Path: utzoo!mnetor!uunet!husc6!bloom-beacon!bu-cs!bzs
From: bzs@bu-cs.BU.EDU (Barry Shein)
Newsgroups: comp.arch
Subject: Re: FORTRAN Horror
Message-ID: <20821@bu-cs.BU.EDU>
Date: 22 Mar 88 17:29:20 GMT
References: <24861@yale-celray.yale.UUCP> <1135@pembina.UUCP> <2596@pdn.UUCP>
Organization: Boston U. Comp. Sci.
Lines: 38
In-reply-to: alan@pdn.UUCP's message of 21 Mar 88 16:41:09 GMT



Yes, I agree that many FORTRAN users are obsessed with performance and
some are good at getting it. Unfortunately many aren't, grad students
in the sciences often have little training in programming and
algorithms and issues like accuracy vs precision. More importantly
they don't know what's important and often don't ask the right
questions.

I remember when I managed the Harvard Chemistry machine, for example,
some very bright chemistry students had a routine for sorting a huge
array of reals. It took something like 20 minutes and brought the 780
(this was several years ago) to its knees. I finally had a look at it,
went across the street to the library, xeroxed the page from collected
algorithms giving a quickersort for Fortran (I think it was
quickersort, it's been a while), typed it in and reduced the run time
to well under a minute, maybe 30 seconds. They had never heard of
collected algorithms and were, needless to say, pleased with the
results. There were other examples, when I worked with the pulmonary
physiologists at Harvard I remember reducing the run-time of a routine
that was run all day on hundreds of data sets from 20 minutes to about
20 seconds on a PDP-11. The programmers (medical physiologists) had
bummed the code a lot but didn't realize what type conversions in the
inner loop of a Hamming filter (convolution) was costing them,
hoisting the code around a little yielded the speed-up.

There's a lot more to it than good code-generators, I would tend to
say that's the least of it. I still wonder why many of those folks who
ran some of the intensive looping I saw ever thought they had anything
more than bit-shift noise left in the resultant data, they didn't seem
to understand the question when I asked and assumed the math in the
machine behaved more or less as it should theoretically, they
understood round-offs a little. I won't even talk about the folks who
talked about the speed of running statistical routines but didn't seem
to understand statistical design and would run test after test,
whatever the hypotheses (hmm, the T-test didn't show anything, let's
try a few F-ratios or ANOVAs and see if that works...)

	-Barry Shein, Boston University