[gmx-developers] About the multi-simulation

[Mark Abraham]

On 16/03/11, Yukun Wang  <wangykoo at gmail.com> wrote:
> Hi all:
>     I want to run a group of simulations which are weakly coupled, like REMD,but with different algorithm. I find there is a data structure called 'gmx_multisim_t' whose definition is below:
>                  typedef struct {
> 
>   int nsim;
>   int sim;
>   MPI_Group mpi_group_masters;
>   MPI_Comm mpi_comm_masters;
>   /* these buffers are used as destination buffers if MPI_IN_PLACE isn't
>      supported.*/
>   mpi_in_place_buf_t *mpb;
> 
> } gmx_multisim_t;
> 
> Then I have three questions:
> 1. If a command line is writen as for GROMACS 4.0 and a seires of input .tpr files is named as 'prefix_0.tpr, prefix_1.tpr ... prefix_9.tpr',
> 
> [...] $ mpirun -np 10 mdrun -s prefix_.tpr -multi (number of steps) (followed by output options)
> Are those simulation are mutually independent? ( Attention no -replex here !)
> 

Yes, in terms of the model physics they implement.


> 2. I know Gromacs from 4.0version have a ablility that each replica can be parallelly simulated. Then I want to know if they are intra-commucation for each replica and inter-communication between replicas.
> 

Yes. The latter is only used for synchronizing writing checkpoints, handling user signals and termination conditions.


> 
> 3. What is the means of  'nsim','sim','mpi_group_masters' and 'mpi_comm_masters'  in the above structure?
> 

The number of simulations, the index of this simulation in the group, and an MPI group and communicator for the set of nodes whose internal rank within each simulation is zero. These handle the inter-simulation communication. IIRC the group is actually redundant.

Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20110316/ee50d001/attachment.html>