Automatic Mapping of Scans and Reductions

Execution of scientiic programs on Distributed Memory Machines yields a real speed-up when communication costs are minimized. This is the aim of a placement function. Our paper focuses on mapping problem and presents an algorithm which computes a multidi-mensional placement function. It automatically distributes both code and data. Moreover some particularities of target machines { such as the provision of preex parallel instructions { are taken into account in order to eeciently exploit the parallelism which is detected under the form of scan operator. La minimisation des communications est une t^ ache primordiale pour d eterminer le placement du code et des donn ees d'un programme scientiique sur les machines parall eles a m emoire distribu ee. Cet article pr esente un algorithme qui calcule automatiquement une fonction de placement multidimensionnelle. De plus, des caract eristiques de la machine cible, telle l'existence d'instructions de type pr eexes parall eles, sont prises en compte aan d'exploiter eecacement le parall elisme d etect e sous la forme d'op erateurs de scans.