Threads WG Meeting 9 03 2019

Participants: Akhil, Manju, Naveen, Min

Notes

• shmem_malloc_with_hint (PR: https://github.com/openshmem-org/specification/pull/259)

  • Q: Usage of HINT_LAT_MEM|HIGH_BW_MEM ?
    • PGAS application level, e.g., stencil
  • Q: Possible implementation ?
    • For example, internally allocate multiple symm heaps at shmem_init, each for different heap kind.
  • Benefit of memory policy: When threads call shmem_malloc, data locality can improve performance (e.g., each thread accesses memory on local socket).
  • Discussed James's questions
  • Reference implementation: OSHMEM
    • HINT_ATOMIC: allocate on NIC memory, fast write, wait(read) may be 1.5x slower

• teams and spaces

  • supported configuration:
    • One space per team
    • N teams per space (e.g., avoid multiple memory registration with different teams for the same buffer in multithreads collectives)
    • N spaces per team (e.g., support SHEAPs with different memory kinds)
  • Team based context: Can a PE use the context to transfer to an external PE of the team ? Current proposal allows it.
  • Implementation must support all three configurations based on the proposed APIs (i.e., shmem_create_mem_space, shmem_team_attach_space)

• Will schedule a WG meeting next week to discuss the MPI interoperability ticket