ methodology that provides guidelines​​ to move from real-life​​​‌ application workflows to representative‌ settings of the physical‌​‌ infrastructure underlying this application​​ in order to accurately​​​‌ reproduce its relevant behaviors‌ and therefore understand end-to-end‌​‌ performance. Understanding end-to-end performance​​ means rigorously mapping the​​​‌ scenario characteristics to the‌ experimental environment, identifying and‌​‌ controlling the relevant configuration​​ parameters of applications and​​​‌ system components, and defining‌ the relevant performance metrics.‌​‌
• Release Contributions:

  Changelog: https://gitlab.inria.fr/E2Clab/e2clab/-/blob/master/CHANGELOG.rst?ref_type=heads​​

  Features (release 1.0.0):

  (i)​​​‌ the configuration of the‌ experimental environment, libraries and‌​‌ frameworks, (ii) the mapping​​ between the application parts​​​‌ and machines on the‌ Edge, Fog and Cloud,‌​‌ (iii) the deployment of​​ the application on the​​​‌ infrastructure, (iv) Edge-to-Cloud network‌ emulation, (v) the automated‌​‌ execution and monitoring, (vi)​​ the application optimization, and​​​‌ (vii) the gathering of‌ experiment metrics.

• News of‌​‌ the Year:

  In an​​ effort coordinated within the​​​‌ PEPR Cloud, we have‌ worked towards adapting E2Clab‌​‌ to run experiment leveraging​​ commercial computing resources provided​​​‌ by Scaleway. Enabling users‌ to occasionally leverage resources‌​‌ provided by Scaleway would​​ give them access to​​​‌ state-of-the art GPU nodes‌ and diversity of computing‌​‌ resources.

  Additional contributions include:​​ - Ongoing experiments with​​​‌ the ECLAT laboratory to‌ provide experimental support to‌​‌ their simulation pipeline -​​ Improved software reliability through​​​‌ testing and usability through‌ easy ssh access to‌​‌ deployed resources - Documented​​ use of the software​​​‌ for new use-cases such‌ as Federated Learning.

  Latest‌​‌ release archive: https://gitlab.inria.fr/E2Clab/e2clab/-/releases/v3.6.0

• URL:​​
  https://­e2clab.­gitlabpages.­inria.­fr/­e2clab/
• Publications:
  hal-04208787,​​​‌ hal-04779813, hal-04698619,‌ hal-02916032, hal-03310540,‌​‌ hal-03269852, hal-03332524,​​ hal-03270129, hal-03338520,​​​‌ hal-03324177, hal-03259975,‌ hal-03409405, hal-03510012,‌​‌ hal-04659211
• Contact:
  Gabriel Antoniu​​
• Participant:
  5 anonymous participants​​​‌

7.1.3 Fives

• Name:
  Simulator‌ for Scheduling on Storage‌​‌ System at Scale
• Keywords:​​
  Simulation, HPC, Distributed Storage​​​‌ Systems
• Scientific Description:
  Development‌ of Fives began in‌​‌ 2023, given the limitations​​ of our previous StorAlloc​​​‌ simulator. At the end‌ of 2023, Fives is‌​‌ still in active development,​​ while its design and​​​‌ initial results are being‌ submitted to a conference‌​‌ in the field.
• Functional​​ Description:

  Fives is a​​​‌ storage resource scheduling simulator‌ for supercomputers based on‌​‌ WRENCH and SimGrid, two​​ state-of-the-art simulation frameworks. In​​​‌ particular, Fives can model‌ a parallel file system‌​‌ such as Lustre, a​​ computing partition, and simulate​​​‌ a set of jobs‌ performing I/O on the‌​‌ resulting HPC system.

  Fives​​ is based on several​​​‌ components. Firstly, as part‌ of the development of‌​‌ this simulator, an abstraction​​​‌ called "Compound Storage Service"​ was proposed to represent​‌ a distributed storage system,​​ and integrated into WRENCH.​​​‌ Within Fives, a job​ model was designed to​‌ represent a history of​​ jobs and submit them​​​‌ to the scheduler present​ in WRENCH. Finally, a​‌ model of an existing​​ supercomputer, Theta at Argonne​​​‌ National Laboratory, and a​ reverse-engineered version of its​‌ Lustre file system were​​ developed in our simulator.​​​‌

  Experiments are underway to​ calibrate and validate Fives.​‌

• Publication:
  hal-04784808
• Contact:
  François​​ Tessier

7.1.4 MOSAIC

• Name:​​​‌
  Merging Operations and SegmentAtion​ for I/O Categorization
• Keywords:​‌
  Categorization, HPC, I/O
• Scientific​​ Description:

  MOSAIC is a​​​‌ Python categorizer that takes​ I/O traces as input​‌ and assigns classes to​​ describe the patterns found​​​‌ inside.

  Those classes form​ a general description of​‌ applications' I/O activity, giving​​ information about the temporality​​​‌ of I/O, whether periodic​ operations occur, and an​‌ estimation of the impact​​ on the metadata servers.​​​‌

  One of MOSAIC's building​ blocks is the automatic​‌ detection of recurring operations.​​ This is achieved with​​​‌ a clustering algorithm that​ groups operations sharing the​‌ same characteristics (duration, I/O​​ amount, etc.) into one​​​‌ single recurring operation.

  MOSAIC​ automatically finds the traces​‌ that were generated by​​ the same program to​​​‌ reduce the number of​ files to be processed​‌ and speed up a​​ system-scale categorization.

  MOSAIC works​​​‌ for now with traces​ from the Darshan monitoring​‌ tool but can be​​ easily extended to fit​​​‌ other trace formats.

  MOSAIC​ was used to process​‌ the 2019 traces from​​ the BlueWaters supercomputer trace​​​‌ dataset (National Center for​ Supercomputing Applications - University​‌ of Illinois).

• Functional Description:​​

  MOSAIC is a tool​​​‌ for categorizing HPC application​ storage activity. It processes​‌ traces containing all application​​ storage operations and assigns​​​‌ classes to describe how​ they are performed.

  MOSAIC​‌ can describe when the​​ activity is performed (when​​​‌ the application starts, at​ the end, throughout the​‌ execution, etc.), find if​​ some operations are recurring​​​‌ (e.g., saving data to​ a file every 10​‌ minutes), and estimate the​​ overhead caused by the​​​‌ metadata operations.

  It can​ analyze large datasets of​‌ I/O traces coming from​​ a supercomputer to find​​​‌ the general behavior of​ the applications that were​‌ carried out on the​​ machine.

• News of the​​​‌ Year:
  Support of file​ temperature, better detection of​‌ periodic behavior and improved​​ performance for very large​​​‌ datasets were implemented in​ 2025. An intermediate data​‌ format, based on the​​ so-called Trace Event Format,​​​‌ was developed for MOSAIC​ to convert traces from​‌ I/O monitoring tools (such​​ as Darshan, Recorder, and​​​‌ so on) to a​ common abstraction.
• Publication:
  hal-04808300​‌
• Contact:
  François Tessier

7.1.5​​ FLAdversary

• Name:
  Emulation of​​​‌ Federated Learning Scenarios with​ Adversarial Clients
• Keywords:
  Federated​‌ learning, Emulation, Adversarial attack​​
• Functional Description:

  Federated Learning​​​‌ (FL) is subject to​ diverse threats from the​‌ Edge of the network​​ where local training runs​​​‌ on widely distributed, heterogeneous​ and volatile resources.

  FLAdversary​‌ provides tools to dynamically​​ introduce adversarial attacks into​​​‌ the FL training phase.​ Different (model and data)​‌ poisoning attacks can be​​ introduced at the client​​ level to emulate adversaries​​​‌ in the FL training.‌ Several defensive strategies are‌​‌ provided as baselines.

• Publication:​​
  hal-04208787
• Contact:
  Gabriel Antoniu​​​‌
• Partner:
  DFKI (German Research‌ Center for Artificial Intelligence)‌​‌

7.1.6 FLDrift

• Name:
  Emulation​​ of Federated Learning Scenarios​​​‌ with Client Drift
• Keywords:‌
  Federated learning, Emulation, Heterogeneous‌​‌ Data
• Functional Description:

  When​​ deploying Federated Learning (FL)​​​‌ on the Computing Continuum,‌ devices are subject to‌​‌ high variations in local​​ data distributions. This limits​​​‌ the capacity of the‌ system to generate a‌​‌ single model optimized for​​ the entire federation of​​​‌ devices.

  FLDrift provides support‌ for various Non-IID scenarios‌​‌ (i.e., introducing concept-drift and​​ label-shift between federated peers)​​​‌ for FL experiments. Several‌ personalization/clustering strategies are provided‌​‌ as baselines.

• News of​​ the Year:
  We implemented​​​‌ several baseline clustering strategies‌ improving personalization in Federated‌​‌ Learning to address client​​ drift. FLDrift proposes 4​​​‌ scenarios to evaluate the‌ performance of clustering approaches.‌​‌ Each scenario introduces a​​ different form of concept​​​‌ drift between client local‌ datasets.
• Publication:
  hal-04779813
• Contact:‌​‌
  Gabriel Antoniu
• Partner:
  DFKI​​ (German Research Center for​​​‌ Artificial Intelligence)

7.2.1 I/O Traces‌​‌

For our IPDPS'25 paper​​ 13, we used​​​‌ traces of I/O activity‌ from four different systems‌​‌ to answer a set​​ of questions about temporal​​​‌ I/O behavior. To focus‌ on realistic workloads, we‌​‌ gathered traces from jobs​​ running over a period​​​‌ of time instead of‌ profiling a limited set‌​‌ of selected applications.

Two​​ of these data sets​​​‌ were Darshan traces available‌ online (from the Intrepid‌​‌9 and Blue Waters​​ systems10), while​​​‌ two others were obtained‌ by us, using file‌​‌ system monitoring tools:

• PlaFRIM​​ (BeeGFS): a 192-nodes experimental​​​‌ platform in Bordeaux, monitored‌ during 26 months (2022–2024).‌​‌
• SDumont (Lustre): the largest​​ supercomputer in Latin America,​​​‌ monitored during 12 months‌ (2020).

The collected file‌​‌ system data was correlated​​ with the batch scheduler​​​‌ logs to obtain two‌ time series of I/O‌​‌ bandwidth per job (for​​ "reads" and "writes"), with​​​‌ a value per second‌ for PlaFRIM and a‌​‌ value every 15 seconds​​ for SDumont. The two​​​‌ datasets as well as‌ all code and instructions‌​‌ on how to reproduce​​ our experiments are provided​​​‌ in Zenodo: https://­doi.­org/­10.­5281/­zenodo.­14965920.‌ As explained in the‌​‌ instructions, additional information can​​ be obtained from https://­github.­com/­tuda-parallel/­FTIO/­tree/­main/­artifacts/­ipdps25​​​‌ for FTIO, and https://­zenodo.­org/­records/­13785395‌ for MOSAIC.

8.1.1 On the Reproducibility‌​‌ Challenges of Federated Learning:​​ Investigating the Gap between​​​‌ Simulation, Emulation and Real-World‌ Deployments

Participants: Cédric Prigent‌​‌, Alexandru Costan,​​ Gabriel Antoniu.

• Collaboration.​​​‌
  This work has been‌ carried out in co-operation‌​‌ with Cédric Tedeschi (University​​ of Rennes, MAGELLAN team),​​​‌ Loïc Cudennec (DGA MI)‌ and Kate Keahey (Argonne‌​‌ National Laboratory), in the​​ framework of the STEEL​​​‌ project of the PEPR‌ CLOUD program and of‌​‌ the UNIFY 2 Associate​​ Team with ANL, associated​​​‌ to teh JLESC international‌ laboratory.

Federated Learning (FL)‌​‌ is an emerging paradigm​​ for decentralized training of​​​‌ Machine Learning models. It‌ has been the subject‌​‌ of a large corpus​​​‌ of research due to​ its innovative approach to​‌ handling sensitive data. A​​ common practice in the​​​‌ FL literature is to​ run simulations on a​‌ single compute node to​​ assess the performance of​​​‌ FL algorithms. While simulation​ enables fast prototyping and​‌ validation of algorithmic concepts,​​ it may face limitations​​​‌ in reproducing the real​ system's performance in heterogeneous​‌ environments such as the​​ Computing Continuum, and particularly​​​‌ on resource-constrained Edge devices.​ Conversely, emulation on distributed​‌ testbeds offers more effective​​ means to accurately reproduce​​​‌ the performance of real-world​ devices. However, to the​‌ best of our knowledge,​​ no prior research has​​​‌ investigated the differences between​ simulation and emulation in​‌ FL experiments. In this​​ work, we study the​​​‌ complementarity of these approaches​ and discuss their respective​‌ challenges, as a first​​ step towards reproducibility of​​​‌ FL experiments. We illustrate​ our study with a​‌ real-life application used as​​ a baseline: an outdoor​​​‌ air quality forecasting framework​ with real-world sensors. Our​‌ results show that simulation​​ can be used to​​​‌ accurately reproduce model performance​ metrics, while emulation can​‌ effectively reproduce the system​​ performance of real-world experiments.​​​‌ Finally, we present a​ set of lessons learned​‌ on the challenges of​​ FL reproducibility and the​​​‌ selection of experimental infrastructures​ for FL experiments and​‌ applications. This work has​​ been published as 16​​​‌.

8.1.2 Evaluating Federated​ Learning Workflows Beyond Simulation:​‌ A Deployment-Aware Methodology

Participants:​​ Mathis Valli, Alexandru​​​‌ Costan, Gabriel Antoniu​.

• Collaboration.
  This work​‌ has been carried out​​ in co-operation with Cédric​​​‌ Tedeschi (University of Rennes,​ MAGELLAN team) and Loïc​‌ Cudennec (DGA MI).

Federated​​ Learning (FL) is often​​​‌ evaluated in simulation, which​ overlooks network variability, system​‌ heterogeneity, and energy costs​​ in geo-distributed settings. We​​​‌ propose a deployment-aware methodology​ that triangulates analytical modeling​‌, simulation, and​​ real-world deployments within a​​​‌ unified FL evaluation framework.​ For a given series​‌ of experimental scenarios, the​​ methodology allows to assess​​​‌ the consistency of performance​ trends across the three​‌ evaluation approaches, quantifying deviations​​ in key metrics such​​​‌ as run time, communication​ overhead, and energy consumption.​‌ This further enables cross-validation​​ of the reliability of​​​‌ multiple measurement tools, highlighting​ discrepancies in commonly reported​‌ metrics such as the​​ energy usage.

The methodology​​​‌ is validated on FL​ workloads by comparing analytical​‌ predictions and simulations against​​ large-scale deployments on the​​​‌ Grid’5000 testbed, spanning 51​ nodes across four geographically​‌ distant sites. By varying​​ key FL components such​​​‌ as aggregation algorithms, client​ sampling rates, and datasets,​‌ we characterize how different​​ FL design choices affect​​​‌ the reliability of the​ three evaluation approaches. Our​‌ findings reveal significant divergences:​​ analytical models accurately capture​​​‌ communication patterns and preserve​ the relative performance of​‌ the scenarios, simulations reflect​​ broad trends but often​​​‌ lead to performance rankings​ of different configurations inconsistent​‌ with those found through​​ actual deployment, while only​​​‌ the latter uncovers hidden​ costs, such as increased​‌ energy consumption due to​​ data imbalances.

This work​​​‌ has been submitted for​ publication to a conference​‌ (currently under evaluation).

8.1.3​​ Supporting SKA data processing​​ workflows with the E2CLab​​​‌ approach to workflow lifecycle‌ management across the continuum‌​‌

Participants: Thomas Badts,​​ Gabriel Antoniu.

• Collaboration.​​​‌
  This work has been‌ carried out in co-operation‌​‌ with Baptiste Besnard and​​ Damien Gratadour (LIRA, Observatoire​​​‌ de Paris)

Tu support‌ automatic deployment, the complete‌​‌ analysis cycle and the​​ optimization of applications on​​​‌ the Computing, we have‌ proposed the E2Clab methodology‌​‌ and its supporting software​​ tool for workflow lifecycle​​​‌ management across the Continuum.‌ We aim to assist‌​‌ the execution of the​​ Karabo pipeline 32 for​​​‌ radioastronomy simulation by enabling‌ reproducible distributed deployments and‌​‌ experiments on academic testbeds.​​ The Karabo pipeline is​​​‌ being developped to support‌ simulation of the future‌​‌ SKA radiotelescope within the​​ ECLAT laboratory. E2Clab also​​​‌ provides the workflow capabilities‌ to run optimization loops‌​‌ over end-to-end experiments and​​ improve parameter discovery and​​​‌ fine-tuning in a complex,‌ cross-disciplinary, stack of software‌​‌ components ranging from distributed​​ computing frameworks to astrophysics​​​‌ simulations.

This collaboration started‌ in 2005 is still‌​‌ in the exploratory stages,​​ further work is expected​​​‌ in the following year.‌

8.1.4 Methodology for Automated‌​‌ IoT Experimentation in Controlled​​ Testbeds Prior to Real-World​​​‌ Deployments

Participants: Elias Del‌ Pozo Punal, Silvina‌​‌ Caino Lores, Thomas​​ Badts, Gabriel Antoniu​​​‌.

• Collaboration.
  This work‌ has been carried out‌​‌ in co-operation with Felix​​ Garcia-Carballeira and Alejandro Calderon​​​‌ from University Carlos III‌ of Madrid.

Several tools‌​‌ and frameworks have been​​ proposed to automate deployments​​​‌ in distributed systems. Infrastructure-as-Code‌ (IaC) approaches such as‌​‌ Ansible, Puppet, Salstack, or​​ Chef are widely used​​​‌ to abstract low-level configuration‌ details. In parallel, some‌​‌ frameworks support experiment description​​ and execution in specific​​​‌ research testbeds, such as‌ the cOntrol and Management‌​‌ Framework (OMF) and the​​ OMF Measurement Library (OML)​​​‌ 27. Despite these‌ advances, existing solutions often‌​‌ remain limited to specific​​ domains or infrastructures, and​​​‌ integrating heterogeneous environments remains‌ a challenge when considering‌​‌ the broader computing continuum,​​ and in particular for​​​‌ IoT deployments. As a‌ result, researchers frequently rely‌​‌ on fragmented tools or​​ manual procedures, which hinder​​​‌ the repeatability and scalability‌ of experiments and ultimately‌​‌ limit the ability to​​ perform consistent pre-deployment validation​​​‌ of IoT systems in‌ controlled environments.

This work‌​‌ proposes a general methodology​​ for automated IoT experimentation​​​‌ and validation in controlled‌ environments. The approach provides‌​‌ a structured workflow for​​ designing, deploying, and executing​​​‌ experiments across different testbeds‌ in a reproducible, scalable‌​‌ manner. It allows researchers​​ to evaluate IoT deployments​​​‌ through controlled simulations and‌ pre-deployment testing, bridging the‌​‌ gap between conceptual design​​ and real-world implementation.

8.2.1‌​‌ Multi-level analysis of the​​ I/O pattern of HPC​​​‌ applications

Participants: François Tessier‌, Théo Jolivel,‌​‌ Jakob Luettgau, Julien​​ Monniot, Gabriel Antoniu​​​‌.

• Collaboration.
  This work‌ has been carried out‌​‌ in close co-operation with​​ Philippe Deniel from CEA,​​​‌ the Inria TADaaM team‌ in Bordeaux within the‌​‌ Exa-DoST project of the​​ PEPR NumPEx program. It​​​‌ also involves a collaboration‌ with Ahmad Tarraf from‌​‌ the Technical University of​​​‌ Darmstadt, Germany.

While the​ ratio of I/O performance​‌ to computing power has​​ declined by a factor​​​‌ of 10 in the​ last decade 11,​‌ the volume of data​​ generated by scientific workflows​​​‌ and applications has significantly​ grown. In some supercomputing​‌ centers for instance, this​​ volume has increased almost​​​‌ 40-fold in ten years.​ This has made access​‌ to storage resources a​​ major bottleneck to scaling​​​‌ up applications.

Several levers​ exist along the data​‌ path to mitigate this​​ burden. For example, optimizations​​​‌ can be applied at​ the I/O library level​‌ or within the application​​ source code to improve​​​‌ I/O performance. At the​ job scheduler level, decisions​‌ can be taken when​​ allocating resources to avoid​​​‌ I/O interference between jobs.​ However, all these optimizations​‌ require a good upstream​​ understanding of application I/O​​​‌ behavior.

In this research​ axis, we are working​‌ on analyzing the I/O​​ behavior of large-scale applications​​​‌ at various levels. The​ thesis that Théo Jolivel​‌ started in October 2024​​ proposes to tackle this​​​‌ question. One approach is​ to exploit public datasets​‌ containing several years of​​ I/O execution traces of​​​‌ applications running on supercomputers.​ We developed multiple methodologies​‌ and tools to pre-process​​ those datasets, extract the​​​‌ relevant data, and analyse​ the data access behavior.​‌ In particular, we extended​​ MOSAIC 79, a​​​‌ categorizer that detects I/O​ patterns from execution traces.​‌ MOSAIC extracts I/O operations​​ contained in I/O traces​​​‌ and assigns classes to​ describe how I/O operations​‌ are performed throughout the​​ execution. The description is​​​‌ based on three distinct​ axes: I/O temporality (when​‌ was data read or​​ written?), access periodicity (are​​​‌ there recurring operations?), and​ metadata overhead (what is​‌ the impact of metadata​​ operations?). This extended version​​​‌ is under submission in​ a conference 21 and​‌ has been presented as​​ a poster during the​​​‌ annual meeting of the​ ExaDoST project 22 (an​‌ updated version of this​​ poster is also under​​​‌ submission for the PASC​ 2026 conference). A complementary​‌ work on the temporal​​ I/O behavior of HPC​​​‌ applications, in collaboration with​ Inria Bordeaux and TU​‌ Darmsdadt, has been presented​​ at IPDPS'2025, an A-rank​​​‌ conference in the field​ 19.

8.2.2 Study​‌ of I/O interference between​​ jobs

Participants: François Tessier​​​‌, Méline Trochon.​

• Collaboration.
  This work has​‌ been carried out in​​ close co-operation with the​​​‌ Inria TADaaM team in​ Bordeaux and Jean-Thomas Acquaviva,​‌ from DDN within the​​ Exa-DoST project of the​​​‌ PEPR NumPEx program.

High-performance​ computing is a key​‌ component for accelerating scientific​​ discovery and innovation by​​​‌ enabling rapid processing of​ complex simulations and large-scale​‌ data analyses. As HPC​​ applications grow in scale,​​​‌ the performance of the​ underlying storage infrastructure, particularly​‌ parallel file systems (PFS),​​ becomes critical. These shared​​​‌ systems distribute data across​ multiple storage targets (OST),​‌ but concurrent access by​​ multiple jobs can lead​​​‌ to interference, reducing performance​ compared to isolated operations.​‌ Interference varies depending on​​ application characteristics, often degrading​​​‌ overall bandwidth and causing​ significant performance variability, sometimes​‌ by orders of magnitude.​​

In the context of​​ Méline Trochon's PhD thesis​​​‌ (CIFRE DDN-Inria) we studied‌ how interference impacts checkpointing,‌​‌ a key fault-tolerance technique​​ in HPC. Checkpointing involves​​​‌ periodically saving application data‌ to persistent storage to‌​‌ recover from failures. As​​ applications handle more data,​​​‌ checkpoint files grow larger,‌ making I/O performance even‌​‌ more crucial. Interference during​​ these operations can severely​​​‌ affect their efficiency, highlighting‌ the need to understand‌​‌ and mitigate its effects.​​

To do this, we​​​‌ launched a large number‌ of experiments with an‌​‌ application that simulates checkpoint​​ phases and one or​​​‌ more applications that simulate‌ interference. Since the checkpoint‌​‌ application has fixed parameters,​​ we looked at how​​​‌ different configurations of interference‌ workloads may or may‌​‌ not affect I/O performance​​ and to what extent.​​​‌ This work is currently‌ being finalized and a‌​‌ paper is expected to​​ be published in 2026.​​​‌ A pre-print is already‌ available online 23.‌​‌ This work will continue​​ in 2026, notably through​​​‌ the development of a‌ simulator that will allow‌​‌ us to test more​​ configurations.

8.2.3 Enabling Efficient​​​‌ Runtime Data Analysis to‌ a Crystal Deformation Simulation‌​‌

Participants: Arthur Jaquard,​​ Silvina Caino Lores,​​​‌ Gabriel Antoniu.

• Collaboration.‌
  This work has been‌​‌ carried out in close​​ co-operation with Laurent Colombet​​​‌ (from CEA DAM) and‌ Julien Bigot (CEA/Maison de‌​‌ la Simulation) within the​​ Exa-DoST project of the​​​‌ PEPR NumPEx program.

Exascale‌ simulations generate massive data‌​‌ volumes that strain I/O​​ and post-hoc analysis. In​​​‌ the framework of Arthur‌ Jaquard's PhD thesis we‌​‌ explore how in-situ analysis​​ as supported by the​​​‌ Damaris in situ middleware‌ can benefit to Coddex,‌​‌ a crystal deformation code,​​ to offload data movement​​​‌ and analysis to dedicated‌ processes. This is achieved‌​‌ by enabling runtime extraction​​ of key diagnostics without​​​‌ writing intermediate files. We‌ evaluated tin hysteresis cases‌​‌ on CEA's INTI cluster​​ (with 14 nodes, 1,728​​​‌ cores) and compare against‌ a ParaView-based post-hoc pipeline.‌​‌ In situ analysis eliminates​​ per-iteration I/O stalls and​​​‌ reduces output time by‌ up to 5x while‌​‌ preserving overall iteration time,​​ with benefits increasing with​​​‌ the number of tracked‌ variables. This work is‌​‌ conducted within the Exa-DoST​​ project of the PEPR​​​‌ NumPEx program, which aims‌ to build the software‌​‌ infrastructure for the first​​ Exascale machine expected to​​​‌ be set up in‌ France (Alice Recoque, Jules‌​‌ Verne project). It has​​ been published as a​​​‌ poster at the SC25‌ conference 24.

8.3.1​​​‌ Result-Scalability: Following the Evolution‌ of Selected Social Impact‌​‌ of HPC.

Participants: Guillaume​​ Pallez.

• Collaboration.
  This​​​‌ work has been carried‌ out in collaboration with‌​‌ Sally Rose Ellingson from​​ the medical college of​​​‌ the University of Kentucky.‌

While the scientific community‌​‌ traditionally relies on various​​ computational metrics to assess​​​‌ the performance of HPC‌ systems –such as the‌​‌ TOP500 list (based on​​ HPL performance), HPCG, Graph500,​​​‌ IO500– these metrics do‌ not capture how HPC‌​‌ contributes to social progress.​​ We propose 11 a​​​‌ novel approach to follow‌ how the growth of‌​‌ HPC systems and the​​​‌ advances of HPC research​ address concrete social challenges.​‌ The uniqueness of these​​ new metrics lies in​​​‌ their ability to not​ only measure the capabilities​‌ of HPC architectures but​​ also to gauge the​​​‌ concrete social advancements achieved​ through their use: it​‌ focuses on the output​​ of the computation instead​​​‌ of its input. Contrarily​ to current measure, it​‌ also promotes the diversity​​ of machines by evaluating​​​‌ the Pareto front created​ between size and result.​‌ We emphasize the need​​ for dynamic, community-driven metrics​​​‌ that can evolve with​ emerging social needs.

8.3.2​‌ Increasing the Lifetime of​​ HPC Machines: Issues, Implications,​​​‌ and Open Challenges

Participants:​ Guillaume Pallez, Robin​‌ Boezennec.

• Collaboration.
  This​​ work has been carried​​​‌ out as a large​ collaboration in Rennes including​‌ two different teams: PACAP​​ and TARAN, as well​​​‌ as with Brice Goglin​ (TADAAM in Bordeaux)

Extending​‌ the lifetime of High-Performance​​ Computing (HPC) machines is​​​‌ becoming an important concern​ for a variety of​‌ reasons. These include the​​ environmental and human costs​​​‌ associated with chip manufacturing,​ the rising demands by​‌ AI workloads, the soaring​​ prices of accelerator chips,​​​‌ political blocks, and delays​ in the delivery of​‌ next-generation supercomputers. As a​​ community, we must reconsider​​​‌ the traditional HPC paradigm​ and explore new strategies​‌ for making existing HPC​​ infrastructure viable for longer​​​‌ periods. In 18,​ we highlight the current​‌ barriers in prolonging HPC​​ machines lifespan and discuss​​​‌ key technical and operational​ challenges towards this goal.​‌

8.3.3 Improving Supercomputer Usage​​ with Aging Awareness.

Participants:​​​‌ Guillaume Pallez, Robin​ Boezennec, Alix Tremodeux​‌.

Lifetime of electronic​​ devices has a critical​​​‌ impact on their environmental​ footprint. In addition, the​‌ high-demand by AI companies​​ of GPU has reduced​​​‌ tremendously their availability for​ supercomputing centers. Consequently, improving​‌ the duration of CPUs​​ and GPUs is becoming​​​‌ a major issue in​ High Performance Computing (HPC)​‌ domain. This contribution 12​​ investigates the optimization of​​​‌ a machine usage before​ a fatal failure and​‌ the trade-offs with performance.​​ The lifetime of computing​​​‌ devices is strongly connected​ with the temperature and​‌ thus with the running​​ frequency. We investigate the​​​‌ node frequency reconfiguration to​ optimize HPC usage. We​‌ estimate the benefit of​​ a dedicated scheduling algorithm​​​‌ compared with a constant​ frequency.

We show that​‌ a correct decision can​​ increase considerably the number​​​‌ of FLOP of a​ machine with a trade-off​‌ in terms of performance.​​ Because aging models are​​​‌ currently inaccurate, we consider​ different models and discuss​‌ the robustness of our​​ algorithms to inaccuracy

8.3.4​​​‌ Priority-BF: a Task Manager​ for Priority-Based Scheduling

Participants:​‌ Guillaume Pallez.

• Collaboration.​​
  This work has been​​​‌ carried out with Ana​ Gainaru and Scott Patkin​‌ (Oak Ridge National Laboratory).​​

The increasing demand for​​​‌ computational resources, particularly in​ High-Performance Computing environments, necessitates​‌ to rethink how we​​ handle job scheduling strategies.​​​‌ In 14, we​ address the challenge of​‌ managing concurrent jobs with​​ differing priorities on overloaded​​​‌ parallel systems, where strict​ QoS constraints are often​‌ difficult for users to​​ define. Our solution relies​​ on a qualitative description​​​‌ of priorities and pulls‌ from two key approaches:‌​‌ the Easy-BF algorithm and​​ the Conservative Backfilling algorithms.​​​‌ This solution improves the‌ response time for high-priority‌​‌ jobs by 50% without​​ affecting the overall system​​​‌ utilization. We show its‌ applicability in several critical‌​‌ scenarios such as High-Performance​​ Computing (HPC) resource management​​​‌ and in-situ computing.

8.3.5‌ Scheduling multiple task-based applications‌​‌ on distributed heterogeneous computing​​ nodes

Participants: Etienne Ndamlabin​​​‌.

• Collaboration.
  This work‌ has been carried out‌​‌ with Bérenger Bramas (CAMUS​​ team, Inria Nancy).

Modern​​​‌ high-performance computing platforms combine‌ extreme parallelism with growing‌​‌ size, complexity, and cost,​​ making inefficient resource usage​​​‌ increasingly critical in terms‌ of performance and energy.‌​‌ Our research addresses this​​ challenge by focusing on​​​‌ the concurrent execution of‌ multiple task-based applications on‌​‌ shared heterogeneous (CPU/GPU) environments.​​ We created load-balancing heuristics​​​‌ to distribute the task‌ graphs over the processing‌​‌ units and designed and​​ implemented RSCHED, an adaptive​​​‌ scheduling framework integrated into‌ the StarPU runtime system‌​‌ 93. RSCHED dynamically​​ reorganizes resource allocation in​​​‌ response to application progress‌ and completion, while jointly‌​‌ optimizing application makespan and​​ resource utilization during concurrent​​​‌ execution. Experimental results real‌ applications show up ,to‌​‌ a 10× reduction in​​ overall makespan compared to​​​‌ consecutive execution, while increasing‌ resource utilization. RSCHED also‌​‌ highlights the benefits of​​ system-level coordination on top​​​‌ of independent application schedulers,‌ compared to unsupervised concurrent‌​‌ execution.

8.4.1 Implementing a​​ Reproducibility Initiative in HPC:​​​‌ Experiences from SC24.

Participants:‌ Guillaume Pallez.

• Collaboration.‌​‌
  This work has been​​ carried out with Sascha​​​‌ Hunold (University of Vienna)‌ and Judith Hill (Lawrence‌​‌ Livermore National Laboratory).

Reproducibility​​ is fundamental to scientific​​​‌ research, but can be‌ particularly challenging in research‌​‌ that involves High Performance​​ Computing (HPC) due to​​​‌ the unique characteristics of‌ supercomputers. Performance-based metrics such‌​‌ as execution time, energy​​ consumption, and throughput further​​​‌ complicate reproducibility, especially on‌ shared systems. In 15‌​‌, we present our​​ experience implementing a reproducibility​​​‌ initiative at SC24, with‌ particular emphasis on changes‌​‌ made compared to prior​​ SC conferences. We outline​​​‌ HPC-specific challenges, describe the‌ measures adopted to address‌​‌ them, and reflect on​​ the limitations of reproducibility​​​‌ badges. Faced with the‌ constraints of the existing‌​‌ badging nomenclature, we discuss​​ our implementation of a​​​‌ reproducibility report, which aims‌ to provide more context‌​‌ about the reproducibility of​​ each paper. We conclude​​​‌ by recommending that the‌ “Artifact Replicable” badge be‌​‌ dropped by HPC conferences​​ at this time, and​​​‌ discuss alternate ways of‌ ensuring replicability evaluation.

8.4.2‌​‌ An In-depth Study of​​ LLM Contributions to the​​​‌ Bin Packing Problem

Participants:‌ Guillaume Pallez.

• Collaboration.‌​‌
  This work has been​​​‌ carried out with Julien​ Herrmann (CNRS).

Recent studies​‌ have suggested that Large​​ Language Models (LLMs) could​​​‌ provide interesting ideas contributing​ to mathematical discovery. This​‌ claim was motivated by​​ reports that LLM-based genetic​​​‌ algorithms produced heuristics offering​ new insights into the​‌ online bin packing problem​​ under uniform and Weibull​​​‌ distributions. In 20,​ we reassess this claim​‌ through a detailed analysis​​ of the heuristics produced​​​‌ by LLMs, examining both​ their behavior and interpretability.​‌ Despite being human-readable, these​​ heuristics remain largely opaque​​​‌ even to domain experts.​ Building on this analysis,​‌ we propose a new​​ class of algorithms tailored​​​‌ to these specific bin​ packing instances. The derived​‌ algorithms are significantly simpler,​​ more efficient, more interpretable,​​​‌ and more generalizable, suggesting​ that the considered instances​‌ are themselves relatively simple.​​ We then discuss the​​​‌ limitations of the claim​ regarding LLMs' contribution to​‌ this problem, which appears​​ to rest on the​​​‌ mistaken assumption that the​ instances had previously been​‌ studied. Our findings instead​​ emphasize the need for​​​‌ rigorous validation and contextualization​ when assessing the scientific​‌ value of LLM-generated outputs.​​

9.1.1​ Associate Teams in the​‌ framework of an Inria​​ International Lab or in​​​‌ the framework of an​ Inria International Program

9.2.1 Visits of​‌ international scientists

9.2.2 Visits​​ to international teams

9.3.1 H2020‌ projects

9.3.2​‌ Collaborations with Major European​​ Organizations

Participants: Gabriel Antoniu​​​‌, Alexandru Costan,​ Jakob Luettgau.

ETP4HPC:​‌ Since 2019, Gabriel Antoniu​​ has served as a​​​‌ co-leader of the working​ group on Programming Environments,​‌ contributing to two successive​​ versions of the Strategic​​​‌ Research Agenda of ETP4HPC.​ Alexandru Costan served as​‌ a member of this​​ working group. Jakob Luettgau​​​‌ served as a member​ of the working group​‌ on Data Storage and​​ I/O. A white paper​​​‌ of this group 25​ was published in 2025.​‌

Exa-DoST​​

Participants: Gabriel Antoniu,​​​‌ François Tessier, Julien​ Monniot, Jakob Luetgau​‌, Etienne Ndamlabin,​​ Silvina Caino Lores,​​​‌ Guilaume Pallez.

Exa-DoST​ project of the NumPEx​‌ PEPR program

• Title:
  Data-oriented​​ Software and Tools for​​​‌ the Exascale
• Duration:
  From​ January 1, 2023 to​‌ April 1, 2030
• Partners:​​
  • Inria
  • CEA
  • CNRS
  • University​​​‌ of Bordeaux
  • Observatoire de​ Paris
  • Observatoire de la​‌ Côte d'Azure
  • Data Direct​​ Networks France (DDN)
• Coordinator:​​​‌
  Gabriel Antoniu (KerData Team,​ Inria)
• Summary:

  The advent​‌ of future Exascale supercomputers​​ raises multiple data-related challenges.​​​‌ To enable applications to​ fully leverage the upcoming​‌ infrastructures, a major challenge​​ concerns the scalability of​​​‌ techniques used for data​ storage, transfer, processing and​‌ analytics. Additional key challenges​​ emerge from the need​​​‌ to adequately exploit emerging​ technologies for storage and​‌ processing, leading to new,​​ more complex storage hierarchies.​​​‌ Finally, it now becomes​ necessary to support more​‌ and more complex hybrid​​ workflows involving at the​​​‌ same time simulation, analytics​ and learning, running at​‌ extreme scales across supercomputers​​ interconnected to clouds and​​​‌ edgebased systems. The Exa-DoST​ project will address most​‌ of these challenges, organized​​ in 3 areas:

  • Scalable​​​‌ storage and I/O;
  • Scalable​ in situ processing;
  • Scalable​‌ smart analytics.

  As part​​ of the NumPEx program,​​​‌ Exa-DoST targets a much​ higher technology readiness level​‌ than previous national projects​​ concerning the HPC software​​​‌ stack. It will address​ the major data challenges​‌ by proposing operational solutions​​ co-designed and validated in​​​‌ French and European applications.​ This will allow filling​‌ the gap left by​​ previous international projects to​​​‌ ensure that French and​ European needs are taken​‌ into account in the​​ roadmaps for building the​​​‌ data-oriented Exascale software stack.​

STEEL

Participants: Gabriel Antoniu​‌, Alexandru Costan,​​ Jakob Luettgau, François​​​‌ Tessier, Mathis Valli​, Thomas Badts.​‌

• Title:
  Secure and efficient​​ daTa storagE and procEssing​​​‌ on cLoud-based infrastructures
• Duration:​
  From June 1, 2023​‌ to 31 August 2030​​
• Partners:
  • Inria
  • CNRS
  • Institut​​​‌ Mines Télécom (IMT)
  • University​ of Bordeaux
  • University of​‌ Rennes
  • INSA Rennes
  • INSA​​ Lyon
• Coordinator:
  Gabriel Antoniu​​​‌ (KerData Team, Inria)
• Summary:​
  The strong development of​‌ cloud computing since its​​ emergence in 2007 and​​ its massive adoption for​​​‌ the storage of unprecedented‌ volumes of data in‌​‌ a growing number of​​ domains has brought to​​​‌ light major technological challenges.‌ In this project we‌​‌ will address several of​​ these challenges, organized in​​​‌ three research directions. The‌ first direction concerns the‌​‌ exploitation of emerging technologies​​ for efficient storage on​​​‌ cloud infrastructures. We will‌ address this challenge through‌​‌ NVRAM-based distributed performance storage​​ solutions, as close as​​​‌ possible to data production‌ and consumption locations (disaggregation‌​‌ principle) and develop strategies​​ to optimize the trade-off​​​‌ between data consistency and‌ access performance. The second‌​‌ direction concerns the efficient​​ storage and processing of​​​‌ data on hybrid, heterogeneous‌ infrastructures within the digital‌​‌ edge-cloud-supercomputer continuum. In many​​ domains (autonomous cars, predictive​​​‌ maintenance, intelligent buildings, etc.)‌ we are witnessing the‌​‌ emergence of hybrid workflows​​ combining simulations, analysis of​​​‌ sensor data flows and‌ machine learning. Their execution‌​‌ requires storage resources ranging​​ from the edge to​​​‌ cloud infrastructures, and even‌ to supercomputers, which poses‌​‌ challenges for unified data​​ storage and processing. The​​​‌ third research direction is‌ dedicated to confidential storage,‌​‌ in connection with the​​ need to store and​​​‌ analyze large volumes of‌ data of strategic interest‌​‌ or of a personal​​ nature. For all of​​​‌ these directions, the project‌ will take into account‌​‌ the need to propose​​ and validate interoperable approaches​​​‌ with a potential for‌ transfer to major French‌​‌ or European industrial players​​ in cloud computing.

ECLAT​​​‌

Participants: François Tessier,‌ Gabriel Antoniu, Théo‌​‌ Jolivel, Jakob Luettgau​​, Thomas Badts.​​​‌

• Title:
  Extreme Computing Laboratory‌ for Astronomical Telescopes
• Duration:‌​‌
  Since May, 2024
• Partners:​​
  • Inria
  • CNRS
  • Université de​​​‌ Rennes
  • Eviden
  • Observatoire de‌ la Côte d'Azur
  • Observatoire‌​‌ de Paris
  • Université Paris-Saclay​​
  • Centrale Supelec
• Coordinator:
  Gabriel​​​‌ Antoniu (KerData Team, Inria)‌
• Summary:
  ECLAT is positioned‌​‌ as a center of​​ excellence dedicated to High-Performance​​​‌ Computing (HPC) and Artificial‌ Intelligence (AI) technologies and‌​‌ techniques applied to astronomical​​ instrumentation. This project brings​​​‌ together sixteen partner laboratories‌ and teams around a‌​‌ common roadmap, aimed at​​ strengthening research and development​​​‌ (R&D) collaborations. The aim‌ is to design and‌​‌ build future cyber-physical systems​​ for astronomy, capable of​​​‌ managing, processing and optimizing‌ gigantic volumes of data.‌​‌

Grid'5000

We are members​​ of Grid'5000 community and​​​‌ run experiments on the‌ Grid'5000 platform on a‌​‌ daily basis.

Inria Exploratory​​ program: Repas

Participants: Guillaume​​​‌ Pallez.

• Project Acronym:‌
  REPAS
• Title:
  New Portrayal‌​‌ of HPC Applications
• Coordinator:​​
  Guillaume Pallez
• Collaboration:
  This​​​‌ is done in collaboration‌ with the team DATAMOVE‌​‌ (Inria Grenoble)
• Duration:
  2022-2025​​
• Summary:
  What is the​​​‌ right way to represent‌ an application in order‌​‌ to run it on​​ a highly parallel (typically​​​‌ exascale) machine? The idea‌ of project is to‌​‌ completely review the models​​ used in the development​​​‌ scheduling algorithms and software‌ solutions to take into‌​‌ account the real needs​​ of new users of​​​‌ HPC platforms.

10.1.1 Scientific events: organisation​​

10.1.2 Scientific events:​‌ selection

10.1.3 Journal‌​‌

10.1.4 Invited talks

• Guillaume‌​‌ Pallez :
  • « Vers​​ un calcul intensif plus​​​‌ sobre », organisé par‌ Laboratoire 1.5
  • “Model (co)-Design‌​‌ and Accuracy for Resource​​ Management in HPC” at​​​‌ Co-Design workshop (Osaka, Jn)‌ co-organized by Jack Dongarra‌​‌ and the Chinese Academy​​ of Science

• François Tessier​​​‌ :
  • "The Difficult Task‌ of Understanding I/O Behavior‌​‌ on Large-scale Systems", Keynote​​ talk at the 3rd​​​‌ NHR Conference, Germany‌

10.1.5 Leadership within the‌​‌ scientific community

• Gabriel Antoniu​​ :
  • Large National project​​​‌ management: Coordinator of ExaDoST,‌ one of the 5‌​‌ targeted projects of the​​ NumPEx PEPR project (started​​​‌ in 2023, budget: 6.2‌ M€). Coordinator of STEEL,‌​‌ one of the 7​​ high-priority projects of the​​​‌ CLOUD PEPR project (started‌ in 2023, budget: 2.8‌​‌ M€).
  • ETP4HPC: Since 2019,​​ co-leader of the working​​​‌ group on Programming Environments,‌ lead co-author of the‌​‌ corresponding chapter of the​​ Strategic Research Agenda of​​​‌ ETP4HPC.
  • International lab management:‌ Executive Director of JLESC‌​‌ for Inria since April​​ 2024 (previously Vice Executive​​​‌ Director). JLESC is the‌ Joint Inria-Illinois-ANL-BSC-JSC-RIKEN/AICS Laboratory for‌​‌ Extreme-Scale Computing. Within JLESC,​​ he also serves as​​​‌ a Topic Leader for‌ Data storage, I/O and‌​‌ in situ processing for​​ Inria.
  • International Working Group​​​‌ management: Co-Leader of the‌ Working group on Data‌​‌ management and Computing Continuum​​ within the InPEX International​​​‌ Post-Exascale Project.
  • Team‌ management: Head of the‌​‌ KerData Project-Team (INRIA-INSA Rennes).​​
  • International Associate Team management:​​​‌ Leader of the UNIFY2‌ Associate Team with Argonne‌​‌ National Lab (2013–2025).
• François​​ Tessier :
  • Work package​​​‌ co-leader with Francieli Boito‌ (Associate Professor, University of‌​‌ Bordeaux) within the NumPEX​​ ExaDoST project.
  • Leader for​​​‌ KerData in the ECLAT‌ joint laboratory.
• Alexandru Costan‌​‌ :
  • Work package leader​​ of WP2 within the​​​‌ PEPR CLOUD STEEL project.‌

10.1.6 Scientific expertise

• Gabriel‌​‌ Antoniu:
  • Evaluator for a​​ Horizon Europe project (HORIZON-CL4-2021-HUMAN-01​​​‌ call)
• Alexandru Costan:
  • Evaluator‌ for several projects submitted‌​‌ to FFPlus, a European​​ initiative highlighting and promoting​​​‌ the adoption of High-Performance‌ Computing (HPC) by SMEs‌​‌ and start-ups across Europe)​​
  • Member of the jury​​​‌ for GDR RSD Prix‌ de thèse, Prix chercheur‌​‌

10.1.7 Research administration

• François​​ Tessier
  • Member of the​​​‌ Commission on Health, Safety‌ and Working Conditions (now‌​‌ called FSS) within the​​ Inria center of Rennes​​​‌
• Guillaume Pallez:
  • Member of‌ the National Commission on‌​‌ Health, Safety and Working​​ Conditions (now called FS)​​​‌
  • Member of the Scientific‌ Board of Inria
• Gabriel‌​‌ Antoniu:
  • Member of the​​ Inria HRS4R Steering Committee​​​‌ (HRS4R: European Human Resources‌ Strategy for Research)

10.2.1 Teaching‌

• Alexandru Costan
  • Bachelor: Software‌​‌ Engineering and Java Programming,​​ 28 hours (lab sessions),​​​‌ L3, INSA Rennes.
  • Bachelor:‌ Databases, 68 hours (lectures‌​‌ and lab sessions), L2,​​ INSA Rennes.
  • . Bachelor:​​​‌ Practical case studies, 24‌ hours (project), L3, INSA‌​‌ Rennes.
  • Master: Big Data​​ Storage and Processing, 28h​​​‌ hours (lectures, lab sessions),‌ M1, INSA Rennes.
  • Master:‌​‌ Algorithms for Big Data,​​​‌ 28 hours (lectures, lab​ sessions), M2, INSA Rennes.​‌
  • Master: Big Data Project,​​ 28 hours (project), M2,​​​‌ INSA Rennes.
• Gabriel Antoniu:​
  • Master (Engineering Degree, 5th​‌ year): NoSQL and Cloud​​ technologies, 21 hours (lectures),​​​‌ M2 level, ENSAI (​École nationale supérieure de​‌ la statistique et de​​ l'analyse de l'information),​​​‌ Bruz.
  • Master: Infrastructures for​ Big Data, 14 hours​‌ (lectures), M1 level, IBD​​ Module, University of Rennes.​​​‌
  • Master: Cloud Computing and​ Big Data, 14 hours​‌ (lectures), M2 level, Cloud​​ Module, MIAGE Master Program,​​​‌ University of Rennes.
• François​ Tessier
  • Bachelor: Computer science​‌ discovery, 15 hours (lab​​ sessions), L1 level, DIE​​​‌ Module, ISTIC, University of​ Rennes.
  • Master: Cloud Computing​‌ and Big Data, 15​​ hours (lectures), M2 level,​​​‌ Cloud Module, MIAGE Master​ Program, University of Rennes.​‌
  • Master (Engineering Degree, 4th​​ year): Storage on Clouds,​​​‌ 5 hours (lecture and​ lab session), M2 level,​‌ IMT Atlantique, Rennes.
• Jakob​​ Luettgau:
  • Master: Cloud and​​​‌ Network Infrastructures (CNI), 4​ hours (lectures), M2 level,​‌ Master Program, University of​​ Rennes.
• Théo Jolivel:
  • Master:​​​‌ Cloud Computing and Big​ Data, 36 hours (lab​‌ sessions), M2 level, Cloud​​ Module, MIAGE Master Program,​​​‌ University of Rennes.
• Mathis​ Valli:
  • Bachelor: Databases, 12​‌ hours (lab sessions), L3,​​ INSA Rennes.

10.2.2 Supervision​​​‌

• Defended PhD theses:
  • Cédric​ Prigent, "Supporting Online Learning​‌ and Inference in Parallel​​ across the Digital Continuum",​​​‌ thesis started in November​ 2021, co-advised by Alexandru​‌ Costan, Gabriel Antoniu and​​ Loïc Cudennec (DGA). Defended​​​‌ on 25 May 2025.​
  • Robin Boezennec, “Reducing HPC​‌ Resource Consumption”, defended on​​ December 10th, 2025, co-advised​​​‌ by Guillaume Pallez and​ Fanny Dufossé (Datamove, Grenoble).​‌ Defended on 10 December​​ 2025.

• PhD in progress:​​​‌
  • Mathis Valli, "Comparative Analysis​ of Federated Learning: Simulations​‌ Versus Real-World Testbeds in​​ dynamic settings", thesis started​​​‌ in April 2023, co-advised​ by Alexandru Costan, Cédric​‌ Tedeschi (Myriads) and Loïc​​ Cudennec (DGA).
  • Théo Jolivel,​​​‌ "Modeling and Simulation of​ Exascale Storage Systems", thesis​‌ started in October 2024,​​ co-advised by François Tessier,​​​‌ Gabriel Antoniu and Philippe​ Deniel (CEA).
  • Arthur Jaquard,​‌ "Dynamic in situ and​​ in transit data analysis​​​‌ for Exascale Computing using​ Damaris", thesis started in​‌ October 2024, co-advised by​​ Gabriel Antoniu, Laurent Colombet​​​‌ (CEA), Silvina Caino-Lores, and​ Julien Bigot (CEA).
  • Méline​‌ Trochon, "Adaptive Checkpoint-Restart System​​ with Knowledge of the​​​‌ Network Load", CIFRE thesis​ started in February 2025,​‌ located at Inria Bordeaux,​​ co-supervised by Francieli Boito,​​​‌ Brice Goglin (TADaaM -​ Inria Bordeaux), Jean-Thomas Acquaviva​‌ (DDN) and François Tessier.​​
  • Serge Meurrens, "Ordonnancement des​​​‌ E/S adapté aux applications​ dans les systèmes HPC",​‌ thesis started in December​​ 2025, located at Inria​​​‌ Bordeaux, co-supervised by Francieli​ Boito, Luan Teylo (TADaaM​‌ - Inria Bordeaux) and​​ François Tessier.
  • Simon Renard​​​‌ , “Data Interfaces for​ Hybrid Quantum-Classical Computational Workflows”,​‌ thesis started on October​​ 2025, co-supervised by Silvina​​​‌ Caino Lores ,Gabriel​ Antoniu and Marc Baboulin​‌ (Inria Paris-Saclay).
  • Alix Tremodeux​​ , “Etude des conséquences​​​‌ du vieillissement sur les​ machines HPC”, thesis started​‌ on September 2025, co-supervised​​ by Guillaume Pallez and​​​‌ Erven Rohou (PACAP -​ Inria Rennes).

• Internships:
  • Remy​‌ Chiv, "Analyse et optimisation​​ des entrées/sorties d'un pipeline​​ de traitement de données​​​‌ pour la radio-astronomie à‌ grande échelle", 5-month Master‌​‌ 2 internship started in​​ May 2025, supervised by​​​‌ François Tessier.

10.2.3 Juries‌

• Gabriel Antoniu :
  • HDR:‌​‌ Towards Better I/O Resource​​ Usage in HPC,​​​‌ Francieli Zanon Boito, Université‌ de Bordeaux, defended on‌​‌ 5 December 2025.
• Alexandru​​ Costan :
  • PhD: Complexity​​​‌ and Algorithmic results for‌ Translocation Distances, Maria‌​‌ Constantinescu, University of Bucharest,​​ defended on 29 May​​​‌ 2025.

International journals​​

• 11 articleS.Sally​​​‌ Ellingson and G.Guillaume​ Pallez. Result-Scalability: Following​‌ the Evolution of Selected​​ Social Impact of HPC​​​‌.International Journal of​ High Performance Computing Applications​‌395April 2025​​, 713-721HALDOI​​​‌back to textback​ to text

International peer-reviewed​‌ conferences

• 12 inproceedingsR.​​Robin Boëzennec, F.​​​‌Fanny Dufossé, G.​Guillaume Pallez and A.​‌Alix Tremodeux. Improving​​ Supercomputer Usage with Aging​​​‌ Awareness.SC Workshops​ '25: Proceedings of the​‌ SC '25 Workshops of​​ the International Conference for​​​‌ High Performance Computing, Networking,​ Storage and AnalysisSustainable​‌ Supercomputing (Workshop of SC25)​​St. Louis, Missouri, United​​​‌ StatesACMNovember 2025​, 1980-1989HALDOI​‌back to text
• 13​​ inproceedingsF.Francieli Boito​​​‌, L.Luan Teylo​, M.Mihail Popov​‌, T.Théo Jolivel​​, F.François Tessier​​​‌, J.Jakob Luettgau​, J.Julien Monniot​‌, A.Ahmad Tarraf​​, A.André Carneiro​​​‌ and C.Carla Osthoff​. A Deep Look​‌ Into the Temporal I/O​​ Behavior of HPC Applications​​​‌.39th IEEE International​ Parallel & Distributed Processing​‌ Symposium (IPDPS)39th IEEE​​ International Parallel & Distributed​​​‌ Processing Symposium (IPDPS)Milan,​ ItalyJune 2025HAL​‌DOIback to text​​
• 14 inproceedingsA.Ana​​​‌ Gainaru, S.Scott​ Klasky and G.Guillaume​‌ Pallez. Priority-BF: a​​ Task Manager for Priority-Based​​​‌ Scheduling.31st International​ European Conference on Parallel​‌ and Distributed ComputingEURO-PAR​​ 2025 - 31st International​​​‌ European Conference on Parallel​ and Distributed ComputingDresden,​‌ Germany2026, 219-232​​HALDOIback to​​​‌ text
• 15 inproceedingsG.​Guillaume Pallez, J.​‌Judith Hill and S.​​Sascha Hunold. Implementing​​ a Reproducibility Initiative in​​​‌ HPC: Experiences from SC24‌.REP 2025 -‌​‌ 3rd ACM Conference on​​ Reproducibility and ReplicabilityVancouver​​​‌ (British Columbia), CanadaJuly‌ 2025, 1-8HAL‌​‌DOIback to text​​
• 16 inproceedingsC.Cédric​​​‌ Prigent, K.Kate‌ Keahey, A.Alexandru‌​‌ Costan, L.Loïc​​ Cudennec and G.Gabriel​​​‌ Antoniu. On the‌ Reproducibility Challenges of Federated‌​‌ Learning: Investigating the Gap​​ between Simulation, Emulation and​​​‌ Real-World Deployments.CCGrid‌ 2025 - IEEE 25th‌​‌ International Symposium on Cluster,​​ Cloud and Internet Computing​​​‌Tromso, Norway2025,‌ 185-194HALDOIback‌​‌ to text

Doctoral dissertations​​ and habilitation theses

• 17​​​‌ thesisC.Cédric Prigent‌. Towards Efficient and‌​‌ Trustworthy Federated Learning on​​ the Computing Continuum.​​​‌INSA de RennesMay‌ 2025HAL

Reports &‌​‌ preprints

• 18 miscR.​​Robin Boëzennec, F.​​​‌Fernando Fernandes dos Santos‌, B.Brice Goglin‌​‌, A.Angeliki Kritikakou​​, G.Guillaume Pallez​​​‌, E.Erven Rohou‌, O.Olivier Sentieys‌​‌ and M.Marcello Traiola​​. Increasing the Lifetime​​​‌ of HPC Machines: Issues,‌ Implications, and Open Challenges‌​‌.2025HALback​​ to text
• 19 report​​​‌F.Francieli Boito,‌ L.Luan Teylo,‌​‌ M.Mihail Popov,​​ T.Théo Jolivel,​​​‌ F.François Tessier,‌ J.Jakob Luettgau,‌​‌ J.Julien Monniot,​​ A.Ahmad Tarraf,​​​‌ A.André Carneiro and‌ C.Carla Osthoff.‌​‌ A deep look into​​ the temporal I/O behavior​​​‌ of HPC applications -extended‌ version.RR-9577Inria‌​‌ & Labri, Univ. Bordeaux​​March 2025, 1-42​​​‌HALback to text‌
• 20 miscJ.Julien‌​‌ Herrmann and G.Guillaume​​ Pallez. An In-depth​​​‌ Study of LLM Contributions‌ to the Bin Packing‌​‌ Problem.October 2025​​HALback to text​​​‌
• 21 miscT.Théo‌ Jolivel, F.François‌​‌ Tessier, J.Jakob​​ Luettgau, G.Gabriel​​​‌ Antoniu and P.Philippe‌ Deniel. A Methodology‌​‌ for System-Scale I/O Pattern​​ Taxonomy for HPC Workloads​​​‌.2025HALback‌ to text
• 22 misc‌​‌T.Théo Jolivel and​​ F.François Tessier.​​​‌ G.Gabriel Antoniu and‌ P.Philippe Deniel,‌​‌ eds. MOSAIC: Automatic categorization​​ of I/O patterns.​​​‌November 2025HALback‌ to text
• 23 misc‌​‌M.Méline Trochon,​​ J.-T.Jean-Thomas Acquaviva,​​​‌ F.Francieli Boito,‌ B.Brice Goglin,‌​‌ F.François Tessier and​​ L.Luan Teylo.​​​‌ On the Impact of‌ Interference from Concurrent Jobs‌​‌ on Checkpointing Performance.​​2025HALback to​​​‌ text

Other scientific publications‌

• 24 inproceedingsA.Arthur‌​‌ Jaquard. Enabling Efficient​​ Runtime Data Analysis to​​​‌ a Crystal Deformation Simulation‌.SC 2025 -‌​‌ International Conference for High​​ Performance Computing, Networking, Storage,​​​‌ and AnalysisSaint Louis,‌ MO, USA, United States‌​‌2025HALback to​​ text
• 25 miscS.​​​‌Sarah Neuwirth, P.‌Philippe Deniel, J.-T.‌​‌Jean-Thomas Acquaviva, M.​​Martin Golasowski, M.​​​‌Michael Hennecke, A.‌Adrian Jackson, T.‌​‌Thomas Leibovici, J.​​Jakob Luettgau and R.​​​‌Ramon Nou. ETP4HPC‌ SRA 6 White Paper‌​‌ - I/O and Storage​​​‌.January 2025HAL​DOIback to text​‌