[mpiwg-tools] Comments on MPI_T Events reading

Martin Schulz schulzm at in.tum.de
Tue Mar 5 14:53:10 CST 2019

Hi all,

Here are the comments from the MPI_T_Events reading.


Connected MPI processes
	-> line 43, page 622, get_info call
	-> need to tweak to include processes before init/after finalize

	same for line 26/27, page 623
	-> requirement to return the same values

line 8, page 626: add “during callback registration”

Line 33, Page 626: rewrite “free-callback”

Line 23, page 627: dropped handler >>callback<< function

Line 30 (or so), page 627: need weasel language in case HW cannot even count dropped events -> best effort

Line 43/44, page 627: wrong sentence, which allocation? Need to clarify

Line 4, page 629: rewrite the sentence with “passes a pointer”

Line 11, Page 629: space before “(pointer)” missing

Line 16/17, Page 629: extent no longer returned

Comment: why MPI_Count
	-> add what the value and behavior of a timestamp is
	-> or add forward timestamp at the bottom of page 629

Line 42, page 630, font for MPI not correct

Line 44, page 630: MPI_T_SOURCE_GET_NUM
	-> need to pull description after the prototypes

Line 7, Page 631, keep “num_sources” on one line

Line 27, Page 631, add “source” / name of argument

Line 39/40, line 17 and 20: page 631, missing type of argument in the prototype description

Discussion on timestamps
	integer can roll-over if we don’t have enough bits
	64 should be enough, but 32 bit can roll over

	add advice to users that problems can happen on true 32 bit platforms 

	should specify what zero is, doesn’t have to be Linux zero time

	should we fix this to a particular time, so we have a fixed point for comparison

	but then the discussion turned away from that due to HW constraints

	still need to add definition of timestamps

	perhaps enough to say that things are comparable within a source

Jeff added two small comments on GitHub
	-> need to look for that

Comment om MPI_T_EVENT_GET_INFO (page 622, 37/38)
	who is allocating the info object
	and should we pass Info or a pointer
	if ignores, should we pass NULL or INFO_NULL
	should probably be NULL

Prof. Dr. Martin Schulz, Chair of Computer Architecture and Parallel Systems
Department of Informatics, TU-Munich, Boltzmannstraße 3, D-85748 Garching
Member of the Board of Directors at the Leibniz Supercomputing Center (LRZ)
Email: schulzm at in.tum.de

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-tools/attachments/20190305/1fad13cf/attachment.html>

More information about the mpiwg-tools mailing list