Notes 2020 11 05

Notes 2020-11-05

Participants

• Marc-Andre Hermanns
• Wesley Bland
• Atul Kulkarni
• Bill Williams
• Martin Schulz
• Michael Knobloch
• Joachim Protze
• Bengisu Elis

Topics

• Stackwalking

  • Expensive to do "every time"
  • Only walk the stack when needed (on-demand)
  • Matter of "quality of implementation"

• MPI should provide the return address

• Using the context object

• What kind of performance expectations do we have

  • Current prototype is OK
  • But on-demand stackwalking is extremely expensive (better ways need to be explored)

• No tool registration after MPI_(Session_)Init

  • Needs a branch in all functions that may be callable before MPI_Init (e.g., Info calls)
  • What about MPI_T?
    • MPI_T_init_thread will start up QMPI.
  • Tools need to be aware that in the sessions model multiple calls can initialize MPI
    • The first time any call with MPI_ prefix
  • Tools cannot do anything but registration inside of bootstrap mechanisms
    • Order of these may compiler/OS dependent (and MPI might not be initialized yet)
      • External library that performs MPI_Session_init before main may break other tools that want to register
        • This is a problem that already exists with PMPI
        • Solution could be a general concept (outside of tools)

• Maybe three-phase system is better

  1. "I am a tool, here is my name, and callbacks for interception setup and initialization"
  2. "Setup all the interceptions"
  3. "Now it's safe to make MPI calls"