All webinars

Modeling Data Centers During a Power Outage

CFD models of three data center designs show how fast rack inlet temperatures rise after a grid outage and what slows the rise.

This webinar uses transient computational fluid dynamics (CFD) to show how rack inlet temperatures change in the minutes between a grid outage and the cooling system recovering. It is for operators and designers working at higher densities and lower Power Usage Effectiveness (PUE), where the margin for this event is smaller. The presenter is the president of Applied Math Modeling, which develops CoolSim.

What the webinar covers

The talk starts with why the problem is getting harder. Designs that lower Power Usage Effectiveness (PUE) use hot aisle containment and higher supply temperatures, which leaves less thermal margin when power is lost.

Case 1 is a non-contained raised floor hall. The model assumes a generator that is running after 10 seconds, fans at full speed by about 30 seconds, and cooling capacity that returns only after 120 seconds. Some racks heat up at once, but the cool air in the room and the underfloor lets the hall ride through the gap for roughly a minute or more.

Case 2 is a high-density hall with hot aisle containment and a fan wall. Cooling data from a supplier shows the fans ramping in about 20 to 25 seconds, and compressor stages returning in sequence after roughly 70 seconds. When the fans stop, the room pressure falls below the contained aisle pressure. Hot air then leaks back through rack rails, blanking panels and seams, and the maximum rack inlet temperature climbs steeply within about 30 seconds.

Case 3 is a ducted supply hall on a chilled water loop, with two scenarios. In the first, the air handlers shut off completely and rack inlets rise within seconds. In the second, the chiller stops but the pumps and fans stay on, and the water already in the loop provides cooling for a couple of minutes.

The talk closes with how such studies are done.

Key points

  • The presenter’s view is that a lower PUE design leaves less margin in a power interruption.
  • In a contained hall the pressure gradient can reverse when the cooling fans stop. The model shows hot air leaking into the cold side and a rapid rise in inlet temperature.
  • Restoring fan airflow quickly matters, because compressor-based cooling needs a staged restart that cannot be rushed.
  • Water already circulating in a chilled water loop can buy time if the pumps and air handler fans stay on backup power. In the example the benefit arrives as steps, as each slug of cool water passes through.
  • Load banks do not behave like servers, so the presenter argues that modeling is the practical way to predict the outcome before it happens.
  • Transient studies need fan and cooling capacity data against time from the equipment supplier. The presenter says that a steady-state model is built first and the transient run is then done on a cluster.