Notes 2016 12 05

• What do we want in MPIR2?

• Want to support multiple DLLs

  • Debuggers are not compiled/linked against multiple MPIs so need to load appropriate library dynamically

• Want an API based mechanism not variable based

• Want to support dynamic processes, not static process table

  • Scalable process table
  • Lazy loading of process table?

• Defined point for detach (finalize)

• MPI thread based - want to support that

• Mpiexec mechanism, co launch MPI applicationswith debugger tools

  • Came with newer machines,
  • 6.1 in MPIR job launch
  • Still need to talk to the starter process

• Debugger needs indication that MPI is ready, all processes are started

  • Now it is MPIR_Breakpoint
  • What about a debugger sets a callback,
  • Or an MPI request that you can wait for (That is the MPI way :-) )

• Define the interface such that you don't have to be a debugger. e.g. spindle

• Multiple tools attach at a time

  • How can we do the dll thing without ptrace attach on starter process

• How do we get around the dll problem?

  • That requires access to the process memory space
  • What about making the interface to MPIR an ABI
  • All of MPIR is already an ABI so it's not breaking from tradition
  • Would only be in the starter process?
  • Well the rest of it is ABI (MSQ) too so …

• Process table

  • Need a notification of whenever it is ready (the process table)
  • And when it has changed due to dynamic processes

• OMPD use as a baseline

• Message queue interface, overly specific on what is returned for routines

  • Jim Dinan said there was a problem, prescribes a queue implementation
  • Slows up OpenMPI need to not dictate implementation
  • We should remember this when redesigning MQS

• Sessions proposal how it interacts

  • Multiple comm worlds? Does it care?
  • How can a debugger ask about a subset?
  • Does the resource manager know about all sessions ?
    • In openMPI apparently so, but maybe not all MPI implemetnations
  • Need a way to query for information, how processes are being logically grouped