Abstract
This paper presents improved learning and optimal cooling models to minimize the power consumption of computer room air conditioners (CRACs) in large data centers. These models consist of a learning model of large data center's thermal system and a CRAC's optimal cooling model. The learning model uses L1-norm regularization for efficient learning. The optimal cooling model involves feedback control of CRACs, linear programming for CRAC's control and server's workload placement based on simple sorting. The proposed models are expected to result in fast learning of the thermal system and large reduction of CRAC's power consumption. Simulation experiments are conducted to evaluate its learning and energy efficiencies. The simulation results indicate that the proposed models resulted in efficient learning of large data center's thermal system and large increase of energy efficiency compared with that of a baseline control model.
