mpi-forum/mpi-issues

Use of 'starting' for processes needs disabiguation

mahermanns opened this issue · 4 comments

Problem

Starting is now a defined term for operations in context of procedure calls. The dynamic chapter uses the term in context of mpiexec and mpirun. We should clean up this ambiguous use of start.

Proposal

Replace the language start by launch where approriate

Changes to the Text

Impact on Implementations

None.

Impact on Users

Clearer use of language in the standard.

References and Pull Requests

Discussion on launch vs. spawn:

  • Spawn adheres to the fact that 'spawned' processes behave slightly different (or have different characteristics)
  • Both words describe how to add MPI processes to an job
  • Processes can be launched on different nodes, but spawned only on the local node (Updated due to discussion)
  • At some point the user can/should not distinguish between how the MPI processes came into being
  • Launching and spawning have an impact on the inter- vs. intra-communicator (can be overcome by merging to an intra-communicator)

@wgropp Could be create some for of a index reference for startup mechanism (see. launch mechanism)

Because cleaning up the index would loose the existing index entries for startup mechanism or startup

  • Processes can be launched on different nodes, but spawned only on the local node

I don't think this is true. One can actually specify the host as info key on MPI_COMM_SPAWN(_MULTIPLE) to specify a host to spawn to (See Section 11.8.4). An on managed resources (e.g. via Slurm) spawning usually works across the available resource allocation (i.e., nodes).

Furthermore, I could not find a place in the standard that enforces spawning to be limited the local node. If there is such a text somewhere in the standard I would appreciate if someone could point me to it.

On the index, you can add a cross reference with the "|" character. Here's an example from the makeindex manual:
\index{at!bat|see{bat, at}} which gives bat, see bat, at (here for a second level index "bat" on "at". I'll have to check that the index processing works for this, but this shouldn't be a problem.