A Unifying Approach of Data and Task Parallelism Based onStructural

Among parallel programming tools, task parallelism and data parallelism are the most common programming models. Task parallelism enables to reach high performances but requires an important programming eeort. In opposite, data parallelism provides programming facilities close to the sequential model at the expense of performances. In this paper, we show that it is possible to propose a language that integrates a task parallel execution model and a data parallel programming model in a unique and unifying framework. This ability depends on a speciic synchronization/communication model relying on structural clock mechanism. We describe an implementation of our model on Cray T3D. It highlights experimental performances that outperform classical MPI programming. Parmi les outils de programmation parall ele, le parall elisme de t^ ache et le parall elisme de donn ees sont les mod eles de programmation les plus courants. Le parall elisme de t^ ache permet d'atteindre de hautes performances mais il n ecessite an eeort de programmation important. A l'oppos e, le parall elisme de donn ees autorise une programmation simple, proche du mod ele s equentiel, au d etriment des performances. Dans cet article, nous montrons qu'il est possible de proposer un langage int egrant dans un cadre unique et uniicateur un mod ele d'ex ecution a parall elisme de t^ ache et un mod ele de programmation a par-all elisme de donn ees. Notre langage repose sur un mod ele de synchronisation/communication particulier utilisant un m ecanisme d'horloges structurelles. Nous pr esentons une implementation de notre mod ele sur Cray T3D. Cette implementation fait appara^ tre des performances qui surpassent celles obtenues a partir d'une programmation en MPI.