[Mpi3-hybridpm] new text for MPI_INIT/MPI_FINALIZE
balaji at mcs.anl.gov
Fri Feb 3 08:51:51 CST 2012
On 02/02/2012 09:57 AM, Douglas Miller wrote:
> I still disagree with the idea that MPI_INIT/FINALIZE is called
> per-MPI-Process. I know that this ticket does not deal with multiple MPI
> processes, but unless this sort of wording is stated correctly, ticket
> #310 won't work.
We should really discuss this in the telecon, rather than on mailing
lists, but each MPI process calling INIT/FINALIZE should still work with
ticket #310. Yes, the implementation will have to do some extra work,
but maybe you can clarify why you think this is not implementable?
> The big brick wall that ticket #310 will run into if we promote the idea
> that "every MPI process has its own MPI_INIT/FINALIZE" is that if there
> are multiple MPI processes then there is no clear way to initialize them
> since there is no argument to MPI_INIT/FINALIZE that specifies the MPI
> Process and there is no way to know what MPI process a thread might
My understanding is that a thread does not represent any MPI process
before it calls MPI_INIT. After it calls MPI_INIT, it represents an MPI
process -- at this point, my expectation is that the MPI library will
add this information to the thread private storage.
More information about the mpiwg-hybridpm