On 09/06/11 16:23, Marcos Veríssimo Alves wrote:
> Jesús,
>
> Thanks for sharing. A few questions:
>
> 1) What mpi are you using? How is it compiled?

   I am using OpenMPI 1.4.3 from the 64-bit version of Debian testing.

> 2) Do you have benchmarks to compare with previous versions of ifort
> (10.x, 11.x) and perhaps other compilers in the same processors?

   I'm afraid I cannot provide any kind of systematic comparison with
ifort 11: although I was able to compile SIESTA with it in the past,
the resulting binary always turned out to be extremely unreliable, with
frequent SEGFAULTs, and thus basically unusable. Luckily, those
problems seem to have disappeared with ifort 12, at least in my case.
   I do have an equivalent version of SIESTA compiled with gfortran for
the same system, using the same OpenMPI framework and the lineal
algebra libraries included in Debian instead of their Intel/MKL
equivalents. The ifort version is noticeably faster, although I tend to
attribute the best part of this speedup to the libraries and not to the
compiler.
   If you are interested in the quantitative results of a specific
benchmark which can be run with both versions of SIESTA in a reasonable
amount of time, I'll be happy to help.

Best,
Jesús

Responder a