Yeah I think that makes sense. It was the comment about mpi_comm_split in the MPI_CART_SUB paragraph in the current standard that was throwing me. I agree that since a process cannot contribute a value (like the 'color' in mpi_comm_split) then there is nothing more we can say. So it just falls into the class of other topology creation operations that may carry over failed processes to the derived communicators. I'll remove the sentence on p561 line 21-22 in the next pass. Thanks, Josh On Tue, Nov 8, 2011 at 3:47 PM, Darius Buntinas <[email protected]> wrote:

In MPI_Cart_sub, processes don't "contribute" to the operation, so the current sentence on p561 doesn't really make sense. �However, I don't think what we did for MPI_Comm_split will work here either. �I think that MPI_Cart_sub should create the communicators as described in Section 7.5.7, which may result in failed processes being included in a new communicator.

E.g., in Example 7.8 on p291, if the process at coordinate (0,2,1) in the topology failed, then if remain_dims = (true, false, true), then the third communicator (i.e., the one containing processes at coordinates (A,2,B), A = 0..1, B = 0..3) would have a failed process (at coordinates (0,1) ).

Maybe the example makes things more confusing. �I think MPI_CART_SUB is already covered in by the sentence starting on line 7 page 561, so we can probably take the MPI_CART_SUB sentence out.

-d

On Nov 7, 2011, at 12:58 PM, Josh Hursey wrote:

...

* I would like some more folks to look over the process topologies chapter. In particular I think we need to fix the wording for MPI_CART_SUB - maybe to better match the wording for mpi_comm_split.

_______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft

-- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey