On 05/18/2013 11:57 AM US Central Time, Jed Brown wrote:

Pavan Balaji <[email protected]> writes:

...

Can you point me to the appropriate location in the standard for this? All I can find is:

"The comm_delete_attr_fn function is invoked when a communicator is deleted by MPI_COMM_FREE or when a call is made explicitly to MPI_COMM_DELETE_ATTR."

I don't think it explicitly says that the first argument to delete_fn is valid, but the standard doesn't make a habit of reiterating when arguments are valid. What would be the purpose of passing a comm argument that is going to be (sometimes?) invalid?

If you think it *should* be acceptable for an MPI implementation to pass an invalid comm to delete_fn, such that compliant user code *must* ignore that argument, then perhaps we should move the discussion to mpi-forum and clarify the wording in the standard.

I'm not saying that it should be acceptable for an MPI implementation to pass an invalid comm. I'm just saying that the current standard doesn't restrict it. And I agree the standard needs to fix this. If you create a ticket, someone from Argonne can present it at the Forum for feedback next month. In the short term, as a principle of least surprise, MS-MPI should fix this, nevertheless. Fab: hopefully you can make a case to your management that this is useful in the "cloud", so you should be allowed to spend time fixing this :-). Jed: can you create a (!STRICT_MPI) only version of this test and send us a patch? I'll include it into our test suite. -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji