[Beowulf] Stroustrup regarding multicore

Robert G. Brown rgb at phy.duke.edu
Wed Aug 27 12:11:46 PDT 2008


On Wed, 27 Aug 2008, Vincent Diepeveen wrote:

> Physics in itself is eating 50%+ of all supercomputer time.

Only 50%?

Comrades, we must work harder!  Build more computers!

It's not for lack of trying -- there are any number of projects any one
of which could use 100% of the aggregate cycles on the planet and still
be cycle starved (according to the researcher:-).

Even CPS people have a few "interesting" (e.g. NP complete) problems
that they work on.

Oh, you mean the SERIOUS question?  About whether or not physics grad
students should be trained computer scientists?

Well, in the best of all possible universes, they'd have brains the size
of basketballs, an utter lack of interest in the society of humans
(especially those of the opposite sex) and mad post-Ph.D. skills in at
least physics, mathematics, computer science, philosophy, and maybe a
few interesting sub-fields like chemistry, biology, neuropsychology, and
statistics.  Maybe even a smattering of game theory and so on.

In spite of their antisocial natures, they'd need to be all sparkly and
presentable at scientific meetings, write like a poet (in their papers
and grant proposals), and schmooze like the best (or worst, hmmm) of
politicians when money or resources were on the line.  Oh, and don't
forget -- they'd need the trait of utter loyalty to their Ph.D. advisor
to the point where they submissively INSIST that he or she be first
author on nearly all of the work that they do on his or her behalf.

And with all that, they'd still need to be serendipitously lucky,
bread-lands-butter-side-up sorts of researchers...

Now in THIS universe (widely recognized as hell by a number of
progressive and enlightened religions) Gerry is dead on the money.  Hope
you can get somebody that actually understands -- or CAN understand,
given time -- the basic physics that you devoutly hope that they'll
learn to master and use constructively to do actual research, and hope
further that they are smart and motivated enough to learn whatever they
need to, including computer science, math, chemistry, etc see list above
in order to successfully complete a project and bring at least SOME
credit back to your program while publishing with you as second, or even
third, author on their papers.

The amazing thing is that even in hell, this works out at least decently
a remarkable number of times, and that there is a quite substantial
chance that over a lifetime of cranking out grad students you'll
encounter one or two that in the fullness of time do indeed resemble in
posse if not in esse that ideal best of all universes grad student, one
that you'll be proud of for the rest of your life.

All kidding aside, think about the alternatives.  Hire somebody who
knows computer science really well but knows no physics or hire somebody
who knows physics really well but knows no computer science.  As
somebody who knows both at least decently -- I may not be up to Perry's
standard of knowledge/experience, but I teach independent study
computing courses to computer science majors and haven't heard any
complaints from students or the department -- I assure you that with the
exception of a small handful of REALLY esoteric problems (some of which
bridge the gap between e.g. information theory, neuroscience, statistics
and even physics) physics is hands down more difficult than computer
science at the levels required to write good code on parallel machines.

Seriously.  No comparison.  I teach physics too, and there are plenty of
ideas I STILL have a hard time understanding even as I try to teach
them.  Feynman said "nobody understands quantum mechanics" and by this I
think that he fundamentally meant that our brains are literally not
wired for a full conceptual understanding of it so we can't no matter
how hard we try.  I would argue that good chunks of relativity theory
come in a close second, and quantum field theory or string theory (which
combine both) are so difficult that they are too difficult for the human
species -- we are literally waiting for supergeniuses to be born or
serendipitous discoveries to be made or massively parallel hard work to
bear fruit as the math is too difficult for anyone to entirely
conceptualize.  Until the next e.g. Ramanujan or Einstein comes along.

No offense to any of the many computer scientists on list.  And I'm not
saying computation is easy; far from it.  But it is essentially
intrinsically linear and classical, conceptually, and the human brain
can manage it, right up to -- yes, you guessed it -- quantum computers.

    rgb

-- 
Robert G. Brown                            Phone(cell): 1-919-280-8443
Duke University Physics Dept, Box 90305
Durham, N.C. 27708-0305
Web: http://www.phy.duke.edu/~rgb
Book of Lilith Website: http://www.phy.duke.edu/~rgb/Lilith/Lilith.php
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977



More information about the Beowulf mailing list