Supercomputers machines, HPC systems in general, embed sophisticated and advanced cooling circuits to remove heat and ensuring the required peak performance. Unfortunately removing heat, by means of cold water or air, costs additional power which decreases the overall supercomputer energy efficiency. Free-cooling uses ambient air instead than chiller to cool down warm air or liquid temperature. The amount of heat which can be removed for-free depends on ambient conditions such as temperature and humidity. Power capping can be used to reduce the supercomputer power dissipation to maximize the cooling efficiency. In this paper we present a power capping approach based on Constraint Programming which enables to estimate at every scheduling interval the power consumption of a given job schedule and to select among all possible job schedules the one which maximizes the supercomputer efficiency.
展开▼
