Re: [AMBER] T-REMD Reproducibility - Cluster v.s Desktop

From: Carlos Simmerling <carlos.simmerling.gmail.com>
Date: Fri, 25 Sep 2015 13:04:13 -0400

I don't know of any way to get them to exactly match unless it's the same
cpu and same compiler, etc . short runs will always diverge. it's not clear
what matching would mean for short runs, since it isn't reliable, just
reproducible.

On Fri, Sep 25, 2015 at 12:58 PM, Thiago Lipinski Paes <thiagopaes.gmail.com
> wrote:

> On Fri, Sep 25, 2015 at 1:17 PM, Carlos Simmerling <
> carlos.simmerling.gmail.com> wrote:
>
> > what do you mean by the results not being the same? the initial MD
> differ,
> > or the overall results (populations etc)?
> > using different compiler or CPU (or # threads for MPI) can cause
> divergence
> > in the MD run fairly quickly. results must be well converged to be
> reliable
> > across different machines (or even across runs on the same machine with
> > different random seeds).
> >
>
> Thanks for the reply.
>
> The overall results differ.
>
> Actually I was testing a small 1ns simulation with 1000 exchanges (each
> 1ps) expecting exactly the same results (energy/rmsd/exchanges), but it not
> seems to be the case.
>
> The fact is that I'm really going to need my simulations to be small as 1ns
> (they are part of a protocol that I intent to use for helping PSP). At
> least on the beginning of the protocol, I'm more interested on sampling
> than convergence.
>
> The initial MD results do not differ... but over time I got different
> sampling.
> Could I use the same # of MPI threads to have the same sampling (desktop vs
> cluster)?
>
>
> Looking at .mdouts I found that MD values are quite similar (at the
> beginning) and that's why I thought the MC metropolis calculation was the
> problem.
>
> Could I monitor the exchange ratios, average energies, # of tunneling
> events and canonical probability distribution of energy to have sure that
> my results are reproducible with an acceptable deviation?
>
>
> --
>
> Thiago Lipinski Paes
> Computer Engineer - Computer Science MSc - Computer Science PhD Candidate
> Laboratório de Bioinformática, Modelagem e Simulação de Biossistemas
> (LABIO)
> Pontifícia Universidade Católica do Rio Grande do Sul (PUCRS)
> Av. Ipiranga, 6681 - Partenon - Porto Alegre/RS - CEP: 90619-900
> Tel: +55 51 33203611 ext.8602
> _______________________________________________
> 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
Received on Fri Sep 25 2015 - 10:30:02 PDT
Custom Search