GPU package

LAMMPS LAMMPS Mailing List Mirror
1

I installed the GPU package according to the instruction given at http://lammps.sandia.gov/doc/accelerate_gpu.html

Everything went smoothly and when I run the command

mpirun lmp_ubuntu -pk gpu 1 -in in.Li-dendritic.nvt

the calculaton works well (though I doubt if it is by GPU) but when I run the command

mpirun lmp_ubuntu -sf gpu -pk gpu 1 -in in.Li-dendritic.nvt

I receive the error:
LAMMPS (16 Mar 2018)
Cuda driver error 999 in call at file ‘geryon/nvd_device.h’ in line 273.
Cuda driver error 999 in call at file ‘geryon/nvd_device.h’ in line 273.
Cuda driver error 999 in call at file ‘geryon/nvd_device.h’ in line 273.
Cuda driver error 999 in call at file ‘geryon/nvd_device.h’ in line 273.

2

I installed the GPU package according to the instruction given at
http://lammps.sandia.gov/doc/accelerate_gpu.html

Everything went smoothly and when I run the command

mpirun lmp_ubuntu -pk gpu 1 -in in.Li-dendritic.nvt

the calculaton works well (though I doubt if it is by GPU) but when I run
the command

​it won't run on the GPU unless you explicitly request pair styles with
names ending in /gpu that input

mpirun lmp_ubuntu -sf gpu -pk gpu 1 -in in.Li-dendritic.nvt

​that will try to find /gpu styles, even if they are not explicitly named,
by appending /gpu to every style name and falling back to ​non-gpu
versions, if gpu variants are not available. please see the LAMMPS manual
for more details on this.

I receive the error:
LAMMPS (16 Mar 2018)
Cuda driver error 999 in call at file 'geryon/nvd_device.h' in line 273.
Cuda driver error 999 in call at file 'geryon/nvd_device.h' in line 273.
Cuda driver error 999 in call at file 'geryon/nvd_device.h' in line 273.
Cuda driver error 999 in call at file 'geryon/nvd_device.h' in line 273.

​what kind of GPU do you have? did you compile for CUDA or OpenCL? what
version of CUDA - if any - did you use? what are the compiler/GPU settings
in your lib/gpu compilation?​

​axel.​

3

Thanks for the clarification, axel and apologies for not providing full information. Makefile.linux looks like

ifeq ($(CUDA_HOME),)
CUDA_HOME = /usr/local/cuda-9.1
endif
NVCC = nvcc

Tesla CUDA

#CUDA_ARCH = -arch=sm_21

newer CUDA

#CUDA_ARCH = -arch=sm_13

older CUDA

#CUDA_ARCH = -arch=sm_10 -DCUDA_PRE_THREE
CUDA_ARCH = -arch=sm_30

I use CUDA 9.1 on Ubuntu 18.04. The GPU is Quadro K2100M

4

Thanks for the clarification, axel and apologies for not providing full
information. Makefile.linux looks like

ifeq ($(CUDA_HOME),)
CUDA_HOME = /usr/local/cuda-9.1
endif
NVCC = nvcc
# Tesla CUDA
#CUDA_ARCH = -arch=sm_21
# newer CUDA
#CUDA_ARCH = -arch=sm_13
# older CUDA
#CUDA_ARCH = -arch=sm_10 -DCUDA_PRE_THREE
CUDA_ARCH = -arch=sm_30

I use CUDA 9.1 on Ubuntu 18.04. The GPU is Quadro K2100M

​please also provide the output of lib/gpu/nvc_get_devices
and try to run the example inputs in bench/FERMI on your GPU.
also, please check whether you have the required permissions to access the
GPU device.

axel.

5

Sorry, my problem was unrelated to compiling. It was resolved by disabling/enabling the GPU. Then, I was able to successfuly run the command

mpirun lmp_ubuntu -sf gpu -pk gpu 1 -in in.Li-dendritic.nvt

And there were 4 GPU processes by lmp_ubuntu. However, it took 34min, exactly the same as non-GPU run. I understand that GPU-acceleration appears for large number of atoms, but I expected to withness a tiny contribution. This is the fun part of playing with the variables as mentioned in the manual. If I made a breakthrough I will share it with the community.

Thanks for your excellent support.

6

Sorry, my problem was unrelated to compiling. It was resolved by
disabling/enabling the GPU. Then, I was able to successfuly run the command

mpirun lmp_ubuntu -sf gpu -pk gpu 1 -in in.Li-dendritic.nvt

And there were 4 GPU processes by lmp_ubuntu. However, it took 34min,
exactly the same as non-GPU run. I understand that GPU-acceleration
appears for large number of atoms, but I expected to withness a tiny
contribution. This is the fun part of playing with the variables as
mentioned in the manual. If I made a breakthrough I will share it with the
community.

​please keep in mind, that you do not have a very powerful GPU with a
rather slow memory transfer speed compared to high-end desktop GPUs.
this results in additional overhead. since the GPU kernels essentially have
to do twice ​the work (for more efficient parallelization), it is much more
challenging to have the GPU acceleration offset the overhead with a
moderate sized, mobile GPU compared to a high-end desktop/workstation GPU.
in addition, it depends a lot on the size of the system (as you noted) and
the specifics of the force field setup (which you have not provided).

thus i strongly suggest you experiment first with the benchmark examples
bundled with LAMMPS before rolling your own. that - and the corresponding
performance numbers published on lammps.sandia.gov - will give you an
assessment of what acceleration is possible with those inputs and then you
can assess what might be possible with your specific setup.

​axel.​