This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).
Heterogeneous architectures with specialized coprocessors are gaining momentum due to compelling performance and energy-efficiency benefits. Heterogeneity introduces two main challenges: how to select the best coprocessor for a particular task, and how to map a task onto a particular accelerator. Programmers should be able to focus on algorithm and software design without worrying about hardware details or sacrificing portability, so rewriting code for every new coprocessor is untenable. While it is possible to schedule and map tasks at compile time, competing tasks and varying workload characteristics mean that effective scheduling must happen at runtime.
The solution proposed with this project is to virtualize the underlying system heterogeneity, which frees programmers from the burden of considering this heterogeneity in their implementation. At runtime, the system software layers schedule tasks to the most effective targets and map the programmer's virtualized task description to that target's architecture. The overarching goal of this project is to design an end-to-end solution to the challenge of heterogeneous code generation and optimization. This project develops the required capabilities. The intellectual merit of this research lies in the novel advances required in the compiler, runtime, and operating system. This work in turn provides broad impact by increasing programmer productivity and providing software tools that enable future research and software development. This work will also train graduate and undergraduate students in cutting-edge computer-systems concepts and design skills, and develop new educational and outreach materials. This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).
