Re: Semantics of MPI Cancel?
W. Saphir (wcs@nersc.gov)
Mon, 10 Mar 1997 08:57:32 -0800
In some sense, this discussion is moot. From a practical point
of view, MPI_Cancel (for sends) is not a part of MPI. Because it
is extremely difficult to implement, it doesn't not work in many
implementations and users can't count on it. Moreover, I have
yet to think of a good reason why an application might want
to cancel a send. When I teach classes in MPI, I tell users
not to use it. If a user were to come to me complaining that
MPI was broken because cancel doesn't work (fortunately none
have, possibly because they've heard my tutorials) I would
suggest they rewrite their program rather than complain to
the vendor. I think it's already unfortunate that vendors have
had to put significant resources into dealing with it.
The point of this is I think the Forum has more pressing things
to worry about at the moment than worrying about cancel (especially,
worrying about the absurd examples that were given). We
should devote our energy to careful readings of the MPI-2
document so that we can avoid this type of problem in the future.
Bill