Top 500 trends

[Robert G. Brown]

On Mon, 25 Nov 2002, Ken Chase wrote:

> > (I'll admit to finding this interesting since I'm one of the people
> > who originally founded Meiko where this technology was originally
> > developed [a few generations ago...], but I have no commercial
> > interest here.)
> 
> Anyone have stats on the top 500 in FLOPS per $?
> 
> Does anyone care? ("no.") Shouldnt we? Isnt that what Beowulf (but
> HPC itself, obviously) is about?

Yes. Yes.  Yes.

Although things like "fastest" almost always have nonlinear cost
premiums associated with them -- the secret is in finding someone with
incredibly deep pockets to pay.  The US Govt does nicely (he says,
bellying up to the trough:-) although larger corporations can sometimes
be convinced to as well.

The really interesting issue is the overall lack of consideration of
Moore's Law in cluster purchase design and long term economic strategy
by the granting agencies.  For example, let's pick a project "X", which
accomplishes some worthy HPC goal, whether it be computing properties of
a quark-gluon plasma, studying critical phenomena, simulating nuclear
devices, computing cosmological evolution.

One thing that is almost NEVER done is to do an actual CBA of the value
of the result.  Sometimes such a result has an obvious and real-dollar
value (rational drug design, weather prediction).  Sometimes the
result's value is subtle -- "increasing our general understanding of the
Universe while incidentally supporting an important component of our
national scientific educational infrastructure in the persons of the
various researchers".  Either way, one thing we (as a society) never do
is ask the question "should we try to (pay to) answer this now or wait
until later".

This is a tough one, because in many cases paying those nonlinear costs
to answer the question "now" makes the answer three, ten, fifty times
more expensive AND results in the actual answer being available only
months sooner than it might have been if the project were postponed.

The reason, of course, is Moore's Law.  It is hard to remember, but
every desktop sold in the US today (presuming that even slow, cheap,
plodgy OTC desktops are at least 1 GHz) are faster than massive,
multimillion dollar vector supercomputers as of a decade or so ago,
systems that were considered controlled exports as potential nuclear
weapons design platforms.  Many is the discussion we had on this very
list back in its first few years out of concern that beowulfs -- which
provided this sort of capability cheaply in an interpolant timeframe --
would be considered controlled exports just like the SP2's they
frequently could outrun.  Fortunately this never QUITE came to pass.

SO, a question that could ONLY be addressed on a $12 million Cray-YMP a
decade plus ago (at the cost of literally millions of dollars in
amortized depreciation on the hardware plus operational time in a shared
facility) can NOW be addressed for a couple of thousand dollars.  Worse,
a project that might have taken ten years of that Cray-time can now be
accomplished on a very modest (<$100K) beowulf in months.  If one were
given $12M to accomplish the task in 1992, with the understanding that
it would take a decade, one could have spent the money like water,
living in total luxury and not doing a lick of work, and started at the
beginning of the year and finished FASTER than that Cray would have done
it if one had committed the entire budget to the Cray.

It would be lovely if time, CBA, and Moore's Law were prominently
associated with project proposal and approval, IF our goal was really to
Probe the Secrets of the Universe most efficiently.  Of course, our REAL
purpose is to Have A Job probing those secrets, and imagine the pain if
I were told "sorry, we can do your Monte Carlo computations now at a
cost of $50K/year for eight years or wait until year seven and complete
them in year eight for $100K total, and you'll just have to wait".
Multiplied by thousands of the nation's top University researchers.

I personally think that we'd be a bit better off as a country if we sort
of split the difference -- recognize that our real goal is supporting
the people and the infrastructure, and spending a bit less on ginormous
supercomputers whose primary benefit is that they provide us with the
answers to questions of interest to (and understandable by) perhaps a
few hundred of the humans on the planet.  That is NOT to say that those
answers aren't important to pursue -- I am a scientist and in hot
pursuit myself -- but I question the wisdom of spending quite as much as
we do to get these answers "right now" without the perspective
associated with Moore's law and the real goal of stretching out human
support anyway.

    rgb

> 
> /kc
> 
> > -- Jim 
> > 
> > James Cownie	<jcownie at etnus.com>
> > Etnus, LLC.     +44 117 9071438
> > http://www.etnus.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