Improving comunication in task-based runtimes for heterogeneous systems

HPC machines in the race for exascale computing are more heterogeneous than ever. The complexity of the systems makes the programming and interoperability of the multiple accelerators a huge task that most of the times the programmer must handle directly. OmpSs-2 Programming model aims to help the programmer handle the heterogeneous systems hiding all the data communication and synchronization efforts and making offloading tasks to different devices easier. This master thesis explores some ways to handle this problems and proposes a framework to add hardware support to OmpSs-2 runtime , without relying in architecture-specific functionalities like unified-memory. The results show that the framework and implementation proposed can increase the performance of the applications when compared with previous versions of the runtime, and is able to easily support the addition of new devices to the OmpSs environment