Abstract
Recently, the rapid development of Internet Service brings the increasing demands of servers and large amount of electricity consumption. The electricity used by the cooling system contributes a major part in the server power budget. Therefore, many researches, relating to thermal design and thermal management, aim to reduce power consumption in servers. In recent years, few studies began to explore low fan power consumption by utilizing fan speed control. To date, the design of optimal fan speed controller is based on the system model. However, server fan cooling system has strong cross-coupling effect that makes the system model hard to be obtained. Moreover, the model-based controller design may hardly be implemented because there exists many different types of servers. Concentrating on power saving, this thesis proposes a decoupled fan control, combining proportional-integral-derivative (PID) controllers and fan weighting settings, which does not need to know the system model in advance. The PID control parameters of unknown system model can be obtained through the on-line tuning using neural network. However, the tuning PID controller is only suitable to the system and needs to be optimized in advance in order to reduce the fan power consumption. The optimization of PID gains is based on low fan power consumption in the transient-state response. Furthermore, fan weighting setting is the effective method to reduce the over-cooling of electronic components due to the cross-coupling effect. It can save fan power usage when the system is in the steady state. A server mockup system simulating a real server was constructed to validate the design of the proposed control system. The experimental results show that the temperature responses satisfy the cooling requirements of the server. Furthermore, up to 14% of a server’s fan cooling power can be saved in the transient-state by optimal PID controllers and the proper fan weighting reduces power use by 16.8% in the steady-state.