[gmx-users] how to improve computing?

[Alex]

This here is pretty useful, as it turns out (наконец-то нормальную 
документацию написали): 
http://www.gromacs.org/Documentation/Acceleration_and_parallelization

Developers tell us that we can get very good performance without 
utilizing all CPU threads. Anyway, if you built gmx with all the default 
flags (-DGMX_BUILD_OWN_FFTW=ON -DGMX_GPU=ON), I would try mdrun with 
something like:

option 1:

-nt 16 -ntmpi 4 -npme 1 -pme gpu -nb gpu -gputasks 0000

option 2:

-nt 8 -pin on -ntmpi 2 -npme 1 -pme gpu -nb gpu -gputasks 00

just to compare performance and see if your GPU is bottlenecking things. 
Depending on the CPU-GPU execution time ratio, if it is <1, you may also 
try not offloading PME's FFT to the GPU (-pmefft cpu). Then try -ntomp, 
etc, etc.

Alex


On 6/13/2018 8:38 PM, Владимир Богданов wrote:
> Hey everyone!
> Could help me somebody, please. I have PC with 8 cores, 16 logical 
> cores and 1 compatible GPU, but gromacs calculates just with 1GPU, 
> 1MPI thread and 8 OpenMP threads. Would it possible for me to use all 
> of 16 logical cores?
> i7-7820X
> Nvidia Titan Xp
> gromacs 2018 CUDA=on
> Ubuntu 16.04
> -- 
> Thanks,
> Vlad
>
>