A Loosely Synchronized Execution Model for a Simple Data-Parallel Language (Extended Abstract)

Data-Parallel languages ooer a programming model structured and easy to understand. The challenge consists in taking advantage of the power of actual architectures by a compilation process allowing to reduce the number and the complexity of synchronizations. In this paper, we clearly separate the synchronous programming model from the asynchronous execution model by the way of a translation from a synchronous Data-Parallel programming language to an asynchronous target language. The synchronous Data-Parallel programming language allows to temporarily mask local computations. The asynchronous target language handles explicit and partial synchronizations through the use of structural clocks. We provide a theoretical framework to prove the correctness of optimisations. R esum e Les langages data-parall eles oorent un mod ele de programmation simple et structur e. La diicult e majeure pour adapter ce mod ele a la puissance des architectures actuelles r eside dans l' elaboration d'un processus de compilation r eduisant le nombre et la complexit e des synchronisations. Dans ce rapport, nous s eparons clairement le mod ele synchrone de programmation du mod ele asynchrone d'ex ecution par l'interm ediaire d'une traduction d'un langage data-parall ele synchrone vers un langage cible asynchrone. Le langage data-parall ele synchrone autorise le masquage temporaire des calculs locaux. Le langage cible asynchrone r ealise des synchronisations explicites et partielles par l'interm ediaire d'horloges structurelles. Nous proposons un cadre th eorique pour prouver la correction des optimisations. At present time, most of distributed architectures are MIMD computers or networks of workstations. The control-parallel programming model is close to their execution model. However, in this model, the programmer has to control several interacting control ows. A control-parallel program is therefore diicult to understand. Moreover this model is a great limitation for scalability and massively parallel architectures. A crucial step has been performed during eighties with the data-parallel programming model, following upon success of SIMD architectures. On a SIMD architecture, the control is centralized and processors execute synchronous actions on their own data. The data-parallel programming model directly corresponds to this execution model. A data-parallel program is a single sequential composition of actions on parallel data. The data-parallel semantics imposes actions to be synchronous on all processors. It involves a single program, hence the control is centralized and very close to a sequential one. Data-parallelism inherit its qualities from classic sequential programming, it is a structured and easy to understand model. A crucial point is to eeciently run …