The wider supercomputing community and the Department of Energy have reached a consensus that any successful exascale computing design will be constrained to 20 megawatts peak power. Less well understood is that power draw varies within and across applications, and any default power setting will provide optimal performance for only a small percentage of applications.
Recent advances in processor design provide the ability to schedule power consumption where it can maximize application performance. Runtime systems have not kept up with these developments. This project develops and evaluates runtime systems that allow maximum performance of any application for a given power bound.
The project investigates all aspects of power — CPU power, memory power, I/O power, GPU/Phi power — and aims to optimize performance under a power bound that accounts for all these sources in an exascale context.
Supported by LLNL under Subcontract B608929. Any opinions, findings, and conclusions expressed are those of the author(s) and do not necessarily reflect the views of LLNL.