Energy Application Profiling and Modeling

Avalon will improve the profiling and modeling of scientific applications with respect to energy consumption. In particular, it will require to improve the tools that measure the energy consumption of applications, virtualized or not, at large scale, so as to build energy consumption models of applications.

Data-intensive Application Profiling, Modeling, and Management

Avalon will improve the profiling, modeling, and management of scientific applications with respect to CPU and data intensive applications. Challenges are to improve the performance prediction of parallel regular applications, to model and simulate (complex) intermediate storage components, and data-intensive applications, and last to deal with data management for hybrid computing infrastructures.

Programming Models and Runtimes

 Avalon will design component-based models to capture the different facets of parallel and distributed applications while being resource agnostic, so that they can be optimized for a particular execution. In particular, the proposed component models will integrate energy and data modeling results. Avalon in particular targets OpenMP runtime as a specific use case and contributes to improve it for multi-GPU nodes.

Application Mapping and Scheduling

Avalon will propose multi-criteria mapping and scheduling algorithms to meet the challenge of automating the efficient utilization of resources taking into consideration criteria such as performance (CPU, network, and storage), energy consumption, and security. Avalon will in particular focus on application deployment, workflow applications, and security management in clouds.

All our theoretical results will be validated with software prototypes using applications from different fields of science such as bioinformatics, physics, cosmology, etc. The experimental testbeds Grid'5000 and SLICES will be our platforms of choice for experiments.

3.4.1 Application Mapping and Software Deployment

Application mapping and software deployment consist in the process of assigning distributed pieces of software to a set of resources. Resources can be selected according to different criteria such as performance, cost, energy consumption, security management, etc. A first issue is to select resources at application launch time. With the wide adoption of elastic platforms, i.e., platforms that let the number of resources allocated to an application to be increased or decreased during its execution, the issue is also to handle resource selection at runtime.

The challenge in this context corresponds to the mapping of applications onto distributed resources. It will consist in designing algorithms that in particular take into consideration application profiling, modeling, and description.

A particular facet of this challenge is to propose scheduling algorithms for dynamic and elastic platforms. As the number of elements can vary, some kind of control of the platforms must be used accordingly to the scheduling.

3.4.2 Non-Deterministic Workflow Scheduling

Many scientific applications are described through workflow structures. Due to the increasing level of parallelism offered by modern computing infrastructures, workflow applications now have to be composed not only of sequential programs, but also of parallel ones. New applications are now built upon workflows with conditionals and loops (also called non-deterministic workflows).

These workflows cannot be scheduled beforehand. Moreover cloud platforms bring on-demand resource provisioning and pay-as-you-go billing models. Therefore, there is a problem of resource allocation for non-deterministic workflows under budget constraints and using such an elastic management of resources.

Another important issue is data management. We need to schedule the data movements and replications while taking job scheduling into account. If possible, data management and job scheduling should be done at the same time in a closely coupled interaction.

7.1.1 Halley

• Name:
  Halley
• Keywords:
  Software Components, HPC
• Scientific Description:
  Halley is an implementation of the COMET component model that enables to efficiently compose independent parallel codes using both classical use/provide ports but also dataflow oriented ports that are used to generate tasks for multi-core shared-memory machines.
• Functional Description:
  Halley transforms a COMET assembly into a L2C assembly that contains some special components that deal with the data flow section. In particular, a dataflow section of COMET generates a "scheduler" L2C component that contains the code that is in charged of creating its tasks.
• Release Contributions:
  In 2024, dynamic workflow support was added. So, for example, the input data size of a meta-task can now depend on the output of a previous meta-task.
• Publications:
  tel-01663718, hal-01518730, hal-01566288, hal-01901806
• Contact:
  Christian Perez
• Participants:
  Jérôme Richard, Christian Perez, Jerry Lacmou Zeutouo

7.1.2 XKBLAS

• Name:
  XKBLAS
• Keywords:
  BLAS, Dense linear algebra, GPU
• Functional Description:

  XKBLAS is yet an other BLAS library (Basic Linear Algebra Subroutines) that targets multi-GPUs architecture thanks to the XKaapi runtime and with block algorithms from PLASMA library. XKBLAS is able to exploit large multi-GPUs node with sustained high level of performance. The library offers a wrapper library able to capture calls to BLAS (C or Fortran). The internal API is based on asynchronous invocations in order to enable overlapping between communication by computation and also to better composed sequences of calls to BLAS.

  This current version of XKBlas is the first public version and contains only BLAS level 3 algorithms, including XGEMMT:

  XGEMM XGEMMT: see MKL GEMMT interface XTRSM XTRMM XSYMM XSYRK XSYR2K XHEMM XHERK XHER2K

  For classical precision Z, C, D, S.

• Release Contributions:
  0.1.x versions: calls to BLAS kernels must be initiate by the same thread that initializes the XKBlas library. 0.2.x versions: better support for libblas_wrapper and improved scheduling heuristic to take into account memory hierarchy between GPUs 0.4.x versions: add support for AMD GPU 0.5.x : better support for AMD GPU (MI250x). Add capacity to clustering GPUs and CPU threads.
• News of the Year:
  New development to clustering both GPUs and CPU threads invoking BLAS kernels in order to reduce communication between resources. Better support for AMD GPU (configuration and performance). Presentation during the MUMPS annual meeting at ANSYS, 2024/06/20. Seminar presentation at EDF, 2024/11/26.
• URL:
  https://­gitlab.­inria.­fr/­xkblas/­versions
• Contact:
  Thierry Gautier
• Participants:
  Thierry Gautier, João Vicente Ferreira Lima

7.1.3 execo

• Keywords:
  Toolbox, Deployment, Orchestration, Python
• Functional Description:
  Execo offers a Python API for asynchronous control of local or remote, standalone or parallel, unix processes. It is especially well suited for quickly and easily scripting workflows of parallel/distributed operations on local or remote hosts: automate a scientific workflow, conduct computer science experiments, perform automated tests, etc. The core python package is execo. The execo_g5k package provides a set of tools and extensions for the Grid5000 testbed. The execo_engine package provides tools to ease the development of computer sciences experiments.
• Release Contributions:
  Release 2.8.1 on October 21, 2024 (list of changes since version 2.7: adapt to changes in oarstat output format (compatibility with old and new output formats), g5k api cache stored as json instead of pickle, support clusters names ending with numbers (eg. abacus-X), canonical_host_name handles interface / kavlan / ipv6 and support cluster names ending with numbers + fix for ifname != ethX (eg. fpgaX), add get_host_interface, extend planning API to allow requests at node level additionally to cluster and site level, spawn process lifecycle handlers in separate threads to avoid blocking + refactoring, handle encoding (py3+) when writing to Process, full redesign of the Processes expect implementation, add get_cluster_queues and get_cluster_jobtypes, add KaconsoleProcess, add substitutions to filenames in stdout/stderr handlers, scp commands in Get / Put as list and shell=False to (securely) handle spaces in path, fix eating 100% of one core iterating through high number of fd to close them in conductor, fix various regexes withinvalid escape sequences, add option in execo_engine for using pty to copy_outputs(), fix corner case in process args handling in Remote)
• URL:
  https://­gitlab.­inria.­fr/­mimbert/­execo
• Contact:
  Matthieu Imbert
• Participants:
  Florent Chuffart, Laurent Pouilloux, Matthieu Imbert

7.2.1 Platform: Grid'5000

Participants: Simon Delamare, Pierre Jacquot, Matthieu Imbert, Laurent Lefèvre, Christian Perez, Jean-Camille Seck, Quentin Assuncao, Cyril Devaux.

Functional Description

The Grid'5000 experimental platform is a scientific instrument to support computer science research related to distributed systems, including parallel processing, high performance computing, cloud computing, operating systems, peer-to-peer systems and networks. It is distributed on 10 sites in France and Luxembourg, including Lyon. Grid'5000 is a unique platform as it offers to researchers many and varied hardware resources and a complete software stack to conduct complex experiments, ensure reproducibility and ease understanding of results.

• Contact: Laurent Lefèvre
• URL: www.grid5000.fr/

7.2.2 Platform: SLICES-FR

Participants: Simon Delamare, Pierre Jacquot, Matthieu Imbert, Laurent Lefèvre, Christian Perez, Jean-Camille Seck, Quentin Assuncao, Cyril Devaux.

Functional Description The SLICES-FR infrastructure aims at providing an experimental platform for experimental computer Science (Internet of things, clouds, HPC, big data, etc. ). This new infrastructure will supersede two existing infrastructures, Grid'5000 and FIT.

• Contact: Christian Perez
• URL: www.slices-fr.eu/

7.2.3 Platform: SLICES

Participants: Simon Delamare, Pierre Jacquot, Laurent Lefèvre, Christian Perez.

Functional Description SLICES is an European effort that aims at providing a flexible platform designed to support large-scale, experimental research focused on networking protocols, radio technologies, services, data collection, parallel and distributed computing and in particular cloud and edge-based computing architectures and services. SLICES-FR is the The French node of SLICES.

• Contact: Christian Perez
• URL: www.slices-ri.eu

8.1.1 Estimating the power consumption of bare metal water-cooled servers

Participants: Maxime Agusti, Eddy Caron, Laurent Lefèvre.

Numerous cloud providers offer physical servers for rental in bare metal paradigm. This mode gives customers total control over hardware resources, but limits cloud providers’ visibility of their usage. Accurately measuring server energy consumption in this context represents a major challenge, as installing physical energy meters is both costly and complex. Existing energy models are generally based on system usage data, which is incompatible with the general privacy policies of bare-metal server contracts. To deal with these problems, it is imperative to develop new approaches for estimating the energy consumption of these servers. This paper presents an original non-intrusive method for estimating the energy consumption of a server cooled by direct-chip liquid-cooling, based on the coolant temperature and the processor temperature obtained via IPMI. Our approach is evaluated on an experiment carried out on 19 bare metal servers of a production infrastructure equipped with physical wattmeters.3

8.1.2 Deep Reinforcement Learning for Energy-efficient Selection of Embedded Services at the Edge

Participants: Laurent Lefèvre, Doreid Ammar, Hugo Hadjur.

Edge computing helps to release the tension at the center of IoT systems' networks, thus reducing the latency, optimizing the bandwidth, and providing new privacy and security solutions, among others. Despite its benefits, edge computing faces unique challenges, including latency, security, and resource constraints. Among these challenges, energy consumption has emerged in the research community, and the global objective is to "do more with less". Researchers explore diverse strategies to enhance sustainability, from hardware optimizations to intelligent algorithms. The quest for energy efficiency and to reduce several impacts aligns with broader efforts to create an environmentally conscious technology landscape. In this paper, we present a task selection model for the edge. We focus on energy consumption and aim to maximize the value given by tasks, all the while minimizing the energy consumed. To do so, we develop a computer environment to simulate outdoor energy-harvesting edge devices, contribute to research reproducibility by recreating a photovoltaic energy harvesting prediction model, and train deep reinforcement learning models to select the best set of tasks at the edge. Our best deep reinforcement learning model, which uses Trust Region Policy Optimization, outperforms our best heuristic and is a robust task selector under varying external conditions.

Some experimental parts of this work occur in the CPER LECO/GreenCube project and some parts are financially supported by aivancity School for Technology, Business & Society Paris-Cachan. This work, within the the Ph.D. of Hugo Hadjur is co-advised by Doreid Ammar (Academic Dean and Professor at aivancity School for Technology, Business & Society Paris-Cachan and external member of Avalon team) and Laurent Lefevre 7.

8.1.3 S-ORCA : A social-based consolidation approach to reduce Cloud infrastructures energy consumption

Participants: Eddy Caron, Laurent Lefèvre, Simon Lambert.

Information and Communication Technologies (ICT) and Data Centres (DC) have nowadays considerable environmental impacts. The number of applications and services hosted in the cloud is significant, and the associated number of infrastructure is steadily increasing. Cloud Service Providers (CSP) need to respond to this growing demand and size their infrastructures accordingly but users misbehaviour and expectations in terms of service quality lead to oversized infrastructures. Infrastructures are sized to meet peak-demand, resulting in poor resource utilization and additional power consumption. But as their behaviour can increase DC energy consumption and environmental footprint, users can also help reducing them. In this paper, we study how users provided with a simple tool can reduce the power consumption of a virtualization cluster in a cloud company. Using a Virtual Machine (VM) shutdown policy, users can directly contribute to the mitigation of the power consumption of the infrastructure. Part of the paper is dedicated to profile the users and understand their behaviour when it comes to powering off their VMs. To further reduce the energy consumption of the cluster, we combine the VM shutdown policy with a simple consolidation heuristic. Simulations show a 23.95% power consumption reduction, with an additional 8.72% reduction thanks to the users. A production implementation was conducted and results in a 12.58% power consumption reduction over one week 8, 23.

8.1.4 Estimating the environmental impact of Generative-AI services

Participants: Adrien Berthelot, Eddy Caron, Mathilde Jay, Laurent Lefèvre.

Generative AI (Gen-AI) represents a major growth potential for the digital industry, a new stage in digital transformation through its many applications. Unfortunately, by accelerating the growth of digital technology, Gen-AI is contributing to the significant and multiple environmental damage caused by its sector. The question of the sustainability of IT must include this new technology and its applications, by measuring its environmental impact. To best respond to this challenge, we propose various ways of improving the measurement of Gen-AI's environmental impact. Whether using life-cycle analysis methods or direct measurement experiments, we illustrate our methods by studying Stable Diffusion a Gen-AI image generation available as a service. By calculating the full environmental costs of this Gen-AI service from end to end, we broaden our view of the impact of these technologies. We show that Gen-AI, as a service, generates an impact through the use of numerous user terminals and networks. We also show that decarbonizing the sources of electricity for these services will not be enough to solve the problem of their sustainability, due to their consumption of energy and rare metals. This consumption will inevitably raise the question of feasibility in a world of finite resources. We therefore propose our methodology as a means of measuring the impact of Gen-AI in advance. Such estimates will provide valuable data for discussing the sustainability or otherwise of Gen-AI solutions in a more transparent and comprehensive way 4, 21 1. This result is a joint work explored during the PhD of Adrien Berthelot 16 co-advised by Laurent Lefevre and Eddy Caron and during the PhD of Mathilde Jay 17 co-advised by Laurent Lefevre and Denis Trystram (UGA).

8.1.5 A Methodology and a Toolbox to Explore Dataset related to the Environmental Impact of HTTP Requests

Participants: Mathilde Jay, Laurent Lefèvre.

EcoIndex has been proposed to evaluate the absolute environmental performance of a given URL using a score ranging from 0 to 100 (the higher, the better). In this article, we make a critical analysis of the initial approach and propose alternatives that no longer calculate a plain score but allow the query to be situated among other queries. The generalized critiques come with statistics and rely on extensive experiments (first contribution). Then, we move on to low-cost Machine Learning (ML) approaches (second contribution) and a transition before obtaining our final results (third contribution). Our research aims to extend the initial idea of analytical computation, i.e., a relation between three variables, in the direction of algorithmic ML computations. The fourth contribution corresponds to a discussion on our implementation, available on a GitHub repository. Along with the paper, we invite the reader to examine the question: What attributes make sense for our problem?, or equivalently, what is a relevant data policy for studying digital environmental impacts? Beyond computational questions, it is important for the scientific community to focus on this question in particular. We currently promote using wellestablished ML techniques because of their potential, which we discuss in this research. However, we also question techniques for their frugality or otherwise. We also want to encourage synergy between technical expertise and business knowledge because this is fundamental for advancing the data project.6

8.1.6 Exploring RAPL as a Power Capping Leverage for Power-Constrained Infrastructures

Participants: Laurent Lefèvre, Vladimir Ostapenco.

Data centers are very energy-intensive facilities whose power provision is challenging and constrained by power bounds. In modern data centers, servers account for a significant portion of the total power consumption. In this context, the ability to limit the instant power consumption of an individual computing node is an important requirement. There are several energy and power capping techniques that can be used to limit compute node power consumption, such as Intel RAPL. Although it is nowadays mainly utilized for energy measurement, Intel RAPL (Running Average Power Limit) was originally designed for power limitation purposes. Some works use Intel RAPL for power limitation in a limited context without full knowledge of the inner workings of this technology and what is done behind the scenes to enforce the power constraint. Furthermore, Intel has not revealed any details about its internal implementation. It is unclear exactly how Intel RAPL technology operates and what effects it has on application performance and power consumption. In this work, we conduct a thorough analysis of Intel RAPL technology as a power capping leverage on a variety of heterogeneous nodes for a selection of CPU and memory intensive workloads. For this purpose, we first validate Intel RAPL power capping mechanism using a high-precision external power meter and investigate properties such as accuracy, power limit granularity, and settling time. Then, we attempt to determine which mechanisms are employed by RAPL to adjust power consumption 9. This work was explored during the PhD of Vladimir Ostapenco 18 co-advised by Laurent Lefevre and Anne-Cécile Orgerie (Magellan team, IRISA).

8.1.7 Many locks, few leverages! Digital sufficiency: the situation is serious but not hopeless

Participants: Laurent Lefèvre.

We are attached to digital technology as a work tool, a research subject, a leisure companion, an interface with the administration, a social bonding tool... it's part of our lives. But on the one hand, it is far from fulfilling its promises in terms of ecology, and on the other, it contributes to locking in a certain choice of society, in which the user loses his autonomy in the face of tools. Yet a digital world that is useful, accessible, repairable, easy to dismantle, recyclable, durable and robust would be a desirable option for everyone. In this article, we review a number of socio-technical, cultural, economic, political and institutional obstacles. The idea here is not to be exhaustive, but to highlight a few structural impediments to the implementation of sobriety-enhancing actions, from the point of view of reducing the negative impacts of digital technology. On the other hand, we also observe some emerging signs that could bring about change. We conclude by highlighting the tensions arising from the locks on the one hand and the levers on the other, as revealing opposing societal visions and value systems.5

8.1.8 Sufficient generative AI: an oxymoron?

Participants: Laurent Lefèvre.

The onslaught of Artificial Intelligence tools is shaking up the technological, financial and economic landscape, creating a bubble of more or less desirable possibilities and echoing the questions posed by the literature. Digital technology has entered a new phase of growth, driven mainly by the development, training and use of deep neural networks and, in particular, generative AI, which are at the root of today's spectacular results. The sheer number of parameters (several hundred billion) and the immense training corpus exploited mean that even the most seasoned of us can be bluffed. Of course, this unbridled growth is not without consequences. While it is difficult to pinpoint the precise impact of AI, mainly due to a lack of data on usage and infrastructure, we do know that it uses hardware, software, energy and human resources on a massive scale. AI tools are growing much faster than studies on their impact, and we are moving forward blindly at a time when the environmental crisis is worsening exponentially. AI is also making inroads into higher education in a variety of ways.14

8.1.9 Fine-grained methodology to assess environmental impact of a set of digital services

Participants: Adrien Berthelot, Eddy Caron, Laurent Lefèvre.

The digital sector does not have a neglectable footprint, contributing to global climate change and overcommitment of planetary boundaries. To face environmental challenges, digital services could provide help and even mitigate other sector environmental impact. But, as the assessment of the environmental impacts of digital services remains difficult, digital services are still uncertain solutions. Global assessments of the digital sector already exist but are still often uncompleted, limited to carbon emission, use phase of the equipment and are made at a too large scale to be applicable. As digital services rely on highly mutualized equipment and infrastructure, it is difficult to assess the cost of a single service. In this paper, we propose a methodology to calculate the complete environmental impact of a service. We take a service provider's perspective on how, and with what data from service hosting, to calculate and reduce the service footprint. From a host point of view, we distribute the footprint between the services, allowing to consider precisely the impacts. With such data on the environmental impacts of the hosted services, we propose strategies to prioritize optimization effort and decommissioning policies. This preliminary work is avaliable in 20.

8.2.1 SkyData: Autonomous Data paradigm

Participants: Eddy Caron, Elise Jeanneau, Laurent Lefèvre, Etienne Mauffret, Christian Perez.

With the rise of Data as a Service, companies understood that whoever controls the data has the power. The past few years have exposed some of the weakenesses of traditional data management systems. For example, application owner can collect and use data to their own advantage without the user's consent. We defined the SkyData concept, which revolves around autonomous data evolving in a distributed system. This new paradigm is a complete break from traditional data management systems. This paradigm is born from the many issues associated with traditional data management systems, such as resells or private information collected without consent, for example. Self managed data, or SKDs, are agents endowed with data, capabilities and goals to achieve. They are free to behave as they wish and try to accomplish their goals as efficiently as possible. They use learning algorithms to improve their decision making and learn new capabilities and services. We introduced how SKDs could be developed and provided some insight on useful capabilities.

In 2024, we explore algorithms that reach a trade-off between data autonomy and cohesion in a given subset of replicas. We propose a deterministic algorithm that ensures cohesion under a costly assumption on the number of failures. Alternatively, we investigate the use of an eventual leader election mechanism in a probabilistic algorithm where a designated leader manages coordination and acts as a communication relay. Compared to the naïve approach of broadcasting new positions to all replicas after each migration, we show experimentally that this approach reduces the loss of cohesion in most scenarios, even without assumption on the number of failures.

8.2.2 Numerics in the Cloud

We have defined some guidelines for critical applications to reduce the arithmetic numerical issue and we provide additional guidelines dedicated to Cloud platform.

Using a simple experiment we shown how the result of a floating-point computation can be affected when the program is compiled and executed in different environments (different processors, with different floating-point extensions and different compiler options), which is to be expected when running applications on a Cloud. Our example is simply based on floating-point summation, which is well known to be “not as easy to compute accurately as it seems” in the literature. However, this experiment is really meant to illustrate the difficulty to guarantee reproducible results, but not to exhibit real accuracy problems. With some care, porting numerical software from a micro-controller to the Cloud, or directly writing applications to the Cloud, can be achieved provided certain recommendations we have defined are followed.

We built an automated, flexible testing framework designed to evaluate the numerical stability of an application across diverse configurations and Cloud platforms. By leveraging advanced DevOps practices, this environment enables:

• The evaluation of numerical stability under varying hardware and software setups, and deployment methods (containerized or native).
• Among the many available Cloud providers, this report focuses on cross-Cloud consistency tests conducted on Grid’5000, AWS, and Azure as a proof of concept.
• The preliminary work for cost-performance analyses to identify optimal Cloud platforms.

We created a scalable pipeline was implemented using tools such as Jenkins, Terraform, Ansible, Docker, and GitLab. The framework supports detailed configuration tests and generates structured outputs for further numerical and cost-effectiveness evaluations. This testing environment forms the backbone of future analytical efforts, enabling accurate and reproducible results across multiple configurations and Cloud environments.

8.2.3 A specialized model and implementation of an actuarial chatbot based on Federated Learning

In this work we focused on developing a language model specialized in the actuarial domain using modern machine learning techniques, including federated learning (FL). The primary objective is to create a chatbot based on large language models (LLMs) capable of meeting actuaries' specific needs in risk modeling, financial forecasting, and other technical tasks.

Training language models is a crucial process where parameters like batch size, learning rate, and optimizers are adjusted to improve performance. Specific fine-tuning techniques, like instruction-based adaptation and alignment with human preferences via Reinforcement Learning from Human Feedback (RLHF), are employed to tailor LLMs to the actuarial field.

In the context of this work, actuarial schools and other financial institutions become clients of the federated system, contributing to the training of a decentralized actuarial model. The central server aggregates updates from local models trained on each client's private data, ensuring data remains secure throughout the process. To optimize performance and reduce communication costs between clients and the server, techniques for compressing updates are applied, including sketched and structured updates.

In this exploratory work, we have gained expertise in LLMs and federated learning.

8.3.1 Measuring and Interpreting Dependent Task-based Applications Performances.

Participants: Thierry Gautier, Romain Pereira.

Breaking down the parallel time into work, idleness, and overheads is crucial for assessing the performance of HPC applications, but difficult to measure in asynchronous dependent tasking runtime systems. No existing tools allow its measurement portably and accurately. The paper 10 introduces POT: a tool-suite for dependent task-based applications performance measurement. We focus on its low-disturbance methodology consisting of task modeling, discrete-event tracing, and post-mortem simulation-based analysis. It supports the OMPT standard OpenMP specifications. The paper evaluates the precision of POT's parallel time breakdown analysis on LLVM and MPC implementations and shows that measurement bias may be neglected above 16µs workload per task, portably across two architectures and OpenMP runtime systems.

8.3.2 Evaluation of mix MPI + Dependent OpenMP task programming models on Fugaku.

Participants: Thierry Gautier, Romain Pereira.

The adoption of ARM processor architectures is on the rise in the HPC ecosystem. Fugaku supercomputer is a homogeneous ARMbased machine, and is one among the most powerful machine in the world. In the programming world, dependent task-based programming models are gaining tractions due to their many advantages: dynamic load balancing, implicit expression of communication/computation overlap, early-bird communication posting. MPI and OpenMP are two widespreads programming standards that make possible task-based programming at a distributed memory level. Despite its many advantages, mixed-use of the standard programming models using dependent tasks is still under-evaluated on large-scale machines. In the paper 11, we provide an overview on mixing OpenMP dependent tasking model with MPI with the state-of-the-art software stack (GCC-13, Clang17, MPC-OMP). We provide the level of performances to expect by porting applications to such mixed-use of the standard on the Fugaku supercomputers, using two benchmarks (Cholesky, HPCCG) and a proxy-application (LULESH). We show that software stack, resource binding and communication progression mechanisms are factors that have a significant impact on performance. On distributed applications, performances reaches up to 80% of effiency for task-based applications like HPCCG. We also point-out a few areas of improvements in OpenMP runtimes.

8.3.3 Handling dynamicity of HPC applications designed by a task-based component model

Participants: Jerry Lacmou Zeutouo, Christian Perez, Thierry Gautier, Romain Pereira.

We extended the Comet component model with the support of dynamic dependencies in its data-flow model. From a meta-task based data-flow, the Comet compiler generates the OpenMP code that will sumbit the tasks as well as the associated dependencies. The limitation of COMET was that these dependencies were to be known when submitting the tasks of all the data-flow. Hence, a task could not depend on a value compute by another task. We have extended the Comet model with the support of dynamic dependencies and have modified accordingly its runtime. A major difficulty was to generate the code that handle those dynamically-known dependencies under HPC constraints. We evaluated the relevance and performance of three models of dependencies (flat, nested, and weak dependencies) provided by OpenMP related runtimes (LLVM, MPC, and OmpSs-2).

8.3.4 New high level programming framework: Experiments with Kokkos

Participants: Thierry Gautier, Gabriel Suau.

The paper 2 describes the implementation of DONUT, a small multi-group S N -DG transport solver that aims at providing efficient and portable sweep kernels on shared-memory architectures for Cartesian and hexagonal geometries. DONUT heavily relies on the Kokkos C++ library for portability and genericity. First encouraging performance results are presented for multicore CPU architectures.

8.3.5 Taskgrind: Heavyweight Dynamic Binary Instrumentation for Parallel Programs Analysis

Participants: Romain Pereira.

Determinacy races are concurrent programming hazards occurring when two accesses on the same memory address are not ordered, and at least one is writing. Their presence hints at a correctness error, particularly under asynchronous task-based parallel programming models. The paper 12 introduces Taskgrind: a Valgrind tool for memory access analysis of parallel programming models such as Cilk or OpenMP. We illustrate the tool's capabilities with a determinacy-race analysis and confront it with state-of-the-art tools. Results show fewer false negatives and memory overheads on a set of microbenchmarks and LULESH, with meaningful error reports toward assisting programmers when parallelizing programs.

8.3.6 ETH4HPC Strategic Research Agenda

Participants: Christian perez.

This white paper 22 is released as part of the ETP4HPC’s Strategic Research Agenda 6. High-performance computing (HPC) applications achieve extreme levels of performance on large-scale systems by utilizing a wide range of tools, including compilers, runtime/middleware, APIs for memory access, debuggers and performance profilers, as well as high-level frameworks and domain-specific languages (DSLs).

Bosch

We have a collaboration with Bosch and AriC (a research team of the LIP laboratory, jointly supported by CNRS, ENS de Lyon, Inria and Université Claude Bernard (Lyon 1)). We conducted a study to provide guidelines for writing portable floating-point software in Cloud environments. With some care, porting numerical software from a micro-controller to the Cloud, or directly writing applications to the Cloud, can be eased with the help of some recommendations. It is in fact not more difficult than porting software from a micro-controller to any general-purpose processor.

CEA

We have a collaboration with CEA INSTN/SFRES / Saclay. This collaboration is based on the co-advising of a CEA PhD. The research of the PhD student (Gabriel Suau) focuses on high performance codes for neutron transport. One of the goal of the PhD is to work on better integration of Kokkos with a task based model.

Octo technology

We have a collaboration with Octo Technology (Part of Accenture). This collaboration is sealed through a CIFRE PhD grant. The research of the PhD student (Adrien Berthelot) focuses on accelerated and driven evaluation of the environmental impacts of an Information System with the full set of digital services

SynAAppS

We have a collaboration with SynAApps (part of Cyril Group). This collaboration is sealed through a CIFRE PhD grant. The research of the PhD student (Simon Lambert) focuses on forecast and dynamic resource provisioning on a virtualization infrastructure.

10.1.1 Participation in other International Programs

10.2.1 Visits of international scientists

10.3.1 Horizon Europe

10.3.2 Other european programs/initiatives

Priority Research Programmes and Equipments (PEPR)

ANR

French Joint Laboratory

Inria Large Scale Initiative

11.1.1 Scientific events: organisation

11.1.2 Scientific events: selection

11.1.3 Journal

11.1.4 Invited talks

• Christian Perez gave a talk about SLICES at 8th Symposium on Interoperability of Supercomputing and Cloud Technologies, SC'24, Atlanta, USA, November 22nd, 2024.
• Laurent Lefevre Perez gave the following talks :
  • "Energy efficiency and environmental impacts of large scale parallel and distributed systems - Exploring frugality at large scale", Seminar in University of Tromso, Norway, November 18, 2024
  • "Beaucoup de verrous, peu de leviers : à quand un numérique plus sobre ?", Françoise Berthoud, Laurent Lefevre, NEC2024: Numérique en Commun[s], Chambéry, September 25, 2024
  • "Contrer les impacts environnementaux des grandes infrastructures numériques : leviers, IA et cycle de vie", Laurent Lefevre, Adrien Berthelot, Mathilde Jay, Festival du Numérique Responsable, Ecole des Mines de St Etienne (online), March 19, 2024
  • "Accueil de groupes de lycéennes à l'Ecole Normale Supérieure de Lyon", Journée internationale des droits des femmes, École Normale Supérieure de Lyon, March 8, 2024
• Thierry Gautier gave a talk "XKBlas report on AMD multi-GPU and its evolutions" for the MUMPS 2024 annual meeting at Ansys, Villeurbanne, 2024-06-20
• Pierre-Etienne Polet gave a talk "Executing MUMPS & XKBlas on Unified Memory Platforms: Methodology and Initial Experiments on Nvidia Grace Hopper" at EDF, Paris, 2024-11-26

11.1.5 Scientific expertise

• Yves Caniou evaluated 2 projetcs for comité d'évaluation of the «CE46 - Calcul haute performance, Modèles numériques, simulation, applications» and «CE25 - Sciences et génie du logiciel - Réseaux de communication multi-usages, infrastructures numériques».
• Eddy Caron was a member of the ANR CE25 comittee and the ANR ASTRID comittee.
• Eddy Caron was a member of the committe PHELMA COS MCF 27 for the LIG (Grenoble).
• Élise Jeanneau was a member of the juries for recruiting CRCN candidates in the Inria Saclay and Inria Grenoble centers.
• Christian Perez evaluated 6 projects for the French Direction générale de la Recherche et de l’Innovation. He was a member of the jury for recruiting CRCN candidates in Inria Bordeaux center.

11.1.6 Research administration

• Eddy Caron is the leader of the SkyData ANR project.
• Élise Jeanneau is a member of the Inria Evaluation Comittee.
• Laurent Lefevre is the leader of the FrugalCloud challenge (défi) between Inria and OVHcloud company. He is a member of the executive board and the sites committee of the Grid’5000 Scientific Interest Group and member of the executive board and the sites committee of the SLICES-FR testbed.
• Christian Perez represents Inria in the overview board of the France Grilles Scientific Interest Group. He is a member of the executive board and the sites committee of the Grid’5000 Scientific Interest Group and member of the executive board of the SLICES-FR testbed. He is a member of the Inria Lyon Strategic Orientation Committee. He is in charge of organizing scientific collaborations between Inria and SKA France.

11.2.1 Teaching

• Licence: Yves Caniou, Algorithmique programmation impérative initiation, 129h, niveau L1, Université Claude Bernard Lyon 1, France.
• Licence: Yves Caniou, Algorithmique et programmation récursive, 54h, niveau L1, Université Claude Bernard Lyon 1, France.
• Licence: Yves Caniou, Programmation Concurrente, 24h and Co-Responsible of UE, niveau L3, Université Claude Bernard Lyon 1, France.
• Licence: Yves Caniou, Réseaux, 12h, niveau L3, Université Claude Bernard Lyon 1, France.
• Licence: Yves Caniou, Systèmes d'information documentaire, 22h, niveau L3, Université Claude Bernard Lyon 1, France.
• Master: Yves Caniou, Responsible of alternance students, 21h, niveau M1, Université Claude Bernard Lyon 1, France.
• Master: Yves Caniou, Sécurité, 24h and Responsible of UE, niveau M2, Université Claude Bernard Lyon 1, France.
• Master: Yves Caniou, Responsible of alternance students, 8h, niveau M2, Université Claude Bernard Lyon 1, France
• Master: Eddy Caron, Distributed System, 20h, M1, École Normale Supérieure de Lyon. France.
• Master: Eddy Caron, Langages, concepts et archi pour les données, 30h, M2, ISFA. Université Claude Bernard Lyon 1
• Master: Eddy Caron, Risques dans les Systèmes et Réseaux - Cloud, 15h, M2, ISFA. Université Claude Bernard Lyon 1.
• Master: Eddy Caron, Service Web et Sécurité, 15h, M2, ISFA. Université Claude Bernard Lyon 1.
• Master: Eddy Caron, Data Mining Avancé: Environnements parallèles et distribués, 12h, M2, ISFA. Université Claude Bernard Lyon 1.
• Licence: Élise Jeanneau, Introduction Réseaux et Web, 36h, niveau L1, Université Lyon 1, France.
• Licence: Élise Jeanneau, Réseaux, 54h, niveau L3, Université Lyon 1, France.
• Licence: Élise Jeanneau, Algorithmique, programmation et structures de données, 24h, niveau L2, Université Lyon 1, France
• Licence: Élise Jeanneau, Architecture des ordinateurs, 24h, niveau L2, Université Lyon 1, France
• Licence: Élise Jeanneau, Réseaux, systèmes et sécurité par la pratique, 23h, niveau L3, Université Lyon 1, France
• Master: Élise Jeanneau, Algorithmes distribués, 45h, niveau M1, Université Lyon 1, France.
• Master: Élise Jeanneau, Réseaux, 6h, niveau M1, Université Lyon 1, France.

11.2.2 Supervision

PhD defended:

• Mathilde Jay. "A Versatile Methodology for Assessing the Electricity Consumption and Environmental Footprint of Machine Learning Training: from Supercomputers to Edge Devices", 2021, Laurent Lefevre (co-dir. Inria. Avalon), Denis Trystram (dir. LIG, UGA), defended 2024, October 15th.
• Adrien Berthelot. The complete environmental footprint of digital usage: a contribution to life-cycle analysis of digital services", 2021, Eddy Caron (dir ENS de Lyon. Inria. Avalon), Laurent Lefevre (co-dir Inria. Avalon), Alexis Nicolas (Octo Technology), defended 2024, November 12th.
• Vladimir Ostapenco. "Modeling, evaluating and orchestrating heterogeneous leverages for large-scale cloud data centers management", 2021, Laurent Lefevre (dir. Inria. Avalon), Anne-Cécile Orgerie (co-dir. Inria. Myriads), Benjamin Fichel (co-dir. OVHcloud), defended 2024, December 18th.

Phd in progress:

• Maxime Agusti. "Observation de plate-formes de co-localisation baremetal, modèles de réduction énergétique et proposition de catalogues", FrugalCloud Inria-OVHCloud collaboration, Feb 2022, Eddy Caron (co-dir. ENS de Lyon. Inria. Avalon), Benjamin Fichel (co-dir. OVHcloud), Laurent Lefevre (dir. Inria. Avalon) et Anne-Cécile Orgerie (co-dir. Inria. Myriads),.
• Simon Lambert. "Forecast and dynamic resource provisioning on a virtualization infrastructure", 2022, Eddy Caron (dir. ENS de Lyon. Inria. Avalon), Laurent Lefevre (co-dir Inria. Avalon), Rémi Grivel (co-dir. Ciril Group).
• Gabriel Suaus. Résolution de l'équation de transport des neutrons sur des architectures massivement parallèles et hétérogènes : application aux géométries hexagonales. Thierry Gautier (dir. INRIA Avalon), Ansar CALLOO (co-dir. CEA), Romain LE TELLIER (co-dir CEA), Remi LE BARON (co-dir CEA).
• Thomas Stavis. "Replay of environmental leverages in cloud infrastructures and continuums", 2024, Laurent Lefevre (dir. Inria Avalon), Anne-Cécile Orgerie (co-dir CNRS, Magellan team, Irisa Rennes)
• Émile Egreteau-bruet. "Analyzing full life cycle of IoT based 5G solutions for smart agriculture", 2024, Laurent Lefevre (dir. Inria Avalon), Nathalie Mitton (co-dir. Inria Lille) and Doreid Ammar (co-dir. Aivancity Inria Avalon)

Eddy Caron was supervisor or co-supervisor of the following internship:

• Maxime Just (PLR 4th year of ENS). A study to provide a set of guidelines for calculation and cost-efficient of numerical algorithms in the cloud.
• Hamza Aabirouche (M2). Étude visant à établir des recommandations pour le calcul et l'optimisation de la stabilité des algorithmes numériques dans le Cloud
• Annour Saad Allamine (M2). SkyISFA : Adaptation d'un modèle spécialisé et mise en {oeœuvre d'un chatbot pour l'actuariat basé sur le Federated Learning.
• Maxime Just (M2). Évaluation et analyse du consensus dans un contexte de données autonomes. Co-supervised with Elise Jeanneau.

11.2.3 Juries

• Eddy Caron was member of the HDR defense committee of Antoine Boutet (INSA. Lyon) : "Privacy issues in AI and geolocation : from data protection to user awareness", Lyon, December 10th, 2024.
• Eddy Caron was PhD reviewer and member of the defense committe of Thomas Bouvier (IRISA Rennes): "Distributed Rehearsal Buffers for Continual Learning at Scale", Rennes, November 4th, 2024.
• Eddy Caron was member of the Phd defense committee of Sylvain Lejamble (Université Mont Blanc Savoie. Annecy): "Conception et mise en œvre d'un système de représentation et de manipulation de la connaissance des Objets Sages (WO) pour la détection de fraudes au sein des systèmes bancaires". Annecy. December 12th, 2024.
• Laurent Lefevre was PhD reviewer of the PhD and member of the defense committe of :
  • Manal Benaissa : "Optimization of the IT and Electrical Sizing Process of a Green, Medium-Sized Datacenter Powered Exclusively by Renewable Energy Sources, Considering Environmental and External Uncertainty", University of Franche-Comté, Besançon, December 10, 2024
  • Vincent Lannurien : "Allocation et placement dynamiques sur ressources hétérogènes pour le cloud serverless", ENSTA Bretagne, Brest, France, November 20, 2024
  • Mohammad Sadegh Aslanpour : "Serverless Edge Computing: Resource Scheduling Algorithms and Software Systems", Monash University, Australia, January 8, 2024
• Laurent Lefevre was member of the defense committee of Klervie Toczé : "Orchestrating a Resource-aware Edge", Linkoping University, Sweden, October 4, 2024
• Christian Perez was reviewer and member of the HDR defense committee of Simon Bliudze: "Rigorous Design of Concurrent Component-Based Software and Systems", Lille, March 8th, 2024.
• Christian Perez was PhD reviewer and member of the defense committe of Ophélie Renaud: "Model Based Granularity Optimization for High Performance Computing Systems in Astronomy", Rennes, October 9th, 2024.

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

• Yves Caniou co-organized the Campus du libre with institutes Université Claude Bernard Lyon 1, Université Lyon 2, Université Lyon 3, INSA, Inria.
• Laurent Lefevre participated in the Pop'Sciences project from University of Lyon for the comic book on with middle and high school students from Lyon and Villeurbanne, April 2024

11.3.2 Participation in Live events

• Laurent Lefevre was panelist in the GreenTech Forum 2024 event: "Avancées et nouvelles technologies pour des datacenters plus responsables / Advances and new technologies for more responsible data centers", GreenTech Forum, Paris, France, November 6, 2024

International journals

• 1 articleA.Adrien Berthelot, E.Eddy Caron, M.Mathilde Jay and L.Laurent Lefevre. Understanding the environmental impact of generative AI services.Communications of the ACMSpecial Issue on Sustainability and Computing2025. In press. HALback to text
• 2 articleG.Gabriel Suau, A.Ansar Calloo, R.Rémi Baron, R.Romain Le Tellier and T.Thierry Gautier. Efficient sweep kernels on shared-memory architectures for the discrete ordinates neutron transport equation on Cartesian and hexagonal geometries.EPJ Web of Conferences302October 2024, 02009HALDOIback to text

International peer-reviewed conferences

• 3 inproceedingsM.Maxime Agusti, E.Eddy Caron, B.Benjamin Fichel, L.Laurent Lefèvre, O.Olivier Nicol and A.-C.Anne-Cécile Orgerie. PowerHeat: A non-intrusive approach for estimating the power consumption of bare metal water-cooled servers.2024 IEEE International Conferences on Internet of Things (iThings) and IEEE Green Computing & Communications (GreenCom) and IEEE Cyber, Physical & Social Computing (CPSCom) and IEEE Smart Data (SmartData) and IEEE Congress on CybermaticsCopenhagen, Denmark2024, 1-7HALback to text
• 4 inproceedingsA.Adrien Berthelot, E.Eddy Caron, M.Mathilde Jay and L.Laurent Lefèvre. Estimating the environmental impact of Generative-AI services using an LCA-based methodology.Procedia CIRPCIRP LCE 2024 - 31st Conference on Life Cycle EngineeringTurin, Italy2024, 1-10HALback to text
• 5 inproceedingsF.Françoise Berthoud, M.Martine Olivi and L.Laurent Lefèvre. Beaucoup de verrous, peu de leviers ! Sobriété numérique : le cas est grave mais pas désespéré.NUDGIS by UBICASTJRES 2024 - Journées réseaux de l'enseignement et de la rechercheRennes, FranceDecember 2024HALback to text
• 6 inproceedingsC.Christophe Cérin, M.Mathilde Jay and L.Laurent Lefèvre. A Methodology and a Toolbox to Explore Dataset related to the Environmental Impact of HTTP Requests.2023 IEEE International Conference on Big Data (BigData) - 3rd International Workshop on Big Data Analytics for SustainabilitySorrento (Naples), ItalyIEEE2024, 1-10HALDOIback to text
• 7 inproceedingsH.Hugo Hadjur, D.Doreid Ammar and L.Laurent Lefèvre. Deep Reinforcement Learning for Energy-efficient Selection of Embedded Services at the Edge.2024 IEEE International Conferences on Internet of Things (iThings)2024 IEEE International Conferences on Internet of Things (iThings)Copenhagen, DenmarkIEEEAugust 2024, 67-74HALDOIback to text
• 8 inproceedingsS.Simon Lambert, E.Eddy Caron, L.Laurent Lefèvre and R.Rémi Grivel. S-ORCA : A social-based consolidation approach to reduce Cloud infrastructures energy consumption.15th IEEE International Conference on Cloud Computing Technology and Science - Cloudcom 2024Abu Dhabi, United Arab EmiratesDecember 2024HALback to text
• 9 inproceedingsV.Vladimir Ostapenco, L.Laurent Lefèvre, A.-C.Anne-Cécile Orgerie and B.Benjamin Fichel. Exploring RAPL as a Power Capping Leverage for Power-Constrained Infrastructures.ICA3PP 2024 - 24th International Conference on Algorithms and Architectures for Parallel ProcessingMacau SAR, China2024, 1-10HALback to text
• 10 inproceedingsR.Romain Pereira, T.Thierry Gautier, A.Adrien Roussel and P.Patrick Carribault. Measuring and Interpreting Dependent Task-based Applications Performances.Parallel Processing and Applied Mathematics 202415th International Conference on Parallel Processing & Applied MathematicsOstrava, Czech RepublicApril 2025HALback to text
• 11 inproceedingsR.Romain Pereira, A.Adrien Roussel, M.Miwako Tsuji, P.Patrick Carribault, M.Mitsuhisa Sato, H.Hitoshi Murai and T.Thierry Gautier. An Overview on Mixing MPI and OpenMP Dependent Tasking on A64FX.International Workshop on Arm-based HPC 2024 (IWAHPCE-2024)Nagoya, JapanACMJanuary 2024, 7-16HALDOIback to text
• 12 inproceedingsR.Romain Pereira, G.George Stelle and P.Patrick Carribault. Taskgrind: Heavyweight Dynamic Binary Instrumentation for Parallel Programs Analysis.International Conference for High Performance Computing, Networking, Storage and Analysis (SC)8th International Workshop on Software Correctness for HPC Applications (Correctness '24)Atlanta (USA), United StatesNovember 2024HALback to text
• 13 inproceedingsT.Thibault Simon, D.David Ekchajzer, A.Adrien Berthelot, E.Eric Fourboul, S.Samuel Rince and R.Romain Rouvoy. BoaviztAPI: a bottom-up model to assess the environmental impacts of cloud services.HotCarbon'24 - 3rd Workshop on Sustainable Computer SystemsSanta Cruz, United StatesJuly 2024HAL

Conferences without proceedings

• 14 inproceedings S.Sylvain Bouveret, E.Emmanuelle Frenoux, L.Laurent Lefèvre, D.Didier Mallarino and D.Denis Trystam. IA générative sobre : un oxymore ? JRES (Journées réseaux de l'enseignement et de la recherche ) 2024 Rennes, France December 2024 HAL back to text

Scientific book chapters

• 15 inbookL.Lucien Ngale, E.Eddy Caron and Y.Yulin Zhang. Fog-Robotics Infrastructures Simulation-Based Sizing Approach.1845Cloud Computing and Services Science : 12th International Conference, CLOSER 2022, Virtual Event, April 27–29, 2022, and 13th International Conference, CLOSER 2023, Prague, Czech Republic, April 26–28, 2023, Revised Selected PapersCommunications in Computer and Information ScienceSpringer Nature SwitzerlandAugust 2024, 188-211HALDOI

Doctoral dissertations and habilitation theses

• 16 thesisA.Adrien Berthelot. The complete environmental footprint of digital usage : a contribution to life-cycle analysis of digital services.Ecole normale supérieure de lyon - ENS LYONNovember 2024HALback to text
• 17 thesisM.Mathilde Jay. A Versatile Methodology for Assessing the Electricity Consumption and Environmental Footprint of Machine Learning Training: from Supercomputers to Edge Devices.Université Grenoble AlpesOctober 2024HALback to text
• 18 thesisV.Vladimir Ostapenco. Modeling, evaluating and orchestrating heterogeneous leverages for large-scale cloud data centers management.Ecole Normal Supérieure de Lyon - ENS LyonDecember 2024HALback to text
• 19 thesisP.-E.Pierre-Etienne Polet. Porting sonar processing chains to heterogeneous architecture : design and evaluation of a moldable task-based programming environment.Ecole normale supérieure de lyon - ENS LYONApril 2024HAL

Reports & preprints

• 20 miscA.Adrien Berthelot, E.Eddy Caron, R.Romain de Laage, L.Laurent Lefèvre and A.Alexis Nicolas. Fine-grained methodology to assess environmental impact of a set of digital services.2024HALback to text

Other scientific publications

• 21 inproceedingsA.Adrien Berthelot, E.Eddy Caron, M.Mathilde Jay and L.Laurent Lefèvre. Towards a multi-criteria evaluation of the environmental footprint of generative ai services.ICT4S 2024 - International Conference on Information and Communications Technology for SustainabilityStockholm, Sweden2024, 1-1HALback to text
• 22 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 2024HALDOIback to text
• 23 inproceedingsS.Simon Lambert, E.Eddy Caron, L.Laurent Lefèvre and R.Rémi Grivel. Evaluating the involvement of users for reducing energy consumption of datacenters in a Cloud company.ICT4S 2024 - ICT for SustainabilityStockholm, SwedenJune 2024HALback to text