back to the issue of cooling

Steven Timm timm at fnal.gov
Wed Apr 23 09:26:38 EDT 2003


> On Tue, 22 Apr 2003, Eray Ozkural wrote:
>
> > Has anybody calculated if the operation of a low-power cluster can amortize
> > the actual price of the system in a couple of years? I'm thinking something
> > like an apple cluster or something, might actually be viable! What about
> > total cost/processing power of the cluster ?
>
> This is the relevant measure, to be sure.  Total cost of ownership with
> a vengeance, per unit of work done, amortized over the life of a
> cluster.  $1 per watt per year for heating and cooling, add cost of
> systems themselves, correct for SPEED of systems (ideally including your
> Amdahl's law scaling hit for using more slower processors!).  I predict
> that the sweet spot is probably Athlon 2400's or possibly 2.4 GHz P4's
> (depending on your code), with fine grained people shifted toward the
> even higher end processors and with some room for Celerons or Durons on
> the low end.  I think the main reason to get transmetas is likely to be
> to get the processing densities, not to save power or money.
>
The problem with the above calculation is that oftentimes the cost
to get the electrical infrastructure into your facility in the
first place is much, much greater than the cost of the electricity
it delivers.  We are spending $560K here at Fermilab to add 250 kVA
of electrical capacity to our floor.  We calculate that the
cost of the electricity to run the machines over 3 years will only
be $50K.

We considered whether to actually put a weighting factor into our
bids so that more electrically-efficient machines would be preferred,
but then when we thought about it, we figured that (1) these
machines are usually slower so you need more of them (2) they
also use up more floor space which isn't free, and (3) within
the same CPU speed class, those machines which use up the most
electricity are also likely to be the ones with the biggest fans
which are the best cooled internally.

Steve Timm


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list