[gmx-developers] Question on P-LINCS

[Berk Hess]

On 02/07/2013 02:10 PM, Anton Feenstra wrote:
> On 05/02/13 17:40, risueno at physik.hu-berlin.de wrote:
>> Dear Gromacs developers
>>
>> In the P-LINCS paper (http://pubs.acs.org/doi/abs/10.1021/ct700200b) 
>> it is
>> stated that P-LINCS can constrain bond angles. However, in the Gromacs
>> manual (http://www.gromacs.org/Documentation/Manual, page 181), I find
>> that P-LINCS is used when there is domain decomposition, but that it
>> "should not be used with coupled angle constraints". So I wonder, is it
>> possible to use P-LINCS with bond angle constraints? If not, why?
>
> The key word here is 'coupled'. P-LINCS works with angle constraints, 
> but not with *coupled* angle constraints. The reason is that the 
> approximation for the matrix inversion assumes a sparse constraints 
> matrix. Having coupled angle constraints breaks this sparsity 
> assumption. (FYI, coupled angle constriants means one has, e.g. four 
> atoms with 5 constraints; three for the bonds (like, i-j, j-k and k-l) 
> and two for the angles (i-k and j-l)).
>
>> In addition, I would like to ask what is the maximum number of cores for
>> an efficient scalling in the parallel run of P-LINCS (perhaps there 
>> exists
>> some benchmark as a function of the total number of constraints).
>
> I don't know that. Berk is the (P-)LINCS expert.
>
This is usually not the right question to ask.
Constraints only contribute a small fraction of the total computational 
cost, so the answer depends a lot on how much other work you have.
Basically P-LINCS scales ok until your spatial domains get smaller than 
the extent of the connected constraints considered in P-LINCS,
which is anyhow a hard limit on the scaling.

Cheers,

Berk