[CP2K:1282] Re: compiler influence question & intro

[Ondrej Marsalek]

dear teo and axel,

thank you for your replies.

On Wed, Aug 6, 2008 at 10:42, Teodoro Laino <teodor... at gmail.com> wrote:
> Let's go to your question now: as long as you do DFT MD you won't see
> too much difference between
> PGI and g95. g95 can be sensibly slow for the classical part but you
> should not see any
> sensible difference on the DFT part.

that is the important part for me.

> If you really want to put some effort in having a fast executable
> then I would suggest you to go
> for the intel compiler+mkl.. It won't be easy.. but at least at the
> end you will have something that both
> for DFT or classical will be extremely faster than g95.

does that mean that precompiled mkl (certainly optimized for intel
processors) is better than let's say gotoblas and scalapack optimized
for the specific platform? the results below would suggest so...

On Wed, Aug 6, 2008 at 11:28, Teodoro Laino <teodor... at gmail.com> wrote:
> As a follow up I'm attaching an example of g95/intel serial build on
> AMD dual-core architecture.
> 1 H2O water molecule 10 MD steps.. it's not the best test for
> benchmarking but  is something that
> runs fast and gives already an idea:
>
> g95:   516 sec
> intel:  391 sec

that looks quite interesting. does the choice of blas implementation
with g95 have a big impact? i have used only gotoblas so far, as i
have had good results with it before.

On Wed, Aug 6, 2008 at 17:20, Axel <akoh... at gmail.com> wrote:
> one more remark. who says that you _have_ to use PGI on AMD
> processors?

no-one (i know of) says that, although one would expect the speed-up
resulting from the exchange of g95 for intel to be bigger on intel
hardware. at the moment pgi is the only commercial compiler available
at our opteron cluster. i could (as far as i know) test the intel
compiler before trying the get someone to buy it, i just wanted to get
a better idea before investing too much effort in it.

> i've been using intel compilers on opterons with _very_ good results.
> you have to check carefully to have a working patchlevel, but that is
> independent o fhte hardware. the key to good and reliable execution
> with intel
> compilers is to not optimize too aggressively.
> with intel 10.1 i get the best performance on DFT codes using:
> -O2 -unroll -pc64 -march=pentium3 -mtune=core2
> works well with both intel core2 and opteron.
> please note that the default settings for intel compilers is to
> optimize for pentium4 (netburst) architecture, which is a very
> different beast than pentium3, opteron, and intel core.

thanks for the insights. seeing the results shown by teo, i will
consider an intel+mkl build and this will certainly be useful.

regards,
ondrej marsalek