Segmentation fault in a loop with intel omp

HI, all

   I met a problem when running Lammps in an openmp mode, with a command : mpirun -np 2 lmp_2014-openmp -sf intel -pk omp 1 <

   Everything goes well until jump to a loop, as showed below : (No problem when use pure mpi )

------> MD Run 11
Step PotEng Pxx Pyy Pzz Temp Press Volume TotEng
     500 -30307.38 12.865555 47.575572 44.224419 41.168762 34.888515 704317.25 -28098.61
     550 -30314.944 9.5099734 48.640318 43.884582 41.153515 34.011625 704235.8 -28106.992
Loop time of 0.633232 on 2 procs (2 MPI x 1 OpenMP) for 50 steps with 18000 atoms
Write the restart file
System init for write_data ...

Makefile.machine (3.45 KB) (2.88 KB)

Mike or Axel may be able to answer this Q.


Mike or Axel may be able to answer this Q.

excessive looping is almost always asking for trouble with accelerated
packages. USER-CUDA and USER-INTEL are the most vulnerable due to the
inherent complexity, GPU and USER-OMP to a lesser degree.

in order to debug this, please run with -sf omp instead of -sf intel
to confirm whether this is triggered by the USER-INTEL or USER-OMP
codepath (both are triggered by -sf intel) and provide as simple as
possible input deck that uses a (small) data file instead of a restart
and has only the minimal set of commands that reproduce the segfault.

segmentation faults are very generic errors and thus it is difficult
to debug them from remote.


Hi Guangrui,

There were some setup/initialization problems in the latest LAMMPS update for intel package (that were my fault).
I sent a fix that Steve hopes to get out as a patch today.

This might also fix your problem. If not, please let me know.

Best. - Mike