Potential Scyld problem
edwardsa at plk.af.mil
Mon Sep 23 18:59:04 PDT 2002
On Mon, Sep 23, 2002 at 03:50:07PM -0700, Alan Grimes wrote:
> > Incidentally, a few weeks ago I asked whether gcc-3.2 would work on
> > SCYLD. I was concerned about the mpi/beowulf libraries. I compiled from
> > sources without any problem and the existing 27bz-7 libraries work
> > without problems for the codes I typically run (GAMESS and QUEST)
Thanks very much for the warning. I am using g77-3.2 because it allows
me to use very large arrays. Per your advice, I'm doing benchmarks
with QUEST and GAMESS. With one result in, g77-3.2 performs better than
g77-2.95 by a factor of 1.21. Gamess is next.
> You should run a comprehensive profile on GCC 3.2 output.
> on the squeak.org ML, very severe performance regressions have been
> noted with GCC 3.1, 3.1.1 and 3.2 with respect to 3.0. (squeak's
> self-benchmark runs 50% slower!!!)
> These regressions only appear when the broken code is excercised...
> using whatever the GCC people have been testing with will only show
> There was also a thread on the GCC ML which descibed such a
> comprehensive test suite.
> Latency is your enemy.
> Bandwidth is your friend.
More information about the Beowulf