2025Activity reportProject-Team‌​‌ROMA

7.1.1​‌ MatchMaker

• Name:
  Maximum matchings​​ in bipartite graphs
• Keywords:​​
  Graph algorithmics, Matching
• Scientific​​​‌ Description:
  The implementations of‌ ten exact algorithms and‌​‌ four heuristics for solving​​ the problem of finding​​​‌ a maximum cardinality matching‌ in bipartite graphs are‌​‌ provided.
• Functional Description:
  This​​ software provides algorithms to​​​‌ solve the maximum cardinality‌ matching problem in bipartite‌​‌ graphs.
• Release Contributions:
  This​​ version includes the initialization​​​‌ algorithm/heuristic 5 and moves‌ files in suitable directories.‌​‌
• URL:
  https://­gitlab.­inria.­fr/­bora-ucar/­matchmaker
• Publications:
  hal-02463717​​, hal-00786548, hal-00763920​​​‌
• Contact:
  Bora Uçar
• Participants:‌
  Ioannis Panagiotas, Bora Uçar,‌​‌ Kamer Kaya, Johannes Langguth​​

8.1.1​​ Fixed-Work vs. Fixed-Time Checkpointing​​​‌ on Large-Scale Failure-Prone Platforms‌

Participants: Quentin Barbut,‌​‌ Lucas Perotin [Vanderbilt University]​​, Anne Benoit,​​​‌ Thomas Hérault [University of‌ Tennessee, Knoxville], Yves‌​‌ Robert, Frédéric Vivien​​.

Consider a High-Performance​​​‌ Computing (HPC) application executing‌ on a large-scale failure-prone‌​‌ platform. The Fixed-Work Checkpointing​​ (FWC) problem consists in​​​‌ minimizing the expected time‌ to execute a fixed‌​‌ amount of work (namely​​ a fraction or the​​​‌ totality of the application).‌ Strategies for the FWC‌​‌ problem have received considerable​​ attention and are well-understood.​​​‌ On the contrary, the‌ dual problem, namely the‌​‌ Fixed-Time Checkpointing (FTC) problem,​​ has been considered only​​​‌ very recently. The FTC‌ problem consists in maximizing‌​‌ the expected work achieved​​ during a fixed amount​​​‌ of time (namely the‌ duration of a reservation‌​‌ granted to the application).​​ This work provides a​​​‌ comparative overview of both‌ problems. First we review‌​‌ existing strategies for the​​ FWC problem and extend​​​‌ them to stochastic checkpoints,‌ i.e., when the checkpoint‌​‌ is no longer a​​ deterministic constant but obeys​​​‌ some probability distribution law‌ instead. Then we provide‌​‌ a comprehensive study of​​ the FTC problem. The​​​‌ problem turns out to‌ be surprisingly difficult, even‌​‌ when restricting to taking​​ one or two checkpoints.​​​‌ We provide a threshold-based‌ heuristic to solve the‌​‌ general instance of the​​ problem with an arbitrary​​​‌ number of checkpoints, and‌ we have to resort‌​‌ to time discretization to​​ provide an optimal strategy.​​​‌ We further extend this‌ latter strategy to stochastic‌​‌ checkpoints.

This work has​​ been published in the​​​‌ International Journal of High‌ Performance Computing Applications 8‌​‌.

8.1.2 Partial Detectors​​ Versus Replication To Cope​​​‌ With Silent Errors

Participants:‌ Anne Benoit, Thomas‌​‌ Hérault [Inria Bordeaux],​​ Yves Robert, Alix​​​‌ Tremodeux.

This work‌ studies an iterative algorithm‌​‌ running on an error-prone​​ platform, where silent errors​​​‌ strike each iteration with‌ some probability. A detector‌​‌ verifies correctness before taking​​ a checkpoint but may​​​‌ fail to detect errors.‌ Specifically, an error at‌​‌ iteration I is detected​​ only after iteration (I​​​‌ - 1) + X,‌ where X follows a‌​‌ bounded probability distribution like​​ a truncated geometric distribution.​​​‌ Intuitively, the error silently‌ amplifies during some iterations‌​‌ before it can be​​ detected at distance X​​​‌ or higher, and there‌ is the risk of‌​‌ missing an error that​​​‌ has struck recently but​ cannot be detected yet.​‌ X is bounded by​​ D, the maximum detection​​​‌ latency. To mitigate undetected​ errors during verification, a​‌ simple strategy keeps two​​ checkpoints and divides the​​​‌ execution into D -​ 1 iteration segments, each​‌ followed by verification and​​ checkpoint. In steady state:​​​‌ (i) if verification succeeds,​ the oldest checkpoint is​‌ erased and replaced; (ii)​​ if it fails, rollback​​​‌ occurs to the oldest​ verified checkpoint. This work​‌ explores whether this scheme​​ outperforms replication and determines​​​‌ the optimal number of​ checkpoints and segment lengths,​‌ both theoretically and via​​ Monte Carlo simulations.

This​​​‌ work was presented at​ the EuroPar 2025 conference​‌ 18.

8.1.3 Fault-tolerant​​ numerical iterative algorithms at​​​‌ scale

Participants: Alix Tremodeux​, Anne Benoit,​‌ Emmanuel Agullo [Inria Bordeaux]​​, Thomas Hérault [Inria​​​‌ Bordeaux; University of Tennessee,​ Knoxville], Luc Giraud​‌ [Inria Bordeaux], Yves​​ Robert.

This work​​​‌ investigates how to protect​ numerical iterative algorithms from​‌ all types of errors​​ that can strike at​​​‌ scale: fail-stop errors (a.k.a.​ failures) and silent errors,​‌ striking both as computation​​ errors and memory bit-flips.​​​‌ We combine various techniques:​ detectors for computation errors,​‌ checksums for memory errors,​​ and checkpoint/restart for failures.​​​‌ The objective is to​ minimize the expected time​‌ per iteration of the​​ algorithm. We design a​​​‌ hierarchical pattern that combines​ and interleaves all these​‌ fault-tolerance mechanisms, and we​​ determine the optimal periodic​​​‌ pattern that achieves this​ objective. We instantiate these​‌ results for the performance​​ analysis of the Preconditioned​​​‌ Conjugate Gradient (PCG) algorithm:​ we report several scenarios​‌ where the optimal pattern​​ dramatically decreases the overhead​​​‌ due to error mitigation.​

This work has been​‌ published in the International​​ Journal of High Performance​​​‌ Computing Applications 13.​

8.2.1 Green​​ Scheduling on the Edge​​​‌

Participants: Joachim Cendrier,​ Rajini Wijayawardana [University of​‌ Chicago], Anne Benoit​​, Yves Robert,​​​‌ Frédéric Vivien, Andrew​ A. Chien [University of​‌ Chicago].

This work​​ aims at designing and​​​‌ evaluating scheduling algorithms that​ minimize carbon cost on​‌ edge platforms. When a​​ job is released to​​​‌ some edge server, difficult​ scheduling questions arise: should​‌ the job be executed​​ on that server? If​​​‌ yes, when? If no,​ which other edge server​‌ should the job be​​ transferred to? Typically, jobs​​​‌ are submitted online, and​ have a deadline to​‌ enforce. Online scheduling problems​​ are already difficult without​​​‌ accounting for different energy​ sources, so one should​‌ not expect any optimal​​ solution. Still, an important​​​‌ research goal is to​ revisit standard algorithms such​‌ as Earliest Completion Time​​ (ECT) and Earliest Deadline​​​‌ First (EDF) in order​ to design and evaluate​‌ carbon-aware variants. This work​​ introduces several new algorithms​​​‌ that use sophisticated scheduling​ policies to efficiently decrease​‌ carbon cost; these algorithms​​ maximize the use of​​ green energy both on​​​‌ local and remote edge‌ servers, by re-evaluating previous‌​‌ decisions whenever needed to​​ accommodate newly released jobs.​​​‌ We provide a comprehensive‌ simulation campaign based on‌​‌ actual platform/job data and​​ carbon traces and report​​​‌ an average gain of‌ 42% over standard approaches.‌​‌

This work was presented​​ at the EuroPar 2025​​​‌ conference 19.

8.2.2‌ Carbon-Aware Workflow Scheduling with‌​‌ Fixed Mapping and Deadline​​ Constraint

Participants: Dominik Schweisgut​​​‌ [Humboldt University Berlin, Karlsruhe‌ Institute of Technology],‌​‌ Anne Benoit, Yves​​ Robert, Henning Meyerhenke​​​‌ [Karlsruhe Institute of Technology]‌.

Large data and‌​‌ computing centers consume a​​ significant share of the​​​‌ world's energy consumption. A‌ prominent subset of the‌​‌ workloads in such centers​​ are workflows with interdependent​​​‌ tasks, usually represented as‌ directed acyclic graphs (DAGs).‌​‌ To reduce the carbon​​ emissions resulting from executing​​​‌ such workflows in centers‌ with a mixed (renewable‌​‌ and non-renewable) energy supply,​​ it is advisable to​​​‌ move task executions to‌ time intervals with sufficient‌​‌ green energy when possible.​​ To this end, we​​​‌ formalize the above problem‌ as a scheduling problem‌​‌ with a given mapping​​ and ordering of the​​​‌ tasks. We show that‌ this problem can be‌​‌ solved in polynomial time​​ in the uniprocessor case.​​​‌ For at least two‌ processors, however, the problem‌​‌ becomes NP-hard. Hence, we​​ propose a heuristic framework​​​‌ called CaWoSched that combines‌ several greedy approaches with‌​‌ local search. To assess​​ the 16 heuristics resulting​​​‌ from different combinations, we‌ also devise a simple‌​‌ baseline algorithm and an​​ exact ILP-based solution. Our​​​‌ experimental results show that‌ our heuristics provide significant‌​‌ savings in carbon emissions​​ compared to the baseline.​​​‌

This work was presented‌ at the ICPP'25 conference‌​‌ 24.

8.2.3 Deadline-Aware​​ Scheduling of Mixed-Criticality Tasks​​​‌

Participants: Maxime Gonthier [University‌ of Chicago], Kyle‌​‌ Chard [University of Chicago]​​, Ian Foster [University​​​‌ of Chicago], Loris‌ Marchal, Frédéric Vivien‌​‌.

High-performance computing centers​​ and cloud providers host​​​‌ a wide variety of‌ workloads, ranging from routine‌​‌ calibration tasks with no​​ strict timing requirements to​​​‌ urgent real-time computations that‌ must be completed within‌​‌ hard deadlines. Traditional approaches​​ reserve resources for high-criticality​​​‌ tasks or preempt and‌ kill lower-criticality tasks when‌​‌ necessary, resulting in wasted​​ compute time and longer​​​‌ turnaround times for lower-criticality‌ tasks. We suggest that‌​‌ a better solution is​​ to interleave the execution​​​‌ of critical and non-critical‌ tasks. We formulate a‌​‌ bi-objective optimization problem: guarantee​​ that all critical tasks​​​‌ meet their deadlines, and‌ minimize the maximum flow,‌​‌ defined as the time​​ a task spends in​​​‌ the system, of non-critical‌ tasks. We introduce a‌​‌ formal model, derive an​​ approximation algorithm and a​​​‌ lower bound, and develop‌ several heuristics based on‌​‌ the approximation framework. Through​​ extensive simulations, based on​​​‌ synthetic and real-world workloads,‌ we show that one‌​‌ of our heuristics reduces​​ the maximum flow of​​​‌ non-critical tasks by up‌ to 14% compared to‌​‌ static resource partitioning.

This​​ work was presented at​​​‌ the ICPP'25 conference 21‌.

8.2.4 Memory-aware Adaptive‌​‌ Scheduling of Scientific Workflows​​​‌ on Heterogeneous Architectures

Participants:​ Svetlana Kulagina [Humboldt University​‌ Berlin], Anne Benoit​​, Henning Meyerhenke [Humboldt​​​‌ University Berlin].

The​ analysis of massive scientific​‌ data often happens in​​ the form of workflows​​​‌ with interdependent tasks. When​ such a scientific workflow​‌ needs to be scheduled​​ on a parallel or​​​‌ distributed system, one usually​ represents the workflow as​‌ a directed acyclic graph​​ (DAG). The vertices of​​​‌ the DAG represent the​ tasks, while its edges​‌ model the dependencies between​​ the tasks (usually data​​​‌ to be communicated to​ successor tasks). When executed,​‌ each task requires a​​ certain amount of memory​​​‌ and if that exceeds​ the available memory, the​‌ execution fails. The typical​​ goal is to execute​​​‌ the workflow without failures​ (i.e., satisfying the memory​‌ constraints) and with the​​ shortest possible execution time​​​‌ (i.e., to minimize its​ makespan). To address this​‌ problem, we investigate the​​ memory-aware scheduling of DAG-shaped​​​‌ workflows on heterogeneous plat-​ forms, where each processor​‌ can have a different​​ speed and a different​​​‌ memory size. We propose​ a variant of HEFT​‌ (Heterogeneous Earliest Finish Time)​​ that, in contrast to​​​‌ the original, accounts for​ memory and includes eviction​‌ strategies for cases when​​ it might be beneficial​​​‌ to remove some data​ from memory in order​‌ to have enough memory​​ to execute other tasks.​​​‌ Furthermore, while HEFT assumes​ perfect knowledge of the​‌ execution time and memory​​ usage of each task,​​​‌ the actual values might​ differ upon execution. Thus,​‌ we propose an adaptive​​ scheduling strategy, where a​​​‌ schedule is recomputed when​ there has been a​‌ significant variation in terms​​ of execution time or​​​‌ memory. The scheduler has​ been closely integrated with​‌ a runtime system, allowing​​ us to perform a​​​‌ thorough experimental evaluation on​ real-world workflows. The runtime​‌ system warns the scheduler​​ when the task parameters​​​‌ have changed, and a​ schedule can be recomputed​‌ on the fly. The​​ memory-aware strategy allows us​​​‌ to schedule task graphs​ that would run out​‌ of memory with a​​ state-of-the-art scheduler, and the​​​‌ adaptive setting allows us​ to significantly reduce the​‌ makespan.

This work was​​ presented at the CCGrid​​​‌ 2025 conference 22.​

8.2.5 A scheduler to​‌ foster data locality for​​ GPU and out-of-core task-based​​​‌ linear algebra applications

Participants:​ Maxime Gonthier [University of​‌ Chicago], Loris Marchal​​, Samuel Thibault [Inria​​​‌ Bordeaux].

Hardware accelerators​ like GPUs now provide​‌ a large part of​​ the computational power used​​​‌ for scientific simulations. Despite​ their efficacy, GPUs possess​‌ limited memory and are​​ connected to the main​​​‌ memory of the machine​ via a bandwidth limited​‌ bus. Scientific simulations often​​ operate on very large​​​‌ data, that surpasses the​ GPU's memory capacity. Therefore,​‌ one has to turn​​ to out-of-core computing: data​​​‌ is kept in a​ remote, slower memory (CPU​‌ memory), and moved back​​ and forth from/to the​​​‌ device memory (GPU memory),​ a process also present​‌ for multicore CPUs with​​ limited memory. In both​​​‌ cases, data movement quickly​ becomes a performance bottleneck.​‌ Task-based runtime schedulers have​​ emerged as a convenient​​ and efficient way to​​​‌ manage large applications on‌ such heterogeneous platforms. We‌​‌ propose a scheduler for​​ task-based runtimes that improves​​​‌ data locality for out-of-core‌ linear algebra computations, to‌​‌ reduce data movement. We​​ design a data-aware strategy​​​‌ for both task scheduling‌ and data eviction from‌​‌ limited memories. We compare​​ this scheduler to existing​​​‌ schedulers in runtime systems.‌ Using StarPU, we show‌​‌ that our new scheduling​​ strategy achieves comparable performance​​​‌ when memory is not‌ a constraint, and significantly‌​‌ better performance when application​​ input data exceeds memory,​​​‌ on both GPUs and‌ CPU cores.

This work‌​‌ has been published in​​ the Journal of Parallel​​​‌ and Distributed Computing 10‌.

8.2.6 Cache Management‌​‌ for Mixture-of-Experts LLMs

Participants:​​ Spyros Angelopoulos [IRL ILLS]​​​‌, Loris Marchal,‌ Adrien Obrecht, Bertrand‌​‌ Simon [CNRS IN2P3].​​

Large language models (LLMs)​​​‌ have demonstrated remarkable capabilities‌ across a variety of‌​‌ tasks. One of the​​ main challenges towards the​​​‌ successful deployment of LLMs‌ is memory management, since‌​‌ they typically involve billions​​ of parameters. To this​​​‌ end, architectures based on‌ Mixture-of-Experts have been proposed,‌​‌ which aim to reduce​​ the size of the​​​‌ parameters that are activated‌ when producing a token.‌​‌ This raises the equally​​ critical issue of efficiently​​​‌ managing the limited cache‌ of the system, in‌​‌ that frequently used experts​​ should be stored in​​​‌ the fast cache rather‌ than in the slower‌​‌ secondary memory. In this​​ work, we introduce and​​​‌ study a new paging‌ problem that models expert‌​‌ management optimization. Our formulation​​ captures both the layered​​​‌ architecture of LLMs and‌ the requirement that experts‌​‌ are cached efficiently. We​​ first present lower bounds​​​‌ on the competitive ratio‌ of both deterministic and‌​‌ randomized algorithms, which show​​ that under mild assumptions,​​​‌ LRU-like policies have good‌ theoretical competitive performance. We‌​‌ then propose a layer-based​​ extension of LRU that​​​‌ is tailored to the‌ problem at hand. Extensive‌​‌ simulations on both synthetic​​ datasets and actual traces​​​‌ of MoE usage show‌ that our algorithm outperforms‌​‌ policies for the classic​​ paging problem, such as​​​‌ the standard LRU.

This‌ work was presented at‌​‌ the EuroPar 2025 conference​​ 16.

8.2.7 Leveraging​​​‌ Expert Usage to Speed‌ up LLM Inference with‌​‌ Expert Parallelism

Participants: Olivier​​ Beaumont [Inria Bordeaux],​​​‌ Raphaël Bourgouin [Inria Bordeaux]‌, Maxime Darrin [IRL‌​‌ ILLS], Loris Marchal​​, Pablo Piantanida [IRL​​​‌ ILLS].

Large language‌ models have become indispensable‌​‌ for many text-processing applications.​​ Their inference, i.e., their​​​‌ use to generate text,‌ is a time-consuming task‌​‌ since tokens have to​​ be generated one after​​​‌ the other, even if‌ the computational load has‌​‌ been reduced by model​​ sparsification, e.g., by using​​​‌ a Mixture of Experts‌ (MoE) models. In the‌​‌ MoE context, a subset​​ of experts is selected​​​‌ at each stage. Note‌ that not all subsets‌​‌ of experts (pairs of​​ experts in most cases)​​​‌ in a given layer‌ have the same probability‌​‌ of being selected. When​​ experts are mapped to​​​‌ different GPUs, there is‌ a risk of load‌​‌ imbalance if the selected​​​‌ experts end up on​ a small number of​‌ GPUs. This work proposes​​ to leverage this heterogeneity​​​‌ in expert usage to​ map experts of popular​‌ subsets onto distinct GPUs,​​ allowing them to be​​​‌ processed in parallel and​ thus reducing the time​‌ needed for inference. Even​​ though this mapping problem​​​‌ is NP-complete, it is​ possible to design simple​‌ greedy strategies that significantly​​ reduce the need for​​​‌ sequential expert processing. Our​ proof-of-concept confirms that our​‌ mapping strategies effectively reduce​​ inference time on the​​​‌ Mixtral model.

This work​ was presented at the​‌ EuroPar 2025 conference 17​​.

8.2.8 Towards Parallel​​​‌ Transformer-Based Large Language Models​ for Fast Inference

Participants:​‌ Gaël Dellaleau [Kog company]​​, Morgan Giraud [Kog​​​‌ company], Loris Marchal​, Félix Wirth [Kog​‌ company].

Transformer-based models​​ have reached quality levels​​​‌ that make them attractive​ for a wide range​‌ of applications, including those​​ demanding faster generation speeds​​​‌ such as multimedia artifact​ generation. However, traditional parallelism​‌ techniques have struggled to​​ meet these speed requirements​​​‌ due to the inherently​ sequential structure of Transformers,​‌ which limits parallel execution.​​ In this work, we​​​‌ have proposed a simple​ modeling to the parallel​‌ execution of Transformer, which​​ identifies synchronization times across​​​‌ inference devices as the​ culprit. Leveraging real world​‌ figures from GPU manufacturers,​​ we highlight the increasing​​​‌ cost of this synchronization​ time with ever more​‌ powerful hardware. In response,​​ we introduce Tensor Parallel​​​‌ Blocks, a naturally parallelized​ architecture. To maintain quality,​‌ we incorporate multiple enhancements,​​ such as delayed inter-block​​​‌ communication, hybrid architecture, and​ different parallelism gateways, which​‌ mitigate potential degradation in​​ model output. We design​​​‌ and validate a dedicated​ training protocol tailored to​‌ this architecture and conduct​​ a preliminary hyperparameter study​​​‌ on 150M models, proving​ hardware-tied architecture can achieve​‌ relevant results. Using insights​​ from these trials, we​​​‌ establish a set of​ rules for scaling the​‌ model to larger size,​​ theorizing a 4.2x speedup​​​‌ in decoding time compared​ to an equivalently sized​‌ LLaMA based model on​​ the target inference platform.​​​‌

This work is detailed​ in a research report​‌ 33 describing the work​​ initiated during the 6​​​‌ month internship of Félix​ Wirth in the ROMA​‌ team and continued in​​ collaboration with Gael Delalleau​​​‌ and Morgan Giraud from​ the Kog company.

8.2.9​‌ Efficient and effective methods​​ for variant selection

Participants:​​​‌ Srinivas Aluru [Georgia Institute​ of Technology], Anne​‌ Benoit, Bora Uçar​​.

Variation graphs succinctly​​​‌ capture genetic variations among​ individuals within a species​‌ or a target population.​​ The use of variation​​​‌ graphs instead of a​ single reference genome is​‌ credited with reducing bias​​ and increasing the accuracy​​​‌ of sequence mapping algorithms.​ However, complete variation graphs​‌ that comprehensively incorporate all​​ genetic variations are often​​​‌ found to be ineffective​ and inaccurate in practice​‌ due to the presence​​ of a combinatorially explosive​​​‌ number of paths in​ the graph that do​‌ not correspond to any​​ observed genome. Thus, a​​​‌ balance is struck in​ carefully selecting a subset​‌ of variants to be​​ incorporated, for which mathematical​​ frameworks have recently been​​​‌ developed. We advance the‌ mathematical framework proposed by‌​‌ Jain et al., where​​ integer linear programming formulations​​​‌ were developed for optimal‌ variant selection. We propose‌​‌ novel graph-based formulations and​​ develop exact and fast​​​‌ algorithms for certain cases,‌ approximation methods for some‌​‌ others, and empirically close​​ to optimal results in​​​‌ all cases. The primary‌ advantage of the algorithms‌​‌ designed here is that​​ they provide near-optimal results​​​‌ at orders of magnitude‌ faster run time of‌​‌ an ILP solver.

This​​ work was presented at​​​‌ the 16th ACM Conference‌ on Bioinformatics, Computational Biology,‌​‌ and Health Informatics 15​​.

8.3.1 Efficient Parallel Sparse‌ Tensor Contraction

Participants: Somesh‌​‌ Singh, Bora Uçar​​.

We investigate the​​​‌ performance of algorithms for‌ sparse tensor-sparse tensor multiplication‌​‌ (SpGETT). This operation, also​​ called sparse tensor contraction,​​​‌ is a higher order‌ analogue of the sparse‌​‌ matrix-sparse matrix multiplication (SpGEMM)​​ operation. Therefore, SpGETT can​​​‌ be performed by first‌ converting the input tensors‌​‌ into matrices, then invoking​​ high performance variants of​​​‌ SpGEMM, and finally reconverting‌ the resultant matrix into‌​‌ a tensor. Alternatively, one​​ can carry out the​​​‌ scalar operations underlying SpGETT‌ in the realm of‌​‌ tensors without matrix formulation.​​ We discuss the building​​​‌ blocks in both approaches‌ and formulate a hashing-based‌​‌ method to avoid costly​​ search or redirection operations.​​​‌ We present performance results‌ with the current state-of-the-art‌​‌ SpGEMM-based approaches, existing SpGETT​​ approaches, and a carefully​​​‌ implemented SpGETT approach with‌ a new fine-tuned hashing‌​‌ method, proposed in this​​ work. We evaluate the​​​‌ methods on real world‌ tensors, contracting a tensor‌​‌ with itself along various​​ dimensions. Our proposed hashing-based​​​‌ method for SpGETT consistently‌ outperforms the state-of-the-art method,‌​‌ achieving a 25% reduction​​ in sequential execution time​​​‌ on average and a‌ 21% reduction in parallel‌​‌ execution time on average​​ across a variety of​​​‌ input instances.

This work‌ has been published in‌​‌ the IEEE Transactions on​​ Parallel and Distributed Systems​​​‌ journal 12. Codes‌ are available here.‌​‌

8.3.2 Semi-Streaming Algorithms for​​ Hypergraph Matching

Participants: Henrik​​​‌ Reinstädtler [Heidelberg University],‌ S.M. Ferdous [PNNL -‌​‌ Pacific Northwest National Laboratory]​​, Alex Pothen [Purdue​​​‌ University], Bora Uçar‌, Christian Schulz [Heidelberg‌​‌ University].

We propose​​ two one-pass streaming algorithms​​​‌ for the $\mathrm{𝒩𝒫}$-hard‌ hypergraph matching problem. The‌​‌ first algorithm stores a​​ small subset of potential​​​‌ matching edges in a‌ stack using dual variables‌​‌ to select edges. It​​ has an approximation guarantee​​​‌ of $\frac{1}{d(‌1+\epsilon )‌‌}$ and requires $𝒪(‌(\frac{n}{\epsilon }){log}^{2}n)‌$ bits of memory, where​​ $n$ is the number​​​‌ of vertices in the​ hypergraph, $d$ is the​‌ maximum number of vertices​​ in a hyperedge, and​​​‌ $ϵ>0$ is​ a parameter to be​‌ chosen. The second algorithm​​ computes, stores, and updates​​​‌ a single matching as​ the edges stream, with​‌ an approximation ratio dependent​​ on a parameter $\mathrm{\alpha ‌}$. Its best approximation​ guarantee is $\frac{1}{(‌2d-\mathrm{1})+2\sqrt{\mathrm{d‌}(d-\mathrm{1})}}$, and it​‌ requires only $𝒪(n)$ memory.

We​​​‌ have implemented both algorithms​ and compared them with​‌ respect to solution quality,​​ memory consumption, and running​​​‌ times on two diverse​ sets of hypergraphs with​‌ a non-streaming greedy and​​ a naive streaming algorithm.​​​‌ Our results show that​ the streaming algorithms achieve​‌ much better solution quality​​ than naive algorithms when​​​‌ facing adverse orderings. Furthermore,​ these algorithms reduce the​‌ memory required by a​​ factor of 13 in​​​‌ the geometric mean on​ our test problems, and​‌ also outperform the offline​​ Greedy algorithm in running​​​‌ time.

This work was​ presented at the ESA​‌ 2025 conference 23.​​ Codes are available here​​​‌.

8.3.3 Algorithms for​ symmetric Birkhoff-von Neumann decomposition​‌ of symmetric doubly stochastic​​ matrices

Participants: Damien Lesens​​​‌, Jérémy E. Cohen​ [CNRS, Lyon], Bora​‌ Uçar.

The classical​​ Birkhoff–von Neumann (BvN) decomposition​​​‌ expresses a given doubly​ stochastic matrix as a​‌ convex combination of permutation​​ matrices. We investigate the​​​‌ BvN decomposition of symmetric​ doubly stochastic matrices where​‌ the permutation matrices in​​ the decomposition are also​​​‌ symmetric, called SymBvn decomposition.​ This decomposition is not​‌ always possible. Two pioneering​​ theoretical works establish the​​​‌ conditions under which such​ a decomposition is possible​‌ using graph terminology. We​​ propose a practical algorithm​​​‌ by combining these two​ works. A simple transformation​‌ converts any given symmetric​​ doubly stochastic matrix, with​​​‌ possibly nonzero diagonal elements,​ to be the adjacency​‌ matrix of an edge-weighted​​ undirected graph. The adjacency​​​‌ matrix of the resulting​ graph admits a SymBvn​‌ decomposition if and only​​ if the given matrix​​​‌ does so. The practicality​ of the proposed algorithm​‌ allows us to implement​​ it, release its source​​​‌ code, and report the​ first set of experiments​‌ ever performed for the​​ SymBvn decomposition. Our experiments​​​‌ suggest that the proposed​ algorithm is as effective​‌ as the state-of-the-art algorithms​​ for the classical BvN​​​‌ decomposition.

This work has​ been accepted to be​‌ published in SIAM Journal​​ on Matrix Analysis and​​​‌ Applications in Dec 2025​ 11. Codes are​‌ available here.

8.3.4​​ SUperman: Efficient Permanent Computation​​​‌ on GPUs

Participants: Deniz​ Elbek [Sabancı University],​‌ Fatih Taşyaran [Sabancı University]​​, Bora Uçar,​​​‌ Kamer Kaya [Sabancı University]​.

The permanent is​‌ a function, defined for​​ a square matrix, with​​​‌ applications in various domains​ including quantum computing, statistical​‌ physics, complexity theory, combinatorics,​​ and graph theory. Its​​​‌ formula is similar to​ that of the determinant,​‌ however unlike the determinant,​​ its exact computation is​​ #P-complete, i.e., there is​​​‌ no algorithm to compute‌ the permanent in polynomial‌​‌ time unless P=NP. For​​ an $n\times \mathrm{n‌}$ matrix, the fastest algorithm‌ has a time complexity‌​‌ of $O\left({\mathrm{2}}^{n-1}\mathrm{n‌}\right)$. Although supercomputers‌ have been employed for‌​‌ permanent computation before, there​​ is no work and​​​‌ more importantly, no publicly‌ available software that leverages‌​‌ cutting-edge, yet widely accessible,​​ High-Performance Computing accelerators such​​​‌ as GPUs. In this‌ work, we designed, developed,‌​‌ and investigated the performance​​ of SUperman, a​​​‌ complete software suite that‌ can compute matrix permanents‌​‌ on multiple nodes/GPUs on​​ a cluster while handling​​​‌ various matrix types, e.g.,‌ real/complex/binary and sparse/dense etc.,‌​‌ with a unique treatment​​ for each type. Compared​​​‌ to a state-of-the-art parallel‌ algorithm on 44 cores,‌​‌ SUperman can be $\mathrm{86}\times$ faster on a​​​‌ single Nvidia A100 GPU.‌ Combining multiple GPUs, we‌​‌ also showed that SUperman​​ can compute the permanent​​​‌ of a $56\times ‌56$ matrix which is‌​‌ the largest reported in​​ the literature.

This work​​​‌ has been recently accepted‌ to be published in‌​‌ Computer Physics Communications 30​​.

8.3.5 Communication Lower​​​‌ Bounds and Optimal Algorithms‌ for Symmetric Matrix Computations‌​‌

Participants: Hussam Al Daas​​ [Rutherford Appleton Laboratory],​​​‌ Grey Ballard [Wake Forest‌ University], Laura Grigori‌​‌ [EPFL], Suraj Kumar​​, Kathryn Rouse [Inmar​​​‌ Intelligence], Mathieu Vérité‌ [EPFL].

In this‌​‌ work, we focus on​​ the communication costs of​​​‌ three symmetric matrix computations:‌ i) multiplying a matrix‌​‌ with its transpose, known​​ as a symmetric rank-k​​​‌ update (SYRK) ii) adding‌ the result of the‌​‌ multiplication of a matrix​​ with the transpose of​​​‌ another matrix and the‌ transpose of that result,‌​‌ known as a symmetric​​ rank-2k update (SYR2K) iii)​​​‌ performing matrix multiplication with‌ a symmetric input matrix‌​‌ (SYMM). All three computations​​ appear in the Level​​​‌ 3 Basic Linear Algebra‌ Subroutines (BLAS) and have‌​‌ wide use in applications​​ involving symmetric matrices. We​​​‌ establish communication lower bounds‌ for these kernels using‌​‌ sequential and distributed-memory parallel​​ computational models, and we​​​‌ show that our bounds‌ are tight by presenting‌​‌ communication-optimal algorithms for each​​ setting. Our lower bound​​​‌ proofs rely on applying‌ a geometric inequality for‌​‌ symmetric computations and analytically​​ solving constrained nonlinear optimization​​​‌ problems. The symmetric matrix‌ and its corresponding computations‌​‌ are accessed and performed​​ according to a triangular​​​‌ block partitioning scheme in‌ the optimal algorithms.

This‌​‌ work has been published​​ in the ACM Transactions​​​‌ on Parallel Computing journal‌ 9.

8.3.6 Minimizing‌​‌ Communication for Parallel Symmetric​​ Tensor Times Same Vector​​​‌ Computation

Participants: Hussam Al‌ Daas [Rutherford Appleton Laboratory]‌​‌, Grey Ballard [Wake​​ Forest University], Laura​​​‌ Grigori [EPFL], Suraj‌ Kumar, Kathryn Rouse‌​‌ [Inmar Intelligence], Mathieu​​ Vérité [EPFL].

In​​​‌ this work, we focus‌ on the parallel communication‌​‌ cost of multiplying the​​ same vector along two​​​‌ modes of a 3-dimensional‌ symmetric tensor. This is‌​‌ a key computation in​​ the higher-order power method​​​‌ for determining eigenpairs of‌ a 3-dimensional symmetric tensor‌​‌ and in gradient-based methods​​​‌ for computing a symmetric​ CP decomposition. We establish​‌ communication lower bounds that​​ determine how much data​​​‌ movement is required to​ perform the specified computation​‌ in parallel. We demonstrate​​ that the communication lower​​​‌ bounds are tight by​ presenting an optimal algorithm​‌ where the data distribution​​ is a natural extension​​​‌ of the triangle block​ partition scheme for symmetric​‌ matrices to 3-dimensional symmetric​​ tensors.

This work was​​​‌ presented at the SPAA​ 2025 conference 14.​‌

10.1.1 Inria associate​‌ team not involved in​​ an IIL or an​​​‌ international program

10.1.2 Participation in other​​​‌ International Programs

10.2.1 Visits​​ of international scientists

10.2.2 Visits to international​​​‌ teams

10.3.1 Horizon Europe​​​‌

Participants: Anne Benoit,‌ Michel Nicolis, Frédéric‌​‌ Vivien.

• HORIZON-INFRA project​​ ODISSEE (2025–2027).
  The ODISSEE​​​‌ project aims to develop‌ innovative technologies and methodologies‌​‌ to process the unprecedented​​ volume of scientific data​​​‌ produced by research infrastructures‌ such as CERN’s HL-LHC‌​‌ and SKAO. It notably​​ plans to develop on-the-fly​​​‌ AI data processing, which‌ is a major challenge‌​‌ in research in the​​ physical sciences, and where​​​‌ the contribution of SLICES‌ RI will be decisive.‌​‌ Coordinated by Damien Gratadour,​​ researcher at the CNRS​​​‌ Laboratory for Instrumentation and‌ Research in Astrophysics, ODISSEE‌​‌ is leveraging the European​​​‌ HPC ecosystem to open​ up a new era​‌ in science, helping to​​ unravel fundamental mysteries such​​​‌ as the nature of​ dark matter. The three-year​‌ ODISSEE consortium brings together​​ 14 partners and 2​​​‌ associate partners from academia​ and industry. F. Vivien​‌ is the head of​​ work-package 5.

ECLAT

Participants: Frédéric​ Vivien.

• Partner Institution(s):​‌
  CNRS, Inria, Eviden, Observatoire​​ de la Côte d’Azur,​​​‌ Observatoire de Paris-PSL
• Date/Duration:​
  2023-
• Summary
  ECLAT is​‌ a joint laboratory gathering​​ 14 laboratories. Its aim​​​‌ is to support the​ French contribution to the​‌ SKAO observatory.

18th Workshop on Scheduling​ for Large Scale Systems:​‌

Loris Marchal , Pablo​​ Piantanida and Yves Robert​​​‌ have organized the 18th​ Workshop on Scheduling for​‌ Large Scale Systems in​​ Montréal in July 2025.​​​‌ Further details can be​ found on the workshop​‌ webpage.

11.1.1 Scientific​​ events: organisation

Bora Uçar​​​‌ was the organizing committee​ co-chair of SIAM Conference​‌ on Applied and Computational​​ Discrete Algorithms (ACDA25) Montréal,​​​‌ Canada. In SIAM conference​ systems this role corresponds​‌ to the general chair.​​

11.1.2 Scientific events: selection​​​‌

11.1.3​​​‌ Journal

11.1.4 Scientific‌ expertise

• Frédéric Vivien was‌​‌ a member of the​​ evaluation committee for the​​​‌ Scientific Evaluation of Forschungszentrum‌ Jülich (FZJ), Germany, in‌​‌ April 2025.
• Frédéric Vivien​​ is an elected member​​​‌ of INRIA commision d'évaluation‌.
• Frédéric Vivien is‌​‌ a member of the​​ scientific council of the​​​‌ IRMIA labex.

11.2.1 Teaching​​​‌

• Master: Anne Benoit, Parallel‌ and Distributed Algorithms and‌​‌ Programs, 39h, M1, ENS​​ Lyon, France
• Master: Suraj​​​‌ Kumar, Loris Marchal and‌ Frédéric Vivien, Resource-aware computations‌​‌ on CPUs and GPUs,​​ 30h, M2, ENS Lyon,​​​‌ France.

11.2.2 Supervision

• Anne‌ Benoit and Frédéric Vivien‌​‌ are co-supervising the PhD​​ of Joachim Cendrier. Joachim​​​‌ works on efficient and‌ environment-friendly scheduling and resource‌​‌ management algorithms at the​​ edge, in collaboration with​​​‌ Andrew Chien from U.‌ Chicago. He is funded‌​‌ by a CNRS -​​ U. Chicago project.
• Anne​​​‌ Benoit and Frédéric Vivien‌ are co-supervising the PhD‌​‌ of Hadi Gholami. Hadi​​ works on resource allocation​​​‌ to jobs running on‌ variable-capacity platforms.
• Anne Benoit‌​‌ is co-supervising the PhD​​ of Svetlana Kulagina with​​​‌ Henning Meyerhenke (Humboldt-Universitat zu‌ Berlin, Germany), as part‌​‌ of the FONDA project,​​ on the execution of​​​‌ large workflows in heterogeneous‌ execution environments.
• Anne Benoit‌​‌ is co-supervising the PhD​​ of Dominik Schweisgut with​​​‌ Henning Meyerhenke (now at‌ Karlsruhe Institute of Technology,‌​‌ Germany), as part of​​ the FONDA project, on​​​‌ algorithms for carbon-aware workflow‌ scheduling.
• Frédéric Vivien and‌​‌ Anne Benoit are co-supervising​​ the PhD of Michel​​​‌ Nicolis. Michel works on‌ predictive maintenance.
• Anne Benoit‌​‌ and Loris Marchal are​​ supervising the PhD of​​​‌ Adrien Obrecht on scheduling‌ task graphs with application‌​‌ to large language models.​​
• Loris Marchal is supervising​​​‌ the PhD of Felix‌ Wirth (Cifre collaboration with‌​‌ the Kog company) on​​ the design of generative​​​‌ models with efficient parallel‌ inference.
• Bora Uçar is‌​‌ co-supervising the PhD of​​ Damien Lesens on nonnegative​​​‌ matrix factorization.

11.2.3 Juries‌

• Anne Benoit was a‌​‌ reviewer of the PhD​​ of Diane Orhan, defended​​​‌ at Bordeaux University on‌ December 9, 2025.
• Anne‌​‌ Benoit was an examiner​​ for the PhD of​​​‌ Roblex Nana Tchakouté, defended‌ at PSL University Paris‌​‌ on December 5, 2025.​​
• Loris Marchal was a​​​‌ reviewer of the PhD‌ of Robin Boezennec, defended‌​‌ at Rennes University on​​ December 10, 2025.
• Loris​​​‌ Marchal was a reviewer‌ of the PhD of‌​‌ Huijun Wang, to be​​ defended at Auckland university​​​‌ (New Zealand) in Spring‌ 2026.
• Frédéric Vivien was‌​‌ a member of the​​ jury for hiring INRIA​​​‌ CRCN and ISFP researchers‌ for the Centre Inria‌​‌ de l'Université de Bordeaux.​​
• Frédéric Vivien was a​​​‌ member of the jury‌ for hiring INRIA senior‌​‌ (DR2) researchers.

International journals

• 8 article​​​‌Q.Quentin Barbut,​ L.Lucas Perotin,​‌ A.Anne Benoit,​​ T.Thomas Herault,​​​‌ Y.Yves Robert and​ F.Frédéric Vivien.​‌ Fixed-Work vs. Fixed-Time Checkpointing​​ on Large-Scale Failure-Prone Platforms​​​‌.International Journal of​ High Performance Computing Applications​‌401August 2025​​, 96-114HALDOI​​​‌back to text
• 9​ articleH. A.Hussam​‌ Al Daas, G.​​Grey Ballard, L.​​​‌Laura Grigori, S.​Suraj Kumar, K.​‌Kathryn Rouse and M.​​Mathieu Verite. Communication​​​‌ Lower Bounds and Optimal​ Algorithms for Symmetric Matrix​‌ Computations.ACM Transactions​​ on Parallel Computing12​​​‌22025HALDOI​back to text
• 10​‌ articleM.Maxime Gonthier​​, L.Loris Marchal​​​‌ and S.Samuel Thibault​. A scheduler to​‌ foster data locality for​​ GPU and out-of-core task-based​​​‌ linear algebra applications.​Journal of Parallel and​‌ Distributed Computing206August​​ 2025, 105170HAL​​​‌DOIback to text​
• 11 articleD.Damien​‌ Lesens, J. E.​​Jérémy E. Cohen and​​​‌ B.Bora Uçar.​ Algorithms for symmetric Birkhoff-von​‌ Neumann decomposition of symmetric​​ doubly stochastic matrices.​​​‌SIAM Journal on Matrix​ Analysis and ApplicationsJanuary​‌ 2025, 1-30In​​ press. HALback to​​​‌ text
• 12 articleS.​Somesh Singh and B.​‌Bora Uçar. Efficient​​ Parallel Sparse Tensor Contraction​​​‌.IEEE Transactions on​ Parallel and Distributed Systems​‌3662025,​​ 1206-1219HALDOIback​​​‌ to text
• 13 article​A.Alix Tremodeux,​‌ A.Anne Benoit,​​ E.Emmanuel Agullo,​​ T.Thomas Herault,​​​‌ L.Luc Giraud and‌ Y.Yves Robert.‌​‌ Fault-tolerant numerical iterative algorithms​​ at scale.International​​​‌ Journal of High Performance‌ Computing Applications2025HAL‌​‌back to text

International​​ peer-reviewed conferences

• 14 inproceedings​​​‌H.Hussam Al Daas‌, G.Grey Ballard‌​‌, L.Laura Grigori​​, S.Suraj Kumar​​​‌, K.Kathryn Rouse‌ and M.Mathieu Verite‌​‌. Minimizing Communication for​​ Parallel Symmetric Tensor Times​​​‌ Same Vector Computation.‌Brief Announcement: Minimizing Communication‌​‌ for Parallel Symmetric Tensor​​ Times Same Vector Computation​​​‌SPAA 2025 - ACM‌ Symposium on Parallelism in‌​‌ Algorithms and ArchitecturesPortland,​​ United StatesJuly 2025​​​‌HALDOIback to‌ text
• 15 inproceedingsS.‌​‌Srinivas Aluru, A.​​Anne Benoit and B.​​​‌Bora Uçar. Efficient‌ and effective methods for‌​‌ variant selection.BCB'25:​​ Proceedings of the 16th​​​‌ ACM Conference on Bioinformatics,‌ Computational Biology, and Health‌​‌ Informatics (BCB 2025)16th​​ ACM Conference on Bioinformatics,​​​‌ Computational Biology, and Health‌ Informatics (BCB 2025)Philadelphie,‌​‌ PA, United StatesDecember​​ 2025HALback to​​​‌ text
• 16 inproceedingsS.‌Spyros Angelopoulos, L.‌​‌Loris Marchal, A.​​Adrien Obrecht and B.​​​‌Bertrand Simon. Cache‌ Management for Mixture-of-Experts LLMs‌​‌.Euro-Par 2025: Parallel​​ Processing. Euro-Par 2025. Lecture​​​‌ Notes in Computer Science‌31st European Conference on‌​‌ Parallel and Distributed Processing​​ (EURO-PAR 2025)15902Dresden,​​​‌ Germany2025, 18–32‌HALDOIback to‌​‌ text
• 17 inproceedingsO.​​Olivier Beaumont, R.​​​‌Raphaël Bourgouin, M.‌Maxime Darrin, L.‌​‌Loris Marchal and P.​​Pablo Piantanida. Leveraging​​​‌ Expert Usage to Speed‌ up LLM Inference with‌​‌ Expert Parallelism.Lecture​​ Notes in Computer Science​​​‌Euro-Par 2025: Parallel Processing‌Dresden, GermanyAugust 2025‌​‌HALback to text​​
• 18 inproceedingsA.Anne​​​‌ Benoit, T.Thomas‌ Herault, Y.Yves‌​‌ Robert and A.Alix​​ Tremodeux. Partial Detectors​​​‌ Versus Replication To Cope‌ With Silent Errors.‌​‌Euro-Par 2025 - 31​​ st International European Conference​​​‌ on Parallel and Distributed‌ ComputingDresden, GermanyAugust‌​‌ 2025HALback to​​ text
• 19 inproceedingsJ.​​​‌Joachim Cendrier, R.‌Rajini Wijayawardana, A.‌​‌Anne Benoit, Y.​​Yves Robert, F.​​​‌Frédéric Vivien and A.‌Andrew A. Chien.‌​‌ Green Scheduling on the​​ Edge.Euro-Par 2025​​​‌ - 31st International European‌ Conference on Parallel and‌​‌ Distributed Computing15900Lecture​​ Notes in Computer Science​​​‌Dresden, GermanySpringer Nature‌ SwitzerlandAugust 2025,‌​‌ 380-394HALDOIback​​ to text
• 20 inproceedings​​​‌A.Adil Chhabra,‌ C.Christian Schulz,‌​‌ B.Bora Uçar and​​ L.Loris Wilwert.​​​‌ Exact Minimum Cuts in‌ Hypergraphs at Scale.‌​‌SIAM Symposium on Algorithm​​ Engineering and Experiments (ALENEX26)​​​‌Vancouver, CanadaJanuary 2026‌HAL
• 21 inproceedingsM.‌​‌Maxime Gonthier, K.​​Kyle Chard, I.​​​‌Ian Foster, L.‌Loris Marchal and F.‌​‌Frédéric Vivien. Deadline-Aware​​ Scheduling of Mixed-Criticality Tasks​​​‌.ICPP '25: Proceedings‌ of the 54th International‌​‌ Conference on Parallel Processing​​ICPP '25 - 54th​​​‌ International Conference on Parallel‌ ProcessingSan Diego, CA,‌​‌ United StatesSeptember 2025​​​‌HALback to text​
• 22 inproceedingsS.Svetlana​‌ Kulagina, A.Anne​​ Benoit and H.Henning​​​‌ Meyerhenke. Memory-aware Adaptive​ Scheduling of Scientific Workflows​‌ on Heterogeneous Architectures.​​CCGrid 2025 - 25th​​​‌ IEEEE International Symposium on​ Cluster, Cloud and Internet​‌ ComputingTromsø, NorwayMay​​ 2025HALback to​​​‌ text
• 23 inproceedingsH.​Henrik Reinstädtler, S.​‌ M.S M Ferdous​​, A.Alex Pothen​​​‌, B.Bora Uçar​ and C.Christian Schulz​‌. Semi-Streaming Algorithms for​​ Hypergraph Matching.Leibniz​​​‌ International Proceedings in Informatics​ (LIPIcs)ESA 2025 -​‌ European Symposium on Algorithms​​Warsaw (POLAND), PolandFebruary​​​‌ 2025HALback to​ text
• 24 inproceedingsD.​‌Dominik Schweisgut, A.​​Anne Benoit, Y.​​​‌Yves Robert and H.​Henning Meyerhenke. Carbon-Aware​‌ Workflow Scheduling with Fixed​​ Mapping and Deadline Constraint​​​‌.International Conference on​ Parallel Processing (ICPP)ICPP​‌ 2025 - 54th International​​ Conference on Parallel Processing​​​‌San Diego (CA), United​ StatesAugust 2025HAL​‌back to text

Reports​​ & preprints

• 25 report​​​‌A.Anne Benoit,​ J.Joachim Cendrier and​‌ F.Frédéric Vivien.​​ Scheduling Jobs Under a​​​‌ Variable Number of Processors​.RR-9582InriaNovember​‌ 2025HAL
• 26 report​​A.Anne Benoit,​​​‌ A. A.Andrew A​ Chien and Y.Yves​‌ Robert. 2nd Workshop​​ on Scheduling Variable CapacityResources​​​‌ for Sustainability.Inria​January 2026HAL
• 27​‌ reportA.Anne Benoit​​, T.Thomas Herault​​​‌, Y.Yves Robert​ and A.Alix Tremodeux​‌. Partial Detectors Versus​​ Replication To Cope With​​​‌ Silent Errors.RR-9581​InriaMarch 2025HAL​‌
• 28 reportR. H.​​Rob H. Bisseling and​​​‌ B.Bora Uçar.​ Optimal partitioning of 2D​‌ meshes for stencil computations​​.RR-9585Inria Lyon​​​‌September 2025HAL
• 29​ reportJ.Joachim Cendrier​‌, R.Rajini Wijayawardana​​, A.Anne Benoit​​​‌, Y.Yves Robert​, F.Frédéric Vivien​‌ and A. A.Andrew​​ A Chien. Green​​​‌ Scheduling on the Edge​.RR-9580InriaMarch​‌ 2025HAL
• 30 report​​D.Deniz Elbek,​​​‌ F.Fatih Taşyaran,​ B.Bora Uçar and​‌ K.Kamer Kaya.​​ SUperman: Efficient Permanent Computation​​​‌ on GPUs.RR-9578​Inria LyonMarch 2025​‌, 1-25HALback​​ to text
• 31 report​​​‌Q.Qi Gao,​ L.Li Han,​‌ S.Sascha Hunold,​​ Y.Yves Robert and​​​‌ F.Frédéric Vivien.​ Coping with Silent Errors​‌ for Workflows of Moldable​​ Tasks.RR-9589INRIA​​​‌June 2025HAL
• 32​ reportA.Alix Tremodeux​‌, E.Emmanuel Agullo​​, A.Anne Benoit​​​‌, L.Luc Giraud​, T.Thomas Herault​‌ and Y.Yves Robert​​. Fault-tolerant numerical iterative​​​‌ algorithms at scale.​RR-9567Inria LyonJanuary​‌ 2025HAL
• 33 report​​F.Félix Wirth,​​​‌ G.Gael Delalleau and​ L.Loris Marchal.​‌ Towards Parallel Transformer-Based Large​​ Language Models for Fast​​​‌ Inference.RR-9573Inria​January 2025, 42​‌HALback to text​​

Software

• 34 softwareK.​​​‌Kamer Kaya, J.​Johannes Langguth, I.​‌Ioannis Panagiotas and B.​​Bora Uçar. MatchMaker​​.May 2025 lic:​​​‌ CeCILL-B.HALSoftware‌ HeritageVCS