Building a data center based on Open Compute Project designs in a relatively hot and humid area like Forest City, North Carolina, presented some interesting challenges. Chief among them, of course, was whether the 100% outdoor air cooling system Facebook debuted in our Prineville, Oregon, facility could operate as efficiently in an environment where the ASHRAE 50-year maximum wet bulb temperature is 21% higher, at 84.5°F instead of 70.3°F.
The design we use in Prineville works great in central Oregon – a high desert plain with hot, dry summers and cool evenings and winters. These are ideal conditions for using evaporative cooling and humidification systems, instead of the mechanical chillers used in more-conventional data center designs. And ASHRAE 50-year design weather maximums and bin weather data show that such a system is sufficient for even the hottest summer days in central Oregon.
When we started looking at Forest City, the bin weather data suggested that refrigeration might not be required, but ASHRAE 50-year design weather maximums suggested otherwise. We ultimately decided to install a direct expansion (DX) coil system in the facility, just in case it might be needed, but it was important to us to find a way to make the free cooling system work — the potential efficiency gains to be found in keeping those DX units switched off were just too great to ignore.
[caption id="attachment_2550" align="aligncenter" width="600"] Psychrometric Chart: Bin weather data plotted on Rutherfordton (approx. 6 miles NW of Forest City) suggests that DX cooling is probably not necessary in a typical year; but ASHRAE MAX WB suggests the opposite.[/caption]
To try to make the free cooling system work in Forest City, we expanded the maximum allowable temperature and humidity conditions for our server environment. Because wet bulb temperatures are higher in western North Carolina than they are in central Oregon, we set the upper end of the server inlet temperature range at 85°F, instead of at 80°F. And because of the higher humidity in North Carolina, we expanded the relative humidity (RH) maximum from 65% RH to 90% RH.
Summer 2012 in the US presented an excellent test of the design. July 2012 was the second hottest month on record in North Carolina, in the third hottest summer ever documented in the continental US. At one point the dry bulb temperature topped 100°F outside the data center.
But despite the record-breaking dry bulb temperatures, we didn't run the DX coils at all this past summer. If you look at the trend data, it shows that when the record hot days occurred, relative humidity was low, allowing the misting system to provide all the needed cooling. The chart for July 1 illustrates this.
[caption id="attachment_2549" align="aligncenter" width="600"] Outside Air DB Temperature: 07/01/12 was one of the hottest days of the summer, with a high of 102° F DB. But it didn’t require DX cooling, because the coincident humidity was 26% RH.[/caption]
We also saw several days of high relative humidity over the course of the summer. High RH is potentially problematic in an evaporative cooling system, as our ability to cool air by adding water decreases as the RH of the incoming air increases. But, as illustrated by the chart for June 25, the dry bulb temperature tended to be low on the days RH was high this summer – meaning that we didn’t need to cool it before sending it into the data hall. (In fact, in some cases we actually added hot return air from the data hall to the supply air, to bring the relative humidity down within 90% RH cap.)
[caption id="attachment_2548" align="aligncenter" width="600"] Outside Air Relative Humidity: The most humid day of this summer was 06/25/12, with a 97% RH. But it didn’t require DX cooling, as the coincident DB temperature was 68°F.[/caption]
So it turns out this design can work in relatively high heat and humidity, and it can work as efficiently as it does at Prineville. In fact, the power usage effectiveness (PUE) for Forest City clocked in at 1.07 this summer, versus Prineville's 1.09 during roughly the same period. It remains to be seen how efficiently the system would perform if both the DB temperature and the WB temperature were to simultaneously exceed the limits we’ve set, but the likelihood of such an event in this region is low, and its impact on annualized PUE would likely be minimal.
[caption id="attachment_2551" align="aligncenter" width="600"] 07/01/12 - 07/07/12: One week’s worth of trend data shows that DX cooling wasn’t required as dry bulb was high, but RH was low enough for evaporative cooling.[/caption]
If you’re interested in learning more about these designs and their performance, be sure to register for the upcoming Open Compute Summit, which will be held in Santa Clara, California, on January 16-17, 2013. We hope to see you there!
Dan is a mechanical engineer at Facebook.