No abstract

In this paper, we consider a server cooling system in a data center with fans and a super-multipoint temperature sensing technology using optimal time-domain reflectometry of optical fiber. The sensing system was developed to visualize the temperature distribution of the room in real time, and it can also be a key technology to control the distribution in order to reduce total power consumption in data centers. In this paper, we first present a concept of the fan control system. The objective of the system is to uniformize the rack air inlet temperature distribution in the presence of heat from each server. Since the control scheme requires a dynamical model of the temperature variations in the room, this paper mainly addresses the modeling. The challenge of the modeling stems from a large volume of data provided by the sensing system. It is computationally hard to directly apply standard identification techniques to the data. We thus present a two-stage reduction scheme of the output dimension using the concept of so-called mutual information. The effectiveness of the proposed scheme is finally demonstrated using real data.

: This paper investigates a heating, ventilation, and air-conditioning (HVAC) system in a data center equipped with a previously developed super-multipoint temperature sensing system. This system is expected to be a key technology for reducing the total power consumption of the HVAC system by controlling the inlet temperature distribution of the servers in real time. For this purpose, we present an overview of our fan-control system based on model predictive control. The main objective of this paper is to identify a dynamical model of temperature variations, in order to predict the future evolution of the distribution. However, the spatially high-density temperature data provided by the sensing system is not suited to the needed model accuracy, and the present modeling problem is differentiated from standard ones. We thus present a systematic scheme for the spatial density reduction of sensors by using spectral clustering and graph theory and associated techniques to acquire the dynamical model. Through simulation with real data, we finally show that the developed model achieves an accuracy of 0.58 degrees Celsius on average.

No abstract

No abstract