[gmx-users] New to Gromacs -- looking for insights

[Mark Abraham]

Simone Pellegrini wrote:
> Thanks for the help.
> 
> On 12/09/2009 05:57 PM, Mark Abraham wrote:
>> Simone Pellegrini wrote:
>>> Dear all,
>>> I am researching optimization for MPI parallel programs.
>>> As I saw Gromacs seems to use MPI and my idea is to test some of the 
>>> program transformations I am aware of to improve the performance.
>>
>> It's already fairly good - be sure to read the GROMACS 4 paper. Up to 
>> about 64 processors on good network hardware, GROMACS runs are usually 
>> quite CPU-bound, so it is not a good candidate for improvement through 
>> any MPI optimization.
> 
> Probably it's working fine with the current MPI standards, but you know, 
> MPI is evolving and some new features will be introduced in MPI 3 and I 
> want to test if gromacs can take advantage of it. :)

OK, well that's slightly different from your original description :-)
Even so, you're going to have to get your hands very dirty understanding
the domain decomposition and PME sections (see paper I referred to
earlier and the section in the manual), which are the ones that do most
of the non-trivial MPI.

>> The MPMD communication pattern with PME at high parallelism is 
>> interesting, though, and might benefit from task layout that is 
>> cunning with respect to the network topology.
>>
>>> The main problem is how should I get started with Gromacs? :) Which 
>>> of the provided programs rely on MPI calls? and furthermore, where 
>>> can I find some significant input files?
>>
>> Only mdrun uses MPI. Look on the web or the GROMACS wiki for some 
>> pointers to tutorials to learn how to construct a workflow that will 
>> lead to a meaningful test.
> 
> thanks, I found a couple of mdb files on line and I am learning from it. 
> Is there any repository with come real world example using gromacs 4? I 
> found some benchmarks based on gromacs 3 and only half of them seem to 
> work with the new gromacs.

Sure, no there's no such repository of which I'm aware. Test sets and
benchmarks are all pretty out of date.

Perhaps it would be a useful thing to have the .mdp+.tpr files for the
GROMACS 4 paper available on the website.

Mark