Notes 2016 03 01

These are the "cleaned up" notes from the Tools WG meet discussing the design for an event interface in MPI_T.

• Buffer context
  • Tool can create multiple buffers
  • Can register different events for different buffer contexts
  • Each buffer context can have different behavior
    • Polling, callback, immediate notification, buffered events bulk notification
  • When does notification occur?
    • Best effort immediately
    • Buffer until N events (about N to prevent buffer overflow, or allocate bigger buffer than N)
      • MPI provides buffer, tool can potentially determine size
      • Tool does not read buffer directly, uses a decoder function
• MPI_T_EVENTS
  • Name
  • Immediate //can notification be iimmediate or not (or "best effort")
  • Timer - not all hardware can figure out wtime(), so it may be hardware-specific timer
    • How do we convert hardware time to wtime?
    • Valid or not valid (maybe not valid before MPI_Init)
• In MPI_T way, we will
  • Query for number of events
  • Iterate over events, get info
  • Events can be classified into categories
• Event struct{
  • Event_id
  • Sequence_id //do we need this? Ask Marc-Andre
  • Timestamp //could be hardware time, could be _T_Wtime
  • Thread_id //what does this mean? Tool provides function "whoami()" returns void * to tid TODO
  • MPI Object bound
  • Void * data? //who allocates this memory? Where is it defined? Who frees it?

    •     //could be an MPI Datatype
      

  • }
• What can you do in a callback?
  • "async signal safe" - who defines this? POSIX? We need to look it up
    • OMPT uses this
    • What about MS Windows. They don't have signals
      • Anh says we can probably use this language, because there are signals (albeit few), the language implies
• Time conversion idea
  • How to convert hardware time stamp into Wtime?
  • There could be an MPIT call that returns now() in hardware time along with current Wtime, so tool can convert
• Can events be dropped?
  • Yes, maybe buffer is full, or maybe it just can't be kept for some reason by implementation
  • What if an MPI Ojbect is freed, is the buffer still valid?
    • Yes, because MPI implementaiton can keep it around until it is no longer needed
    • You may want to decode events in that buffer even after it is freed

These are pictures of the white board from the meeting. Picture 1 Picture 2