[AMBER] Fwd: NVE GPU question

From: Joseph Baker <bakerj.tcnj.edu>
Date: Thu, 22 Jan 2015 14:20:58 -0500

This didn't go through the first time because I wasn't subscribed yet.
Ross, some answers to your questions below. Thanks!

On Thu, Jan 22, 2015 at 1:48 PM, Ross Walker <ross.rosswalker.co.uk> wrote:

> Hi Joe,
> I've moving this over to the AMBER mailing list as I think this is of
> interest to others and they maybe able to offer more suggestions than me.
> My initial guess is that this may be due to the way PME is working with
> ionic liquids. Are the systems definitely neutral in all cases? The
> behavior of PME maybe different if the system is not neutral.
In the amber output files the "sum of charges" line says that
     Sum of charges from parm topology file = -0.00020404 (pure IL system)
          Sum of charges from parm topology file = -0.00102021 (waters +
So they look to be effectively charge neutral.

> My second suspicion is that the average velocities might be higher for
> ionic liquids - have you tried at different temperatures.

I have not tried at different temperatures. One of the preparation
protocols tried was that the system was annealed up to 500 K and then
brought back down to 300 K (in NPT using langevin thermostat and monte
carlo barostat). Then left to run at 300 K in NPT for a bit before
switching to NVE simulations.

> In the case of a protein in TIP3P water with Shake you have no atoms that
> are affectively unconnected to anything else. So your waters have an
> effective mass of 18 per molecule and your protein has a huge per molecule
> mass. The waters are less than sodium for example in mass but there may be
> hydrogen bonding effects, or their VDW are different such that they are
> moving less. This is just a stream of conciousness - I am not sure why
> you'd see the difference but I could image there being more effective loss
> of precision somewhere due to either - charge charge interaction being
> larger on average - bigger +'ve and -'ve numbers that cancel out for
> example. This sort of thing would certainly effect energy conservation
> within the SPFP precision model.
> There might also be more integration error going on if the particles are
> moving faster although that would imply a higher overall temperature. Does
> reducing the timestep help?
So, this is a series of runs that were done of a pure IL system (the one
from the Amber tutorial with just gaff charges) with 3984 atoms in NVE
starting from the end of just a short annealing protocol, and the timestep
was changed to see the influence

dt = 0.0001
drift = 1.8 x 10^-4 kT/ns/dof

dt = 0.005
drift = 2.9 x 10^-4 kT/ns/dof

dt = 0.001
drift = 3.6 x 10^-4 kT/ns/dof

dt = 0.002 (with ntc=ntf=2)
drift = 1.0 x 10^-2 kT/ns/dof

To get these drifts I just did a simple linear regression through the
entire NVE Etot vs time dataset using xmgrace.

(I can give more details on how each of these were run if it is useful at
some point)

I also tried to check the energy drift for another ammonium salt system (in
water, 20917 atom system) that my collaborator I'm working on these ILs
with had from some previous work, and there the drift in NVE looked much

dt = 0.001
drift = -2.0 x 10^-5 kT/ns/dof

He did not just use gaff for the ammonium salt charges, so could this just
be an issue of needing to do a more refined charge fitting?

> Have you tried DPFP to see how it does with that?

I did try DPFP at a 2 fs step (same as the pure IL system described above),
and there I got

dt = 0.002
drift = 8.8 x 10^-3 kT/ns/dof

So it was a little better than the 2 fs timestep in SPFP

> Anyone else want to chime in?
> All the best
> Ross
Thanks for the input.


> On Jan 20, 2015, at 3:26 PM, Joseph Baker <bakerj.tcnj.edu> wrote:
> Hi Ross,
> I hope you've been doing well. I've got a question for you in relation to
> trying to compare some energy drifts to your paper on explicit solvent PME
> in GPU amber. My protein simulations have NVE drift that is spot on (5 x
> 10^-6 kT/ns/dof for 1 fs steps, and 1.3 x 10^-5 kT/ns/dof for 2 fs step),
> so that's been very nice (the code really does great for energy
> conservation in the protein NVE sims).
> However, I've tried to dabble a bit with some ionic liquids (currently am
> testing the one that is in the amber ionic liquid tutorial that uses gaff
> parameters). There I've tried a whole ton of different timestep, cutoff,
> equilibration protocol combinations, and the best I can seem to get is
> around a couple x 10^-4 kT/ns/dof (these are pure ionic liquid simulations,
> but I get a similar result in a simulation where I sprinkled the ionic
> liquid inside a box of water, actually there I get a bit worse, about 10^-3
> kT/ns/dof). I realize that this is still not very large at all, but do you
> have any reason to believe that pmemd.cuda with ionic liquids is going to
> do worse in NVE?
> If you've got a spare minute at some point for an opinion on this, it
> would be appreciated!
> Thanks,
> Joe
> --
> Joseph Baker, PhD
> Assistant Professor
> Department of Chemistry
> C101 Science Complex
> The College of New Jersey
> Ewing, NJ 08628
> Phone:(609) 771-3173
> Web:http://bakerj.pages.tcnj.edu/
> <https://sites.google.com/site/bakercompchemlab/>
AMBER mailing list
Received on Thu Jan 22 2015 - 11:30:03 PST
Custom Search