Notes 2018 09 20

MPI_T Events Reading, 20.9.2018

• Line 22, mpit.tex: missed MPI macro

• get info: do we need to say array of handles and array of integer

• get info: optional info object, valid before Init?

• get info: return the same, even for different info objects?

• Page 598, bottom: array of displacement(s) / inconsistent

• discussion on adding fixed sized types to MPI_T events

  • needed for efficient implementation
  • but let's not make them too large
  • comment: should look at all types, but make conscious decision on which to add full generality may be hard

• Page 599, line 19: is >>the<< ... pointer

• Page 599, line 28: typo / variable

• handle alloc: we describe the index differently as with get info

• Page 600, line 45: IN

  • same problem with info

• Page 601, line 15 typo (event registration handle, not communicator)

• Page 601, description of arguments for cb

  • just point to CB prototype and dont repeat (that would also be consistent with the rest of the standard, need to double check)

• Page 601, line 29.5:

• page 601, table, >>a<< restricted set

• Page 601, line 48, extend caption to extend hierarchy

• Page 601/602, discussion on signal safety / safety levels

  • consequences for MPI functions
  • need new list of functions that can be called if asnyc flag is set
  • need discussion on what has to be the case for thread safety
  • query option for routines whether they are safe to call?
  • turn things around and have CBs for different levels
    • delay if the right handler is not available
    • or drop it and report in dropped handler
    • but tools could do that itself

• Bill: would it be good to have a ¨shared" level that allows to communicate more was very important for IBM LAPI in threaded environments

• Page 601, last constant may be too long

• Page 601, thread safe row in table should be replaced with a description of expected behavior

• Page 602: make function list as a table

• Page 602, line 24: handle should be IN only

• Page 602, line 33: runtime -> implementation

• Page 602, line 35: add clarification on timing (talk about raised on completeness)

• Page 602, line 45: runtime -> implementation

• Page 603, line 17, such no -> no such

• Page 603, line 18, rephrase, make clear that we don't cancel ongoing CBs

  • rather say no more CBs will be called

• Page 603, discussion on dropped count

  • does this have to be accurate?
  • is this actually always possible?
    • example: network packets

• General comment: replace source with event source

• add advice to implementors on sources and ordering and how they are intended to be used also state that we want as few sources as possible for an implementation

• Question on the ordering of text,

  • should sources be put first?
  • or split the sources part and at least move the first part up?
  • at early text, add page ref to be able to jump forward

• Page 604, line 10, something is missing, rewrite

• Page 604, line 10-13 should be streamlined and adjusted to other parts

• Page 604, line 24: back to the callback (?)

• Page 605, line 37: replace with ¨timestamp captured during invocation¨

• Bill: suggests to add a small glossary type section upfront

• Page 606, line 17, make clear that this can only be called in that callback

• Should all nums in MPI_T use MPI_Count

  • general type and symmetry discussion

• Page 606, line 32, text missing for source get num

• Page 607, source get info

  • description of index is not parallel to other get info calls
  • line 2, binding misses name

• Page 607, line 21: id -> index

• Page 607, line 31, boilerplate text missing

• Page 607, check on ad-hoc if this is the right hyphen, probably should take it out and replace with \

  • may also need to be italics as in the other parts of the standard

• terms.tex, new functions as macros -> turn into table?

• tools-3.tex, line 19: macro for MPI

  • also drop ¨library¨

• General comment: do we need a way to identify how to print a variable of an MPI type, like Count

  • Bill: problem once we leave the default type system

• Strawpoll: Source vs. stream

  • Conflicts with P2P on either case
    • Should be Event Source
    • general sentiment was source