Re: Gfortran vs Ifort when using MPI

["Johan Seland" <johan dot seland at gmail dot com>]

Just to keep you all informed.

I have made some progress during my instrumentation today. It seems
like the vectorization or FP-handling is not blame.

The deciding factor seems to be which Fortran compiler that was used
in compiling the MPI library.  In other words, when using ifort to
compile MPI, the simulator can be compiled using gfortran and the
speeds are on par with ifort compiled simulator.

I will continue to investigate.

On Wed, Jul 9, 2008 at 9:38 AM, Johan Seland <johan.seland@gmail.com> wrote:
> Thanks for all the tips.
>
> As a first test, there was no difference when using march=native and
> ftree-vectorize. Furthermore I suspect fvect-cost-model is a 4.3
> feature only?
> When executed through mpirun, my binary is not dumping profile
> information (when compiled with -pg), so I will now start to manually
> instrument my code.
>
> I keep you informed.
>
> On Wed, Jul 9, 2008 at 9:04 AM, IainS <developer@sandoe-acoustics.co.uk> wrote:
>>
>> On 8 Jul 2008, at 21:25, FX wrote:
>>
>>>> Depending on the system, using the right -march makes a big difference.
>>>> Try -march=native (not supported on all targets). (If the compiler thinks
>>>> that there is no SSE, -march=native can help a lot; if I recall correctly,
>>>> for x86 the settings are rather conservative.
>>>
>>> For a 50x-100x performance difference, I suppose there is something
>>> *awfully* wrong going on, like wrong alignment or NaNs all over the place. I
>>> suggest a) checking the results, b) roughly profiling the gfortran code
>>> ("gfortran -pg" and using gprof) to get an idea where it spends so much
>>> time. It'll be much easier than us shooting in the dark.
>>
>> One thing you might check for in your profiling is presence of, and
>> different handling of, denormals - that can make huge difference on x86.
>> Iain
>>
>
>
>
> --
> Regards Johan Seland
>



-- 
Regards Johan Seland