The second option will be slower. Apart from the fact that the MPI implementation needs to manage a large number of communicators, in most implementations this option will disable shared memory communication between the slave and master. -- Pavan

On May 11, 2015, at 7:17 AM, Roy, Hirak <[email protected]> wrote:

Hi MPICH team,

I have an application which is mostly using master-slave kind of architecture. One master and <n> number of slaves. The slaves do not communicate between themselves and only sends/receives messages to/from master.

I have two ways I can setup the application 1> Static : Master and slaves are all in the same MPI World 2> Dynamic : Each slave is connected to master using a MPI_Comm and the communicator is created using MPI_Comm_connect/MPI_Comm_accept. Essentially master has <n> communicators.

The way master works is : it uses MPI_WaitAny kind of method to check messages from slave.

Could you please tell me whether <2> can be significantly slow with respect to <1> if all the slaves send lots of small messages? Is there some kind of buffering overhead in <2> for <n> communicators and which can be solved using some configuration.

Thanks, Hirak _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss