Accessing neighbor's neighbor list in MPI multi-process runs

LAMMPS LAMMPS Mailing List Mirror
1

Dear LAMMPS users,

In MPI multi-process simulation, when I want to add a local density dependent term in force calculation,

F_ij = … + B_ij * (rho_i + rho_j) * wd(r_ij) e_ij

where rho_i = sigma_{j != i} wp(rho_ij) for r<rc

I may not be able to correctly calculate rho_j, since the neighbor particle j’s neighbors may not reside in the same partition with particle i.

So I assume LAMMPS by default does not guarantee the neighbor’s neighbor list accessible in MPI mode. I am wondering if such algorithm is possible to be implemented in LAMMPS? Please let me know if I can refer to any existing piece of code for dealing with this situation.

Thanks in advance for any comment!

Best regard,

2

Dear LAMMPS users,

In MPI multi-process simulation, when I want to add a local density
dependent term in force calculation,

F_ij = ... + B_ij * (rho_i + rho_j) * wd(r_ij) e_ij

where rho_i = sigma_{j != i} wp(rho_ij) for r<rc

I may not be able to correctly calculate rho_j, since the neighbor particle
j's neighbors may not reside in the same partition with particle i.

So I assume LAMMPS by default does not guarantee the neighbor's neighbor
list accessible in MPI mode. I am wondering if such algorithm is possible to
be implemented in LAMMPS? Please let me know if I can refer to any existing
piece of code for dealing with this situation.

it is already available and as far as i can tell used in the pair
styles airebo, bop, comb3, lcbop, reax/c, and snap.
if you read the respective code, you'll see, that you have to modify
the neighbor list request to set the ghost flag to 1 and also tell the
pair style through setting ghostneigt to 1 in the respective
constructor that you need the neighbors of ghosts.

axel.

3

Cool, thanks a lot for your prompt reply — I found those files in the MANYBODY folder. Will take a hard look at them.

4

One more minor question that is related:

I am not quite clear how atom->nmax is calculated, and its arithmetic relation to atom->nlocal and atom->nghost. Here I suppose a “p p p” system with predefined number of atoms.

Thanks,

5

One more minor question that is related:

I am not quite clear how atom->nmax is calculated, and its arithmetic
relation to atom->nlocal and atom->nghost. Here I suppose a "p p p" system
with predefined number of atoms.

atom->nmax is the largest value of atom->nlocal + atom->nghost that
has existed across all subdomains (i.e. MPI processes) during the
LAMMPS run so far. it can only grow, never shrinks.

axel.

6

FYI, variables like this are documented in the comments of the header
files where they are declared. so in atom.h you have:

  // atom counts

  bigint natoms; // total # of atoms in system, could be 0
                                // natoms may not be current if atoms lost
  int nlocal,nghost; // # of owned and ghost atoms on this proc
  int nmax; // max # of owned+ghost in arrays on this proc
  int tag_enable; // 0/1 if atom ID tags are defined
  int molecular; // 0 = atomic, 1 = standard molecular system,
                                // 2 = molecule template system

7

Your formula for density dependence looks a lot like EAM,

in which case you do not really need neighbors of ghost atoms.

Instead you should look at how pair_eam.cpp does communication

of quantities computed for atoms owned by other procs and stored by ghost atoms.

Steve

8

Thanks for the advice! I have been following the way done in pair_eam.cpp for calculating the local density for ghost atoms.

Best regard,