multiprocessor jobs

Dear Sirs,

I would like to ask if there are any known issues with the commands
'dump .. cfg' and 'dump_modify .. sort' in multiprocessor jobs, as
compared to single processor runs. I attach a sample control file for a
peridynamics simulation which crashes if more than 1 processor and the
above commands are used. I have encountered similar problems with some
reaxc simulations (not tested without dump commands).
Thank you,

in.peri (3.87 KB)

Dear Sirs,

I would like to ask if there are any known issues with the commands
'dump .. cfg' and 'dump_modify .. sort' in multiprocessor jobs, as

not yet. there were recent modifications to support using MPI-IO for
certain styles of dump files. these kind of changes can sometimes lead
to unexpected side effects.

compared to single processor runs. I attach a sample control file for a
peridynamics simulation which crashes if more than 1 processor and the
above commands are used. I have encountered similar problems with some
reaxc simulations (not tested without dump commands).

there are two very important pieces of information missing here: a)
which is the exact version of LAMMPS that you are using? and b) what
architecture/platform are you running on and how was the code
compiled.

axel.

Hello,

the 17Dec13 lammps version with openmpi architecture was used. The code
was compiled by the computing center staff.
Thanks,

Osvalds

i can confirm that there is a problem, even with the latest version of
the sources. we are currently shaking out problems like this that may
have been introduced with the recent changes, so your submission of a
test case is much appreciated. please have a little patience while we
try to track this down.

axel.

ok. i found the bug and have a workaround for it and will send it to
steve ASAP for inclusion into the next LAMMPS patch. you need to use
peridynamics, dump_style cfg and dump_modify sort to trigger it.

please check your inputs if you can find other cases that trigger
problems that do not fall into this category.

thanks,
     axel.

thanks to Axel for finding the issue - the
fix will be in the next patch - I fixed it in
a different way, so be sure to test it for your problem

Steve

Hi Steve,

We ended up fixing it by adding a conditional in Dump::write() around line 313 for the “peri” atom style in dump.cpp. Another conditional based on the process rank was added in DumpCFG::write_header() around line 149 in dump_cfg.cpp, to make sure that only one MPI process writes the header. It seems to work fine for Osvalds’ script.

Thank you all for your help in the matter; we’ll make sure to do some testing and patch the code with your fix when it becomes available.

Pierre-Yves