Impact of communication times on mixed CPU/GPU applications scheduling using KAAPI

David Beniamine 1
1 MOAIS - PrograMming and scheduling design fOr Applications in Interactive Simulation
Inria Grenoble - Rhône-Alpes, LIG - Laboratoire d'Informatique de Grenoble
Abstract : High Performance Computing machines use more and more Graphical Processing Units as they are very efficient for homogeneous computation such as matrix operations. However before using these accelerators, one has to transfer data from the processor to them. Such a transfer can be slow. In this report, our aim is to study the impact of communication times on the makespan of a scheduling. Indeed, with a better anticipation of these communications, we could use the GPUs even more efficiently. More precisely, we will focus on machines with one or more GPUs and on applications with a low ratio of computations over communications. During this study, we have implemented two offline scheduling algorithms within XKAAPI's runtime. Then we have led an experimental study, combining these algorithms to highlight the impact of communication times. Finally our study has shown that, by using communication aware scheduling algorithms, we can reduce substantially the makespan of an application. Our experiments have shown a reduction of this makespan up to $64\%$ on a machine with several GPUs executing homogeneous computations.
Type de document :
Mémoires d'étudiants -- Hal-inria+
Distributed, Parallel, and Cluster Computing [cs.DC]. 2013
Liste complète des métadonnées

Littérature citée [22 références]  Voir  Masquer  Télécharger

https://hal.inria.fr/hal-00924020
Contributeur : David Beniamine <>
Soumis le : jeudi 23 janvier 2014 - 14:37:39
Dernière modification le : jeudi 11 octobre 2018 - 08:48:03
Document(s) archivé(s) le : jeudi 24 avril 2014 - 10:30:47

Fichier

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

Identifiants

  • HAL Id : hal-00924020, version 1

Collections

Citation

David Beniamine. Impact of communication times on mixed CPU/GPU applications scheduling using KAAPI. Distributed, Parallel, and Cluster Computing [cs.DC]. 2013. 〈hal-00924020〉

Partager

Métriques

Consultations de la notice

254

Téléchargements de fichiers

305