Skip to content
Howard Pritchard edited this page Jul 12, 2021 · 1 revision

#07/12/21 webex notes

Attending: Dan Holmes, Howard Pritchard, Thomas Hines, Martin Schreiber

Picking up with modified RM/application request/response (dataflow) diagram. This also includes Dan's original state transition style diagram. Distinguish between control flow - which is not being illustrated here - and data flow. Agreement that Martin's dataflow diagram and Dan's state transition diagram are different ways of looking at the problem of dynamic resource allocation/returning.

one-sidedness vs send/recv (pt2pt) semantics for exchanging resources between RM and application. send/recv semantics works better from the app point of view, but the RM would probably map better to the request/response (one-sided) approach.

Discussion of requests from an application possibly "conflicting". Considering an application using multiple independent sessions.

Look back at MPI 5.0 resource changes v4. Do these diagrams capture the first table in the document? No looks like we're missing some things, e.g. consensus maker. Coercive, coopertaive - maps to our previous speed diagram. Notification of change requests - we're not capturing that in the diagrams yet. Event vs. polling. Except for some MPI_T functions, MPI is primarily a polling interface. Do some table clean up as part of the conversation.

Clone this wiki locally