home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.edu:1422 comp.lang.fortran:3299 comp.lang.misc:2818 comp.arch:9103 sci.math:10735
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!sdd.hp.com!wupost!usc!elroy.jpl.nasa.gov!ames!lll-winken!diego.llnl.gov!miller
- From: miller@diego.llnl.gov (Patrick Miller)
- Newsgroups: comp.edu,comp.lang.fortran,comp.lang.misc,comp.arch,sci.math
- Subject: Re: Scientists as Programmers
- Message-ID: <134918@lll-winken.LLNL.GOV>
- Date: 31 Aug 92 16:28:36 GMT
- References: <1992Aug25.154501.8654@colorado.edu> <1992Aug26.192410.6523@ultb.isc.rit.edu> <1992Aug27.154823.583@alchemy.chem.utoronto.ca> <BtpAIn.EE5@mentor.cc.purdue.edu> <34742@cbmvax.commodore.com> <1992Aug31.133811.3626@crd.ge.com> <1992Aug31.144045.11416@hubcap.
- Sender: usenet@lll-winken.LLNL.GOV
- Followup-To: comp.edu
- Organization: Lawrence Livermore National Laboratory
- Lines: 41
- Nntp-Posting-Host: diego.llnl.gov
-
- |> davidsen@ariel.crd.GE.COM (william E Davidsen) writes:
- |> ... To do this even competently is a skill which
- |> conspicuously eludes most people who don't write code for a living, and
- |> to do this with consummate skill requires training, practice, and a
- |> natural talent given to only a few people.
-
- In steve@hubcap.clemson.edu ("Steve" Stevenson) writes:
- |> My experience is that most computer science types would not know the best
- |> numerical algorithm because they don't understand the science. The problem
- |> continues to be that scientific codes are meant to deal with nature and
- |> that takes someone who understands the problem. ...
- ------------------------------------------------------------
- Both writers have a point. Computer science types can produce better,
- more elegant code because they understand the nature of _computing_,
- while physicists, chemists, etc... can write better codes because only
- they understand the nature of the _science_.
-
- This has worked acceptably well so far. We were been able to teach
- physicists to write in FORTRAN in the 60's, to eliminate GOTO's in the
- 70's, and to write vectorizable codes in the 80's. All required skill
- and training, but non-computer scientists were able to learn these
- skills or sufficiently powerful compilers were able to cover their
- lack. Now, however, we need to teach these same physicists how to
- parallelize their codes for massively parallel computers. This is
- much harder. I spent hours trying to convince a computational chemist
- that race conditions are bad and altogether too easy to generate.
- FORTRAN 90 makes some strides in that direction by introducing
- functional vector semantics, but no variant of FORTRAN (IMHO) has
- adequately addressed many of the other problems (data layout, aliasing
- problems, etc...). Here, we will need to see changes in the language
- physicists use because massively parallel computers EXPOSE TOO MUCH OF
- THE NATURE OF THE COMPUTER -- and their use requires much more
- knowledge of how to manipulate them. We will need to create some
- synergy between language designers and physical scientist so that we
- don't need to pair a computer scientist with physicist (or cross-train
- them).
-
- --
- Pat Miller Computing Research Group
- patmiller@llnl.gov Lawrence Livermore Natl Labs
- Opinions are my own, and don't necessarily reflect the views of the Lab.
-