Thermal anomaly detection in datacenters

摘要

The high density of servers in datacenters generates a large amount of heat, resulting in the high possibility of thermally anomalous events, i.e. computer room air conditioner fan failure, server fan failure, and workload misconfiguration. As such anomalous events increase the cost of maintaining computing and cooling components, they need to be detected, localized, and classified for taking appropriate remedial actions. In this article, a hierarchical neural network framework is proposed to detect small- (server level) and large-scale (datacenter level) thermal anomalies. This novel framework, which is organized into two tiers, analyzes the data sensed by heterogeneous sensors such as sensors built in the servers and external sensors (Telosb). The proposed solution employs a neural network to learn about (a) the relationship among sensing values (i.e. internal, external, and fan speed) and (b) the relationship between the sensing values and workload information. Then, the bottom tier of our framework detects thermal anomalies, whereas the top tier localizes and classifies them. Our solution outperforms other anomaly-detection methods based on regression model, support vector machine, and self-organizing map, as shown by the experimental results.