[gmx-developers] pencil decomposition

hess at sbc.su.se hess at sbc.su.se
Thu Sep 2 09:53:37 CEST 2010


Hi,

The 2D PME decomposition has a large effect on performance on a large
number of nodes, on a small number the effect is small or slightly negative.
There is no effect on the results and mdrun chooses it automatically,
which can be done relatively independently of architecture, as Roland said.
So the "normal" user does not need to know about it.
It might be useful though to add one or two lines on this in the pdf manual.
I am at a workshop right now, so I don't know if I have time to do this
before the 4.5 manual release.

Berk


> On Wed, Sep 1, 2010 at 9:34 PM, Mark Abraham
> <mark.abraham at anu.edu.au>wrote:
>
>> Hi,
>>
>> I note the new pencil decomposition for PME with interest. Presumably
>> that's only for the 3DFFT part?
>
> yes pencil decomposition refers to the 2D decomposition for the FFT.
>
> Under what circumstances does it get used? Under what circumstances does
> the
>> old decomposition get used?
>
> it is always used if the number of pme nodes (all nodes if no separate pme
> nodes) is larger than the nodes in the X direction of the PP
> decomposition.
>
>
>> Does the user need to know anything to use it effectively on different
>> hardware? (e.g. BlueGene vs Intel+Infiniband)
>>
> We didn't notice any difference in optimal settings between Cray and
> Infiniband. We haven't tested BlueGene. Feedback is welcome.
>
> Should any of this get a mention in the manual?
>
> Not sure what would be helpful.
>
> Roland
>
>
>
>> --
>> gmx-developers mailing list
>> gmx-developers at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-developers
>> Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-developers-request at gromacs.org.
>>
>
>
>
> --
> ORNL/UT Center for Molecular Biophysics cmb.ornl.gov
> 865-241-1537, ORNL PO BOX 2008 MS6309
> --
> gmx-developers mailing list
> gmx-developers at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-developers
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-developers-request at gromacs.org.




More information about the gromacs.org_gmx-developers mailing list