InKS, a programming model to decouple performance from semantics in simulation codes

Olivier Aumage 1, 2, * Julien Bigot 3, * Ksander Ejjaaouani 3, 4, * Michel Mehrenberger 5, 4, *
* Auteur correspondant
2 STORM - STatic Optimizations, Runtime Methods
LaBRI - Laboratoire Bordelais de Recherche en Informatique, Inria Bordeaux - Sud-Ouest
4 TONUS - TOkamaks and NUmerical Simulations
IRMA - Institut de Recherche Mathématique Avancée, Inria Nancy - Grand Est
Abstract : Existing programming models lead to a tight inter-leaving of semantics and computer optimization concerns in high-performance simulation codes. With the increasing complexity and heterogeneity of super-computers this requires scientists to become experts in both the simulated domain and the optimization process and makes the code difficult to maintain and port to new architectures. This report proposes InKS, a programming model that aims to improve the situation by decoupling semantics and optimizations in code so as to ease the collaboration between domain scientists and expert of high-performance optimizations. We define the InKS language that enables developers to describe the semantic of a simulation code with no concern for performance. We describe the implementation of a compiler able to automatically execute this InKS code without making any explicit execution choice. We also describe a method to manually specify these choices to reach high-performance. Our preliminary evaluation on a 3D heat equation solver demonstrates the feasibility of the automatic approach as well as the ability to specify complex optimizations while not altering the semantic part. It shows promising performance where two distinct specifications of optimization choices in InKS offer similar performance as existing hand-tailored versions of the solver.
Type de document :
Pré-publication, Document de travail
2017
Liste complète des métadonnées

https://hal-cea.archives-ouvertes.fr/cea-01493075
Contributeur : Julien Bigot <>
Soumis le : lundi 20 mars 2017 - 20:12:15
Dernière modification le : samedi 26 mai 2018 - 01:18:22

Fichier

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

Licence


Distributed under a Creative Commons Paternité - Partage selon les Conditions Initiales 4.0 International License

Identifiants

  • HAL Id : cea-01493075, version 1

Citation

Olivier Aumage, Julien Bigot, Ksander Ejjaaouani, Michel Mehrenberger. InKS, a programming model to decouple performance from semantics in simulation codes. 2017. 〈cea-01493075〉

Partager

Métriques

Consultations de la notice

962

Téléchargements de fichiers

135