Notes 2020 10 01

MPI Forum Tools WG Meeting 2020-10-01

• Marc-Andre Hermanns
• Bengisu Elis
• Bill Williams
• Christ Chambreau
• Michael Knobloch
• Joachim Protze

Topics

• Open Questions on QMPI

  • Do we want to allow PMPI and QMPI Applications to be used at the same time?
    • We have to (for the implementation -> QMPI tools will take long to be released)
    • We will loose the return address to the application
      • Could we portably stack walk to this address?
        • Hard, because we don't know how many indirections we have
    • The PMPI symbols will act as shim layer to distinguish between calls to PMPI (from a PMPI tool) and the last layer in the QMPI stack
      • QMPI layer has context point pre-pended
        • Pre-pended because of MPI_Pcontrol()
          • Replace it with something more reasonable?
          • Does it make sense in a QMPI context?
  • How to handle the context?
    • The Tool context is a void *
    • Opaque handle?
      • Address of the calling function (at least)
      • If you have getter/setter functions it's easier to add more stuff later
      • How would a tool generated on of these?
  • When trying to call the next func
    • Ask: What is the next func
    • Ask: What is the next context
      • Maybe tool can query context itself by using its known tool registration ID
        • How would a wrapper (i.e. send) which instance (i.e. ID) it is called from
      • Function could be: "here is my tool ID, give me the next func and context"
    • Then: call func with context
    • Could we hide some of this manual query work behind some magic MPI call?

• Tool Registration

  • See Joachim's proposal

• Should we allow tools to be "bad citizens" by skipping other tools in the stack?

  • It could be a feature?
  • Where is the difference between skipping and branching?