With power consumption becoming a critical—if not the critical
supercomputer operating constraint, future systems must adhere to
strict power limits. For example, in the US, exascale systems cur-
rently have a target power consumption—set by the Department
of Energy—of 20 MW, other agencies world-wide have set simi-
lar limits. We anticipate that total machine power will be divided
across multiple simultaneous jobs, with each job being allocated a
power bound and a set of nodes. This motivates the development
of techniques for optimizing performance under a strict power con-
straint for multi-node, multi-core systems. The problem at hand is
then to choose a per-processor configuration (a number of cores and
frequency/voltage [DVFS] state) for a given application executing
on a particular number of nodes.