Hybrid Parallel Programming with MPI/StarSs
Jesús Labarta, Marjanović Vladimir, Eduard Ayguadé, Rosa M. Badia, Mateo Valero · Advances in parallel computing · 2012
This paper describes how hybrid MPI/StarSs programming provides an efficient way to program future clusters of multicore nodes. The use of the programming model results in achieving overlap between communication and computation while maintaining a simple source code structure. Additional benefits include tolerance to limited bandwidth availability and tolerance to OS noise. This paper presents three different implementations of this approach due to the requirements of three different supercomputing platforms: MareNostrum, Jugene and Juropa.