Skip to Main Content
Spacecraft on-board platforms will soon contain high performance off-the-shelf computing components, as the only way to deal with applications which require High Performance Computing (HPC) capabilities. Heterogeneous architectures, give the possibility to profit from specific features of such components, by combining them on the same computing platform. Such architectures integrate different multicore CPUs with many-core accelerators (GPGPUs) so that parts of the same application can efficiently execute on the CPU while other parts can profit more from the GPU features. The problem stands in the way how to program these heterogeneous architectures. In this paper we introduce the OmpSs Programming Model as a framework to program heterogeneous architectures. For benchmarking purposes, we use a Synthetic Aperture Radar (SAR) application, which we parallelized with OmpSs for Symmetric Multi-Processor (SMP) architectures and for hybrid SMP/GPU environments. We also compare the results obtained with OmpSs with the ones obtained with OpenMP in SMP environments and it turns out that OmpSs outperforms OpenMP. When applying OmpSs to the SAR code for hybrid architectures, OmpSs does not offer better performance than CUDA, but it offers support for heterogeneous architectures and it also increases programmer's productivity.
Date of Conference: 25-28 June 2012