Thermal anomaly prediction in data centers

In recent years, the demand for data centers has seen tremendous growth. Simultaneously, power densities have increased resulting in greater chances of thermal anomalies - situations where the temperature at a location exceeds the safety threshold for equipment placed there. In this paper, we explore some techniques for predicting such thermal anomalies so that preemptive steps can be taken to address them. Four such techniques - a simple threshold method, a moving averages-based method, an EWMA-based method, and, a machine learning technique called naïve Bayesian classifier - are tried on three months of temperature sensor data obtained from a real data center. The initial results are encouraging and the naïve Bayes method performs better than the others, although the false positive rate needs to be improved.