[Beowulf] Simple Cluster
camm at enhanced.com
Mon Dec 15 17:00:14 EST 2003
Greetings! You may be interested in Debian's atlas setup. We have
several binary packages which depend on a virtual blas2 and lapack2
package, which can be provided by either the reference libraries or a
variety of atlas provided versions with various ISA instructions
supported. For example, on i386, we have sse, sse2, and 3dnow builds
in addition to the 'vanilla' x86 build. As you know, the isa
instructions are only one of many factors affecting atlas tuning.
They are the key one, however, in a) determining whether the lib will
run at all on a given system, and b) that delivers the lion's share of
the performance. The philosophy here is to provide binaries which
give factors of 2 or more of performance gain to be had, while making
it easy for users to get the remaining 10-20% by customizing the
package for their site. 'apt-get -q source atlas; cd atlas-3.2.1ln;
fakeroot debian/rules custom' gives one a tuned .deb for the running
We need to get newer versions of the lib uploaded, but otherwise it
works great. 'Almost' customized performance automatically available
to R, octave,.... without recompilation.
"Robert G. Brown" <rgb at phy.duke.edu> writes:
> On Mon, 15 Dec 2003, Andrew Latham wrote:
> > I have a small 9 node p133 cluster. It works.
> > What does the list think about the idea of developing software on the
> > smaller(mem) and older systems. I have one so I am bias but I do see that
> > developing software that can handle 64meg of ram on a P586 system would lend to
> > tighter and more efficant code. I am not trying to sell the P133 systems, only
> > thinking about good code for them would be really nice(fast) on a Xeon or
> > better. I already know this could spark a discussion on busses and chipsets and
> > processors. Just thinking
> More likely a discussion on balance. I actually think that developing
> on small clusters is good, but I'm not so sure about small REALLY old
> systems. The problem is that things like memory access speed and
> pipelining change so much across processor generations that not only are
> the bottlenecks different, the bottlenecking processes have different
> thresholds and are in different ratios to the other system performance
> determiners. Just as performance on such a cluster would not be
> terribly good as a predictor of performance on modern cluster from a
> hardware point of view, it isn't certain that it would be all that great
> from a software point of view.
> My favorite case study to illustrate the point is what I continue to
> think of as a brilliant piece of code -- ATLAS. Would an ATLAS-tuned
> BLAS built on and for a 586 still perform optimally on a P4 or Opteron?
> I think not. Not even close. Even if ATLAS-level tuning may be beyond
> most programmers, there are issues with stride, cache size and type, and
> for parallel programmers the relative speeds of CPU, memory, and network
> that can strongly affect program design and performance and scaling.
> So I too have a small cluster at home and develop there, and for a lot
> of code it doesn't matter as long as one doesn't test SCALING there.
> But I'm not sure the code itself is any better "because" it was
> developed there.
> Although given that my beer-filled refrigerator is just downstairs, it
> may be...;-)
> > --- Jim Lux <James.P.Lux at jpl.nasa.gov> wrote:
> > > Outstanding, Ranjit...
> > > Great that you wrote up a page describing how you did it, too!! Especially,
> > > describing the problems you encountered (i.e. slot dependence for network
> > > cards..)
> > >
> > > So now you can say you built your own supercomputer. How cool is that.
> > >
> > > Jim
> > >
> > >
> > > At 01:54 PM 12/15/2003 +0000, Ranjit Chagar wrote:
> > > >Hi robert/jim,
> > > >
> > > >Well I built a cluster just for the hell of it. And as you said, before the
> > > >flames start, it was built just to see what I could do, built from cheap
> > > >PCs, just for the fun of it. They are 133Mhz PII and I built mine following
> > > >the instructions from pondermatic. Okay, so in this day and age that is old
> > > >hat, and so is my system but I enjoyed building it and enjoy playing around
> > > >with it. And then, being stupid myself, I wrote out instructions so that I
> > > >could did it again cause I will be the first to admit my memory isn't that
> > > >good.
> > > >
> > > >Full details at http://homepage.ntlworld.com/ranjit.chagar/
> > > >
> > > >Robert - if you have any questions let me know.
> > > >
> > > >Jim - I dont mean for this email to sound bad but my english sometimes is
> > > >taken wrong. I mean to say that you can do it if you want.
> > > >
> > > >Best Regards, Ranjit
> > >
> > > James Lux, P.E.
> > > Spacecraft Telecommunications Section
> > > Jet Propulsion Laboratory, Mail Stop 161-213
> > > 4800 Oak Grove Drive
> > > Pasadena CA 91109
> > > tel: (818)354-2075
> > > fax: (818)393-6875
> > >
> > > _______________________________________________
> > > Beowulf mailing list, Beowulf at beowulf.org
> > > To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> > =====
> > /---------------------------------------------------------------------------------------------------\
> > Andrew Latham -LathamA - Penguin Loving, Moralist Agnostic.
> > What Is an agnostic? - An agnostic thinks it impossible to know the truth
> > in matters such as, a god or the future with which religions are concerned
> > with. Or, if not impossible, at least impossible at the present time.
> > LathamA.com - (lay-th-ham-eh) - lathama at lathama.com - lathama at yahoo.com
> > \---------------------------------------------------------------------------------------------------/
> > _______________________________________________
> > Beowulf mailing list, Beowulf at beowulf.org
> > To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> 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
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
Camm Maguire camm at enhanced.com
"The earth is but one country, and mankind its citizens." -- Baha'u'llah
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