Re: [ESPResSo] ESPResSo Compilation

[Axel Arnold]

On Monday 23 November 2009 04:57:32 pm Salvador Herrera Velarde wrote:
> Hi All,
>
> Thanks a lot for your last suggestions. I check my restarting simulations
> and they look fine, time and parameters are according to the last snapshot.
> Besides, if i restart the simulation in my desktop it runs without trouble.
>
> After talking with the admistrator of our cluster we think that the problem
> is related with my ESPREsSo compilation.  After removing the optimization
> flags, both NPT and the restarting simulations run without problems, but
> they are very slow.
>
> I am trying different flags for the compilation, e.g,
> /configure --with-myconfig=config-NoDPD.h CC=pathcc CFLAGS="-Wall -O3
> -OPT:Ofast -fno-math-errno -TENV:simd_zmask=OFF -TENV:simd_imask=OFF
> -TENV:simd_omask=OFF -I/opt/hpmpi/include
> -I/opt/sharcnet/fftw/3.2.1/pathscale/include"
> LDFLAGS="-L/opt/hpmpi/lib/linux_amd64 -lmpi
> -L/opt/sharcnet/fftw/3.2.1/pathscale/lib"
>
>  but in my last trial I got the following messages:
> /communication.c:2169: warning: type-punning to incomplete type might break
> strict-aliasing rules /ghosts.c:423: warning: type-punning to incomplete
> type might break strict-aliasing rules /p3m-charges.c:846: warning:
> type-punning to incomplete type might break strict-aliasing rules
> /ewald.c:346: warning: type-punning to incomplete type might break
> strict-aliasing rules /fft.c:1308: warning: type-punning to incomplete type
> might break strict-aliasing rules
>
> A similar warning appears in nsquare,  layered, mmm2d, elc, maggs and
> rattle.
>
> Could you help to clarify these warning messages?

The warnings seem to occur everywhere where MPI is used, so probably this is a 
problem between your compiler and the mpi-headers. As it looks, you want to 
use the PathScale-compilers, that we have no experience with (although the 
warning is a typical gcc-warning, you might want to check that configure 
really uses the compiler you intend).

This warning is normally  issued when the compiler sees type-casts to 
incomplete pointers, which at the places you indicate doesn't happen on 
Espresso's side. However, I don't see how these warnings should cause 
problems - MPI cannot handle overlapping memory anyways, so the memory in 
MPI-operations is never overlapping. The warning is still issued, since the 
compiler cannot know that; and the warning is only issued at higher 
optimization levels, when these aliasing optimizations are actually switched 
on.

Axel

-- 
Dr. Axel Arnold
Schloss Birlinghoven
53754 St. Augustin, Germany
Tel. +49 2241 14 2575