Worlds and Universes: MPI_FINALIZE

[cavemanloverboy]

The Universe type calls MPI_FINALIZE() upon being dropped, but it is possible to hold a universe.world() after dropping a Universe and the internode communications methods are impl'd on the world. This is a problem.

I recently wrote a program that dropped a universe (unintentionally) and then tried to used the world but ran into issues since MPI_FINALIZE had been called. I wonder if it is possible to enforce that a universe outlive world through the lifetime system by, for example, requiring that world hold some reference to a dummy field on a universe.

[cavemanloverboy]

Another solution would be to have some unit struct type Finalize which is held in an Arc<T> and held by the Universe and SystemCommunicator::world(). You could have the special Drop impl for Finalize which calls MPI_FINALIZE. Then, the final drop of Arc<Finalize> would call this whether it is Universe or the SystemCommunicator::world().

[jedbrown]

MPI doesn't have a concept of a universe as an object, and even the size (MPI_UNIVERSE_SIZE) is just an optional attribute (need not be present) on MPI_COMM_WORLD. I think we should handle this with PhantomData lifetimes so that communicators can't outlive the universe. Could you make a pull request that adds a reproducer in which your communicator outlives universe (and thus errors)?

[cavemanloverboy]

Yea I’ll submit one later this week. I have a busy schedule these next three days.

[jedbrown]

Cool. There's also relevant discussion in #32.

[SolidTux]

Are there any updates on this?