Présentations plénières

Laura Grigori

Mercredi 25 juin

Solving large scale problems through randomization

In this talk we discuss recent progress in using randomization for solving large scale linear algebra problems as solving linear systems of equations or eigenvalue problems. We discuss first sketching techniques that allow to embed large dimensional subspaces while preserving geometrical properties and their parallel implementations. We then present randomized versions of processes for orthogonalizing a set of vectors and their usage in the Arnoldi iteration. We discuss associated Krylov subspace methods for solving large scale linear systems of equations and eigenvalue problems. The new methods retain the numerical stability of classic Krylov methods while reducing communication and being more efficient on modern massively parallel computers.

Laura Grigori is a Full Professor in applied and computational mathematics at EPFL and PSI and Chair of high performance Numerical algorithms and simulations. From 2013 to 2023 she has lead Alpines group, a joint group between INRIA and J.L. Lions Laboratory, Sorbonne University, in Paris. She is a SIAM Fellow, the recipient of an ERC Synergy Grant, and the recipient of the SIAM SIAG on Supercomputing Career Prize 2024. Her field of expertise is in numerical linear, multilinear algebra, randomized algorithms, and high performance scientific computing for challenging applications ranging from astrophysics to molecular simulations. For her work on communication avoiding algorithms she was awarded with her co-authors the SIAM Siag on Supercomputing Best Paper Prize 2016 for the most outstanding paper published in 2012-2015 in a journal in the field of high performance computing. She was the chair of PRACE Scientific Steering Committee, the Program Director of the SIAM special interest group on supercomputing, January 2014 - December 2015, then the Chair of this interest group, January 2016 - December 2017. She is currently the chair of the SIAM interest group on numerical linear algebra.

Romain Rouvoy

Jeudi 26 juin

Cloud computing: sharing is caring

En quelques années seulement, les infrastructures de cloud computing sont devenues le moteur du déploiement massif de services numériques dans notre société. Ces infrastructures, qui prônent la virtualisation, favorisent la consolidation des ressources physiques dans des data centers de plus en plus imposants. Cependant, il reste de nombreux défis à relever pour réduire l’empreinte environnementale des services en ligne qu’elles hébergent. Dans cet exposé, je présenterai les travaux que nous menons depuis plusieurs années au sein de l’équipe Spirals pour évaluer l’efficacité de ces infrastructures logicielles et matérielles. Au-delà des paradoxes que soulèvent ces indicateurs, je présenterai également quelques contributions systèmes que nous avons pu proposer au fil des ans pour repousser les limites de la consolidation traditionnelle des ressources, tout en évoquant leurs propres limites.

Romain Rouvoy est Professeur des Universités au sein du Département d’Informatique de l’Université de Lille et membre de l’équipe-projet Spirals, associée au CRIStAL (UMR CNRS 9189) et au centre Inria de l’Université de Lille. Ses travaux de recherche portent sur les problématiques systèmes et génie logiciel relatives à l’impact environnemental du numérique. En particulier, il s’attache à lutter contre les différentes sources de gaspillage de ressources logicielles et matérielles, même si cela implique de remettre en question le mythe de la performance à n’importe quel prix.

Arthur Perais

Vendredi 27 juin

A Day in the Life of Instruction Fetch

Modern general purpose high performance processors can execute around ten instructions each cycle, but can only do so when they know what instructions to execute. As a result, significant research and engineering is spent ensuring that the correct instructions are brought to the processor execution units fast enough. In this talk, we will first review what the modern instruction fetch hardware looks like, highlighting the considerable progress made since the 5-stage MIPS pipeline. Notably, so-called “datacenter” workloads and their ever increasing instruction footprint have caused a major shift in organization in the last decade or so. Then, we will present two techniques that accelerate instruction delivery in specific scenarios: i) Workloads with high Instruction Level Parallelism and ii) Workloads with hard to predict conditional branches (e.g., sorting).

Arthur Perais has been a CNRS researcher at TIMA in Grenoble since 2020, working on general purpose microarchitecture techniques often based on speculation. From 2017 to 2020, he was part of the performance modeling team at Qualcomm Datacenter Technologies (Raleigh, US) then at Microsoft (Raleigh, US), working on custom ARM microarchitectures. He obtained his PhD in 2015 from Université de Rennes 1.