[Gllug] Best option for a lot of compute power

[Nix]

On Sat, 12 Jun 2004, John Hearns bespoke:
> On Sat, 2004-06-12 at 00:14, Nix wrote:
>> On Mon, 07 Jun 2004, John Hearns said:
>> > As regards hyperthreading, we normally switch it off on compute nodes.
>> > For HPC type applications, the concensus seems to be it has no
>> > advantage. But as always YMMV - worth benchmarking with/without HT with
>> > your applications.
>> 
>> When the next GCC release happens, that may change.
>> 
>> (Quite a number of autovectorization patches are landing... :) )
> 
> That's interesting.
> 
> Most of our customers use the Portland or Intel compilers, in preference
> to gcc.
> And if you are talking Fortran 90 there is really no open source
> alternative.

... but there will be. When tree-ssa was merged, g77 finally died the
death, and gfortran replaced it. It's by no means fully functional yet,
but it *is* an actively-maintained (and -developed) f95 frontend. (How
many of the weird and wonderful g77 extensions it will ever support is
an open question.)

> I'm no compiler expert, but the Intel stuff seems impressive.
> It has a feature where you run your code, collect stats, then
> re-optimize.

Bah, GCC's had that for, ooh, eight years; see -fprofile-arcs and
-fbranch-probabilities. (You can even build GCC with profile-directed
feedback; see `make profiledbootstrap'.)

Now that tree-ssa's merged, we've got a decent optimizer infrastructure
and sane intermediate language to put on top of it; a good few of the
GIMPLE optimizers can use profile-arcs feedback info (jump threading
being the largest one I can think of right now).

(Scheduling can already benefit, but that's an RTL-level optimization;
it can benefit in 3.3+, I think.)

-- 
`If you believe in strong AI, then death is no longer a mystery,
 but merely a lack of adequate backups.' --- Steven McDougall
-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug