Improving the reliability and the performance of CAPE by using MPI for data exchange on network

van Long Tran (1, 2, 3) , Eric Renault (1, 2, 3) , Viet Hai Ha (4)

1 TSP - RS2M - Département Réseaux et Services Multimédia Mobiles
2 R3S-SAMOVAR - Réseaux, Systèmes, Services, Sécurité
3 CNRS - Centre National de la Recherche Scientifique
4 College of Education [Hue] (Hue University)

CAPE - which stands for Checkpointing Aided Parallel Execution - has demonstrated to be a high-performance and compliant OpenMP implementation for distributed memory systems. CAPE is based on the use of checkpoints to automatically distribute jobs of OpenMP parallel constructs to distant machines and to automatically collect the calculated results on these machines to the master machine. However, on the current version, the data exchange on networks use manual sockets that require time to establish connections between machines for each parallel construct. Furthermore, this technique is not really reliable due to the risk of conflicts on ports and the problem of data exchange using stream. This paper aims at presenting the impact of using MPI to improve the reliability and the performance of CAPE. Both socket and MPI implementations are analyzed and discussed, and performance evaluations are provided

Mots clés: CAPE; OpenMP; MPI; High performance computing; Parallel programming

Domaines: Informatique [cs]

van Long Tran, Eric Renault, Viet Hai Ha. Improving the reliability and the performance of CAPE by using MPI for data exchange on network. MSPN 2015 : First International Conference on Mobile, Secure and Programmable Networking, Jun 2015, Paris, France. pp.90 - 100, ⟨10.1007/978-3-319-25744-0_8⟩. ⟨hal-01377623⟩