Dynamic provisioning of hybrid storage resources.

While for years high-performance computing (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 cloud-type infrastructures. This migration has been mainly motivated by the cloud's ability to perform data analysis tasks efficiently.

From an I/O and storage perspective, this means having to deal with two very different worlds: the world of cloud computing, where direct access to resources is extremely limited due to a very high level of abstraction, and the world of on-premise supercomputers offering a low level approach requiring tight user control. The abstraction layer of clouds also allows storage, network and computing resources to have a certain elasticity and to be exclusively allocated.

In this context, we propose to converge these two worlds by exploring ways to provide storage resources distributed across hybrid HPC/cloud infrastructures to complex scientific workflows combining simulation and data analysis.

To do so, we continue our recently started work on scheduling algorithms dedicated to storage resources and implemented in a storage-aware scheduler developed in the team (simulator and scheduler). We also start a new research line focused on the abstraction of storage resources in order to provide a unified interface allowing to query any type of storage on an hybrid infrastructure.

I/O Orchestration over hybrid infrastructures.

On hybrid infrastructures, in the same way as the amount of generated data increases, the need for persistence has stepped up. A broad variety of large-scale scientific applications and workflows in scientific domains such as materials, high energy physics or engineering have massive I/O needs.

On a HPC system, for instance, it is typically estimated that around 10 % to 20 % of the wall time of this class of applications is spent in I/O. In addition, in the case of workflows running on hybrid infrastructures, these I/O are extremely varied and are no longer restricted to a single system but are spread across complex architectures.

To take advantage of the capability of current systems and hope to leverage future ones, improving the I/O performance is decisive. The complexity of both the federation and the different underlying systems implies having a strong knowledge of the workloads' I/O behavior and adopting a topology-aware approach for data movement orchestration.

We focus our effort on two research lines here. We first model the I/O behavior from the application and workflow's point of view. The parameters influencing I/O performance may be as diverse as the data size, the data model (multidimensional arrays, meshes, etc.), the data layout (array of structures, structure of arrays, etc.) or the access frequency.

The impact of each characteristic on I/O performance will be evaluated with benchmarks and real applications on the different systems of an hybrid infrastructure (HPC, cloud and edge later on) and an I/O workload model will be proposed.

Then, while using this I/O characterization, we focus our effort on data aggregation, taking into account the underlying topology, which consists of selecting a subset of intermediate resources to collect data before moving it from/to the destination (a storage system or a data processing system in case of in-transit workflows, for instance). This technique has several advantages: it increases the I/O bandwidth by reading or writing larger chunks of data, it highly reduces the number of concurrent streams to the destination and it minimizes the network contention.

Supporting repeatable, replicable and reproducible automatic deployments across the continuum.

As communities from an increasing number of scientific domains are leveraging the Computing Continuum, a desired feature of any experimental research is that its scientific claims are verifiable by others in order to build upon them. This can be achieved through repeatability, replicability, and reproducibility (3 Rs).

E2Clab is a first step towards enabling these goals and, to the best of our knowledge, it is the first platform to support the complete analysis cycle of an application on the Computing Continuum. We plan to further consolidate E2Clab in order to make it a promising platform for future performance optimization of applications on the Edge-to-Cloud Continuum through reproducible experiments.

Specifically, we plan to focus on three main directions: (1) develop new, finer grained abstractions to model the components of the entire data processing pipeline across the continuum (from data production to permanent storage) and allow researchers to trade between different costs with increased accuracy; (2) enable built-in support for other large-scale experimental testbeds, besides Grid'5000, such as Vagrant and Chameleon and ultimately provide a community driven tool for large scale experimentation; and (3) develop a benchmark for processing frameworks within the Computing Continuum atop E2Clab.

Many exciting research questions could then be explored leveraging such an enhanced deployment and optimization tool, especially in domains like machine and deep learning: how to improve the convergence speed of distributed algorithms (i.e., gradient descent) to reach good accuracy quickly? how to appropriately partition a model based on the capability of different cloud or edge devices?

Continual learning and inference in parallel across the Computing Continuum.

As neural network architectures and their training data are getting more and more complex, so are the infrastructures that are needed to execute them sufficiently fast. Hyperparameter setting and tuning, training, inference, dataset handling are operations that are all putting a growing pressure on the underlying compute infrastructure and call for novel approaches at all levels of the workflow, including the algorithmic level, the middleware and deployment level, and the resource optimization level.

Our goal is to address the following specific research questions: how can the various possible deployment options of complex AI workflows on the available underlying infrastructure impact performance metrics? how can this infrastructure be best leveraged in practice, potentially through seamless integration of supercomputers, clouds, and fog/edge systems?

We will focus on the middleware and the deployment level. Our objective is to investigate various deployment strategies for complex AI workflows (e.g., potentially combining continual training, simulations and inference, all in parallel and in real-time) on hybrid execution infrastructures (e.g., combining supercomputers and cloud/fog/edge systems).

Efficient federated learning in heterogeneous and volatile environments.

The latest technological advances in hardware accelerators like the GPUs enable the execution of machine and deep learning tasks on large volumes of data in a time that has become reasonable. Embedded systems make it possible to deploy some inference tasks as close as possible to the operational context. One of the major challenges of these heterogeneous distributed systems lies in the ability to have relevant data in a given place and at a given time.

One approach is to rely on the recent privacy-preserving Federated Learning paradigm that leverages the edge devices for training. However, such solutions raise some major challenges related to system and statistical heterogeneity, energy footprint and security.

Our goal is to identify and adapt such emerging approaches resulting from the Computing Continuum in order to respond to the problems of distribution of computations and processing, particularly in the case of workflows involving AI. This exploratory topic has concrete application use-cases such as with the smart autonomous vehicles or military and civilian warning systems.

Towards unified data processing techniques for hybrid simulation/analytics workflows executed across potentially hybrid CPU/GPU infrastructures.

In the high-performance computing area (HPC), the need to get fast and relevant insights from massive amounts of data generated by extreme-scale computations led to the emergence of in situ/in transit processing. In the Big Data area, the search for real-time, fast analysis was materialized through a different approach: stream-based processing. A major challenge is the joint use of these techniques in a unified data processing architecture.

Preliminary work already started within the "frameworks" work package of the HPC-Big Data Inria Challenge. It is also a core direction of our team's involvement in the ACROSS H2020 EuroHPC project. A typical scenario considered in ACROSS consists in executing hybrid workflows combining simulations and (potentially learning-based) analytics running concurrently.

The challenge is to integrate both stream and in-situ/in-transit processing tasks in the targeted workflows, leading to a decrease in execution times for data-intensive/deep learning like HPC simulations and modeling workloads. In particular, we will introduce programmatic support for on-demand data analytics on platforms that were traditionally used only for simulations. This new type of workflow (combining simulations with data analytics) could help anticipate the future behavior of the simulated systems.

Analyzing and exploiting stored data jointly with simulated data can provide a richer tool for much deeper interpretation of the targeted systems, enabling more reliable, transparent and innovative decision making. To this purpose, Damaris will be extended to support asynchronously Big Data analytics plugins, to enable in-situ and in transit analysis of simulation data, then to support hybrid (stream-based and batch-based) in transit data analysis. These new, hybrid workflows will allow on one hand to reduce the simulation time (by pre-analyzing some parts of the results locally, in-situ) and on the other hand to use simulations to train proxy models for optimization.

In the EUPEX EuroHPC Project, one goal is to introduce cross-application optimizations for data-driven stream parallel applications. This will rely on Damaris to orchestrate transfers, by leveraging various storage capabilities to provide scalable asynchronous I/O and non-intrusive in situ and in transit data processing on the data nodes. This provides another motivation to adapt Damaris to support workflows and Big Data analytics plugins by enabling in-situ and in-transit analysis of stream data.

Finally, as a piece of software considered for the Inria Exascale Software Task, in collaboration with the CEA, we plan to investigate new types of scenarios for hybrid CPU/GPU machines, where simulations could trigger on-demand analytics potentially run on GPU hardware.

Social impact.

One of our target applications is Early Earthquake Warning. We proposed a solution that enables earthquakes classification with an outstandingly perfect accuracy. By enabling accurate identification of strong earthquakes, it becomes possible to trigger adequate measures and save lifes. For this reason, our work was distinguished with an Outstanding Paper Award — Special Track for Social Impact at AAAI-20, an A* conference in the area of Artificial Intelligence. This result has been highlighted by the Le Monde journal in its edition of December 28, 2020, in a section entitled: Ces découvertes scientifiques que le Covid-19 a masquées en 2020. This collaborative work continued beyond 2020.

Environmental impact.

As presented in Section 4, we are partners with CybeleTech in the framework of the EUPEX EuroHPC project. CybeleTech is a French company specialized in precision agriculture. Within the framework of our collaboration, we propose to focus our efforts on a scale-oriented data management mechanism targeting two CybeleTech use-cases. They address irrigation scheduling for orchards and optimal harvest date for corn, and their models require the acquisition of large volumes of remote data. The overall goal is to improve the accuracy of plant growth models and improve decision making for precision agriculture. The underlying approach in favour of precision agriculture has also encountered some criticism 1.

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.

  In 2023, in the context of the ACROSS EuroHPC project, we added an interface for Damaris to enable asynchronous analytics, in particular to support Dask (www.dask.org), a Python-based library for scalable analytics. Dask offers a suite of useful distributed analytic methods using familiar Python-like interfaces, similar to NumPy and Pandas. Our proposed Python interface has enabled access to the suite of Python based visualization libraries and Damaris has been successfully tested with new options for in situ visualization.

  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 (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:
  v1.10 Improved Dask data sharing in damaris4py library. v1.11 adds better support for variables not written in every iteration and a C API function for introspection on what Damaris plugins are available. v1.12 adds Debian / Redhat compatible packages generation for distribution, enables the initialization of Damaris with an XML object containing the configuration, and adds the ability of Damaris configuration creation programmatically to ease integration with other libraries (PDI for instance).
• News of the Year:
  In 2024, the support for Dask data sharing was enhanced. In addition, in the context of the NumPEX project, we started the development of the PDI / Damaris plugin to enable a running integration of the Damaris library with PDI (https://pdi.dev/main/), an interface for data access, read and write data from HDF5 files, in-situ visualization, and more.
• URL:
  https://­project.­inria.­fr/­damaris/
• Contact:
  Gabriel Antoniu
• Participants:
  Jean Etienne Ndamlabin Mboula, Gabriel Antoniu, Lokman Rahmani, Luc Bouge, Matthieu Dorier, Orçun Yildiz, Hadi Salimi, Joshua Bowden
• 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 to make E2Clab the most complete framework for edge-to-cloud experiments, a support for energy monitoring (Kwollect) on the Grid'5000 platform has been added, paving the way for energy-aware infrastructure optimization and other experiments.

  Additional contributions include: - Ongoing experiments on AI workflows on the computing-continuum and data provenance for explainable AI models. - Adapting existing provenance data capture to work with Flowcept, a runtime provenance data exporter for ML workflows developed at ORNL. - An improved documentation and CI pipeline has been implemented to ensure better reliability and stable behaviour across versions with improved code quality and throughout testing. - Improved logging files in experiment artifacts to facilitate the software's user-friendliness for members of the scientific community.

  Latest release archive: https://gitlab.inria.fr/E2Clab/e2clab/-/releases/v3.3.1

• 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
• Participants:
  Thomas Badts, Daniel Rosendo, Gabriel Antoniu, Alexandru Costan, Mathieu Simonin

7.1.3 StorAlloc

• Keywords:
  Simulation, HPC, Distributed Storage Systems
• Functional Description:

  StorAlloc is a simulator of a job scheduler dedicated to heterogeneous storage resources. It allows to model storage infrastructures, to simulate their partitioning and allocation, and to evaluate various scheduling algorithms.

  In practice, StorAlloc takes a storage request as input, which represents the presumed storage requirements of a job executed on a HPC system. It then offers to select some fitting storage resources, to be used by the client job. Storage resources are defined by the users, thanks to a YAML format with storage nodes and disks. Their selection happens by mean of an algorithm, also chosen by user (either from predefined algorithms, or user-developed). During simulation, various metrics are stored by StorAlloc all along the processing of storage requests, and eventually written to file when the simulation ends. Components of StorAlloc are independent and communicate through messages. They are easily extensible and new components may also be added.

• URL:
  https://­github.­com/­hephtaicie/­storalloc
• Publications:
  hal-03922866, hal-03878252, hal-03683568
• Contact:
  François Tessier
• Participants:
  Julien Monniot, François Tessier, Gabriel Antoniu

7.1.4 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

7.1.5 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:
  All features of the first version of MOSAIC were developed in 2024.
• Publication:
  hal-04808300
• Contact:
  François Tessier

7.1.6 Neomem

• Keywords:
  Machine learning, Continual Learning, High performance computing
• Scientific Description:

  Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), 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 has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs to achieve high accuracy, short runtime, and scalability. It leverages a set of buffers (local to each GPU) and uses several asynchronous techniques for updating these local buffers in an embarrassingly parallel fashion, all while handling the communication overheads necessary to augment input mini-batches (groups of training samples fed to the model) using unbiased, global sampling.

  We further propose a generalization of rehearsal buffers to support both classification and generative learning tasks, as well as more advanced rehearsal strategies (notably dark experience replay, leveraging knowledge distillation). We illustrate this approach with a real-life HPC streaming application from the domain of ptychographic image reconstruction. We run extensive experiments on up to 128 GPUs of the ThetaGPU supercomputer to compare our approach with baselines representative of training-from-scratch (the upper bound in terms of accuracy) and incremental training (the lower bound). Results show that rehearsal-based continual learning achieves a top-5 validation accuracy close to the upper bound, while simultaneously exhibiting a runtime close to the lower bound.

• Functional Description:
  Training neural networks with continuously generated data poses challenges related to forgetting previously acquired knowledge, a phenomenon known as "catastrophic forgetting." An effective solution involves replaying certain previously observed data to maintain associated knowledge. Neomem implements this approach, aiming to achieve excellent predictive performance at the cost of a slight increase in training time. Our approach allows for the utilization of dozens of GPUs, making it applicable to scientific simulations within the high-performance computing (HPC) community.
• News of the Year:
  We introduced a set of abstractions that do not impose any particular constraints on the data shape of training samples retained in the rehearsal buffer, allowing to accommodate a large range of deep learning workloads and rehearsal strategies. Besides, we introduce the concept of annotated tuples of tensors to serve representative training samples and their associated states conveniently to the AI runtime. Such an approach (1) addresses the need to support more rehearsal strategies (notably strategies leveraging knowledge distillation) while (2) transparently augmenting minibatches produced by data pipelines of data-parallel CL training instances.
• Publication:
  hal-04600107
• Contact:
  Alexandru Costan
• Participants:
  Thomas Bouvier, Alexandru Costan, Gabriel Antoniu

7.1.7 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)

7.1.8 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)

8.1.1 Provisioning storage resources for HPC and Cloud systems

Participants: François Tessier, Julien Monniot, Gabriel Antoniu.

• Collaboration.
  This work has been carried out in close co-operation with Henri Casanova (University of Manoa, HI, USA)

One of the recent axis we are developing in the context of high-performance data access concerns the provisioning of storage resources. The way these resources are accessed on supercomputers and clouds opposes a complex low-level vision that requires tight user control (on supercomputers) and a very abstract vision that implies uncertainty for performance modeling (on clouds). Nevertheless, taking full advantage of all available resources is critical in a context where storage is central for coupling workflow components. Our goal is then to make heterogeneous storage resources distributed across HPC+Cloud infrastructures allocatable and elastic to meet the needs of I/O-intensive hybrid workloads.

This is the context of Julien Monniot's thesis (thesis started in October 2021). He explores the techniques for scheduling storage resources on large-scale systems through simulation. The modeling of storage infrastructures and the evaluation of storage-aware scheduling algorithms are the main contributions of this work.

The work started in 2022 around a first proof-of-concept called StorAlloc 55, 54, 56 evolved in 2023 and 2024 into Fives (Simulator for Scheduling on Storage Systems at Scale), a new simulator implemented with WRENCH 43, a state-of-the-art simulation framework. Fives not only reproduces the results of StorAlloc, but goes far beyond it. Thanks to this simulator, we were able to model a Lustre parallel file system (both hardware and software, which we reverse-engineered), as well as a supercomputer mimicking Theta, a 11 PFlops HPC system at Argonne National Laboratory. Using a job abstraction layer we devised, Fives enabled us to simulate several weeks of job execution on Theta from an I/O point of view (Darshan traces). The aim of this simulation was to calibrate our simulator with a view to subsequently predicting the Lustre's I/O performance on any other dataset. This work was accepted and presented in an A-rank conference in the field 27. Our paper was a Best Paper Award nominee.

8.1.2 Data profiling and benchmarking of computational workflows beyond the classical Computing Continuum

Participants: Silvina Caino Lores, Gabriel Antoniu.

• Collaboration.
  This work has been carried out in cooperation with Elaine Wong and Travis Humble (Oak Ridge National Laboratory, USA)

Quantum Computing (QC) systems are increasingly explored as the next high-impact extension to the Computing Continuum, in particular with integration into supercomputers like Joliot-Curie at GENCI (France) and cloud environments like Amazon Web Services. As part of our leadership roles in the Workflows Community Initiative and workflow-specific venues like the WORKS workshop, we have identified that the successful interoperability between these systems in the Computing Continuum will depend on middleware able to interact with heterogeneous hardware technologies (e.g., CPU, GPU, TPU, FPGA, quantum processing unit (QPU)) and their associated software stacks and data management methods 36. In the context of QC, immediate approaches towards integrating QC into the wider Computing Continuum are hybrid and involve interaction with classical systems 41. This leads to complex open challenges on how to combine multiple programming models in a single application with workflow steps combining quantum and classical processing in a domain-agnostic manner.

In our previous work, we explored avenues of refinement for the QPU in the context of many-tasks management in order to understand the value it can play in linking QC with HPC 25. Notably, current works on the integration of QC into classical computing ecosystems focus on the interoperability and performance of the algorithms without considering data-oriented optimizations (e.g., data encoding, arrangement, locality, or mapping to high-level data abstractions), and workflow-specific challenges like task-resource mapping are rarely explored 60. We are conducting exploratory work on profiling and characterising data access and transfer patterns in small scenarios of variational quantum algorithms 44–a feasible near-term approach to QC integration– with the goal of developing new workflow benchmarks.

8.2.1 Investigating the Gap between Simulation, Emulation and Real-World Deployments for Reproducible Federated Learning Experiments

Participants: Cédric Prigent, Alexandru Costan, Gabriel Antoniu.

• Collaboration.
  This work has been carried out in close co-operation with Kate Keahey, University of Chicago / Argonne National Laboratory, who supervised the internship of Cédric Prigent.

A common practice in the Federated Learning (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 was accepted and presented in an A-rank conference in the field 28. This paper was also Best Paper Award nominee.

8.2.2 Formalisation of workflow provenance for trustworthy and explainable AI

Participants: Silvina Caino-Lores, Alexandru Costan.

• Collaboration.
  This work is part of an ongoing collaboration with Renan Souza and Rafael Ferreira da Silva (Oak Ridge National Laboratory, USA).

Artificial Intelligence (AI) is driving scientific discovery and economic growth in all kinds of application domains while impacting from routine daily tasks to societal-level challenges. However, research communities, industry players and social actors are expressing increasing concern about the potential ethical and practical implications of the pervasive presence of AI. Of particular concern are the explainability of AI, or making AI's decision-making process understandable, and transparency of AI, ensuring clarity in AI's design, data and operation. Working towards advancing explainability and transparency of AI is currently a priority, essential for responsible and trustworthy AI applications.

In our work, we have formalised the AI lifecycle and developed a vision of workflow provenance as a tool for data-to-insights, identifying key research priorities and open challenges 29. Provenance refers to the origin or history of data and models, capturing the lineage of their creation, modification, and usage. Provenance in the form of metadata captured at runtime during the execution of AI workflows provides a detailed record of data sources, processing steps, and model configurations, ensuring transparency and traceability throughout the AI lifecycle. A challenging aspect of working with AI workflows is that today there are no comprehensive formalisms able to capture the complexity and relationships in workflow and model provenance data. Our ongoing work towards the definition of ontologies and taxonomies for AI workflow provenance data aims to fill this gap and serve as a theoretical foundation for developing tailored provenance data management systems tailored for the different stakeholders involved in AI applications. In addition, we are exploring new analysis and visualisation techniques to facilitate the integration and inspection of heterogeneous and complex provenance data.

8.2.3 Efficient Resource-Constrained Federated Learning Clustering with Local Data Compression on the Edge-to-Cloud Continuum

Participants: Cédric Prigent, Alexandru Costan, Gabriel Antoniu.

• Collaboration.
  This work has been carried out in close co-operation with Loïc Cudennec (DGA), who is co-advising the PhD thesis of Cédric Prigent, and with DFKI in the context of the ENGAGE Inria-DFKI project.

Federated Learning (FL) has been proposed as a privacy-preserving approach for distributed learning over decentralized resources. While it can be a highly efficient tool for large-scale collaborative training of Machine Learning (ML) models, its efficiency may be strongly impacted by a high variability in data distributions among clients 20. Clustered FL tackles this problem by grouping clients with similar data distributions and training personalized models. Despite increasing model accuracy for federated peers, existing clustering approaches overlook system and infrastructure constraints leading to sustainability problems for resource-constrained devices.

In 28, we introduce a new method for resource-constrained FL clustering. We leverage pre-trained autoencoders to compress client data into low dimensional space and build lightweight embedding vectors used to cluster federated clients. A randomized quantization approach specifically secures the client embedding vectors against data reconstruction. Extensive experiments using a multi-GPU testbed with multiple scenarios introducing concept drift between clients demonstrate the generalitity of our approach to personalized FL. While each of the baselines encounters performance degradation in at least one of the scenarios, our strategy demonstrates top efficiency in all of them.

8.2.4 Efficient Data-Parallel Continual Learning with Asynchronous Distributed Rehearsal Buffers

Participants: Thomas Bouvier, Alexandru Costan, Gabriel Antoniu.

• Collaboration.
  This work has been carried out in the context of a JLESC project (Contnual Learning at Scale) in close co-operation with Bogdan Nicolae (Argonne National Laboratory- ANL, USA), who serves as a technical advisor for the PhD work of Thomas Bouvier.

Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), 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 has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs to achieve high accuracy, short runtime, and scalability.

This year we introduced a set of abstractions that do not impose any particular constraints on the data shape of training samples retained in the rehearsal buffer, allowing to accommodate a large range of deep learning workloads and rehearsal strategies. Besides, we introduce the concept of annotated tuples of tensors to serve representative training samples and their associated states conveniently to the AI runtime. Such an approach (1) addresses the need to support more rehearsal strategies (notably strategies leveraging knowledge distillation) while (2) transparently augmenting minibatches produced by data pipelines of data-parallel CL training instances 24, 17

8.2.5 Comparative Analysis of Federated Learning: Simulations Versus Real-World Testbeds

Participants: Mathis Valli, Alexandru Costan, Gabriel Antoniu.

• Collaboration.
  This work has been carried out in close co-operation with Loïc Cudennec (DGA). Mathis Valli PhD is co-supervised by Cédric Tedeschi (Myriads team).

Federated Learning (FL), while a significant advancement in decentralized machine learning, is predominantly assessed through simulations. These simulations often fail to capture the complexities and unpredictability of real-world environments. They usually overlook factors like network variability, device heterogeneity, and real-time operational challenges, leading to a gap between theoretical efficiency and practical applicability.

Addressing these limitations, our current work emphasizes the deployment of FL models on real-world testbeds. By moving beyond purely simulation-based analyses, we aim to better understand how factors such as bandwidth variability and device heterogeneity influence both model convergence and system-level behavior. Specifically, our experiments focus on monitoring performance metrics like model accuracy, training time, communication overhead, and energy consumption to evaluate how FL reacts to realistic conditions. By comparing these real-testbed outcomes with simulation results, our study intends to highlight the disparities and refine the practical applicability of FL models, bridging the gap between theory and practice.

We conduct these experiments on a heterogeneous set of nodes in Grid’5000, using E2Clab2's orchestration tools to manage and record resource usage data. Through continuous observation and data collection, we are gaining insights into how real operational challenges—like transient node failures or resource contention—affect FL training processes. Our current emphasis remains on quantifying these impacts rather than deploying self-adaptive mechanisms.

While this phase of our research focuses on understanding and quantifying the influence of system constraints, we have also highlighted the potential benefits of adapting FL systems in real time. Our publication 31 presented the case for incorporating dynamic adaptation, suggesting that an adaptive approach could help FL frameworks cope more effectively with evolving operational conditions. Ultimately, we envision robust FL solutions that can autonomously fine-tune their parameters to meet the demands of dynamic environments, thus bridging the remaining gap between controlled simulations and practical, large-scale FL deployments.

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

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

• Collaboration.
  This work has been carried out in close co-operation with the Inria TADaaM team in Bordeaux and 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 2, 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 introduced MOSAIC, a categorizer that detects I/O patterns from execution traces. MOSAIC extracts I/O operations contained in Darshan 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 work has been published in one of the major workshops at the Supercomputing conference 26 while a complementary work in collaboration with Inria Bordeaux and TU Darmsdadt has been accepted at IPDPS, an A-rank conference in the field 37.

8.3.2 Topology and affinity-aware data aggregation

Participants: François Tessier.

• Collaboration.
  This work is the conclusion of several years of collaboration with Emmanuel Jeannot (TADaaM team at Inria Bordeaux) and Venkatram Vishwanath (Argonne National Laboratory) as part of a JLESC project.

Over the years, moving data from application to the storage system has become more and more challenging. While the amount of data to manage has drastically increased, the capability of the HPC architectures to absorb this burden has reduced. In order to limit concurrency or non-contiguous accesses on file systems, a preliminary phase of data aggregation is often necessary before moving data. In the context of I/O, the two-phase I/O algorithm is a common method which involves selecting a subset of the processing entities, called aggregators, to accumulate contiguous pieces of data (aggregation phase) before writing/reading them to/from the storage system (I/O phase).

As part of a long-term JLESC collaboration, we have been focusing on optimizing data movement on large-scale systems, especially through data aggregation techniques for I/O intensive applications. For that purpose, we designed and developed the TAPIOCA library. TAPIOCA is a C++ I/O library based on the two-phase I/O scheme mentioned above and performing scalable architecture-aware data aggregation. This library features several key innovations: a RDMA-based implementation reducing the data movement cost between data producers and the aggregators, a way to capture the data model and the data layout of the application to optimize the I/O scheduling, and a model and an objective function computing an architecture-aware placement of the aggregators. The work on TAPIOCA has been published in the FGCS journal 21.

8.3.3 Scalable asynchronous I/O and in-situ processing with Damaris

Participants: Joshua Bowden, Etienne Ndamlabin, Gabriel Antoniu.

• Collaboration.
  This work has been done in colaboration with Atgeirr Rasmussen (SINTEF) and his team within the framework of the EuroHPC H2020 ACROSS project.

As large-scale simulations can take a long time to run, and require significant high-performance computing resources, we investigate how asynchronous I/O and in situ processing can help improve the performance, scaling, and efficiency of workflow executions. Within the ACROSS project we consider the use case povided by the OPM Flow software, used for a Carbon Sequestration simulation. In 2023 we started to investigate how the Damaris approach developed by the KerData team could be leveraged by OPM Flow to provide asynchronous analytics, in particular to support Dask (www.dask.org), a Python-based library for scalable analytics. Dask offers a suite of useful distributed analytic methods using familiar Python-like interfaces, similar to NumPy and Pandas. Our proposed Python interface has enabled access to the suite of Python based visualization libraries and Damaris has been successfully tested with new options for in situ visualization. This work was continued in 2024 with the goal to improve Damaris's support for Dask data sharing.

The EuroHPC ACROSS project has supported this work and the results are benefiting the OPM Flow simulation software which integrates Damaris in a public release. The developed capabilities within Damaris are now studied in collaboration with CEA within the NumPEx exploratory PEPR project.

8.3.4 Qualitatively Analyzing Optimization Objectives in the Design of HPC Resource Manager

Participants: Robin Boëzennec, Guillaume Pallez.

• Collaboration.
  This work has been done in colaboration with Fanny Dufosse from the INRIA Datamove team.

A correct evaluation of scheduling algorithms and a good understanding of their optimization criteria are key components of resource management in HPC. In this axis, we discuss bias and limitations of the most frequent optimization metrics from the literature. We provide elements on how to evaluate performance when studying HPC batch scheduling. We experimentally demonstrate these limitations by focusing on two use-cases: a study on the impact of runtime estimates on scheduling performance, and the reproduction of a recent high-impact work that designed an HPC batch scheduler based on a network trained with reinforcement learning. We demonstrate that focusing on quantitative optimization criterion ("our work improves the literature by X%") may hide extremely important caveat, to the point that the results obtained are opposed to the actual goals of the authors. Key findings show that mean bounded slowdown and mean response time are hazardous for a purely quantitative analysis in the context of HPC. Despite some limitations, utilization appears to be a good objective. We propose to complement it with the standard deviation of the throughput in some pathological cases. Finally, we argue for a larger use of area-weighted response time, that we find to be a very relevant objective.

8.3.5 Allocation Strategies for Disaggregated Memory in HPC Systems

Participants: Robin Boëzennec, Guillaume Pallez.

• Collaboration.
  This work has been done in colaboration with Fanny Dufosse and Danilo Carastan-Santos from the INRIA Datamove team.

In this axis we consider scheduling strategies to deal with disaggregated memory for HPC systems. Disaggregated memory is an implementation of storage management that provides flexibility by giving the option to allocate storage based on system-defined parameters. In this case, we consider a memory hierarchy that allows to partition the memory resources arbitrarily amongst several nodes depending on the need. This memory can be dynamically reconfigured at a cost. We provide algorithms that pre-allocate or reconfigure dynamically the disaggregated memory based on estimated needs. We provide theoretical performance results for these algorithms. An important contribution of our work is that it shows that the system can design allocation algorithms even if user memory estimates are not accurate, and for dynamic memory patterns. These algorithms rely on statistical behavior of applications. We observe the impact on the performance of parameters of interest such as the reconfiguration cost.

8.3.6 Scheduling distributed I/O resources in HPC systems

Participants: Guillaume Pallez.

• Collaboration.
  This work has been done in colaboration with the TADAAM Inria team.

Parallel file systems cut files into fixed-size stripes and distribute them across a number of storage targets (OSTs) for parallel access. Moreover, a layer of I/O nodes is often placed between compute nodes and the PFS. In this context, it is important to notice both OST and I/O nodes are potentially shared by running applications, which may lead to contention and low I/O performance.

Contention-mitigation approaches usually see the shared I/O infrastructure as a single resource capable of a certain bandwidth, whereas in practice it is a distributed set of resources from which each application can use a subset. In addition, performance measured in practice does not scale proportionally with the proportion of OST used. Indeed, depending on their characteristics, each application will be impacted differently by the number of used I/O resources.

We conducted 22 a comprehensive study of the problem of scheduling shared I/O resources — I/O nodes, OSTs, etc — to HPC applications. We tackled this problem by proposing heuristics to answer two questions: 1) how many resources should we give each application (allocation heuristics), and 2) which resources should be given to each application (placement heuristics). These questions are not independent, as using more resources often means sharing them. Nonetheless, our two-step approach allows for simpler heuristics that would be usable in practice.

In addition to overhead, an important aspect that impacts how “implementable” algorithms are is their input regarding applications’ characteristics, since this information is often not available or at least imprecise. Therefore, we proposed heuristics that use different input and studied their robustness to inaccurate information.

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

• Sarah Neuwirth:
  Sarah Neuwirth is a Research Professor at Johannes Gutenberg University Mainz (JGU), Germany. She visited the team in December 2024 and gave a seminar entitled "Toward Explainable I/O for HPC Systems".
• Frederic Suter:
  Frederic Suter is a Senior Research Scientist at Oak Ridge National Laboratory, USA. He visited KerData in December 2024 and gave a talk about his work on data management across large-scale workflows.

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, the latest one being published in 2024. 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.

Exa-DoST

Participants: Gabriel Antoniu, François Tessier, Julien Monniot, Joshua Bowden, 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.

• 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

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

• François Tessier:
  • Talk at the monthly I/O Seminar of Inria Bordeaux about the provisioning of heterogeneous storage resources on supercomputers.
• Silvina Caino-Lores:
  • Unified Data Abstractions for Scientific Workflow Composition in the Computing Continuum, Driving Scientific Workflows from the Data Plane Minisymposium at SIAM PP 2024 (Baltimore, USA).
• Jakob Luettgau:
  • Talk at the Per3S workshop on Performance and Scalability of Storage Systems in Paris, France
• Guillaume Pallez:
  • "How to evaluate HPC"; Invited Talk at CCDSC'24

10.1.5 Leadership within the scientific community

• Gabriel Antoniu:
  • Large-wingspan 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 (latest edition published inin 2024).
  • 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.
  • Bilateral Inria-DFKI project management: French coordinator of the ENGAGE project (2022-2024).
  • Team management: Head of the KerData Project-Team (INRIA-INSA Rennes).
  • International Associate Team management: Leader of the UNIFY Associate Team with Argonne National Lab (2019–2022).
• François Tessier:
  • Work package co-leader with Francieli Zanon-Boito (Associate Professor, University of Bordeaux) within the NumPEX ExaDoST project.
• Alexandru Costan:
  • Work package co-leader with René Schubotz(DFKI) within the ENGAGE Inria-DFKI project.
  • Work package leader within the PEPR CLOUD STEEL project.

10.1.6 Scientific expertise

• Guillaume Pallez:
  • Member of the Inria Scientific Board
• Gabriel Antoniu:
  • Evaluator for a Horizon Europe project (HORIZON-CL4-2021-HUMAN-01 call)
  • 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)

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)
• 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, 20 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.
• Silvina Caino-Lores:
  • Master: Processing Artificial Intelligence and Machine Learning Workloads at Scale, 9 hours (lectures), M2 level, Big Data Storage and Processing Infrastructures Module, Cloud and Network Infrastructures Master Program, EIT Digital School, ISTIC, University of Rennes.
  • Master: Processing Artificial Intelligence and Machine Learning Workloads at Scale, 9 hours (lectures) and 15 hours (tutorials), M1 level, Artificial Intelligence Master Program, ISTIC, University of Rennes.
• Thomas Bouvier:
  • Master: Stream Processing, 12 hours (lectures and lab sessions), M2 level, INSA Rennes.
  • Master: Database optimizations, 30 hours (lab sessions), M1 level, ISTIC, University of Rennes.
• Cédric Prigent:
  • Master: Cloud Computing and Big Data, 36 hours (lab sessions), M2 level, Cloud Module, MIAGE 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.
• Arthur Jaquard:
  • Master: Processing Artificial Intelligence and Machine Learning Workloads at Scale, 13.5 hours (tutorials), M1 level, Artificial Intelligence Master Program, ISTIC, University of Rennes.

10.2.2 Supervision

• PhD:
  • Thomas Bouvier, "Supporting Continual Learning across the Computing Continuum", thesis defended in November 2024, co-advised by Alexandru Costan and Gabriel Antoniu.
  • Julien Monniot, "Enabling accurate simulation of HPC storage systems: methodology and practical techniques", thesis defended in December 2024, co-advised by François Tessier and Gabriel Antoniu. "
  • Alexis Bandet, thesis defended in December 2024, co-advised by Guillaume Pallez and Francieli Zanon Boito.

• PhD in progress:
  • 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)
  • 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).
  • Robin Boezennec, thesis started in November 2022, co-advised by Guillaume Pallez and Fanny Dufossé (Datamove, Grenoble).

• Internships:
  • Théo Jolivel, "An Abstraction Layer for I/O Characterization of Large-Scale Applications", 5-month Master 2 internship started in March 2024, co-advised by François Tessier and Guillaume Pallez.
  • Hugo Thay, "Analysis of the I/O Access Pattern of a Radio Astronomy Application", 10-week Master 1 internship started in May 2024, supervised by François Tessier.
  • Nicolas Vincent, "Computational Storage Programming Paradigms and Execution Engines", 7-month Master 1 project internship started in September 2024, supervised by Jakob Luettgau.
  • Alix Trémondeux, "Refurbished HPC", 5-month post-M2 internship started in September 2024, supervised by Guillaume Pallez.

10.2.3 Juries

• Alexandru Costan:
  • GDR RSD - Member of the juries for the PhD award and Young/Senior Researcher award
• Silvina Caino-Lores:
  • PhD thesis: New techniques to build and manage agnostic workflows for the processing of digital products, Dante Sanchez-Gallegos, University Carlos III of Madrid, Spain

10.3.1 Productions (articles, videos, podcasts, serious games, ...)

• Silvina Caino-Lores:
  • Featured profile in the 2024 edition of the Women Leaders of the 21st Century publication by the IT Innovation Foundation (Spain).

International journals

• 16 articleR.Robin Boëzennec, F.Fanny Dufossé and G.Guillaume Pallez. Qualitatively Analyzing Optimization Objectives in the Design of HPC Resource Manager.ACM Transactions on Modeling and Performance Evaluation of Computing Systems2024, 1-28In press. HAL
• 17 articleT.Thomas Bouvier, B.Bogdan Nicolae, A.Alexandru Costan, T.Tekin Bicer, I.Ian Foster and G.Gabriel Antoniu. Efficient Distributed Continual Learning for Steering Experiments in Real-Time.Future Generation Computer SystemsJuly 2024, 1-19HALDOIback to text
• 18 articleD. I.Daiane I Dolci, J. R.James R Maddison, D. A.David A Ham, G.Guillaume Pallez and J.Julien Herrmann. checkpoint_schedules: schedules for incremental checkpointing of adjoint simulations.Journal of Open Source Software9March 2024, 1-4HALDOI
• 19 articleY.Yishu Du, L.Loris Marchal, G.Guillaume Pallez and Y.Yves Robert. Improving Batch Schedulers with Node Stealing for Failed Jobs.Concurrency and Computation: Practice and Experience36122024, 1-36HALDOI
• 20 articleC.Cédric Prigent, A.Alexandru Costan, G.Gabriel Antoniu and L.Loïc Cudennec. Enabling Federated Learning across the Computing Continuum: Systems, Challenges and Future Directions.Future Generation Computer Systems160June 2024, 767-783HALDOIback to text
• 21 articleF.François Tessier, V.Venkatram Vishwanath and E.Emmanuel Jeannot. Adding topology and memory awareness in data aggregation algorithms.Future Generation Computer Systems159October 2024, 188-203HALDOIback to text

International peer-reviewed conferences

• 22 inproceedingsA.Alexis Bandet, F.Francieli Boito and G.Guillaume Pallez. Scheduling distributed I/O resources in HPC systems.30th International European Conference on Parallel and Distributed Computing 26 - 30 August 2024 Madrid, Spain 30th International European Conference on Parallel and Distributed ComputingMadrid, SpainAugust 2024HALback to text
• 23 inproceedingsR.Robin Boëzennec, D.Danilo Carastan-Santos, F.Fanny Dufossé and G.Guillaume Pallez. Allocation Strategies for Disaggregated Memory in HPC Systems.HiPC 2024 - 31st IEEE International Conference on High Performance Computing, Data, and AnalyticsBengalore, IndiaIEEE2024, 1-11HAL
• 24 inproceedingsT.Thomas Bouvier, B.Bogdan Nicolae, H.Hugo Chaugier, A.Alexandru Costan, I.Ian Foster and G.Gabriel Antoniu. Efficient Data-Parallel Continual Learning with Asynchronous Distributed Rehearsal Buffers.CCGrid 2024 - IEEE 24th International Symposium on Cluster, Cloud and Internet ComputingPhiladelphia (PA), United States2024, 1-10HALDOIback to text
• 25 inproceedingsS.Silvina Caino-Lores, D.Daniel Claudino, E.Eugene Dumitrescu, T. S.Travis S Humble, S. L.Sonia Lopez Alarcon and E.Elaine Wong. Rethinking Programming Paradigms in the QC-HPC Context.WAMTA 2024 - 2nd International Workshop on Asynchronous Many-Task Systems and Applications14626Lecture Notes in Computer ScienceKnoxville, TN, United StatesSpringer Nature SwitzerlandMay 2024, 84-91HALDOIback to text
• 26 inproceedingsT.Théo Jolivel, F.François Tessier, J.Julien Monniot and G.Guillaume Pallez. MOSAIC: Detection and Categorization of I/O Patterns in HPC Applications.SC24-W: Workshops of the International Conference for High Performance Computing, Networking, Storage and AnalysisPDSW 2024 - 9th International Parallel Data Systems WorkshopAtlanta, United States2024, 1-7HALDOIback to text
• 27 inproceedingsJ.Julien Monniot, F.François Tessier, H.Henri Casanova and G.Gabriel Antoniu. Simulation of Large-Scale HPC Storage Systems: Challenges and Methodologies.HiPC 2024 - 31st IEEE International Conference on High Performance Computing, Data, and AnalyticsBangalore, India2024, 1-11HALback to textback to text
• 28 inproceedingsC.Cédric Prigent, M.Melvin Chelli, A.Alexandru Costan, L.Loïc Cudennec, R.René Schubotz and G.Gabriel Antoniu. Efficient Resource-Constrained Federated Learning Clustering with Local Data Compression on the Edge-to-Cloud Continuum.HiPC 2024 - 31st IEEE International Conference on High Performance Computing, Data, and AnalyticsBengaluru (Bangalore), India2024, 1-11HALback to textback to textback to text
• 29 inproceedingsR.Renan Souza, S.Silvina Caino-Lores, M.Mark Coletti, T. J.Tyler J Skluzacek, A.Alexandru Costan, F.Frédéric Suter, M.Marta Mattoso and R. F.Rafael Ferreira da Silva. Workflow Provenance in the Computing Continuum for Responsible, Trustworthy, and Energy-Efficient AI.e-Science 2024 - 20th IEEE International Conference on e-ScienceOsaka, JapanIEEESeptember 2024, 1-7HALDOIback to text
• 30 inproceedingsA.Ahmad Tarraf, A.Alexis Bandet, F.Francieli Zanon Boito, G.Guillaume Pallez and F.Felix Wolf. Capturing Periodic I/O Using Frequency Techniques.IPDPS 2024 - 38th IEEE International Parallel & Distributed Processing SymposiumSan Francisco, United States2024, 1-13HAL
• 31 inproceedingsM.Mathis Valli, A.Alexandru Costan, C.Cédric Tedeschi and L.Loïc Cudennec. Towards Efficient Learning on the Computing Continuum: Advancing Dynamic Adaptation of Federated Learning.FlexScience'24: Proceedings of the 14th Workshop on AI and Scientific Computing at Scale using Flexible Computing InfrastructuresFlexScience 2024 - 14th Workshop on AI and Scientific Computing at Scale using Flexible Computing InfrastructuresPisa, ItalyACM2024, 42-49HALDOIback to text

Scientific books

• 32 bookS.Sai Narasimhamurthy, N.Nico Mittenzwey, F.Fabrizio Magugliani, M.Marc Duranton, C.Craig Prunty, P.Pascale Rossé-Laurent, M.Manolis Marazakis, P.Paul Carpenter, G.Gabriel Antoniu, S.Sarah Neuwirth, P.Philippe Deniel, D.Dirk Pleiter, U.-U.Utz-Uwe Haus, E.Erwin Laure, A.Andreas Wierse, T.Tobias Becker, R.Robert Haas, M.Michael Malms, H.-C.Hans-Christian Hoppe, V.Valeria Bartsch, S.Sagar Dolas, O.Ondrej Vysocky, M.Maria Perez, A.Andy Forrester, K.Kristel Michielsen and E.Estela Suarez. ETP4HPC's SRA 6 - Strategic Research Agenda for High Performance Computing in Europe.ETP4HPC Strategic Research AgendaZenodo2024HALDOI

Reports & preprints

• 33 miscA.Alexis Bandet, F.Francieli Boito and G.Guillaume Pallez. Prediction and Interpretability of HPC I/O Resources Usage with Machine Learning.2024HAL
• 34 reportA.Alexis Bandet, F.Francieli Zanon Boito and G.Guillaume Pallez. Allocation and Placement Algorithms for Scheduling Distributed I/O Resources in HPC Systems.RR-9549Inria Bordeaux; Inria RennesMay 2024, 1-27HAL
• 35 reportC.Clément Barthélemy, F.Francieli Zanon Boito, E.Emmanuel Jeannot, G.Guillaume Pallez and L.Luan Teylo. Implementation of an unbalanced I/O Bandwidth Management system in a Parallel File System.RR-9537InriaJanuary 2024HAL
• 36 reportR.Rafael Ferreira da Silva, D.Deborah Bard, K.Kyle Chard, d. W.de Witt Shaun, I.Ian Foster, T.Tom Gibbs, C.Carole Goble, W.William Godoy, J.Johan Gustafsson, U.-U.Utz-Uwe Haus, S.Stephen Hudson, S.Shantenu Jha, L.Laila Los, D.Drew Paine, F.Frédéric Suter, L.Logan Ward, S.Sean Wilkinson, M.Marcos Amaris, Y.Yadu Babuji, J.Jonathan Bader, R.Riccardo Balin, D.Daniel Balouek, S.Sarah Beecroft, K.Khalid Belhajjame, R.Rajat Bhattarai, W.Wes Brewer, P.Paul Brunk, S.Silvina Caino-Lores, H.Henri Casanova, D.Daniela Cassol, J.Jared Coleman, T.Taina Coleman, I.Iacopo Colonnelli, A. A.Anderson Andrei Da Silva, D.Daniel de Oliveira, P.Pascal Elahi, N.Nour Elfaramawy, W.Wael Elwasif, B.Brian Etz, T.Thomas Fahringer, W.Wesley Ferreira, R.Rosa Filgueira, J.Jacob Fosso Tande, L.Luiz Gadelha, A.Andy Gallo, D.Daniel Garijo, Y.Yiannis Georgiou, P.Philipp Gritsch, P.Patricia Grubel, A.Amal Gueroudji, Q.Quentin Guilloteau, C.Carlo Hamalainen, R.Rolando Hong Enriquez, L.Lauren Huet, K.Kevin Hunter Kesling, P.Paula Iborra, S.Shiva Jahangiri, J.Jan Janssen, J.Joe Jordan, S.Sehrish Kanwal, L.Liliane Kunstmann, F.Fabian Lehmann, U.Ulf Leser, C.Chen Li, P.Peini Liu, J.Jakob Luettgau, R.Richard Lupat, J.Jose M. Fernandez, K.Ketan Maheshwari, T.Tanu Malik, J.Jack Marquez, M.Motohiko Matsuda, D.Doriana Medic, S.Somayeh Mohammadi, A.Alberto Mulone, J.-L.John-Luke Navarro, K. W.Kin Wai Ng, K.Klaus Noelp, B.Bruno P. Kinoshita, R.Ryan Prout, M.Michael R. Crusoe, S.Sashko Ristov, S.Stefan Robila, D.Daniel Rosendo, B.Billy Rowell, J.Jedrzej Rybicki, H.Hector Sanchez, N.Nishant Saurabh, S. K.Sumit Kumar Saurav, T.Tom Scogland, D.Dinindu Senanayake, W.Woong Shin, R.Raul Sirvent, T.Tyler Skluzacek, B.Barry Sly-Delgado, S.Stian Soiland-Reyes, A.Abel Souza, R.Renan Souza, D.Domenico Talia, N.Nathan Tallent, L.Lauritz Thamsen, M.Mikhail Titov, B.Benjamin Tovar, K.Karan Vahi, E.Eric Vardar-Irrgang, E.Edite Vartina, Y.Yuandou Wang, M.Merridee Wouters, Q.Qi Yu, Z.Ziad Al Bkhetan and M.Mahnoor Zulfiqar. Workflows Community Summit 2024: Future Trends and Challenges in Scientific Workflows.Oak Ridge National Laboratory, USAOctober 2024HALDOIback to text
• 37 miscF.Francieli Zanon 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.January 2025HALback to text

Other scientific publications

• 38 miscP.Paul Carpenter, G.Gabriel Antoniu, M.Manuel Arenaz, O.Olivier Aumage, J.Jakub Beránek, A.Alfredo Buttari, A.Alexandru Costan, S.Sonja Happ, V.Venkatesh Kannan, C.Christian Perez, A.Antonio Peña, A.Alberto Scionti, X.Xavier Vigouroux and P.Paolo Viviani. ETP4HPC SRA White Paper - Programming Environment.December 2024HALDOI
• 39 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 2025HALDOI