[Beowulf] newbie

[Gus Correa]

Chris Samuel wrote:
> ----- "John Hearns" <hearnsj at googlemail.com> wrote:
> 
>> A commercial compiler - Intel / Pathscale /Portland
>> will always give you better performance from an application.
> 
> The feedback from a number of our users on our Barcelona/
> Shanghai cluster is that GCC beats the Intel compilers
> hands down [1] and that the PGI ones don't tend to hold
> a significant margin over GCC either.
> 
> Interestingly using the -march=amdfam10 option in GCC 4.3.3
> to emit Barcelona specific optimisations produces real
> application code that runs slower than the standard
> optimisation options (confirmed by a couple of users).
> 

Hi Chris, list

Thanks for the info!
I have been blindly using "-march=amdfam10" (and -O3)
out of the gcc 4.1.2 man page,
simply because it is recommended for Barcelona
(and hopefully also the Shanghais we have).
What would be the right level of optimization then?

> I've built GCC 4.4.0, but we've not tested with it
> (or announced it to users) yet.
> 
> [1] - Not surprising, it's hardly their home turf. ;-)

On Opteron Shanghai the Intel 10.1 compilers seem to reject
any architecture-dependent optimization flag above -xW (SSE and SSE2),
although the processor has SSE3, etc, and would qualify for
higher levels of optimization, or not?
Is this deliberate?

> 
> cheers,
> Chris

Thank you,
Gus Correa
---------------------------------------------------------------------
Gustavo Correa
Lamont-Doherty Earth Observatory - Columbia University
Palisades, NY, 10964-8000 - USA
---------------------------------------------------------------------