Dear all,
We are trying to use Nvidia GPU to accelerate lammps with eam potential. The part of the code as following:
units metal
boundary p p p
package gpu force/neigh 0 3 1
atom_style atomic
newton off
read_data CuAl-1024000.dat
mass 1 26.982
mass 2 63.546
region center block -80 80 -80 80 -80 80 units box
group center region center
dump 1 all xyz 10000 M01.xyz
dump 2 center xyz 10000 M0X_1.xyz
pair_style eam/alloy/gpu
pair_coeff * * AlCu.eam.alloy Al Cu
There are four K20m GPUs on the workstation. The problem is that the job will stop without any error information after it runs some steps. It seems the job is still attached there, but GPUs don't work.
We tried several times, and it crushed each time with different steps. even though the job runs well, the utilization of GPUs is only about 55%~65% for each GPU. Then we removed the region and group
command, as well as dump 2. The job run well without any that weird problem occurred. Could someone can explain and help us solve this problem? How can fully use the GPUs? Thanks very much.
Best,
Lili
Dear all,
We are trying to use Nvidia GPU to accelerate lammps with eam potential. The part of the code as following:
units metal
boundary p p p
package gpu force/neigh 0 3 1
atom_style atomic
newton off
read_data CuAl-1024000.dat
mass 1 26.982
mass 2 63.546
region center block -80 80 -80 80 -80 80 units box
group center region center
dump 1 all xyz 10000 M01.xyz
dump 2 center xyz 10000 M0X_1.xyz
pair_style eam/alloy/gpu
pair_coeff * * AlCu.eam.alloy Al Cu
There are four K20m GPUs on the workstation. The problem is that the job will stop without any error information after it runs some steps. It seems the job is still attached there, but GPUs don't work.
We tried several times, and it crushed each time with different steps. even though the job runs well, the utilization of GPUs is only about 55%~65% for each GPU. Then we removed the region and group
command, as well as dump 2. The job run well without any that weird problem occurred. Could someone can explain and help us solve this problem? How can fully use the GPUs? Thanks very much.
it is practically impossible to give any advice based on such vague information.
you need to provide:
- which version of LAMMPS you are using. FWIW: the GPU library was
very recently updated.
- what platform you are running on, which version of the CUDA toolkit
and nvidia driver you are using.
- output of nvc_get_devices
- information about compile flags, particularly, if you compiled the
gpu library for single, mixed, or double precision math.
- detailed information about the input. posting a part of the input is
worse than not posting any input at all.
- if you want somebody to look into it seriously, you have to reduce
the input to the smallest possible system that reproduces the bad
behavior and post it here. if you
as for increasing utilization, you can ramp up the number of MPI
tasks. utilization also depends on system size. if your system is too
small, there is not much gain. 4 GPU/node is a *lot* for kepler
generation GPUs.
axel.
axel.