[gmx-users] mixed, double precision

[Mark Abraham]

Hi,

On Mon, Feb 20, 2017 at 12:58 PM Peter Stern <peter.stern at weizmann.ac.il>
wrote:

> I think so.  There is not much meaning to the input/output precision since
> no coordinates are more than single precision.


Technically, they'll be written in double precision by a double-precision
build. Whether there's meaningful information in those significant digits
is another matter.

  The purpose of double precision is to prevent computational errors
> arising from accumulated roundoff errors, for example, in normal mode
> analysis where computations involving large matrices are required.


Indeed.


>   But by the same token, such a series of runs may not be reproducible.
>

Note that by design, very few trajectories from mdrun are reproducible. The
sampling as you approach the ergodic limit should be reproducible, but the
additional computational cost imposed by running mdrun in its reproducible
mode is not worth it, particularly given other modelling limitations like
under sampling, fixed point-charges, or cut-off artificialities.


> Regards,
> Peter
>
> -----Original Message-----
> From: gromacs.org_gmx-users-bounces at maillist.sys.kth.se [mailto:
> gromacs.org_gmx-users-bounces at maillist.sys.kth.se] On Behalf Of Joel
> Markgren
> Sent: Monday, February 20, 2017 1:09 PM
> To: gromacs.org_gmx-users at maillist.sys.kth.se
> Subject: [gmx-users] mixed, double precision
>
> Gooday dear Gromacs users,
> I have access to a 5.1.4 version of gromacs compiled in double precision
> on a cluster.
>
> I equilibrated a couple of samples for NVT and NPT and used following
> commands: (the ... means that the rest of the input commands are not
> relevant for the question).
>
> gmx_mpi grompp ... -o nvt.tpr
>
> gmx_mpi_d mdrun -v -deffnm nvt
>
> gmx_mpi grompp ... -c nvt.gro  -o npt.tpr
>
> gmx_mpi_d mdrun -v -deffnm
>
> I realized after the runs that I forgot to specify that I wish to use the
> grompp commands in double precision (gmx_mpi_d grompp).
>
> Have I understood it correctly, that when grompp receives input from a run
> made in double precision, it will produce an output in mixed precision, if
> not specified to do it in double precision?


Yes, the precision of the build determines how mdrun operates.


> When the mdrun gets the input which is in mixed precision it will still
> make the run in double precision even if the input is in mixed.
>
> Have I lost some of the precision between the runs, with risks of energy
> drifts?
>

That's equivalent to a tiny "mistake" in one of the MD coordinate update
stages. That is going to be a much smaller perturbation than if (say) the
v-rescale thermostat chose a different set of random numbers. So I expect
no evidence of increased drift. By all means, re-run one of the studies
using both full double precision, and your inadvertent mixed protocol, to
see what you see. (But be aware that might expose that there is
undersampling...)

Also consider the thoughts at
http://biorxiv.org/content/early/2016/10/24/083055

Mark


> I see it's a bit messy but I hope you understand. Please let me know if
> you need more info or clarifications to answer my question.
>
> Thanks in advance for taking you time to answer my questions.
>
> Regards
>
> Joel Markgren
> PhD student in Biology
>
> Sveriges lantbruksuniversitet
> Swedish University of Agricultural Sciences
>
> Department of Plant Breeding
> PO Box 101, Sweden 23053 Lund
> Visiting address: Sundsvägen 10, 23053 Lund
> Mobile: +46 706101128 <070-610%2011%2028>
> joel.markgren at slu.se, www.slu.se<http://www.slu.se/> [
> https://internt.slu.se/PageFiles/81117/SLU_Logo_mail.png]
>
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.
>