HOMPI: A Hybrid Programming Framework for Expressing and Deploying Task-Based Parallelism

This paper presents HOMPI, a framework for programming and executing task-based parallel applications on clusters of multiprocessors and multi-cores, while providing interoperability with existing programming systems such as MPI and OpenMP. HOMPI facilitates expressing irregular and adaptive master-worker and divide-and-conquer applications avoiding explicit MPI calls. It also allows hybrid shared-memory / message-passing programming, exploiting fully the availability of multiprocessor and multi-core nodes, as it integrates by design with OpenMP; the runtime infrastructure presents a unified substrate that handles local threads and remote tasks seamlessly, allowing both programming flexibility and increased performance opportunities.