PDI, an approach to decouple I/O concerns from high-performance simulation codes

Abstract : High-performance computing (HPC) applications manipulate and store large datasets for scientific analysis, visualization purposes and/or resiliency. Multiple software libraries have been designed for interacting with the parallel file system and in some cases with intermediate storage levels. These libraries provide different level of abstraction and have been optimized for different purposes. The best I/O library for a given usage depends on multiple criteria including the purpose of the I/O, the computer architecture or the problem size. Therefore, to optimize their I/O strategies, scientists have to use multiple API's depending on the targeted execution. As a result, simulation codes contain intrusive and library dependent I/O instructions interwoven with domain instructions. In this paper, we propose a novel interface that transparently manage the I/O aspects and support multiple I/O libraries within the same execution. We demonstrate the low overhead of our interface and perform experiments at scale with the Gysela production code.
Type de document :
Pré-publication, Document de travail
2017
Liste complète des métadonnées

Littérature citée [26 références]  Voir  Masquer  Télécharger

https://hal.archives-ouvertes.fr/hal-01587075
Contributeur : Corentin Roussel <>
Soumis le : mercredi 13 septembre 2017 - 16:01:44
Dernière modification le : vendredi 29 septembre 2017 - 11:24:17

Fichier

paper.pdf
Fichiers produits par l'(les) auteur(s)

Licence


Distributed under a Creative Commons Paternité 4.0 International License

Identifiants

  • HAL Id : hal-01587075, version 1

Collections

Citation

Corentin Roussel, Kai Keller, Mohamed Gaalich, Leonardo Bautista Gomez, Julien Bigot. PDI, an approach to decouple I/O concerns from high-performance simulation codes. 2017. 〈hal-01587075〉

Partager

Métriques

Consultations de la notice

95

Téléchargements de fichiers

13