[gmx-users] Is it possible to run implicit solvent simulations in parallel?

[Mark Abraham]

On 4/05/2011 11:23 PM, Justin A. Lemkul wrote:
>
>
> Ozlem Ulucan wrote:
>>
>> Dear Gromacs Users,
>>
>> I have been trying to simulate a protein in implicit solvent. When I 
>> used a single processor by setting -nt to 1 , I did not encounter any 
>> problem.  But when I tried to run the simulations using more than 1 
>> processor I get the following error.
>>
>> Fatal error:
>> Constraint dependencies further away than next-neighbor
>> in particle decomposition. Constraint between atoms 2177--2179 evaluated
>> on node 3 and 3, but atom 2177 has connections within 4 bonds 
>> (lincs_order)
>> of node 1, and atom 2179 has connections within 4 bonds of node 3.
>> Reduce the # nodes, lincs_order, or
>> try domain decomposition.
>>
>>  I  set the lincs_order parameter in .mdp file to different values. 
>> But it did not help.  I have some questions regarding the information 
>> above.

See comments about lincs_order in 7.3.18. Obviously, only smaller values 
of lincs_order can help (but if this is not obvious, please consider how 
obvious "it did not help" is :-))

>>  1) Is it possible to run implicit solvent simulations in parallel?
>>
>
> Yes.
>
>>  2) As far as I know gromacs uses domain decomposition as default. 
>> Why does in my simulations gromacs use the particle decomposition 
>> which I do not ask for.
>>
>
> Without seeing the exact commands you gave, there is no plausible 
> explanation. DD is used by default.

Not quite true, unfortunately. With the cutoffs set to zero, the use of 
the all-against-all GB loops is triggered, and that silently requires 
PD. It should write something to the log file.

>
> -Justin
>
>> Any suggestions are appreciated very much.
>>  I am ussing gromacs-4.5.4 with charmm force field and the OBC 
>> implicit solvent model. If you need further informations, probably a 
>> run input file, let me know.

A run input file would have helped me avoid guessing above about those 
cutoffs :-)

The real issue is that not all systems can be effectively parallelized 
by a given implementation. How many processors and atoms are we talking 
about? If there's not hundreds of atoms per processor, then parallelism 
is not going to be worthwhile.

Mark