about the scale of lammps-gpu

Hello,Mike Brown

My name is Wang。There are two gpu cards on my computer,tesla C2075 and C1060.I want to know the maximum scale of the lammps-gpu.Because when I use the gpu/morse on my computer,the scale is only 400000.Can you tell me some methods to improve my scale.At first,we want to use tersoff,but,it is not suppored at present.When will the tersoff be added to the lammps-gpu.Here is the input- file,I don’t know whether there are some problems in it.

newton off

package gpu force/neigh 0 1 1

units metal

boundary s s p

atom_style atomic

region box1 block 0 200 0 200 0 200 units box

region tool sphere 220 210 100 15 units box

region box union 2 box1 tool units box

create_box 4 box

lattice diamond 5.431

create_atoms 1 region box1 units box

lattice diamond 3.5667

create_atoms 1 region tool units box

pair_style morse/gpu 8.025

pair_coeff * * 0.17733 1.53744 3.7598 8.025

pair_coeff 1 4 0.435 4.6487 1.9475 8.025

pair_coeff 2 4 0.435 4.6487 1.9475 8.025

pair_coeff 3 4 0.435 4.6487 1.9475 8.025

pair_coeff 4 4 3.68 2.20 1.54 8.025

neighbor 2.0 bin

neigh_modify delay 5

#neigh_modify exclude type 1 1

#neigh_modify exclude type 1 2

#neigh_modify exclude type 1 3

#neigh_modify exclude type 1 4

mass 1 28.09

mass 2 28.09

mass 3 28.09

mass 4 12

region 1 block 0 200 0 20 INF INF units box

region 2 block 0 20 20 200 INF INF units box

region 3 block 20 40 20 200 INF INF units box

region 4 block 40 200 20 40 INF INF units box

region 5 block 40 200 40 200 INF INF units box

region boundary1 union 2 1 2 units box

region heng1 union 2 3 4 units box

group 11 region boundary1

group 22 region heng1

group 33 region 5

group 44 union 22 33

group 55 region tool

set group 11 type 1

set group 22 type 2

set group 33 type 3

set group 55 type 4

compute 3 33 pe/atom

compute 4 33 ke/atom

compute 5 33 coord/atom 3.0

compute chang1 22 temp

velocity 22 create 293.0 5812775 temp chang1 units box

compute chang2 33 temp

velocity 33 create 293.0 5812775 temp chang2 units box

fix 1 all nve

fix 2 11 setforce 0.0 0.0 0.0

fix 3 22 temp/rescale 10 293.0 293.0 10.0 1.0

fix 4 55 setforce 0.0 0.0 0.0

fix_modify 3 temp chang1

velocity 55 set -2 0 0 units box

velocity 44 set 0.0 0.0 0.0 units box

dump 1 33 custom 5000 dump.atom id xs ys zs c_3 c_4 c_5

dump mydump all custom 100 dump-nvt.lammpstrj id type x y z vz vy vz

thermo 20

#thermo_modify temp chang1

timestep 0.001

run 2000

Forwarding this to Mike.
What does scale = 400K mean?
There is a tersoff in the USER-CUDA package
you can try.

Steve

Hi Andy

you question is not very clear. Do you mean with scale what the largest system is you can run on a single C2075? Pairing the C1060 and the C2075 is nothing you should try to do. I am also not sure whether the GPU package does support that considering that they are from two totally different GPU Generations. That said, the number of atoms per GPU is usually limited by the size of the neighborlist. There is no artificial limit on the number of atoms you can put on a GPU otherwise (except maybe for the integers only supporting values up to about two billion, limiting your total number of atoms maybe to something like 500 million or so - but I doubt that anyone tested that).

The USER-CUDA package has already Tersoff support btw. and you can run simulations with a bit more than 2,000,000 atoms with it.

Christian

-------- Original-Nachricht --------

Hi, I am not sure what you mean by "scale" here. Can you explain?

I do not yet have an eta of 3-body potentials for gpu.

Thanks. - Mike