[gmx-users] Re: Segmentation fault, mdrun_mpi

[Ladasky]

Update:


Ladasky wrote
> 
> Justin Lemkul wrote
>> Random segmentation faults are really hard to debug.  Can you resume the
>> run 
>> using a checkpoint file?  That would suggest maybe an MPI problem or
>> something 
>> else external to Gromacs.  Without a reproducible system and a debugging 
>> backtrace, it's going to be hard to figure out where the problem is
>> coming from.
> Thanks for that tip, Justin.  I tried to resume one run which failed at
> 1.06 million cycles, and it WORKED.  It proceeded all the way to the 2.50
> million cycles that I designated.  I now have two separate .trr files, but
> I suppose they can be merged.
> 
> I don't know whether my crashes are random yet.  I will try re-running
> that simulation again from time zero, to see whether it segfaults at the
> same place.  If it doesn't, then I have a problem which may have nothing
> to do with GROMACS.

I just tried exactly that, a re-run of the same structure.  This time, it
ran without stopping, from time zero to 2.50 million cycles!  No crash at
1.06 million cycles this time.

Unless GROMACS is using some random number generator which affects the
outcome of repeated simulations (and I think that the only time that random
number generation would be needed would be when initial velocities are
generated, which was done during the earlier equilibration step), I will
conclude that my simulation conditions are indeed acceptable, and that
sometimes the software just behaves badly.

Is that a common occurrence?

I could write a script which just automatically restarts my simulations
provided that they 1) ran for a decent number of cycles and b) exited with a
segmentation fault error.  I could then have the script check in after a few
minutes to make sure that they haven't crashed again, and soldier on.
 



--
View this message in context: http://gromacs.5086.n6.nabble.com/Segmentation-fault-mdrun-mpi-tp5001601p5001879.html
Sent from the GROMACS Users Forum mailing list archive at Nabble.com.