Experimental Characterization of Data Centers

With many data center air flow and heat transfer investigations in the literature are computational in nature, high-fidelity analyses requires computational models to be experimentally verified. For example, the fidelity of standard turbulence models implemented in commercially available CFD software applications can be verified by detailed full-field velocity measurements using particle image velocimetry (PIV). For a thermal analysis, temperatures obtained from CFD/HT studies can be validated with data obtained using a three dimensional grid-based temperature measurement system. The measurements can potentially be used to identify the limitations of CFD modeling and quantify the error incurred with common simplifications. Such issues included using a lumped resistance to model perforated tiles and the appropriate boundary conditions to accurately model the CRAC units. Also, there are also some simplified models developed to study different failure modes of data center.

Predictive Modeling for Adaptive Data Centers

Since data centers are mission-critical facilities, any modifications for improving energy efficiencies must not compromise operational reliability. Therefore, determining optimal energy efficiencies require fast and high-fidelity modeling frameworks. Two modeling frameworks being investigated are proper orthogonal decomposition (POD) and Wavelets. These frameworks involve lower dimensional modelings of high-dimensional processes and significantly reduces computation times involved to such an extent that they can deliver near-real-time predictions which facilitate automation and virtualization in data center operations.

Automation and Control Strategies

Several control strategies are being developed to account for the multiple length-scale nature of a data center (i.e. from the chip level to the facility level). Due to the dynamic nature of the compute loads in a data center, particularly with the substantial growth in cloud computing, control algorithms for either the cooling-resource allocation or the compute-load allocation are being explored through the use of linear controllers such as P-controller and PID controllers, and non-linear controllers such as adaptive controllers.