Cooling anomaly detection for servers and datacenters with Naïve ensemble
Cong Li · 2016
We propose a novel approach to predictive analysis of potential cooling failures in servers and datacenters in which unsupervised anomaly detection is performed in multi-dimensional temperature sensor data. A naïve and obviously invalid independence assumption is employed to model the probability distribution. We provide a theoretical justification demonstrating that the approach relies on correctly comparing the probabilities estimated rather than accurate probability estimation. The approach is also justified empirically in simulation experiments for two different predictive failure analysis scenarios: identifying potentially worn-out fans based on the server component temperature sensor data and identifying computer room air-conditioner failures before hotspots arise based on server inlet temperature sensor data.