[gmx-users] Re: GROMACS PME on BLUEGENE

[Mark Abraham]

On 4/03/2010 3:54 AM, Xevi Biarnes Fontal (SISSA) wrote:
> Dear Mark,
>
> I found a topic in the gromacs-users mailing list regarding your
> troubles in running gromacs with PME on BlueGene/P
> (http://lists.gromacs.org/pipermail/gmx-users/2009-June/042421.html)
>
> did you solve the problem?

Please keep such discussions on the mailing list. I might not have the 
expertise or time to help, and if someone can help, the answer should be 
archived for all to find. Imagine if you'd not being able to find 
anything with searches :-)

> I am now facing exactly the same problem. If I turn on PME I get the
> execution stacked and a core file written out. If I don't request for
> PME (i.e. Cut-Off) the execution runs without problems.

I wasn't having any problem with BlueGene, if you read that email 
exchange carefully. Jakob was. He apparently never shared his solution, 
if any.

Try running on one CPU (mdrun -np 1 -exe blah -args blahblah) to see if 
the problem is generic, or parallel-related.

Read in your BlueGene documentation how to use addr2line to probe the 
stack dump. You'll need to compile a version of mdrun with debugging 
enabled (add "-g" to any CFLAGS="..." on your configure command line) 
for this detective work to succeed. That will tell you the function and 
line number where the problem occurs, which will likely lead to a 
solution. Complain to IBM that this is a 1960s solution :) Obviously, 
once it's fixed, get rid of the debugging version.

> If you solved the problem, can you tell me which steps did you follow.
>
> I compiled gromacs-4.0.7 with FFTW3.1.2 for the backend and with a
> modified version of the same libraries for the frontend. (I am using the
> configure options explained in
> http://www.gromacs.org/index.php?title=Download_%26_Installation/Installation_Instructions/GROMACS_on_BlueGene)

If you're using BG/P, note Puetz's comments there about fixing MPI 
libraries and such.

Mark