LIGGGHTS in parallel

Submitted by andrea.pasquali on Mon, 04/11/2011 - 09:30

Hi,
I'm trying LIGGGHTS in parallel. For example:
mpirun -np 8 lmp_openmpi < in.case
When it starts it prints:
1 by 1 by 8 processor grid
The problems is that sometime it works well, sometime it crashes with this error:
ERROR on proc 0: Failed to allocate -1738615408 bytes for array neigh:binhead

For example it works well with:
communicate multi vel yes
units si
neighbor 0.001 bin
neigh_modify delay 0
variable radius1 equal 0.005 # m

It doesn't work with:
communicate multi vel yes
units si
neighbor 0.003 bin
neigh_modify delay 0
variable radius1 equal 0.0025 # m

I have 3 questions:
1) Is it possible to split the grid in different way? For example 2x2x2 instead of 1x1x8? What is the best?
2) To run in parallel is the "communicate multi" appropriate? Is the skin (in "neighbor") related with the radius of particles? Do I need "check yes" in "neigh_modify"?
3) Running in parallel is not so faster than single cpu.

Thanks for any help

Andrea

Forums:

alice's picture

alice | Mon, 04/11/2011 - 10:52

Hi Andrea,
i think i can help with questions 1 and 3:
It is possible to split the computational domain in any way you want, just add the command "processors x y z" to you input script, where x, y and z are the number of domains you want to get in the very direction. (e.g. processors 2 2 2 instead of processors 1 1 8) I think this should do what you are aiming for.
and concerning 3: the parallelization runs very efficiently, so it really makes sense to use that feature...
Cheers,
Alice

andrea.pasquali | Mon, 04/11/2011 - 11:07

Hi Alice,
thank you for you reply!
Do you know also what is "ERROR on proc 0: Failed to allocate -2141728528 bytes for array neigh:binhead" related to?

Thanks again

Andrea

alice's picture

alice | Mon, 04/11/2011 - 11:26

Hi,
I was confronted with the same error too, and it disappeared when i changed the skin in the neighbor-command (neighbor skin style). This parameter affects the rebuilding of the pairwise neighbor lists, the bigger it is, the less often the lists are rebuilt. It depends on the units of your simulation. (for details see the documentation)
Cheers,
Alice

andrea.pasquali | Mon, 04/11/2011 - 11:47

Hi Alice,
yes I though it depends on skin value.
I'm thinking if a correlation exists between radius of particles and skin to set the simulation file correctly the first time only.
I saw if I decrease the radius I have to increase the skin.
However I will try.

Regards

Andrea