On Mon, Oct 24, 2022, 张兴赫 via AMBER wrote:
>
>FIrst one, the fundamental of QM/MM calculation, so far as what I
>understand, is to replace classical force field (eg. gaff2, etc) with
>quantum mechanical calculation. So for Gaussian/AMBER simulation, Gaussian
>will calculate the energy and force for the QM part, which is all that
>Gaussian do in simulation. Sander or pmemd will take care of the evolution
>of coordinates and velocities of the entire system, no matter the QM part
>or MM part, using certain algorithm, eg., leap-frog, verlet, etc. Do I make
>some mistakes, or are there any details that I missed?
Above sounds correct.
>
>Nest, for parallel calculation, I got confused. Under the &gau namelist,
>the Amber manual indeed mentioned "num_threads". So for example, if I
>assign 12 cpu to sander.MPI, i.e., command "mpirun -np 12 sander.MPI xxx",
>and set num_threads=6, then for qm part, Gaussian will use 72 cpu, or
>merely 6 cpu? And the number of cpu requested is the sum of sander.MPI and
>Gaussian, or Gaussian just takes a portion of cpu from what I assign to
>sander?
For "most" QM/MM simulations with ab initio codes, the QM part dominates the
calculation. So there is likely no benefit to running the MM part in
parallel, and may be some disadvantages.
Try a short simulation with sander (not sander.MPI) and num_threads = 1
in the &gau namelist. Then increase num_threads and see what sorts of
timings you get. (The gaussian code doesn't know anything about how many
MPI threads you requested for sander.)
If you really need parallel sander in QM/MM, others on the list will have to
chime in. Or, just run some simple experiments on your system, starting
with 2 MPI threads.
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Oct 24 2022 - 06:00:03 PDT
