I assume you mean something like mca_coll_foo_init_query() for your initialization function. And I'm guessing you're exchanging some sort of address information for your network here?

What I actually did in my collective component was use PML's modex (module exchange) facility, defined in ompi/mca/pml/base/pml_base_module_exchange.h. I think all of the BTLs use this, so look there if you would like to see examples.

The fact that this is associated with the PML is a bit of a historical artifact; it is perfectly reasonable to use the modex stuff from a coll component.

Another facility you might be interested in is ORTE's DSS (data pack/unpack) and RML (remote messaging). The BTLs tend to use this for out-of-band communication necessary for lazy connection establishment. The headers you want are orte/dss/dss.h and orte/mca/rml/rml.h. Again, many of the BTLs use this, so look there for examples.

Hope this helps!

Andrew

Georg Wassen wrote:
Hello,

I'm working on a collective component and need point-to-point communication during module-initialization. As BTL is initialized prior to the collectives, I tried to use send and recv like MPI_Send/_Recv do:

err = MCA_PML_CALL(send(buf, size, MPI_CHAR, to_id,
        COLL_SCI_TAG,  MCA_PML_BASE_SEND_STANDARD, comm));

This works for a total of 32 Bytes (8 times 4 Bytes) to each receiver. If I increase the number of processes, recv just doesn't return after 8 recv-calls. (Further, this communication is extremely slow.)

When I try to use MPI_Send/_Recv themselves, an error "MPI not yet initialized" occurs.

Long story short: is there a way to communicate during mca_coll_*_module_init between different processes? (I don't want to use TCP/IP-sockets while Open MPI should be able to do this more portable.)

Thanks for your help,
Georg Wassen.
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel

Reply via email to