Charm++ parallel programming system

[Phil Miller <mille121 AT illinois.edu>]

Hi Roberto,

The load balancing database is only recording point-to-point
communication, while collective communication (such as MPI_Bcast) is
excluded. I believe that is because the communication edges added by
collectives are not very informative from a load-balancing
perspective. Collective algorithms are distributing the communication
such that every rank will effectively participate to a similar extent,
or at least the underlying processors will.

I've just tested various programs that do point-to-point
communication, and they do record communication edges as expected.

Phil

On Fri, Aug 16, 2013 at 8:09 PM, Roberto de Quadros Gomes
<rqg.gomes AT gmail.com>
 wrote:
> Hi,
>
> I'm using a simple code written in AMPI for test and it have a loop like
> that:
> "
> i=100;
> do
>     {
>         i--;
>         res = MPI_Bcast(&sum, sizeof(int), MPI_INT, 0, MPI_COMM_WORLD);
>         MPI_Barrier(MPI_COMM_WORLD);
>         if( res != MPI_SUCCESS)
>         MPI_Finalize();
>        a--;
>        if( a <= 0 )
>        {
>          MPI_Barrier(MPI_COMM_WORLD);
>          MPI_Migrate();
>          a = 5;
>        }
>    }while( i > 0);
> "
> When LB strategy is called, LDStats stats->n_comm always = 0. I believed
> this value would be 5 or larger than 5 because the MPI_Bcast function is
> there.
> I had a problem like that when I tried read a pe_speed,  however I found out
> I needed use +LBTestPESpeed on command line. Is there any argument like this
> for LBStats to count each communication event?  Or am i doing confusing?
>
> _______________________________________________
> charm mailing list
> charm AT cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/charm
>