Skip to Main content Skip to Navigation

Expression et optimisation des réorganisations de données dans du parallélisme de flots

Abstract : Embedded systems designers are moving to multicores to increase the performance of their applications. Yet multicore systems are difficult to program. One hard problem is the expression and the optimization of data reorganizations. We would like to propose a compilation chain that : 1) uses a simple high-level syntax to express data reorganizations within a parallel application ; 2) ensures the deterministic execution of the program (critical in an embedded context) ; 3) optimizes and adapts the program to the target's constraints. To address point 1) we propose a high-level language, SLICES, describing data reorganizations through multidimensional slicings. To address point 2) we show that it is possible to compile SLICES to a dataflow language, SJD, that is built upon the Cyclostatic Data-Flow formalism and therefore ensures determinism. To address point 3) we define a set of transformations that preserve the semantics of SJD programs. We show that a subset of these transformations generates a finite space of equivalent programs. This space can be efficiently explored with an heuristic that selects the program variant more fit to the target's constraints. Finally, we evaluate this method on two classic problems : reducing memory and reducing communication costs in a parallel application.
Document type :
Complete list of metadatas

Cited literature [78 references]  Display  Hide  Download
Contributor : Pablo de Oliveira Castro <>
Submitted on : Saturday, March 26, 2011 - 11:12:02 PM
Last modification on : Monday, February 10, 2020 - 6:12:33 PM
Long-term archiving on: : Saturday, December 3, 2016 - 11:24:43 PM


  • HAL Id : tel-00580170, version 1




Pablo de Oliveira Castro Herrero. Expression et optimisation des réorganisations de données dans du parallélisme de flots. Informatique [cs]. Université de Versailles-Saint Quentin en Yvelines, 2010. Français. ⟨tel-00580170⟩



Record views


Files downloads