Section: New Results

Runtime System Interoperability with StarPU

Parallel HPC applications increasingly build on multiple parallel libraries, which results in interferences if the parallel entities in the application and in the libraries it uses access computing resources in an uncoordinated manner. A set of resource management APIs has therefore been designed within the context of H2020 project INTERTWinE (see http://www.intertwine-project.eu/developer-hub/resource-manager), and implemented in the StarPU task-based runtime system developed by Team STORM, as well as in the OmpSs/Nanos 6 task-based runtime system developed at the Barcelona Supercomputing Center (BSC). It enables StarPU and OmpSs to interoperate within an application, along multiple scenarios such as nested interoperability, with a host runtime system executing parallel tasks over a guest runtime system, or concurrent interoperability, with several runtime systems dynamically sharing computing resources over the application lifespan.