GPU for granular simulations

LAMMPS LAMMPS Mailing List Mirror
1

Dear Developers,

I am running a granular frictionless short range repulsion simulation:

pair_style gran/hooke 1 0 1 0 0 0
pair_coeff * *

We have installed LAMMPS GPU on our CUDA cluster which is:

GPU : Nvidia Tesla m2050 (1.15 GHz, 448) * 4EA

Each machine has 4 cores, so I need to use MPI to get the full machine:

Set parallel environment (pe) to OpenMPI (opmi) and number of cores required.

#$ -pe gpu_openmpi 4

and then I use the following command to run the code:

/home/program/lammps/lmp_gpu -sf gpu -pk gpu 1 -in in.file

However, I get the following error:

/home/program/lammps/lmp_gpu: symbol lookup error: /opt/intel//impi/5.0.1.035/intel64/lib/libmpifort.so.12: undefined symbol: MPI_UNWEIGHTED

Is this because pair_style gran/hooke apparently has no GPU speed up?

Or it is the installation problem on GPU?

I see in the net that people have already used GPU for granular simulations, but style hooke has no explicit GPU extension.

In advance, I appreciate your kind help!

2

Dear Developers,

I am running a granular frictionless short range repulsion simulation:

pair_style gran/hooke 1 0 1 0 0 0
pair_coeff * *

We have installed LAMMPS GPU on our CUDA cluster which is:

GPU : Nvidia Tesla m2050 (1.15 GHz, 448) * 4EA

Each machine has 4 cores, so I need to use MPI to get the full machine:

# Set parallel environment (pe) to OpenMPI (opmi) and number of cores
required.
#$ -pe gpu_openmpi 4

and then I use the following command to run the code:

/home/program/lammps/lmp_gpu -sf gpu -pk gpu 1 -in in.file

this is a pointless enterprise. there currently is no GPU acceleration
for any granular pair style in LAMMPS. your gpu suffix flag has no
effect.

However, I get the following error:

/home/program/lammps/lmp_gpu: symbol lookup error:
/opt/intel//impi/5.0.1.035/intel64/lib/libmpifort.so.12: undefined symbol:
MPI_UNWEIGHTED

Is this because pair_style gran/hooke apparently has no GPU speed up?

no. this looks like you have an incompatible version of the _intel_
(!!) mpi runtime library in your LD_LIBRARY_PATH environment, that
differs from the version your executable was compiled with. this
conflicts with your claim, that you were using OpenMPI.

Or it is the installation problem on GPU?

it is a compilation/installation problem.

I see in the net that people have already used GPU for granular simulations,

but they must have used a different software or not shared their
implementation with the LAMMPS developers.
whatever is written "somewhere in the internet" is irrelevant here.
what matters is what the LAMMPS manual says: whatever is not mentioned
is not there; no feature is added to LAMMPS without documenting it.

axel.

3

Dear Axel,

Thanks for your kind response.

You mean there is no support for granular pair forces by GPU LAMMPS?

4

Dear Axel,

Thanks for your kind response.

You mean there is no support for granular pair forces by GPU LAMMPS?

where does it say there is?

axel.