[lammps-users] Performance about lammps on GPU


please always copy the list.

yes, this looks like you are getting the same behavior that i've seen earlier.
i have not looked into the code, but it might be that the fix GPU is ignoring
the offset or GPU index range parameter.


it also would be more flexible to handle this in a similar way as NAMD does it,
i.e. have single string that is then turned into a list, e.g. 0,1,2,3
or 0-3 or 1,3,2,4,
the latter is which is what i need for our 4-GPU nodes to get the best GPU to
processor mapping (since our machines have two southbridges, and you want
to tie the GPUs to the processor that hosts the southbridge where the individual
GPU is attached to).


2011/3/15 Yangpeng Ou <[email protected]>:

Hi Yangpeng,

Yes, glancing at the code, I believe that you have found a bug - the code will split work between multiple GPUs correctly, but is always using 0 as the first GPU index. I will look at a fix this afternoon.

Regarding parallel performance, I would _guess_ that you should see good parallel performance on a single desktop. Try running lammps on all available cores, even if this is more than the number of GPUs - otherwise the PPPM computation run on the CPU can dominate and limit performance.

Axel -

I will add a more sophisticated ordering option to the todo list... :slight_smile:

- Mike