[Beowulf] RE: Capitalization Rates - How often should you replace a cluster? (resent - 1st sending wasn't posted ).

Lux, James P james.p.lux at jpl.nasa.gov
Wed Jan 21 17:22:58 PST 2009




On 1/21/09 5:12 PM, "Robert G. Brown" <rgb at phy.duke.edu> wrote:

> On Wed, 21 Jan 2009, Lux, James P wrote:
>
>> Sure, it's lowball, but even if you multiply by 5 ($5/sq ft), it's still
>> less than the utility costs...
>>
>> And, really, the question originally was whether in 3 years you've paid as
>> much for infrastructure and utilities than you did on the box, and I think
>> that it's certainly in that area.
>
> Absolutely.
>
> It's just that in the colo facilities that I've seen, they charge more
> by the U or by the rack than by ft^2.  Not that I have sweeping
> experience in this arena -- I've priced out one commercial server center
> and looked extensively at a state sponsored center in the RTP -- the
> latter is quite reasonable (but partially subsidized for academic use).
> They seemed considerably higher "per ft^2" if you fill a rack on a per-U
> basis.  I seem to recall $50-100/month per 1U or 2U server, something
> like that (including power/ac) which is commensurate with your estimate
> but clearly charges more for "volume" rather than floor space per se.
>
> Either way, you quite rapidly get to where this cost for N nodes will
> pay for a single node Nx as powerful when you're paying out anywhere
> from 20% to 33% of the node budget per year.  As little as two years,
> more often the 3-5 discussed, and by 6 the node needs to have "free"
> power and housing or it starts not being worth it period.
>
>    rgb



I think there's a B-school thesis in this topic. It sure would be
interesting, especially if you had a decent modeling tool that would let you
plug in cost of money, funding profiles, etc.

Jim





More information about the Beowulf mailing list