Re: [AMBER] Fwd: Periodicity in MD simulations

From: Oliver Kuhn <oak.amber.googlemail.com>
Date: Thu, 21 Jul 2011 10:56:18 +0200

Dear Ignacio and others,

I have found out that the oscillation artifacts stem from using

mpirun -np 4 pmemd -O ...

instead of

mpirun -np 4 pmemd.MPI -O ...

The first command creates four equal threads all writing to the same
outputfiles I think, and I wonder why the results are not totally messed up,
but MMPBSA calculation works fine on this data (giving comparable results to
uncorrupted data).

The newer (purple) trajectories have been run using only a single processor
and do not show oscillations. (Other simulation sets calculated on GPU also
do not show any correlations or oscillations)

I have no idea of the details so far. I'll try to find out.


Dear Amber Developers,

could you please assert that it is not possible to run 'mpirun -np >1
pmemd'.
This is a really beasty thing.

Greets,
Oliver




2011/7/20 Ignacio J. General <ijgeneral.gmail.com>

> Dear Oliver,
>
> I have read your series of questions and self-answers :-), but I am
> still not convinced that what you see is an artifact. In your first
> message, you said that the simulations in your plot were all started
> from the SAME initial structure. If that is the case, then I strongly
> believe your plot is what you should expect. To decide on this point,
> I think what you should do is to plot the energies when you add
> simulations started from DIFFERENT initial structures. Do you have, or
> can make such a plot?
>
> Ignacio
>
> On Wed, Jul 20, 2011 at 3:51 AM, Oliver Kuhn <oak.amber.googlemail.com>
> wrote:
> > I'll stop talking to myself now on this list ;-), please do not think
> about
> > it.
> > It must be something sander related, I can specify now.
> > I'll come back when I have some more information and some more specific
> > questions.
> >
> > ca
> > Olli
> >
> > ---------- Forwarded message ----------
> > From: Oliver Kuhn <oak.amber.googlemail.com>
> > Date: 2011/7/20
> > Subject: Fwd: [AMBER] Periodicity in MD simulations
> > To: AMBER Mailing List <amber.ambermd.org>
> >
> >
> > Hi again,
> > OK. Now I have found a difference. New simulations are calculated using
> > Amber11, old ones with Amber10.
> > I diffed the input files, its all the same but different random seeds.
> >
> > Are there any significant algorithmic differences in pmemd.MPI for
> Amber10
> > to 11 that could lead to such an effect (attached file)?
> > I'm sorry for this unspecific question, but I do not know where the
> > correlations come from, so I can't specifiy any further.
> >
> > Greets,
> > Olli
> >
> >
> >
> > Sorry, new simulations were not on GPU but also CPU. Maybe I can find
> > differences in the setup.
> >
> > Olli
> >
> >
> >
> >
> > Hi Daniel,
> > thanks for the for the 'sieve' keyword explanation. Sounds reasonable.
> > And yes, Carlos is right, in the end, I did not pose my question right.
> > When I read these words "problem of periodicity", I wondered if there is
> a
> > known technical issue bringing up periodicity in MD simulations, but your
> > explanation asures me that there is something else meant by that, closely
> > related to something real.
> >
> > Regarding Ignacios explanation,
> > I first thought of some global oscillation of the complex molecule and
> > looked if I can see that in the first PCA mode and I also tried
> clustering
> > high and low energy snapshots in the hope to find some hint for that but
> > could not find anything.
> >
> > In the end I had and have the impression that it is something technical.
> > Yesterday I was was very surprised when I had added 10 new 'independent'
> > simulations to the 21 I had, and this time, these are completely
> > uncorrelated - and there is a technical difference, this time I had
> > calculated on GPUs (I used the SPDP).
> >
> >
> > At the moment, I have no more concrete question on that topic, I'm
> looking
> > closer to my data. Thank you for any help.
> >
> > Regards,
> > Oliver
> >
> > I attach one new plot with CPU simulations in "green" and GPU simulations
> in
> > "purple" (this time BOND energies).
> >
> >
> > 2011/7/19 Daniel Roe <daniel.r.roe.gmail.com>
> >
> >> Hi,
> >>
> >> I assume you're talking about the "cluster" command in ptraj?
> >>
> >> The "problem of periodicity" in clustering with the "sieve" keyword
> >> can be illustrated with this simple example. Suppose you have a
> >> trajectory in which you have a protein structure that has some sort of
> >> periodic motion to it, let's say a domain motion which provides access
> >> to an active site. For the purposes of this example lets say the
> >> "open" conformation occurs about every 10 steps, and you choose to
> >> cluster with a sieve of 10 - this means that on the first pass you are
> >> effectively clustering based on a trajectory composed of entirely
> >> "open" conformations, which is probably not desirable. However, if you
> >> sieve randomly you are probably going to get a better representation
> >> of the overall trajectory.
> >>
> >> Clustering experts feel free to add to this or correct me but this is
> >> how I understand "sieve" and periodicity.
> >>
> >> -Dan
> >>
> >> On Mon, Jul 18, 2011 at 2:17 PM, Oliver Kuhn <oak.amber.googlemail.com>
> >> wrote:
> >> > Dear Amber Users and Developers,
> >> >
> >> > I have read in the AmberTools Manual that the 'sieve' command can be
> set
> >> to
> >> > random to avoid THE POTENTIAL PROBLEM OF PERIODICITY.
> >> > What exactly is meant by that?
> >> > I am asking because in the course of MMPBSA calculations I have done
> >> several
> >> > 'independent' simulations using different intial random seeds.
> >> > Sometimes simulations are totally uncorrelated, but sometimes they are
> >> all
> >> > correlated together (for several nanoseconds). This can be seen in
> >> internal
> >> > energies, e.g. angles (attached - 21 * 2ns simulation time
> HIV-Protease
> >> drug
> >> > complex, blockwise averaged).
> >> > I have read about synchronization effects of Langevin thermostat, but
> >> that
> >> > is not what I observe. In my simulations I have the SAME starting
> >> structure
> >> > and different seeds, not different structures and same seeds.
> >> > I'm very puzzled about these correlations and have no idea what these
> >> are.
> >> > What I know about the nature of correlations: Either simulations are
> >> > correlated with pearson coefficient around 0.4 or uncorrelated with
> >> around
> >> > 0.0 and nothing inbetween.
> >> > My random seeds range from 01 02 and so on up to 21.
> >> >
> >> > Maybe someone can give me a hint for paper.
> >> >
> >> > Regards,
> >> > Oliver
> >> >
> >> > --
> >> > Oliver Kuhn, Department of Bioinformatics,
> >> > Center for Medical Biotechnology, University of Duisburg-Essen,
> >> > Universitätsstr. 1-5, 45141 Essen, Germany
> >> > phone +49 201 183-3121, oliver.kuhn.uni-due.de
> >> >
> >> > _______________________________________________
> >> > AMBER mailing list
> >> > AMBER.ambermd.org
> >> > http://lists.ambermd.org/mailman/listinfo/amber
> >> >
> >> >
> >>
> >> _______________________________________________
> >> AMBER mailing list
> >> AMBER.ambermd.org
> >> http://lists.ambermd.org/mailman/listinfo/amber
> >>
> >
> >
> >
> > --
> > Oliver Kuhn, Department of Bioinformatics,
> > Center for Medical Biotechnology, University of Duisburg-Essen,
> > Universitätsstr. 1-5, 45141 Essen, Germany
> > phone +49 201 183-3121, oliver.kuhn.uni-due.de
> >
> >
> >
> > --
> > Oliver Kuhn, Department of Bioinformatics,
> > Center for Medical Biotechnology, University of Duisburg-Essen,
> > Universitätsstr. 1-5, 45141 Essen, Germany
> > phone +49 201 183-3121, oliver.kuhn.uni-due.de
> >
> >
> >
> > --
> > Oliver Kuhn, Department of Bioinformatics,
> > Center for Medical Biotechnology, University of Duisburg-Essen,
> > Universitätsstr. 1-5, 45141 Essen, Germany
> > phone +49 201 183-3121, oliver.kuhn.uni-due.de
> >
> >
> >
> > --
> > Oliver Kuhn, Department of Bioinformatics,
> > Center for Medical Biotechnology, University of Duisburg-Essen,
> > Universitätsstr. 1-5, 45141 Essen, Germany
> > phone +49 201 183-3121, oliver.kuhn.uni-due.de
> >
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber
> >
> >
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>



-- 
Oliver Kuhn, Department of Bioinformatics,
Center for Medical Biotechnology, University of Duisburg-Essen,
Universitätsstr. 1-5, 45141 Essen, Germany
phone +49 201 183-3121, oliver.kuhn.uni-due.de
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jul 21 2011 - 02:00:03 PDT
Custom Search