Skip to content
Howard Pritchard edited this page Jun 27, 2022 · 1 revision

#06/27/22 webex notes

Attending: Howard Pritchard, Dan Holmes, Martin Schreiber(briefly), Thomas Hines, Dominik Huber, Martin Schulz, Aurelien Bouteiller

Agenda items

Review Dominik's work and updates. The code he sent out to the mail list has not been discussed earlier this month. Smaller chunks - have to inform newly started resources what pst they are part of. Assume there are some primary ranks in the process set. Discuss some of the particular functions present in the test code. For example Howard asks about MPI_Session_Test_Exists.

Dominik walks us through a short summary of new functionality they are proposing.

Dan has concerns about a function to set pset info key/vals by the application. Would this need to be handled globally? I.e. would this need to be handled collectively.

Dan describes how we were thinking to handle new processes in a sessions model - parent group in children, and children group in the spawner processes. We had not considered the case where the runtime could fire up new proceses directly and notify (somehow) the existing application processes. Consider a set of processes that buy in to having RT fire up new processes. We could probably treat these as parent processes.

Dominik also reviews this process set operation. Dan thinks we will need a process set from group type functionality. Dan argues that these process set operations methods could be done using group operations in conjunction with this process set from group functionality. Case of only one or a subset of group members know about some new resources (processes for a new group). Discuss the existing port functionality and compare to the tag notion in MPI_Comm_create_from_group. Can we avoid use of intercommunicators for a session spawn method? Likely if we had this process from group functionality.

How did we handle multiple child process sets in the original discussions of an MPI exec type function.

We need to discuss if there's ever a case where an application wants to "broadcast" a process set to the RT or other processes?

Dominik will give a fuller presentation of this work on 1 August.

Clone this wiki locally