Skip to content
Howard Pritchard edited this page Jan 10, 2022 · 1 revision

#01/10/22 webex notes

Attending: Howard Pritchard, Thomas Hines, Dan Holmes, Isaias Urena, Trupeshkumar Patel, Brian Smith, Martin Schreiber, Martin Schulz, Ralph Castain

Agenda items

  • Dominik Huber gives a presentation about his thesis
  • Issue Dan found recently that needs to be fixed in MPI 4.1

Dominik presentation discussion

Based on work of Jan Feck. Aurelian asks are there guarantees MPI_Session_get_res_change returns same delta_pset if called by multiple processes. Dan thinks one would need two broadcasts. One within the RTE, and once in the application. Do we need another input parameter to this function? Current proposal uses a MPI_RC_BOUND_PSET key to handle this case.

Discuss the example on page 59 of Dominik's thesis. Some concerns about whether the first call to MPI_Session_get_res_change will return something consistent for delta_pset and incl. Dan suggests a different way to interpret data_pset with it being different between the newly created processes and the already existing processes. Example of the 3 originally running and two new ones added. Later is a delta_pset with size 3 and the formet see a delta_pset of size 2. Discuss how this could also be used inside MPI_Init.

Links:

https://www.martin-schreiber.info/data/student_projects/2021_MA_dominik_huber.pdf

TODOS

Clone this wiki locally