Open questions from 5/21 reading

[nspark]

Extracting items from #124, the 5/21 reading raised some specific questions about various teams features that need discussion by the Teams WG. These are listed here so as not to be lost amidst the sea of editorial comments from the reading.

• Team Management Routines
  • Team Objects and Multithreading Within a PE
  • Should only collective team routines require serialized accesses from multiple threads?
  • shmem_team_config_t
  • Can this structure have vendor-specific extension fields?
  • What is the configuration of SHMEM_TEAM_WORLD?
  • Suggestions re: default and user-provided values
    • Specify that the default value of num_contexts be implementation defined
    • Provide a value indicating the user is requesting an "unspecified" number of contexts
    • Explicitly specify whether shmem_team_get_config returns the "unspecified" sentinel value or the actual value for which the team was configured
  • Suggest: field that means team only supports collectives
  • shmem_team_destroy
  • Require destruction of contexts on a team before that team's destruction or use reference counting on the team itself. (see #127 for draft changes)
    • Suggest: The team could automatically destroy shared contexts, but not private contexts.
    • Suggest: Add that shmem_finalize destroys all teams and associated shared contexts. Clarify return value
  • shmem_team_split_strided
  • Clarify return value
    • If new_team is created successfully, is a zero value returned to all PEs in parent_team or only those in new_team?
    • If new_team cannot be created by any PE (e.g., due to resource constraints), do all PEs return a nonzero value?

[nspark]

Other thoughts:

• Should shmem_team_n_pes return a size of zero for SHMEM_TEAM_INVALID? (Personally, I like 0 vs. -1 for team size, since we already specify that a valid team always has at least one PE in it.)
• @gmegan Why was "It is invalid to pass SHMEM_CTX_DEFAULT to this routine." removed from the notes for shmem_ctx_destroy?
  • Edit: I think the answer is that "shmem_ctx_destroy destroys a context that was created by a call to shmem_ctx_create or shmem_team_create_ctx", and the default context was not created by the userland API.

[nspark]

For the return value of shmem_team_split_strided, I think I'd prefer that:

• Upon successful creation of new_team, zero is returned to all PEs in new_team, and a nonzero value is returned to all other PEs.
• Upon a failure to create the child team, a nonzero value is returned to all PEs in parent_team.

That is, any PE that calls shmem_team_split_strided and has new_team updated to equal SHMEM_TEAM_INVALID also gets a nonzero return value..