Notes 2016 05 05

Participants

• Jean-Baptiste Besnard
• Ralph Castain
• John DelSignore
• Daniel Everett
• Kathryn Mohror
• Bob Munch
• Soren Rasmussen
• Jeff Squyres
• Anh Vo

Notes

PMIX Debugger Launch

• Our plan from last time was to collect requirements from the resource manager for launching debuggers
  • All of us forgot! We all promised to do it for next time.
• For the wiki page: https://github.com/mpiwg-tools/tools-issues/wiki/Process-Acquisition-Interface
• What do we want from resource manager?
  • Ability of an application to ask for a debugger to be launched against it
  • Ability for a debugger to attach to a job
  • Do we need ability of a debugger to be co-launched with an application?
    • Not sure, that seems straightforward and not really part of PMIX

MPIR 2 New Interface

• Anh talked from his slide deck
• New function MPIR_Finalize
  • Want to be able to deallocate resources, also could unset "being debugged" state of application so MPI implementation can stop tracking message queue information, etc
• What does this mean in light of sessions where you could initialize and finalize repeatedly?
  • There might be a point where a debugger is attached to an MPI process but MPI is not initialized anymore
  • Don't think debugger cares
  • As long as when a session starts it does the "right things" the tools should be okay
  • Don't know what to do about the dynamic case where new processes come into play
  • This is especially true for the MQD interface, but for MPIR it shouldn't matter
  • The debugger doesn't care if a process calls MPI Init/Finalize, just only cares if the process is in the proctable
• Don't process specific callbacks need a context? How does the debugger specify which process it is talking about?
  • Anh made notes on this -- we'll talk about it next time
  • What about scalability? Do we have to have a context for every MPI process?
    • No, we're talking about starter processes, not MPI processes
• Get world size query
  • What does "world" mean in context of sessions
  • Jeff says current understanding is that there is one world that you (as MPI process) belong to
  • What about processes that join or leave worlds?
  • Gets tricky. In connect/accept model, who is the starter process?
  • Suggest to change this query to get proctable size to disassociate it from the concept of "world" which is too confusing
• What are the processes in a threaded MPI (like MPC)?
  • Mapping multiple MPI processes onto a single OS process
• For next time, we need to talk more about this threading issue
  • We'll look to see what was done for OMPD, maybe we can copy it
• At face to face, we'll talk more about the consequences of Sessions for MPIR and MQD