Previously, I co-authored a semitherm paper on the use of our HVAC CFD tool, FloVENT, on the design of a Mentor Graphics data center. Recently I have seen this topic republished in the data center journal with the title “EDA Vendor takes its Own Advice on Data Center Design”. I find this topic interesting, so I thought I’d blog about this and break from my typical CFD blog topics on household applications.
To give you the background, over the years, to support Mentor’s growing needs, individual offices had created their own small data center rooms. Mentor saw a way to improve capacity, efficiency and redundancy by consolidating these small data centers into a large data center. It just so happened that around this time, Mentor Graphics had just acquired Flomerics, so we now had some internal capabilities to do CFD analysis on this project using FloVENT.
I’ll skip a lot of the details that have already been covered in the semitherm paper and published elsewhere, and just focus on the design progression. To start with, the initial design concept was for a raised floor design. These are very common in exisitng data centers. By having the air conditioned air put into the raised floor plenum, you theoretically allow it achieve a uniform pressure distribution and it allows you to control where the air is supplied to the equipment using perforated floor tiles. Noticed I said theoretically. In our case, we stared with a 2 foot raised floor design, but besides the extra cost of $35 a square foot (~$100,000 total), the FloVENT simulation showed there would be pressure problems and the raised floor plenum had to be bigger. Increasing the raised floor height to 3 feet would increase the cost from $35/sqft to $40-45/sqft. In comparison, the initial cost of ductwork for a ventilation system was only $10/sqft, so the raised floor design was abandoned.
The next design contender was a dropped ceiling design, where a chimney system would duct the hot exhaust air from the cabinets to the dropped ceiling plenum. Here, the air would be allowed to achieve a uniform pressure before being ducted to the roof top air conditioning units. There were 2 main issues with this design. A simple FloVENT simulation indicated the plenum would need to be 10 ft deep, but as the building was only going to be 20 ft high, this would drastically cut the data center air volume which would have a negative impact on the supply air uniformity. You can see from this from the 4 FloVENT result images below. If the supply air is not uniform, then you are going to have racks that are either getting hot air, which is bad, or not enough air, also bad. The 2nd issue related to fire safety, as a return plenum of this size would need to be metal lined to meet fire codes, and the data center could be shut down if the fire marshall was ever not happy with it. So this design was also abandoned.
We also looked at hot aisle containment, where the area that the cabinets exhaust to is sealed with doors. The issue with this design is maintanence of the equipment. OSHA has regulations about how long someone can work in a area at a certain temperature. As some servers have a 54 deg F temperature rise, the contained hot aisle would have temperatures in above 110 degF. The OSHA regulations for this harsh environement made this option not cost efficient.
The other design option was a chimney system, so that a technician would only be exposed to some exhaust air but would be working in the cooler supply air in the room. Because this averaged air temperature was substantially cooler, the OSHA regulations were not detrimental to the design. Below is an image of multiple cabinets being serviced at once. FloVENT showed that even when the equipment was being serviced, 2/3 of the exhaust air continued to travel through the chimney system. This was the winning design concept, to utilize exhaust chimneys to duct the hot exhaust air away from the servers.
Even with that though, there was still plenty of FloVENT analysis to do to iron out the design details. One example of this was the design of the collector duct. Basically, the chimneys from the equipment feed into a collector duct, which then feeds a return plenum, that supplies air to the roof top air conditioning units. What FloVENT found was if the return plenum was at one side (left in the image below), it causes a substaintial pressure gradient along the collector duct, so that the cabinets at the opposite end had air flow issues. This was corrected by moving this connection to the middle of the collector duct, which substantially reduced the pressure gradient some servers had to work against.
I hope this helped show a real life design that utilized CFD analysis as a intrigral part of the design process. Without using FloVENT to analyze this data center, we would have not came up with a efficient design. For those of you interested in more details on this project, the title for the semitherm paper was “Data Center Design Using Improved CFD Modeling and Return on Investment Analysis”. You can download a copy of the paper from the IEEE website.