Home / server / Questions / 595135
In Process
Pavel
Asked: 2014-05-15 09:58:03 +0800 CST

Water-Cooled Rack Door: does it work well with Blades or GPU nodes?

  • 772

We're considering to equip out server farm with water cooled rack doors, similar to

since this is a pretty specific type of hardware, it's hard to find any information on those things that is independent from vendors.

In the most simple setting, we would buy a single read-door-cooled rack connected to a Coolant Distribution Unit (CDU) that provides about ~20 kW of cooling capacity. In this rack we would like to deploy

  • option 1: 3 Blade enclosures (x16 nodes) or
  • option 2: 2 enclosures + several GPU nodes

Both options will produce about 17 kW of heat. All dimensionality questions aside (water/room temperature, water flow, pressure, etc), does anyone of you have experience with such cooling rack doors and could share how well they works in a room that has approx. 20°C air temperature? The vendors promise t_out = t_in, but I'm curious whether it holds true in case of a really hot hardware. My concerns include:

  • the heat is not uniformly distributed over the door area, since the blades have those hot spots where the fans are blowing lots of hot air
  • how strong does a cooling door slow down the air flow? isn't there a chance of air congestion?
  • is it possible for the CDU to adapt to changes in the room temperature? on a hot summer day, the room temperature may increase by a few degrees, so that it would be nice to have t_out < t_in.

The alternative is to set up a proper water-cooled rack that is completely closed and that pumps chilled air in front of the computers. however, those are typically somewhat more expensive. that's why I'm interested in your opinions on the read-door-cooling in the first place. Especially in case of very hot computers.

Thanks!

rack

1 Answers

  1. In my experience yes the doors work very well. I have only worked with the doors from IBM on their iDataPlex racks but they work great.

    • 1