High performance checksum computation for fault-tolerant MPI over InfiniBand
Résumé
With the increase of the number of nodes in clusters, the probability of failures and unusual events increases. In this paper, we present checksum mechanisms to detect data corruption. We study the impact of checksums on network communication performance and we propose a mechanism to amortize their cost on InfiniBand. We have implemented our mechanisms in the NEWMADELEINE communication library. Our evaluation shows that our mechanisms to ensure message integrity do not impact noticeably the application performance, which is an improvement over the state of the art MPI implementations.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...