During the last decade, parallel processor architectures have become a powerful tool to deal with massively-parallel problems that require High Performance Computing (HPC). The last trend of HPC is the use of heterogeneous environments, that combine different computational power units, such as CPU-cores and GPUs. Performance maximization of any GPU parallel implementation of an algorithm requires an in-depth knowledge about its underlying architecture, becoming a tedious task only suited for experienced programmers. In this paper we present TuCCompi, a multi-layer framework that not only transparently exploits heterogeneous systems, but automatically tunes the GPU capabilities by choosing the optimal values for their configuration parameters, using the kernel characterization provided by the programmer. This model is very useful to tackle problems characterized by independent , high computational-load tasks with none or few communications , such as embarrassingly-parallel problems. We have evaluated TuCCompi in different, real-world heterogeneous environments using the APSP problem as a case study.
