Hi,
performance of MPI_Comm_spawn in the v1.8/v1.10 series is known to be poor,
especially compared to v1.6
generally speaking, I cannot recommend v1.6 since it is no more maintained.
that being said, if performance is critical, you might want to give it a
try.
I did not run any performance measurement with master, especially since we
moved to PMIx,
that might be worth a try too
Cheers,
Gilles
On Sunday, April 3, 2016, Emani, Murali <ema...@llnl.gov> wrote:
> Hi all,
>
> I am trying to evaluate the time taken for MPI_Comm_spawn operation in the
> latest version of OpenMPI. Here a parent communicator (all processes, not
> just the root) spawns one new child process (separate executable). The
> code I¹m executing is
>
> main(){
> {
> Š..
> // MPI initialization
> Š..
> start1 = MPI_Wtime();
> MPI_Comm_spawn(³./child", MPI_ARGV_NULL,1, MPI_INFO_NULL, 0,
> MPI_COMM_WORLD, &inter_communicator, MPI_ERRCODES_IGNORE );
> End = MPI_Wtime();
>
> printf(³ spawn time: %f², (end-start));
> MPI_Barrier(inter_communicator); // spawn is collective, but still want to
> ensure it using a barrier
> ..
> ..
> // MPI finalize
> }
>
>
> In child.c
> main(){
> {
> Š..
> // MPI initialization
> Š..
>
> MPI_Comm_get_parent(&parentcomm); // gets the inter-communicator
> MPI_Barrier(parentcomm);
> ..
> ..
> // MPI finalize
> }
>
> My observation is that the spawn time is very high (almost 80% of the
> total program execution time). It increases exponentially with the number
> of processes in the parent communicator. Is this method correct, and is
> the MPI_Comm_spawn operation expensive.
> I have also tried MPI_Comm_spawn_multiple but it still measures the same
> time.
>
> Could kindly someone guide me in this issue.
>
> Thanks,
> Murali
>
>
>
> _______________________________________________
> users mailing list
> us...@open-mpi.org <javascript:;>
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
> Link to this post:
> http://www.open-mpi.org/community/lists/users/2016/04/28871.php
>
