[gmx-users] changing cutoffs with gmx tune_pme

[Leandro Bortot]

Dear users,

     I'm doing some tests with gmx tune_pme to optimize the setup for my
system. In the help options I saw that I could change the coulomb and LJ
cut-offs in addition to varying the number of PME ranks. So I did that to
test.

     My system runs faster when rcoulomb and rvdw are small. Specifically =
0.5 nm in my tests.
     When I compare test runs for the same number of PME ranks I see that
fourier-nx, fourier-ny and fourier-nz change with the cut-off, but
fourierspacing remains constant.

     My question is: is actually running the production run with such small
cutoffs justifiable?
     I think it is not and that this would break the forcefield (AMBER99).
I would like to hear what you think about that.


     Here is some data in case it is helpful:

For the same number of cores and PME ranks:

Performance = 4.688 ns/day
   rcoulomb                       = 1
   rvdw                               = 1
   fourierspacing                = 0.12
   fourier-nx                     = 320
   fourier-ny                     = 120
   fourier-nz                     = 120
   pme-order                      = 4
   ewald-rtol                     = 1e-05
   ewald-rtol-lj                  = 0.001

Performance = 2.060 ns/day
   rcoulomb                       = 1.5
   rvdw                              = 1.5
   fourierspacing                 = 0.12
   fourier-nx                     = 208
   fourier-ny                     = 80
   fourier-nz                     = 80
   pme-order                      = 4
   ewald-rtol                     = 1e-05
   ewald-rtol-lj                  = 0.001

Performance = 8.248 ns/day
   rcoulomb                       = 0.5
   rvdw                           = 0.5
  fourierspacing                 = 0.12
   fourier-nx                     = 640
   fourier-ny                     = 240
   fourier-nz                     = 240
   pme-order                      = 4
   ewald-rtol                     = 1e-05
   ewald-rtol-lj                  = 0.001


Best regards,
Leandro