On developing a fast, cost-effective and non-invasive method to derive data center thermal maps

Ongoing research has demonstrated the potential benefits of thermal-aware load placement in data centers to both reduce cooling costs and component failure rates. However, thermal-aware load placement techniques have not been widely deployed in existing data centers. This is mainly because they rely on a thermal map or profile of the data center, the derivation of which is an interruptive process to the data center operation. We propose a noninvasive solution of producing a thermal map; it consists of training a neural network with observed data from actual data center operation. Our results show that gathering the data and selecting a training set is a fast process, while the neural network with no hidden layers achieves the lowest mean squared error.