[mpich-discuss] Non-blocking collectives

[Jed Brown]

Rajeev Thakur <thakur at mcs.anl.gov> writes:

> I think the Wait is allowed to wait until others call Wait. 

Do you read this as allowing an implementation to defer completion until
a _matching_ MPI_Wait is called, or merely until the application
re-enters MPI somehow (sufficiently many times, or blocking)?  I had the
latter interpretation, which would allow non-blocking operations on
different communicators to be used without risking deadlock.

If the intent was really that the implementation can block until a
_matching_ MPI_Wait is called, then I think the standard should clarify
this point because it risks deadlock in a natural use case.

> See the Advice to Users on 197:27-29:
>
> "Users should be aware that implementations are allowed, but not
> required (with exception of MPI_IBARRIER), to synchronize processes
> during the completion of a nonblocking collective operation."
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20130629/7e2f166e/attachment.sig>