problem running on multi- nodes

Hi users,

I am able to run lammps version on multinode for several systems,

But for a new prepared system with following parameters,

pair_style hybrid/overlay lj/cut 14.0 lj/charmm/coul/long 14.0 14.5 hbond/dreiding/lj 4 4.3 4.4 130.0
pair_modify tail yes
bond_style harmonic
angle_style harmonic
dihedral_style hybrid charmm opls
improper_style cvff
kspace_style pppm 0.00001

timestep 2.5

run_style respa 3 2 2 bond 1 dihedral 2 pair 3

I am able to run serially, but not able to run on more no. of processors, but often pops out with following

PPPM initialization …
G vector (1/distance)= 0.0835161
grid = 12 12 12
stencil order = 5
estimated absolute RMS force accuracy = 0.00282613
estimated relative force accuracy = 8.51081e-06
using double precision FFTs
3d grid and FFT values/proc = 968 108
Setting up run …
APPLICATION TERMINATED WITH THE EXIT STRING: Segmentation fault (signal 11)

Instead of respa, If i use verlet, its running fine.

As I understood from documentation, lj/charmm/coul/long support respa.

Can anyone help me?

Thanks

Karteek Kumar

HI users,

I have tested files with same parameters across different clusters,

pair_style hybrid/overlay lj/cut 14.0 lj/charmm/coul/long 14.0 14.5 hbond/dreiding/lj 4 4.3 4.4 130.0 with respa is not running in parallel. but with verlet its not complaining.

In two of the clusters,
job ran on 6 proc for once, but if i kill and resubmit its giving error ( this may be looking stupid to you all, but it happened)

Need for help

Thanks
Karteek Kumar

i don't think you can have charmm dihedrals at a different respa level
than the corresponding pair potential.
you have a crazy force field setup anyway. there are likely problems
galore. how are you going to validate this mess?

axel.

Thanks Axel garu,

I have set up the system, so as to take the advantage of having 1-4 interactions to part of the system (A) and ignore for the rest(B). ( with the help of lj/charmm/coul/long)

I have tested 1-4 interactions for pure system (A), first via special bonds and second one via dihedral weighting factor, both giving same energy.

Now, Respa levels are modified to be evaluate dihedral and pair at same level, this has solved the problem.

But still, I have fullish question, how can it ran serially?

Thanks
Karteek Kumar

Thanks Axel garu,

I have set up the system, so as to take the advantage of having 1-4
interactions to part of the system (A) and ignore for the rest(B). ( with
the help of lj/charmm/coul/long)

I have tested 1-4 interactions for pure system (A), first via special
bonds and second one via dihedral weighting factor, both giving same
energy.

Now, Respa levels are modified to be evaluate dihedral and pair at same
level, this has solved the problem.

But still, I have fullish question, how can it ran serially?

lots of things run (note, that i don't say "work") in serial that
crash in parallel. often it is about data that is not initialized,
because it has not been communicated yet. in serial, this is not
needed.

axel.