notes 2017 06 14

14th June 2017 Meeting Notes (Face-to-face, MPI Forum, Aachen Germany)

Allocating receive and freeing send

• New topic: How are these operations impacted by datatype?
  • How are LB and UB interpreted in the context of allocating memory in Arecv? LB can be after the start of data and UB can be before the end.
  • General consensus that datatypes with absolute addresses (e.g. relative to MPI_BOTTOM) don’t make sense for these operations.
  • In MPI_Fsend, a pointer returned by MPI_Alloc_mem must be passed as the source argument.
    • The datatype must be constructed so this pointer can be passed. For example, if a subset (e.g. vector of a larger matrix) of a larger allocation is to be sent and then the full buffer freed, the datatype must be relative to the start of the matrix, not a simple vector starting at the row to be transmitted.
  • In MPI_Arecv, how do you allocate memory for a buffer that is only partially filled in by the transfer (e.g. a subarray datatype)?
  • Result of this discussion was to suggest that, as a first step, prohibit the use of derived datatypes in Arecv/Fsend

Info assertions

• Reviewed list of potential info assertions.
• Priority and threading level / order emerged as potentially interesting for future discussion.

FP16 built-in types

• Adopted by this WG, to avoid creating a new one.