HydEE, vers un protocole de recouvrement arrière hiérarchique pour les machines exascales. De l'exploitation du déterminisme des émissions dans les protocoles de recouvrement arrière

Abstract

The move towards exascale super-computers requires new fault tolerance solutions. Regarding parallel message passing applications, existing rollback-recovery protocols are not suited. To be able to deal with very large scale applications and high failure rate, a protoTechnique et science informatiques 2 Technique et science informatiques col should be able to confine failures consequences to a small subset of the processes, while providing good failure free performance, and logging a limited amount of data, especially in memory. To fulfill these needs, we propose HydEE, a hierarchical rollback-recovery protocol that combines coordinated checkpointing and message logging. HydEE leverages the senddeterminism of scienfitic parallel applications to tolerate multiple failures without relying on a stable storage. Our experiments show that for most applications, saving less than 15% of the messages payload in memory is enough to limit the rollbacks after a failure to less than 15% of the processes. MOTS-CLÉS : Calcul haute performances, applications parallèles, MPI, tolérance aux fautes, recouvrement arrière, protocoles hiérarchiques