The goal of this project is to study dynamically adapting compiler-translated MPI applications in a distributed system. In previous work we have created a system for dynamic adaptation and tuning of computational applications. The current study includes, deploying a translator that converts OpenMP parallel applications directly into MPImessage passing programs, and investigating new methods to enable dynamic, adaptive optimization and tuning in diverse architectures. The architecture space explored includes homogeneous multiprocessors, clusters, and heterogeneous systems, such as computational Grids. The project includes: selecting compilation parameters in the OpenMP-to-MPI translator, developed previously, that are amenable to dynamic adaptation, and turn these parameters into compiler options; adaptively optimize these parameters with a perfomance tuning system; and study the potential for dynamic adaptation through analytic and measured evaluation of the overheads and gains of performing adaptation during the program execution. An important part of the evaluation will be to test the methods pursued on applications that exhibit potential for such adaptation. As part of the project a workshop on Compiler-driven Adaptive Execution Systems will be organized, to convene researchers active in the field of dynamic and adaptive Compilation, and address a broad range of topics from success stories of compiler-driven adaptive execution to the potential of such methods to the enabling technology.
