Experimental Validation of the COMPACT Code in Data Centers

Abstract

This paper summarizes the comparison of predictions by a compact model of air flow and transport in data centers to temperature measurements of an operational data center. The simplified model and code package, referred to as COMPACT (Compact Model of Potential Flow and Convective Transport), is intended as an alternative to the use of time-intensive full CFD thermofluidic models as a first-order design tool, as well as a potential improvement to plant-based controllers. COMPACT is based on potential flow and combined with an application of convective energy equations, using sparse matrix solvers to seek flow and temperature solutions. Full-room solutions can be generated in 15 seconds on a commercially available laptop, and an accompanying graphical user interface has also been developed to allow quick configuration of data center designs and analysis of flow and temperature results. Experiments for validation of the model were conducted at the HP Labs data center in Palo Alto, CA, which is in a traditional configuration consisting of inlet floor tiles feeding cold air between two rows of multiple server racks. Subsequently, air exits either through ceiling tiles or direct room-return to CRAC units located on the side of the room. Temperatures were recorded at multiple points along entering and exiting flow faces within the room, as well as at various points in cold and hot aisles, and are presented and compared to model predictions to assess their accuracy. Areas of greater and lesser accuracy are analyzed and presented, in addition to conclusions as to the strengths and weaknesses of the model. For some cases, the average predicted temperature along in-flowing rack faces was within one degree of the average measured temperature. However, the differences in temperature are not evenly distributed. The most pronounced variations between the model and room measurements were located in areas above server racks where recirculation was shown to most likely occur. In these areas, the predicted temperature was higher than experimental values; this can likely be attributed to the absence of buoyancy effects in the simplified potential flow model. Adaptations of the model and its configuration standards for more accurate temperature distributions are proposed, as well as investigations into the effect on temperature comparisons to idealized model output by unaccounted heat sources or flow phenomena.