-
Notifications
You must be signed in to change notification settings - Fork 1
2023 01 09 webex
#01/09/23 webex notes
- Sessions and parallel debuggers
- MPI_Session_init – do we need something more precise than “lightweight”
- Proposal from Tools working group (contingent on Marc-Andre Hermmans availability)
- Future meeting agenda items
What is the state of MPIR - is it dead? Marc-Andre thinks the effort is stalled at this point. Marc-Andre thinks we should add a note about to the document. Howard will see about amending the document to state this only is applicable to the world process model.
MPI_Session_init needs to have local semantics - Dan does say there's lot of freedom though on how to implement.
Proposal from Tooks working group - Joachim Jenke drives. Resize mpi://world process set.
See google doc - https://docs.google.com/presentation/d/1jWTvik9NjY_xzCyomlzw4RCGkwBRzU4wc876kXJaxy0/edit?usp=sharing
Using sessions to redefine the number of processes available to the application.
Some use cases:
- tools that want to take some processes before world gets used
- system vendors who want to "reorder" processes for optimization reasons
- FT libraries that want to reserve some spare processes.
MPI_Comm_spawn is, as always, problematic.
Tool currently intercepts application calls involving MPI_COMM_WORLD and handles the shrinkage there.
Concern about missing out on collective optimizations for MPI_COMM_WORLD.
Dan walks through intercepting mpi://world process set name when used in MPI_Group_from_session_pset.
Also discuss if there's a way to catch MPI_COMM_WORLD arguments without having to do a check at every invocation of functions that take comm argument.
Dan brings up process set versioning the WG has been considering. Problems of consensus on versioned process set names to create groups and hence communicators. Considering ramifications of a process set version update mechanism. Do we need to revisit isolation of sessions?