2025 03 24 webex

Agenda - 3/24/25

Update on “dynamic processes with pset” proposal

Martin = Martin Schreiber

Discussion

• Martin posts link to dyn res in HPC workshop at Europar - https://sites.google.com/uji.es/dynreshpc-workshop/home

• Martin+Dominik discusses a dictionary proposal.

  • setup
  • remove
  • data publish
  • data lookup

• envisioned as a "global" scope in some sense.

• Howard points out we need to consider how this would related to the publish interface.

• Sonja and Roberto ask about synchronicity model.

• Locality information required

• What would happen if an entry already exists?

• For MPI Forum:

  • Expected question: Why do we need to do this because there's a name publishing functionality in it.

• Roberto Rocco:

  • Is this dictionary local? Or accessible by all processes?
  • Answer by Dominik: Yes, all processes within a job can access it.

• Local to MPI Sessions? Howard: It should be independent of Sessions.

• Howard: Describe which error handler should handle it.

• Sonja:

  • Local or global? Hint about how it's implemented.
  • Is it required to initialize MPI?
    • Possible solution: At least one MPI Session needs to be used, or MPI Initialize needs to run.
    • Another solution: A reference counter could clean up everything.

• Only non-blocking variants of the dictionary?

  • Howard: In the interest of simplicity, having only the non-blocking variants would be good.

• Michael Raymond (HPE): Overlap with pvars and cvars:

  • Querying MPI and getting status on it.

• lookup dictionary entry: Add that we need to wait until the information arrives.

• Howard: data_lookup

  • uses input-only and output arguments: one with the keys and potentially other information, and one for the output as key-value pairs.

• Multiple PSets are required as the output of the operation. Array of strings for PSets or are there alternatives?

  • A simple solution would be to have “;” as a delimiter and all PSet URIs in a single string separated by the delimiter.
  • Sonja: How about using an array of indices in the PSet List?
    • Martin: List would get too large after many resource changes, and we like to avoid frequent MPI session destroy / buildup => They don’t exist in the PSet list, which can be used for initialization.
    • Howard: MPI spawn multiple where an array of strings is used: “CHARACTER(LEN=), INTENT(IN) :: array_of_commands()”. But this works only for input.