bari at onelabs.com
Mon Jun 4 21:36:06 PDT 2001
Bob Drzyzgula wrote:
> I suppose that there's always mineral oil...
> That site hasn't been updated in years, but it's
> kind of an interesting idea.
> Nonetheless, I think that there's one aspect to all
> this that may be missing from this discussion. The
> big problem, as I see it, is not how to get the heat
> out of the enclosure, but rather how to get it out
> of the building.
> Suppose that you had a rack filled with 40 1U cluster
> nodes, and that each node burned juice at about 100W;
> In this case, your rack will be burning up about 4000W,
> or around 13,600 BTU/hr. One ton of chiller capacity
> can remove 12,000 BTU/hr, so just for that rack you'll
> need over a ton of air conditioning. If you *don't* have
> more than a spare ton of AC, the temperature in your
> rack is going to go up steadily. It doesn't matter
> how good your heat sink, heat pipe or whatever, is --
> if the heat generated by your cluster isn't removed
> from the data center, no amount of forced airflow through
> the rack will do any good.
> I think that most 1U cases will do OK as long as
> the ambient temperature stays within tolerable levels.
> But feed them hot air to cool with and you have
> problems. Somewhere I read -- perhaps it was on this
> list? -- that some entrepreneurs were proposing to
> set up a massive web hosting facility on the North
> Slope of Alaska, largely to reduce cooling costs...
If you're already in Alaska it would seem to make sense. Locating next
to a lake would also do the trick if your demands were that high.
Using forced convection along with sensible conduction cooling
techniques allow for even large clusters to be used in areas where the
ambient temperatures may be 40 C. So you could still run your TFLOP
machine with the windows open in Phoenix during mid July. Though running
it in Fairbanks at 10 C would bring up the MTBF a notch :-)
More information about the Beowulf