Objectifs de l'équipe et activités de recherche
Objectifs
- Etudier et développer des méthodes numériques et des logiciels génériques pour un usage optimal (performance et fiabilité) des calculateurs scalaires, vectoriels et parallèles lors de la résolution de systèmes d'équations et de problèmes d'optimisation.
- Promouvoir le calcul à haute performance à travers la formation avancée et les activités de consultance.
Activités de recherche
- Algèbre linéaire:
- solveurs linéaires directs creux
- solveurs linéaires itératifs et préconditionnement
- problèmes aux valeurs propres
- noyaux de calcul et librairies de recherche pour les calculateurs à haute performance
- Systèmes non-linéaires et optimisation
- Calcul qualitatif
Formation avancée
- Organisés par le CERFACS: cycles de formation et workshops annuels, séminaires et visites de chercheurs, implication dans l'enseignement.
A scientific paper on algorithmics among the most frequently downloaded.
The Society for Industrial and Applied Mathematics (SIAM) has implemented a new feature that provides a list of the SIAM journal articles that are most often downloaded. We are pleased to congratulate Pavel Jiranek, postdoctoral fellow in the Parallel Algorithms team: his joint paper with Miroslav Rozloznik and Martin Gutknecht "How to Make Simpler GMRES and GCR More Stable " (SIMAX, vol. 30-4, pp 1483-1499, 2008) was the fifth most downloaded article from the "SIAM Journal Matrix Analysis and its Applications" (SIMAX) in January 2009. SIMAX is one of the premier journals in the field of numerical linear algebra.
Success at the Stckholm PRACE petascale summerschool - 68 billion equations solved
During the PRACE Petascale Summer School held in Stockholm on August 26th-29th, Xavier Pinel and Xavier Vasseur from the Algo Team successfully ported a three-dimensional Helmholtz solver to the IBM Blue Gene P at Juelich, Germany and the CRAY XT4 at Espoo, Finland. They showed almost perfect scalability on up to 65,536 cores both in the weak sense (when the computation per processor is kept constant) and in the strong sense (when the same problem is run on an increasing number of processors) on academic model problems. They were able to solve a linear system of 68 billion unknowns in less than 3 hours setting a new benchmark for the solution of large sparse indefinite systems. The PhD thesis of Xavier Pinel is funded by TOTAL and cosupervised by Serge Gratton (CNES/CERFACS) and Xavier Vasseur. This work is in collaboration with Henri Calandra at TOTAL.
