One of the critical problems in reaching the exascale computing goal is power. Exascale systems have a target power constraint of 20 megawatts, yet today's petascale systems — with performance at least two orders of magnitude below prospective exascale systems — consume around 5 megawatts. Hardware improvements alone will not bridge this gap.
We are developing a run-time system called Conductor to address the power issue. The overall goal is to produce near-optimal application performance under a prescribed power bound. Conductor carries this out by allocating power both between and within nodes:
- Power Scheduling — Addresses the inter-node case, allocating power across nodes to maximize aggregate performance.
- Power Gating — Addresses the intra-node case by powering off individual components in a fine-grain manner beyond what architectures currently provide.
- User Annotations — Allows users to assist Conductor in cases where power scheduling and gating alone are insufficient.
Conductor is essential to achieving exascale performance on nontrivial applications and will help push toward the exascale goal — an important national priority that will impact many application domains.
Supported by the National Science Foundation (NSF) under grant no. CNS-1216829. Any opinions, findings, and conclusions expressed are those of the author(s) and do not necessarily reflect the views of NSF.