[Beowulf] Re: /. Cooler room or cooler servers?
mathog at mendel.bio.caltech.edu
Mon Apr 11 11:10:11 PDT 2005
> David Mathog wrote:
> > Convert 16KW to horsepower and find 16000 * .00134 => 214 hp.
> 16KW / (746W/HP) = 21.45 HP
Kevin is correct, the decimal point slipped one position in
my post. Oops. So a 16KW rack is in the range where if it were an
automobile motor it probably would be air cooled. As others
have pointed out, motors air cool better with ambient temp air
than do computers because the difference in temp between the heat
source and the cooling medium is much greater.
We've discussed water cooling. What about a phase change based
system? Specifically, some coolant which is liquid until it hits
the CPU and then boils, with cooling taking advantage of the latent
heat of vaporization, which is likely to be much larger than the
heat capacity of a normal liquid coolant. I'm not talking about
the little units that go on each CPU and radiate out the back
of each case, but rather a common system where all the tubes from
the heat sinks join up outside the case and go off to one big condensor.
There are numerous substances that boil at around the right temperature,
although pentane (bp 36C) has some obvious disadvantages. Most of
the freon based refrigerants I've seen boil at -30C or so, and that's
actually a bit cooler than needed, since going that route would
lead to condensation problems on the return pipes. Surely there must
be heavier ones that boil at higher temps, and all of these would
be insulators, which is good if leaks occur. Water could
be used if the vapor pressure in the gas phase section was very low.
Pretty sure that's how the little heat pipes in the Shuttle mini
computers function. For wiring many machines to a common condensor
probably better though if the gas phase was at around 1 atmosphere
so that a leak somewhere wouldn't mess up the cooling so dramatically,
at least until all the coolant had boiled off.
One might imagine a relatively simple system with a
"room coolant level" (RCL) below which every CPU must be located.
Distribution pipes from the bottom of the main coolant tank
(at room temp) gravity feed all the CPUs. Effectively they
are all submerged. A second pipe out the "top" of each cooling
block goes upward, carrying the evaporated coolant, and these join
together somewhere above RCL where they enter a condensor,
presumably outside the room, and the condensed coolant comes back
in to the top of the coolant tank. If proper geometry is maintained
no pumps would be needed. Of course the condensor could be
located in the machine room if an existing chilled water
line is available.
The trickiest thing to my mind would be keeping the pressure
in the gas phase part of the loop low enough that each
CPU remains "immersed". If the vapor pressure goes too high
the coolant would be pushed down below the heat sink at which
point effective cooling would cease.
mathog at caltech.edu
Manager, Sequence Analysis Facility, Biology Division, Caltech
More information about the Beowulf