Sessions: need to clarify limitations when mixing the sessions model with dynamic process model #434
Description
Problem
The only mechanism currently available for the spawned processes to do so is via MPI_COMM_GET_PARENT. However, this function can only be called after MPI is initialized using the world model approach in the spawnees. Thus, although from the spawner's perspective, processes can be spawned using communicators created using the world or sessions model, from the spawnee's perspective the world model must be used both to allow the spawner's MPI_COMM_SPAWN to complete as well as to obtain the intercommunicator returned from MPI_COMM_GET_PARENT.
Proposal
The existing text implies this asymmetry, especially in the description of MPI_COMM_GET_PARENT, but explicit text documenting this asymmetry would be a significant improvement. This clarifying text should be included in a MPI 4.1 standard.
Changes to the Text
Text needs to be added to the standard to clarify limitations in the usage of the dynamic process model - in particular use of MPI_COMM_SPAWN and MPI_COMM_SPAWN_MULTIPLE to start new processes, and the procedure available for the spawned processes to establish communication with the spawnee processes.
Impact on Implementations
None -- no additional or removed functionality.
Impact on Users
Additional clarity.
References and Pull Requests
Related to PR: https://github.com/mpi-forum/mpi-standard/pull/618