Abstract
Meeting the power requirements for future Exascale computing machines is one of the major challenges. Power consumption in a high performance computing (HPC) data center comprises machine power and cooling power. Earlier work has focused on minimizing machine power consumption but not much attention has been paid to minimize the power spent on cooling.
We propose a dynamic frequency and voltage scaling (DVFS) based technique ensuring all cores in a job run below a user defined temperature threshold. This control comes at the cost of execution time penalty, which is amplified in tightly coupled parallel applications because a single slow-running core can slow the entire application down. We develop a strategy that minimizes this penalty. We present results from a 160 cores cluster with a dedicated air conditioning unit. Not only can we constrain the average core temperatures, we are also able to reduce chances of hot spots.