[gmx-users] Re: stepsize too small: help to solve problem in GROMACS installation

[Anna Marabotti]

Dear Mark,
many thanks for your suggestions. At present, the problem I reported here
seems to be solved using the gcc442 compilers, we are doing a lot of checks
on the machine and we will take into account all your observations.
Many thanks also for the correction on the web site.

Anna
------------------------------

Message: 4
Date: Thu, 27 May 2010 17:28:48 +1000
From: Mark Abraham <mark.abraham at anu.edu.au>
Subject: Re: [gmx-users] Re: stepsize too small: help to solve problem
	in	GROMACS installation
To: Discussion list for GROMACS users <gmx-users at gromacs.org>
Message-ID: <fbd7c308e524.4bfeabd0 at anu.edu.au>
Content-Type: text/plain; CHARSET=US-ASCII

----- Original Message -----
From: Anna Marabotti <anna.marabotti at isa.cnr.it>
Date: Thursday, May 27, 2010 16:55
Subject: [gmx-users] Re: stepsize too small: help to solve problem in
GROMACS installation
To: gmx-users at gromacs.org

> Dear Justin, dear all,
> sorry for not having seen the warning before asking the question about
> compilers (just a little justification, however: the warning 
> appears only on
> the Download page, not on the Installation page; I went directly 
> to the
> installation page and I did not notice it).

Fixed.
 
> The only other possibility we have is to try compiling GROMACS 
> with gcc-442
> compiler. This time I checked more carefully (at least, I hope 
> so!) on the
> site and it seems to me that no problems are reported. For your 
> experience,is it a good choice or we can expect other problems? 
> In this last case, what
> do you suggest as the best compiler for Gromacs?

GROMACS performance is pretty insensitive to compiler version so long as you
are using assembly-language inner loops (e.g. always true by default on
x86). Most of the compute time is spent there, so most of the rest is
irrelevant.

The main side issue is the FFT performance for PME. Make sure you're using
either vendor FFT libraries, or suitably well-compiled FFTW. Even there,
FFTW with gcc -O5 (making sure the SSE(n) optimization is used) was only a
hair behind icc+mkl in some tests I did a year ago.

Mark