Threads WG Meeting 9 10 2019

Participants: Manju, Naveen, Min

Agenda

• Read MPI and OpenSHMEM Interoperability proposal

Note

• Decided to read the simplified version. PR: https://github.com/minsii/openshmem-specification/pull/2
• Comments and discussion:
  • [1 MPI Interoperability]
    • The last sentence in paragraph 2 "In such a case, internal interference may occur." is ambiguous. Try to delete or explain it.
  • [1.2 Dynamic Process Creation]
    • Try to make stronger advice for users. "users should avoid using OpenSHMEM and MPI dynamic process model in the same program." -> "users should not use..."
  • [1.4 Mapping Process Identification Numbers]
    • The example uses collective routines. Change is needed once the teams proposal is added.
    • It might be useful to define a function for OpenSHMEM/MPI_COMM_WORLD id mapping, thus the runtime can simply return ids (e.g., read from process manager) if it supports both models.
      • Concerns: for standalone OpenSHMEM implementations, it has to call MPI functions inside the helper function. Also, CrayMPI does process rank reordering. So, it is better to leave it as a vendor extension or a helper function created by the user program.
  • [1.5 RMA Programming Models]
    • Try to make stronger requirement for users. E.g., user should not use both RMA models in the same program.
  • [Other]
    • It might be useful to define an interface that hints runtime that the program will use both MPI and OpenSHMEM, or only OpenSHMEM. E.g., an environment variable. Min will think about the possible implementation and benefit.
• Will continue the reading for Interoperability proposal next week