[gmx-users] Compiling with ICC: advantages and -if yes- a suitable protocol? (it seems I can't)

[ms]

On 12/10/10 00:33, Mark Abraham wrote:
>> I have heard (read: read on random blogs here and there) that on
>> Intel compiling GROMACS with icc instead of gcc can bring up to
>> 50% performance improvement.
>
> If you are referring to this post (http://kent-vandervelden.blogspot.com/2010/08/optimization-of-gromacs-407.html) then his comparison is based on use of FFTPACK, and is at best misleading. I've posted a follow-up comment that is awaiting approval. He didn't cite relative performance for gcc+mkl, for starters.

Among others, yes :)

> GROMACS performance is dependent on
> a) non-bonded kernels (which are coded in assembler for the large majority of platforms for which icc exists, and so compiler version is largely irrelevant)

I use mostly tabulated functions for the model I'm developing...

> b) FFT performance (only for PME, and gcc+FFTW3 is only a few percent behind icc+MKL in my tests)

OK, good.

> c) a whole pile of MD algorithmic effects (but time spent in these is negligible compared with those above)
> You can see this immediately in the timing breakdown reported at the end of the mdrun .log file.

Yep, I know. Thanks!

>> Since I always used gcc compiled GROMACS, I'd like to know:
>> - Is this true?
>
> No.

Great. So I can stop bothering and keep my happy gcc :)

Thanks!
Massimo