COTS cooling

Robert G. Brown rgb at
Wed Apr 24 06:50:38 PDT 2002

On Tue, 23 Apr 2002, Robert B Heckendorn wrote:

> We don't have to pay for the cooling but the cost of the installation
> of cooling is being used as an argument to cut corners on the machine
> itself.  :-( So I would love to get the cost of the installation of
> cooling down.
> One of the responses to my mail said:
> "We just purchased ~150 dual AMDs, and are cooling them with 4 Fujitsu
> ceiling-mounted air-conditioners: about 50kW of AC cost us about $25k,
> which is about 10% of the cost of the machines."
> This sounds like COTS cooling to go with our COTS machines.  :-) It
> has the nice feature that if one AC goes out the others keep running.
> It is also nice in that half a dozen 125KBTU/hr units in the ceiling
> would seem to handle a fairly large load and all machines for the next
> 4 years of expansion.
> 450W/dualnode * 3.4BTU/hr/W * 400 nodes  = 612K BTU/hr
> Does anyone else comments on this scheme (pros or con)?
> Is anyone doing anything like this?

An AC consists of two separate components.  One is the heat
exchanger/blower/ductwork, which generally lives "in" the space being
cooled. To remove a lot of heat, one has to move a lot of air over a lot
of cold surface, so whether you use one large blower or several smaller
ones, you have to move the same volume of air over the same area cooled
the same amount, and one ALSO has to locate the ducting so that cold air
flows out, gets pulled through your systems, and then goes into the
return efficiently.  Otherwise your room will have hot spots and cool
spots, and hot-spot nodes will fail.  You can feel the heat just walking
past banks of nodes in our room, but there is always a feel of cold air
going past you towards the heat as well.

The other is the chiller, the part that actually takes the heat from the
room, squeezes it out (literally) into the outside air, and returns
cold-something (water, coolant, whatever) to the in-room heat exchanger.
Window unit ACs combine the two into a single package -- in central AC
and building AC they almost always are distinct.

In a centralized operation, the chillers might be located far from the
room.  They may have constraints on WHERE they can be located (typically
on the roof, for example, but likely only on certain parts of the scarce
roof real estate).  Getting unplanned insulated high volume pipes
through a big building made out of steel reinforced concrete is
nontrivial, getting power to the roof for new chillers is nontrivial,
putting the (heavy) chillers on the roof where they won't just fall
through onto the people working below is nontrivial.  Nontrivial =
expensive.  Then there is a wide range of ways that people/organizations
"bill" for this sort of construction -- beware creative accounting.

So sure, it might cost a relatively small amount to install a relative
large number of relatively cheap chillers and heat exchangers IF your
room e.g. has an outside wall with a preexisting window or ductwork to
the outside, there is plenty of room outside for a concrete pad and
wiring to support the chillers, and so forth.  OTOH, if you are in the
basement (we are) of a large building with an existing chiller delivery
system and "have" to add capacity or upgrade capacity to our existing
chiller farm (as does Bob, clearly, and likely many others) and get
"billed" according to how they account for the cost (which may add in
all sorts of administrative, architectural, engineering expenses and not
just the cost of the hardware) and if your room is fairly small and has
relatively low ceilings (precluding lots of ceiling mounted heat
exchangers) this might not work.

YMMV.  Some sites might be able to do things more cheaply than others,
and it wouldn't surprise me at all to see a factor of 4 difference in
cost from one end to the other.  Ours was quite expensive, but it gives
us a pretty good node density in a relatively small space, and SPACE is
very, very "expensive" in our building (which we "share" with math while
both departments try to grow).


Robert G. Brown	             
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at

More information about the Beowulf mailing list