Threads WG Meeting 10 30 2018 - openshmem-org/specification GitHub Wiki

(Thanks Anshuman for the notes)

Agenda

  • OpenSHMEM and MPI interoperability (Min)

Attendees

  • Manjunath Gorentla (Mellanox)
  • Naveen (Cray)
  • Anshuman (NVIDIA)
  • Min Si (ANL)
  • Md Rahman (Intel)
  • Dave (Intel)

Notes

  • Interop slides presented by Min
  • New API routines shmem_initialized and shmem_finalized
  • Example shows when shmem_finalized is used : when MPI_Finalize causes SHMEM calls to be no longer available, shmem_finalized should return 1 in that case
  • MPI_finalized, _initialized already defined in MPI spec
  • Manju : shmem_initialized thread safety, multiple threads may call shmem_init
  • rank/PE mapping : no mapping provided, user should query rank and PE separately for portable program
  • Benefit of providing mapping is limited because needs all implementations to support which seems difficult
  • Progress polling : required
  • New APIs : initialized, finalized, query_interop (shmemx?)
  • Manju : if property is implementation specific, then shmemx_ makes sense But if properties are unified across implementations, then shmem_
  • Manju : concern over progress polling So can query_interop say no interop
  • Manju’s suggestion : shmemx_ can be added in appendix