🚀 go-pugleaf

From article <170500013@uiucdcsb>, by robison@uiucdcsb.cs.uiuc.edu:
> Some friends and I conjecture that a lot of supercomputer time 
> is spent running poor algorithms.  Anyone have any empirical
> evidence or anecdotes?

I think you're right....

A friend of mine (NOT a FOAF) has a half-dozen of these fancy
wall clocks with little plaques on them praising him for saving
the company (a NASA contractor, so indirecly the government)
some number of dollars.  The company's cost-reduction reward
program includes several gifts, of which the clock represents
the highest level.  Needless to say, the bean counters won't
let him have any of the lesser gifts even though he has all
the clocks he can use....

These come from walking by a bunch of programmers working on
some "trivial" program and making small changes that effect
significant changes in run time.

One example, which is fresh in my mind as I had to endure
a retelling of it last night over some suds:  A certain
trivial algorithm was to be run over some data.  It performed
an operation analogous to averaging - the details are unimportant.
In addition to being naively written with respect to numerical
issues (got the wrong answer) it had this other problem...
It took over two hours to run for something like 600,000
records...   Turns out the dataset was in some binary format
and these clowns were turning it into card images and reading
it with FORTRAN formatted input.  Recoding to read the original
binary data cut the runtime to a much more reasonable figure,
like maybe two minutes.

Since they were planning to run the program "several thousand"
times over the next few months this amounted to something
over a million dollars in cost savings...

Its sad but people who do good work are far outnumbered
by the marching morons, to the extent that management in
many industries thinks software is _supposed_ to be difficult
and expensive and not work very well.   sigh.

Oh - another anecdote - it seems this programmer had been
submitting version after version of this little program
to the cray... like for a month.  It just kept getting
the wrong answer, and he couldn't figure out why.  My friend
says "What answer is it giving you?"   "Zero."  "Hmmm, where
is it calculating it?"  "Right here."

"Well, if you change that 1 to a 1.0 it won't do that any more..."

If I had ever worked around a super computer myself I could
probably give you some first hand anecdotes - I sure have enough
from the micro and mini realm.
-- 
Steve Nuchia	    | [...] but the machine would probably be allowed no mercy.
uunet!nuchat!steve  | In other words then, if a machine is expected to be
(713) 334 6720	    | infallible, it cannot be intelligent.  - Alan Turing, 1947