Open tonyskjellum opened 6 years ago
This is the initial proposal text.
The plenary discussion of the Collective Communication chapter in Portland and, in particular, Ticket #78 led to the recommendation that MPI_COMM_IDISCONNECT be defined here. We agreed but decided to keep it also in Ticket #78, because it has broader usability than DPM. See discussion above.
Here is the current prospective chapter for reading in June, 2018 in Austin, subject to adding, amending, and improving between now and end of May, 2018.
Added PR for this Ticket.
This is the public copy.
We did the first reading attempt on June 13, 2018 at the Austin meeting. We found some small issues and quality improvements that are needed. We will make these updates and present a new reading in Barcelona.
Updated text in advance of San Jose meeting: mpi32-report-ticket81-24sep18.pdf
This may be mooted by "Sessions 2.0" proposal(s) we expect to be offered for MPI-5.0.
For now, we marked for MPI-5.0 work.
Problem
Dynamic process management functions are long running, and nonblocking versions are needed. We seek also to orthogonalize MPI-4 with nonblocking variants of blocking collective and long-running operations where appropriate.
Proposal
The nonblocking dynamic process management functions will be added to the standard:
This set of functionality is complementary to Tickets #78 and #82 .
Changes to the Text
These functions are to be added:
Note that Ticket #78 contains also has MPI_COMM_IDISCONNECT, which is of broader use than just dynamic process management and should be voted with those constructors/destructors. Whichever ticket passes first, the other will delete MPI_COMM_IDISCONNECT.
Because MPI_COMM_JOIN is on the path to deprecation via Ticket #13 , no MPI_COMM_IJOIN is proposed. If that changes, we would add MPI_COMM_IJOIN to this proposal in due course. =
Impact on Implementations
Describe changes that implementations will be required to make here.
Impact on Users
Describe the changes that will impact users here.
References
The associated PR is at: mpi-forum/mpi-standard#49