Notes 2015 12 03

Participants

• Jean-Baptiste Besnard
• John DelSignore
• Marc-Andre Hermanns
• Michael Knobloch
• Kathryn Mohror
• Soren Rasmussen
• Jeff Squyres
• Anh Vo

Notes

Forum preparation discussion

• Anything else we need to talk about before Forum readings?
• Anh will present the MPIR being_debugged ticket: F-Issue #5
• Jeff or Martin will present the MPI_T code bug ticket: F-Issue #12
• For folks that want to join via webex, Jeff will set it up and put it on the MPI Forum agenda page

MPIR line numbers, F-Issue #6

• Marc-Andre created a PDF using the new line-numbering package and put it in the ticket
• Notable differences between this and MPI Standard line numbering
  • No line numbering for figures, figure captions
    • People don't think this matters since figures have numbers already
  • Equations only get one line
    • Agree this doesn't matter either since derivations are broken into different eqns anyway
  • Best thing is that numbering actually paired with document text, so it is easy to know what line you are talking about instead of the current MPI Standard way
• Everyone agrees the new line numbering is best
• We'll prepare this for reading at next Forum meeting since it's too late for December

Old Trac tickets

• No one looked at old trac tickets to see what we need to still transfer over to github
• We'll put this on the agenda for next week's meeting

MPIR Topics, The Future

• Last meeting we had a good overview discussion of open topics for debuggers in MPI
  • Handle Interface, Support for non-traditional (e.g., threaded) MPI implementations, Type information interface for MPIR, MPIR "fixes"
• What of these do we want to focus on moving forward?
• Consensus is that they are all important
  • MPIR needs to be updated such that it is a first class API instead of simply hard-coded knowledge of symbols and structure/layout
  • MQD is a much better interface than MPIR. Should we start with it?
  • Perhaps focus on Handles Interface and figure out what we want with it, and use that model for revamping MPIR?
    • Concern is that focusing on a specific functionality may cause us to miss something important needed for the higher level MPIR
    • Also, Handles was based on MQD interface. MQD interface has some outstanding issues that need to be addressed
      • Generally, the basic process model is much better understood than it was 20 years ago
      • Need better definition of "image" which is based on the idea of identical, static binaries. Now we have highly dynamic, modular systems; heterogeneity.
      • Address space concepts for threads and processes are broken, need revamping
  • We need to redesign the basic models to get the terminology right
  • Handles Interface is still malleable, so whatever changes end up in MPIR-2 can go into it
• Consensus is to put focus into new MPIR interface and fix the real problems, instead of applying partial fixes
  • Want to include support for threaded MPIs, dynamic processes, future resource managers, etc
  • Let's try to get some others involved, Matt L, Dong A, Ralph C
  • We'll need to keep in mind a migration path for the new interface, i.e. support the "old way" and the "new way" concurrently

PERUSE

• We haven't talked about this in a while, don't want to drop it
• Have been waiting for Nathan H to join, since he's wanting to move the Open MPI support for PERUSE into whatever we come up with
• Nathan will be at the Forum meeting, so we'll put this on the agenda

Message Ordering/Piggybacking

• Brief discussion of tools problem with maintaining exact message ordering for replay with thread_multiple
• Is this a use case for revisiting the piggybacking problem?
• We'll come back to this