Interaction between MPI and TCP in grids - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Rapport (Rapport De Recherche) Année : 2008

Interaction between MPI and TCP in grids

Résumé

As MPI applications are more and more resource consuming, they need to be executed on grids. The communications on the WAN interconnecting clusters mostly use TCP which suffers from WAN features: high latency, sharing between users, bandwidth smaller than the aggregate bandwidth of the nodes. In this report, we first study the interaction between MPI and TCP on grids. We show why the nature of MPI traffic raises problems while using TCP on WAN links. TCP's loss detection and congestion control mechanism can both slow down the application. Then, we propose MPI5000, a transparent applicative layer between MPI and TCP, using proxies to improve the execution of MPI applications on a grid. Proxies aim at splitting TCP connections in order to detect losses faster and avoid to return in a slowstart phase after an idle time. Finally, we test our layer on Grid'5000, the French research grid, using MPICH2. The results on the NPB (NAS Parallel Benchmarks) validate our architecture that reduces the number of idle timeout and the number of long-distance retransmissions for certain benchmarks, namely BT, SP and LU benchmarks. Using MPI5000, these applications can decrease their execution time by 35%, 28%, and, 15% respectively.

Mots clés

Fichier principal
Vignette du fichier
RR-6945.pdf (265.38 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Loading...

Dates et versions

inria-00389836 , version 1 (29-05-2009)

Identifiants

  • HAL Id : inria-00389836 , version 1

Citer

Ludovic Hablot, Olivier Glück, Jean-Christophe Mignot, Romaric Guillier, Sébastien Soudan, et al.. Interaction between MPI and TCP in grids. [Research Report] RR-6945, INRIA. 2008, pp.26. ⟨inria-00389836⟩
137 Consultations
239 Téléchargements

Partager

Gmail Facebook X LinkedIn More