Skip to Main content Skip to Navigation
Theses

Athapascan-1 : vers un modèle de programmation parallèle adapté au calcul scientifique

Abstract : Parallel computers offer an interesting alternative for the applications of scientific computation, which need large resources of calculation and memory. However, the effective programming of these machines is often difficult and the obtained implementations are generally not easily portable. We propose in this thesis a parallel programming model allowing simple, portable and efficient programming of parallel applications. This model is based on an explicit decomposition of the application into tasks which communicate through objects in a shared memory. The semantics of the access to the shared data is quasi sequential and precedences between the tasks are implicitly defined to respect this semantic. We present in a first part the implementation of this parallel programming model in the C++ Athapascan-1 interface. An analysis at run-time execution of the dependences of data among tasks makes it possible to extract the data flow and thus the precedences between the tasks to be run. Scheduling algorithms suitable to the application and to the target machine are then used. We also show how, on a distributed architecture, the knowledge of the data flow between the tasks can be used by the system to reduce the communications and to effectively manage the distributed shared memory. This parallel programming model and its implementation in the Athapascan-1 interface are then validated on various architectures and various applications of linear algebra, in particular the two-dimensional Cholesky factorization. The simple programming of these applications in this interface, and the obtained results (for example, we outperform ScaLapack for dense Cholesky factorization on a machine with 60 processors) confirm the validity of our approach.
Complete list of metadatas

Cited literature [29 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00004825
Contributor : Thèses Imag <>
Submitted on : Wednesday, February 18, 2004 - 2:04:11 PM
Last modification on : Friday, November 6, 2020 - 4:39:45 AM
Long-term archiving on: : Wednesday, September 12, 2012 - 1:30:29 PM

Identifiers

  • HAL Id : tel-00004825, version 1

Collections

Citation

Mathias Doreille. Athapascan-1 : vers un modèle de programmation parallèle adapté au calcul scientifique. Modélisation et simulation. Institut National Polytechnique de Grenoble - INPG, 1999. Français. ⟨tel-00004825⟩

Share

Metrics

Record views

407

Files downloads

1176