Measurement and Management Technologies for Data Centers - Modeling and Analytics
The MMT modeling architecture distinguishes between measurement, base and dynamic models. The differentiation results from the fact that "changes" in the data center occur on very different time scales. Measurement models are obtained from very detailed and high resolution assessment data, which might only be available once a year or even less frequent. Alternatively, detailed CFD data sets could be leveraged although actual measurements are generally preferred. The measurement model should be thought of as an initial starting point defining room dimensions, power delivery systems, IT layout, initial power consumption and air flow of servers etc. Measurement models use rendering techniques to display detailed heat maps and pin point hotspots etc. helping operators to understand the specific environment of the data center. New base models have to be generated if new equipment is added, racks get moved, which might happen as often as once or twice a week. This requires solving partial differential equations (PDE), which often results in very long calculation times. We are developing algorithms and techniques to drastically accelerate these models by leveraging the fact that the availability of more measurement data can be “traded” against the complexity of the PDE description. The deviation from the latest base model - due to daily fluctuations such as changes in server utilization (and thus power levels) or throttling of fans - will be handled by fast ( 10 s) dynamic models. For example, we use knowledge based models and trends, which are obtained from both measurements and physics models and then combine these trends from a knowledge base with statistical kriging techniques.
One
Two Three