[mpiwg-sessions] notes from sessions (Enclaves) reading at 12/5/18 forum
Pritchard Jr., Howard
howardp at lanl.gov
Wed Dec 5 20:03:03 CST 2018
Notes from the Sessions reading at 12/5/18 forum:
Dan gives some background about PR #55 as is. He says we’ve got enough changes that aren’t ticket 0 that this really shouldn’t be called a reading.
We’ve added 7271 lines and subtracted 1004. A lot of this is moving around. Move all the getting processes started into chapter 10.
Wesley asks about dynamic being incompatible with peer to peer – Dan says no, can be mixed. Tony is getting us confused. How to do Sessions
outside of MPI? Looking for using sessions concepts for more general job startup. Aurelian asks about how we get groups from nothing – Dan
says we aren’t getting groups from nothing – hence the runtime process set notion.
Martin says we should be more careful about Process vs MPI Process – but this is in places we’ve not changed for sessions in most cases.
Some discussion of the use of the term Session and “Peer to Peer” - should we use some others?
List of things brought up during the reading/plenary
* Martin wants to drop “unless the application explicitly calls MPI_INIT” in page 416, lines 46-47. Add valid as an adjective to MPI COMM WORLD in this section.
* We need to fix the second paragraph of 10.1 to better integrate the dynamic process model back into the 3 models for process creation
* fix 10.1 subtitle!
* page 417 line 22 add Window or File
* water down line 45 on page 416 or remove
* question about MPC-MPI MPI process is a thread model from Aurelian and Martin. We need to think about this – maybe add a blurb in the pset definition
* Maybe move last paragraph in 10.4 up to intro part of chapter and reword a bit. remove backward compatible?
* swap 10.4.1 with 10.4.2.
* page 418 line 1 remove use of “runtime” in that first, todo: decide if we want to mandate or highly recommend mpi://UNIVERSE and mpi://comm type shared.no examples.
* make clearer what’s going on with lines 4-5, what exactly is a pset. define semantically?
* some discussion of groups and getting groups from psets
* key is mandatory for all process sets - Jeff Squyres thinks this needs more verbiage.
* add some more discussion about the flags bit mask to MPI Session init.
* fix f08 binding for MPI_Session_init flags should be TYPE(MPI_Flags), check other f08 bindings too where INTEGER should be replaced with a TYPE
* could we use info arg to MPI_Session_init rather than flags
* use tools cvar/pvar approach for psets and their lengths
* replace MPI_Errhandler with MPI_Session_errhandler in multiple places
* add text for thread level support
* Aurelian asks about thread safety for things like data type manipulation, etc. This needs to go with the info ops before MPI Init discussion.
* an advice to users about sessions and how implementations may use them – Dan talks about isolation and multiplexing, grouping of communicators.
* should we call Sessions Enclaves?
* pre-defined info keys
* options naming for sessions? Nathan asked about this.
* communicator inherits uri as its name idea
* need to add local to the description of MPI_Session_init
* be consistent with capitalization of session/enclave Martin says put MPI in front of Enclave or Session
* function name for MPI_Session_init line 46 on page 418
* make session handle in out to MPI_Session_finalize
* need a lot more text – use analogous not equivalent, should we add a drop-all-resources type of call that frees all resources associated with a session. Collective. Files, windows, communicators, groups get leaned up
* MPI Session get num sets replace pset_name with npsets or something like this,
* same issue with get_psetlen
* could we replace all this with an in/out info object
* definitely need to keep the figure
* rework chapter describing MPI_SESSION_GET_NTH_PSET about room for NULL terminator etc. Check MPI T and info
* in the code example title add a reference to chapter six functions.
* question about where the example would block
* verbiage about session finalize associated with the example
* uniqueness of uri (Jeff says don’t use this!!!) across sets of processes, suggestion and advice to users for MPI_Create_comm_from_group gov://llnl thingy
* Tony is talking about doing state transitions.
* should we rename MPI_GROUP_FROM_SESSION_PSET . maybe return MPI GROUP EMPTY if not valid pset name?
Office 9, 2nd floor Research Park
TA-03, Building 4200, Room 203
Los Alamos National Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the mpiwg-sessions