This project will develop an application-driven runtime scheduling substrate for adaptive and irregular computations on high-performance computing platforms which combine hardware multithreading, cache-coherent symmetric multiprocessing, and message-passing. Unstructured mesh generation and refinement will be used as a working example of an adaptive and irregular application with variable and unpredictable computation and communication patterns. Also, parallel adaptive mesh generation is an important building block for many large-scale scientific and engineering simulations like crack propagation, an end-to-end adaptive application we will use to evaluate our approach. The proposed substrate will allow the programmer to use message passing plus a single address space to achieve communication between application-specific decompositions and leverage compiler and runtime support to exploit multilevel and irregular parallelism within SMP nodes. In order to achieve this goal, the re-structuring of both the application and the runtime support system is proposed. A compiler-driven runtime scheduling engine for fine-grain parallelism exploitation and tighter integration of message-passing with hardware multithreading and cache-coherence will be developed.
