Intra node parallelization of MPI programs with OpenMP

The availability of multiprocessors and high performance networks ooer the opportunity to construct CLUMPs (Cluster of Multiprocessors) and use them as parallel computing platforms. The main distinctive feature of the CLUMP architecture over the usual parallel computers is its hybrid memory model (message passing between the nodes and shared memory inside the nodes). Some of the primary issues to address for the CLUMP are: 1) to be able to execute the existing programs with few modiications 2) to provide some programming models coherent with the performance hierarchy of the data movements inside the CLUMP 3) to limit the eeort of the programmer while ensuring the portability of the codes on a wide variety of CLUMP conngurations. We investigate an approach based on the MPI and OpenMP standards. The approach consists in the intra-node parallelization of the MPI programs with an OpenMP directive based parallel compiler. The paper presents a detailed study of the approach in the context of the biprocessor PC CLUMPs. It provides three contributions. First, it evaluates the ability of bipro-cessor PCs to eeectively provide a speed up over single processor PCs in the context of shared memory parallel programs. Second, it investigates the method to transform MPI parallel programs in order to execute them on a CLUMP. Third, it presents the performance evaluation of this method applied on the NAS parallel benchmarks executed on a cluster of biprocessor PCs.