Notes 2016 08 11

Participants

• Jean-Baptiste Besnard
• Ralph Castain
• John DelSignore
• Marc-Andre Hermanns
• Matthias Moeller
• Kathryn Mohror
• Joachim Protze
• Soren Rasmussen
• Anh Vo

Notes

MPI Handles

• Joachim Protze and Matthias Moeller are working on a project on MPI Handles
• Presented slides describing approach
• Going to create a specification for handle introspection from scratch and then compare that with our existing Handles Interface prototype
• Plan to use a PMPI wrapper tool as the initial implementation with the goal that this is an easy way to prototype a new interface
• Later, this could be implemented inside an MPI implementation
• Advantage of PMPI approach is that the library could be used on top of any MPI implementation, won't need the MPI to support it
  • Frees debugger of having any MPI specific knowledge
• Disadvantage is that dynamic linking with wrappers is not easily done on all platforms, e.g. Microsoft
  • Won't be able to get all handle information, e.g. context ids that are internal to implementation. Does that matter?
  • Would mean that you couldn't debug a performance tool layer, i.e. can't have 2 PMPI tools at the same time currently

PMIx vs MPIR - what do we want?

• Do we want to define a new MPIR interface if PMIx already provides everything we want?
• Is there an advantage to having a layer defined on top of PMIx?
• We need to get more community input on this, beyond just this group
• Should bring this up at the forum
• Also reach out to people by email
• We think our consensus though is that PMIx is for starting jobs, discovering what is running and where and attaching debuggers/tools, and that MPI interface is for MPI specific debugging, like MPI handles and message queues