SRC09. Enabling a Data-Centric Model on the Open Community Runtime
Student: Sri Raj Paul (Rice University)
Supervisor: Bala Seshasayee (Intel Corporation)
Abstract: Achieving exascale performance requires addressing challenges arising from the complexity of exascale architecture, its power constraints, as well as application complexity. Exascale architectures, expected to consist of millions of heterogeneous cores and extremely non-uniform hierarchical memory, have to optimally distribute applications, data and parallelize computation. Dynamic task-based execution models hold promise in achieving this, as they help express fine-grained parallelism, along with decoupling computation and data from underlying resources. Open Community Runtime (OCR) is a community-led effort to explore various asynchronous task-parallel runtime principles that can support a broad range of higher-level programming constructs. Legion is a data-centric programming model in which the runtime extracts task-based parallelism from programs, freeing the developer from having to explicitly express them. In this poster we describe our efforts to run Legion programs on top of OCR as their underlying runtime, to combine the parallelism extracted by Legion with the flexibility provided by OCR.
Two-page extended abstract: pdf