fix 0 all gpu force/neigh 1 1 -1.0

LAMMPS LAMMPS Mailing List Mirror
1

Hi,

I have two gpus in a linux box and I did nvc_get_devices and both are
seen as Device 0 and Device 1 respectively. I tried to run two
different simulations at the same time and changed the input files of
each run as "fix 0 all gpu force/neigh 0 0 -1.0" and "fix 0 all gpu
force/neigh 1 1 -1.0" hoping to assign one simulation to Device 0 and
the other to Device 1. This doesn't seem to work. And, If I ran the
simulation with "fix 0 all gpu force/neigh 1 1 -1.0", the gpu used by
the run is Device 0 and not Device 1. I used nvidia-smi to track GPU
utilization. My cuda version is 3.2.

Is there a way to effectively choose which GPU the simulation will use?

Thanks

Jan-Michael

2

Hi,

I have two gpus in a linux box and I did nvc_get_devices and both are
seen as Device 0 and Device 1 respectively. I tried to run two
different simulations at the same time and changed the input files of
each run as "fix 0 all gpu force/neigh 0 0 -1.0" and "fix 0 all gpu
force/neigh 1 1 -1.0" hoping to assign one simulation to Device 0 and
the other to Device 1. This doesn't seem to work. And, If I ran the
simulation with "fix 0 all gpu force/neigh 1 1 -1.0", the gpu used by
the run is Device 0 and not Device 1. I used nvidia-smi to track GPU
utilization. My cuda version is 3.2.

Is there a way to effectively choose which GPU the simulation will use?

do you use the latest version of LAMMPS?

there was a bug in the gpu code and i think it has been fixed since.
you can search the mailing list archives.

axel.

3

Hi Jan-Michael

I think the fix for the issue you are running into is here: http://lammps.sandia.gov/threads/msg19380.html

BTW, You shouldn't run into this issue with the GPU library update from my site. If you do, let me know.

- Mike

Jan-Michael Carrillo wrote:

4

Mike,Steve and Axel;

I updated the Apr 29,2011 version using gpu_package.tar.gz from Mike's
website. The change in lib/gpu/pair_gpu_device.cpp for the variable
my_gpu was done during the update. This is similar to the link that
Axel showed me. The update now assigns the gpu correctly and my
simulations are running.

However, when I downloaded the May 2 version of lammps from the
sandia website (not the patch) and I tried to compile the gpu library,
I get an error: " No rule to make target `nv_kernel_def.h' needed by
`pair_gpu_atom.o'" when I type the command make -f Makefile.linux in
the /lammps/lib/gpu directory. I suspect that the May 2 version was
not packaged properly.

-Jan Michael

5

I'll let Mike answer this and make any changes he suggests.

Steve

6

There were some missing files in that patch (my fault) which Mike
alerted me to. So an addendum 4May patch came out yesterday.
See if that fixes this problem?

Steve