liquid nitrogen cooling a possibility?

Robert G. Brown rgb at phy.duke.edu
Tue Jun 12 11:40:02 PDT 2001


On Tue, 12 Jun 2001, Alan Ward wrote:

> Liquid nitrogen seems a bit nasty (cf the Octool reference), but what about CO2 at low pressure? Ai'nt expensive
> either. Though perhaps not at all practical :-)

I think your final observation is the apropos one, Alan;-)

We seem to be recapitulating our just (finally!) finished discussion on
fancy "wet" cooling systems, only we've gone from moderately insane to
positively demented.

Let me see if I can wrap this up.

  a) The best way to build a cluster out of commodity computers is to
engineer those commodity components so that their operating environment
is precisely according the manufacturer's recommendation as otherwise
the system's behavior becomes unpredictable (in the sense of not
certified by the manufacturer and quite possibly fundamentally unknown).
The only way to find out if a system would run at all at temperatures
well below spec is to cool one down there and run EXTENSIVE tests.
Anyone with the time, money and inclination to do so is welcome to,
otherwise why waste time with speculation?

  b) High quality heatsink/fan units are themselves commodity items that
are easily matched with CPUs and are not particularly expensive.  High
quality chassis cooling fans and power supplies ditto.

  c) The technology for cooling air (the ambient air of the room in
which the computer is kept) is simple, cheap, plentiful and robust.  The
technology for delivering any sort of wet cooling or exotic cooling to a
CPU is (at best and/or by definition) complex, rare, expensive, and
fragile.

This >>strongly<< suggests that the best way to engineer anything from a
beowulf to a gaming machine is to to use a non-overclocked CPU with a
high quality heatsink/fan in a high quality case with a good power
supply and adequate airflow for cooling, kept in a room with sufficient
air conditioning to keep the ambient air at a stable (+/- 2 degree C)
temperature in the 15-25C range (which typically suffices to keep CPUs
in the 30-40C range).  It >>strongly<< suggests that deviating too far
from this simple recipe will be expensive and an overwhelming
cost/benefit loss.

I will spare everybody a lecture on the laws of thermodynamics and the
theoretical bound on the coefficient of performance of a refrigerating
unit, but suffice it to say that it is high (an A/C is "efficient") if
and only if the temperature of the cold reservoir (where heat is
absorbed) is as close as possible to the temperature of the hot
reservoir (where that heat plus the work done by the compressor is
rejected).  >>All<< refrigeration costs -- operational and hardware and
maintenance -- go up, generally quite nonlinearly and significantly, as
this temperature differential is increased, and the "refrigerators" that
produce liquid nitrogen are very expensive to purchase, to operate, and
to maintain.  Ditto, frankly, for anything designed to deliver e.g. CO_2
at -20C, although not as much.

The humble air conditioner, on the other hand, can be purchased for a
few hundred dollars at Home Depot or installed for a few thousand
dollars in multiton capacities by service persons readily available on a
contract basis in any city or town in the US and most cities of any size
around the world.  They are reliable and relatively cheap to operate and
maintain (as they typically maintain only a 10-30C differential between
the hot and cold reservoir depending on climate and time of year).  One
needs a >>damn good reason<< to opt for any other technology.

So please allow me to reiterate -- folks interested in liquid cooling of
beowulfish systems for "fun" should feel free to play right on through
as their time, budget, and inclination suggest.  Folks interested in
the idea if and only if it makes some sort of economic cost/benefit
sense, well (in my ever so humble opinion) it not only doesn't, it
doesn't even come close enough to be taken seriously.

This isn't intended to be flames.  I openly admit that I could be
mistaken.  However, if somebody really thinks that I am, could they
PLEASE do the serious and requisite legwork of working out even a
first-order approximation of the marginal engineering cost per
node/system and the associated performance "benefit" thus attained over
straight A/C and a decent heatsink and START their next communication in
the thread with that comparison.  At least then we'll have something
concrete to discuss.

   rgb

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







More information about the Beowulf mailing list