-
Notifications
You must be signed in to change notification settings - Fork 1
2017 10 02 webex
- Brian Smith
- Dan Holmes
- George Bosilca
- Rich Graham
- Wesley Bland
- Kathryn Mohror
- Howard Pritchard
Dan starts a discussion about the local/global slides Martin Schulz had presented at the Forum. We can't use session (its already being used by the MPI T API), what do we want to call it now? Could we re-use the endpoint terminology. Dan thinks in some ways, the concepts are very close. The proposed endpoint API hasn't changed much in several years.
Two issues currently that have impeded making progress on endpoints are:
- There is no open-source implementation of Endpoints today.
- There is also push-back from certain parties that you would not need endpoints if MPI_THEAD_MULTIPLE could be better.
Dan also discusses the "set of processes" term. Maybe we need something better. How about an "MPI Team" instead? OpenMP has terminology for similar concepts - OpenMP Teams.
Kathryn is concerned about reuse of the endpoints proposal. She suggests separating of concerns. How you're going to use runtime support/information vs the support itself, and the terminology questions. George brings up the Netloc effort presented at the Chicago/ANL EuroMPI in September. Dan sees Sessions as giving the user info about what's available, with netloc/hwloc providing things like distance, topology, etc. between these resources.
Other items. Dan says Ralph Castain presented PMIx 2 at the last EuroMPI. If this takes over the MPI world, we need to work closely with the PMIx folks to make sure that MPI implementations can get what they need for a Sessions-like capability. We need to make sure we have conversations with other potential providers of this type of infrastructure to underpin Sessions (or whatever it morphs in to).
- all think about how to map Sessions to existing endpoints proposal.
- all - work on better terminology for the "set of processes" concept.