On Sun, Oct 2, 2016 at 2:02 PM, Elisa Pieri <elisa.pieri90.gmail.com> wrote:
> Thank you for your help,
>
> using the CPU code, I didn't get any error: the potential energy lowered a
> lot on the first steps and the density increased.
>
> But now I have a different problem: I started the pH-REMD and they are
> incredibly slow! Almost stuck.
> I did this calculation on the system without membrane (32000 atoms) and I
> got 1.19 ns/day on 12 nodes (144 cores in total) for 12 pH-REMD.
> I used the same input, and the same hardware for the system WITH membrane
> (41000 atoms) and I'm getting 78 ps in the last 24 hours. The only things I
> changed are the modules I loaded, choosing gcc49 instead of gcc47 and
> upgrading to python 2.7.11. Here is the input for pH1:
>
> &cntrl
> imin=0, ! No minimization
> irest=1, ! Restart the simulation
> ntx=5, ! Coordinates and velocities from coordinate file
> ntxo=2, ! Format of the final coordinates NetCDF
> ntpr=1000, ! Every ntpr steps, energy is printed
> ntwx=1000, ! Every ntwx steps, the coordinates are written
> nstlim=100, ! Number of steps between exchange attempts
> dt=0.002, ! time step in ps
> ntt=3, ! Langevin dynamics
> tempi=300, ! Initial temperature
> temp0=300, ! Reference temperature
> gamma_ln=5.0, ! collision frequency
> ig=-1, ! random seed
> ntc=2, ! SHAKE bonds involving hydrogen are constrained
> ntf=2, ! bond interactions involving H-atoms omitted
> cut=8, ! nonbonded cutoff
> iwrap=1, ! restart coordinate and trajectory files "wrapped" into a
> primary box
> ioutfm=1, ! Binary NetCDF trajectory
> icnstph=2, ! CpHMD run in explicit solvent
> ntcnstph=50, ! number of steps between attempting protonation state
> changes
> solvph=1, ! solvent pH
> ntrelax=100, ! number of solvent relaxation steps
> saltcon=0.1, ! salt concentration for the GB calculations
> numexchg=10000, ! number of times to attempt exchanging
> ntwr=10000, ! Every ntwr steps, the restrt file is written
> /
>
> Do you have any idea of the reason of this slowness? Indeed, equilibrating
> the system using the mdin give by Jason on his wikidot tutorial, I got 15
> ns/day on 2 nodes (24 cores).
pH-REMD simulations are a lot slower than normal simulations for several
reasons:
1. The protonation state change attempts need to be done in implicit
solvent, meaning that you need to run a GB calculation on all atoms that
are *not* water molecules
2. Following any successful protonation state change attempts, a
"relaxation" dynamics of ntrelax steps is run (only solvent atom positions
are updated)
3. Replica exchange attempts are synchronized, which means if one replica
does 5 more "relaxation" steps than other replicas (because it had more
successful protonation state moves), all of the other replicas need to wait
for that slow replica to finish before performing exchanges.
In my simulations, I set ntcnstph=100 -- you can set this even higher to
get better performance (but not too high or you won't sample protonation
states enough). But when you add membranes, the protonation state moves
(in GB) gets even more expensive. If the non-water part of the system gets
large enough, then this part of the calculation becomes rate-limiting (I'm
not sure if that's what is happening here or not).
I've never tested this model with membranes, and there are quite a few
ways that it could not work well at all. The use of GB in a membrane
system is particularly problematic, and the lack of PBC in the protonation
state change attempts could introduce serious artifacts in the protonation
state sampling depending on exactly what set of periodic images are used in
the GB calculation for each of the membrane molecules.
For a single protein this is not a problem, but for a system with multiple
"molecules" it could be a big one.
HTH,
Jason
--
Jason M. Swails
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Oct 03 2016 - 06:30:02 PDT
