gpu along with respa

LAMMPS LAMMPS Mailing List Mirror
1

Dear all,

Is it possible to use lammps gpu-package with respa. I observed usage of respa slowing down the computation compared to without respa case.

thanks in advance for your comments.

with regards

RAVI KUMAR REDDY A

2

Dear all,

Is it possible to use lammps gpu-package with respa. I observed usage of respa slowing down the computation compared to without respa case.

yes, it is possible. how do you determine that the simulation is slowing down? and what kind of a system with which force field and what respa settings are you running?

axel.

3

Dear axel,

Please find the input files in the attachment.

I did two simulations for the same system one without using respa 0.5 fs as timestep and one one with respa with outer timestep of 1.0 fs.

For 0.1 ns run of both simulations, case I has completed in 52 min whereas case II took 55 min. I expect case II to be relatively faster …
Is it because of an-optimized respa levels I used ?

input_gpu.txt (1.25 KB)

input_respa_gpu.txt (1.31 KB)

4

Dear axel,

Please find the input files in the attachment.

I did two simulations for the same system one without using respa 0.5 fs as timestep and one one with respa with outer timestep of 1.0 fs.

For 0.1 ns run of both simulations, case I has completed in 52 min whereas case II took 55 min. I expect case II to be relatively faster …

neither of those runs lasts for 0.1 ns. also, your choice of r-RESPA settings makes no sense.
can you please also create and post inputs and log files for both test runs with the “thermo” and “thermo_style” commands removed and the number of steps reduced by a factor of 100?
this way the files are not too long and that would provide more reliable information than your say-so.

axel.

5

Dear axel,

I have modified the input files as you mentioned and ran those short simulations again.
please find the input along with log files in the attachment.

Thanks and regards
RAVI KUMAR REDDY A

input_gpu.txt (1.26 KB)

input_gpu_respa.txt (1.31 KB)

log_gpu.txt (4.71 KB)

log_gpu_respa.txt (4.92 KB)

6

thanks. i will have a closer look.
there is one more check, that you could do:
please compare the performance of verlet vs respa for the same setup without using the GPU. do you get the same performance relation or something else?

thanks,
axel.

7

it looks to me, you are not gaining much performance, because of your choice of timestep and respa levels.
you should be able to do an outer time step of 2fs on kspace, pair and dihedral and then use 0.5fs for angle/bond and fix spring/self.
you could gain a little more by using doing kspace only every other step and thus have a time step of 4.0fs and
use: run_style respa respa 3 2 4 angle 1 dihedral 2 pair 2 kspace 3

for run_style verlet, you should try out using the non-GPU pppm style that may be run concurrently with pair on the GPU and thus be a little faster.

axel.