[Beowulf] First 96-Node Transmeta Desktop Cluster Ships
Robert G. Brown
rgb at phy.duke.edu
Mon May 9 16:21:32 PDT 2005
On Mon, 9 May 2005, Andrew Piskorski wrote:
> On Wed, May 04, 2005 at 08:20:13PM -0500, Alan Grimes wrote:
> > > http://www.rocketcalc.com/
> > BLEH!
> > http://www.iwill.net/product_2.asp?p_id=90&sp=Y
> Your comment doesn't make much sense. An 8-way Opteron box is very
> nice if you actually need shared memory, but those 8xx series Opterons
> are likely to be a LOT more expensive then the 2xx chips used in
> duals. Rocketcalc is selling a 12 cpu box using 6 dual motherboards,
> so its hardware at least has a real chance to be substantially
> There's also the whole "ready to use compute appliance" thing, which a
> bare-bones box from Iwill obviously is not. It will be interesting
> however to see what different sorts of HPC-suitable bare-bones Opteron
> boxes the various vendors might offer. Like, say, Supermicro, now
> that they're finally doing Opterons.
Or (to give a plug to a pretty decent vendor) the many folks LIKE
Penguin who sell ready-to-rack U1 duals for very reasonable prices,
preinstalled (if you like) with ready to use linux or Scyld. Dual
opterons are available at a not-horribly-large $/FLOP premium compared
to el cheapo vanilla boxes (which often have the best P/P on the basis
of pure aggregate clocks or theoretical FLOPS). Fast WIDE memory really
DOES make a major league difference in processor throughput.
Then there is the AMD-64, for people who want an el cheapo box with
quite a lot of the benefit of an opteron for a vanilla box price.
pricewatch has motherboard/CPU combos available for from $144 (clearly a
mass market price) up to just under $1000 for a 928 FX. The 2.44 GHz
one I have upstairs cost about $600 to build out of parts with 512 MB of
memory six months ago, and would be even cheaper now.
There isn't a thing wrong with getting turnkey clusters, and for cluster
novices getting a turnkey operation from somebody who will really
support you can easily be worth the cost. But you WILL pay for it. In
clustering, DIY is cheapest IF you have the mad skills required and some
opportunity cost time to dispose of. DIY from a quality vendor is
second cheapest and just right for most larger organizations with a real
staff with mad skills cuz that's what they do. Turnkey is probably best
for smaller labs or groups with very specific needs and no slave or
cheap labor in the form of graduate students, local linux hackers, local
cluster gurus, etc to rely on.
However, I'm >>hoping<< that as CWM's "cluster agenda" project takes
root, we'll end up being able to focus open source energy on making
cluster installation and management even easier with complete and fairly
standard fully open toolsets. Warewulf and other related projects are
(I think) the key to DIY clustering's future.
Up to now, clustering has been an expert's game in an expert-friendly
operating system. It will probably always be that to some extent, but
it could be far easier than it is and things like warewulf are what will
get it there.
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