Automatic parallelization in 4.3?

[Janne Blomqvist]

On Tue, Oct 24, 2006 at 09:49:24AM +0200, Tobias Burnus wrote:
> Hi,
> 
> Craig Powers wrote:
> >> I found one post by Zdenek Dvorak regarding auto-parallelization, but
> >> it doesn't mention Fortran AFAICT.  See
> >> http://gcc.gnu.org/ml/gcc-patches/2006-09/msg01243.html
> >
> > One comment -- for my use of Fortran in a cluster environment, I
> > actually prefer to avoid use of additional processor cores because
> > they are controlled resources managed by the cluster's master
> > scheduler.

=> So you should use a batch scheduler that understands multithreaded
programs in addition to the usual serial and MPI jobs. It ought to be
possible in SGE, but like everything else in SGE it's bound to be a
gigantic kludge.

>  So if I were to use gfortran to build my group's code, I
> > would prefer to be able to turn something like this off easily,
> > preferably without compromising other optimizations.
> To turn this off is easy: Don't use the -ftree-parallelize-loops=n (n =
> number of threads) option.

*boggle* The number of threads to use is hardcoded? Why not do it like
OpenMP? Too much overhead for small loops?

> (I don't know how well -ftree-parallelize-loops will work, but the
> autoparallelization in ifort made one of my programs actually slightly
> slower (on a dual-core AMD) than a single process [I don't know how
> typical this is]. 

FWIW, I had a similar experience a few years ago when I tried the
autoparallelization on XL Fortran/AIX. It was *a lot* slower than just
using MPI over shared memory (ideological purity be damned).

Still waiting for the magic parallelization fairy, I suppose..


-- 
Janne Blomqvist
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 185 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20061024/7d529d00/attachment.sig>