Power At Scale: Addressing HPC Power Issues on Massively Parallel Machines
The wider supercomputing community in general and the Department of Energy in particular 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 by necessity provide optimal performance for only a small percent of applications. Instead of being limited to a fixed design, 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. In this project we propose to do the research necessary to develop and evaluate runtime systems that will allow maximum performance of any application for a given power bound.
We will be investigating all aspects of power---for example, CPU power, memory power, I/O power, GPU/Phi power. Exascale computing will involve optimizing performance under a power bound that involves all of these power sources.
People
- Faculty:
- David Lowenthal
- Postdoc:
- Aniruddha Marathe
- Graduate Student:
- Tapasya Patki
This material is based upon work supported by LLNL under Subcontract B608929. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of LLNL.