Hamiltonian REX: different exchange probability with -nb gpu

User discussions
1

GROMACS version: 2019.4
GROMACS modification: Yes (PLUMED 2.5.7 patched)

Dear all,
I am testing the hamiltonian replica exchange for a small RNA tetraloop system (NVT condition, hydrogen mass repartition 4 fs time step). As a part of testing, I am using the same scaled topology for four replicas to check exchange probability. The test run was performed for 400 ps. However, when I use the “-nb gpu” option with mdrun, I get a slightly low exchange rate. The simulations are repeated and the results are consistent.
Also, replex (250) is set as multiple of nstlist (25) and dynamic load balancing is also off (-dlb no).

Please, can you suggest whether using “-nb gpu” could affect energies and thus exchange probability in general?

mdrun command line:
mpirun -np 4 gmx_mpi mdrun -v -deffnm md -plumed plumed.dat -replex 250 -hrex -multidir dir_0 dir_1 dir_2 dir_3 -dlb no -nb gpu -nsteps 100000 >& log

With “-nb gpu”:
dir_0/md.log:Repl average probabilities:
dir_0/md.log-Repl 0 1 2 3
dir_0/md.log-Repl .96 .84 .97
dir_0/md.log-Repl number of exchanges:
dir_0/md.log-Repl 0 1 2 3
dir_0/md.log-Repl 190 172 192
dir_0/md.log-Repl average number of exchanges:
dir_0/md.log-Repl 0 1 2 3
dir_0/md.log-Repl .95 .86 .96

without “-nb gpu”:

…/No_nb_gpu/dir_0/md.log:Repl average probabilities:
…/No_nb_gpu/dir_0/md.log-Repl 0 1 2 3
…/No_nb_gpu/dir_0/md.log-Repl .96 .97 .97
…/No_nb_gpu/dir_0/md.log-Repl number of exchanges:
…/No_nb_gpu/dir_0/md.log-Repl 0 1 2 3
…/No_nb_gpu/dir_0/md.log-Repl 198 192 197
…/No_nb_gpu/dir_0/md.log-Repl average number of exchanges:
…/No_nb_gpu/dir_0/md.log-Repl 0 1 2 3
…/No_nb_gpu/dir_0/md.log-Repl .99 .96 .99

2

Hi,
Calculating the nonbonded interaction on gpu or not should not affect the results.
If you are running on the same machine and GROMACS is compiled accounting for gpu the non-bonded interactions are calculated on GPU with and without explicitly add the option -nb gpu).
My guess is that the difference you observed are due too not enough sampling.
Best regards
Alessandra

3

Hi Alessandra,
Thanks for your reply. I agree that I should observe the same results.
I observed these results by repeating these tests, so I was a bit worried.
I will run longer simulations and will post the outcome here.

Thanks,
Mandar Kulkarni