[mpi3-coll] Updates after first reading
Torsten Hoefler
htor at cs.indiana.edu
Mon Dec 29 18:42:01 CST 2008
Hello,
I updated the document with all comments that I received at the last
meeting. I created tickets for the more compplicated issues. I assigned
four reviewers to each ticket and ask them to drive the discussion (the
people who were most active in the discussion at the last meetings).
However, everybody should review the changes and review the tickets. I
am open for comments and suggestions.
The new document is at [1] and in the wiki [2] and the LaTeX diff [3].
[1]: http://www.unixer.de/sec/nbc-proposal-12-29-2008.pdf
[2]: https://svn.mpi-forum.org/trac/mpi-forum-web/attachment/wiki/NBColl/nbc-proposal-12-29-2008.pdf
[3]: http://www.unixer.de/sec/nbc-proposal-12-29-2008.diff
The detailed list of changes:
1) p1:18-41 - added all MPI_Ixxx calls to 5.1
2) p3:11 - "callers" changed to caller's
3) p3:22 - added new sentence (proposed by Jesper): "A blocking
operation is complete as soon as the call returns. a
Nonblocking call requires a separate completion
operation" -> ref. to Sec 3.7
4) p3:23 - "call" changed to "operation"
5) p3:26 - "function" changed to "operation"
6) p5-7:all - added non-blocking versions (also to the next
paragraphs)
7) p49:38-42 - added semantic advantages
8) p50:4-6 - removed send buffer restrictions! "All associated send
buffers should not be modified and all associated receive
buffers should not be accessed between init and completion"
-> opened separate ticket (need to clarify send and
receive buffers!?)
#81 - reviewers: Erez, Darius, Jeff, Keith
9) p49:34 - added comma: communication"," which is eventually
10) p49:41-42 - system resources: asked Jeff for details (change P2P too)!
#82 - reviewers: Jeff, Rich, Bronis, Keith
11) p50:7-10 - dropped "Advice to implementors" completely
12) p50:11-12 - replaced regexp by "all request test and completion
functions (e.g., MPI_Wait)"
13) p50:13-16 - mention that mixing with other requests as possible
(I/O, NB P2P)
14) p50:19-21 - rationale against MPI_Request_free, offline to ML (I
updated it)
#83 - reviewers: Jesper, Rolf, Bill, Erez
15) p50:19-28 - combine rationales and put them after the paragraph
16) p51:22 - add rational why cancel is not supported -> semantically
unclear
17) p50:22 - point-to-point -> collective =)
18) p50:17 - new paragraph before "Coll operations"
19) p50:22-23 - rationale: "consistent for blocking call + additional
benefit is HW support"
20) p50:31 - "in the same order per communicator"
21) p50:37 - have to -> could
22) p50:36 - only need -> would be
23) p50:40 - advice for users: "this is different than for P2P"
24) p51:1 - "will be ignored" -> ask Jeff (change text and advice to
users)
#84 - Jeff, Jesper, Rolf, Bronis
25) p51:9 - extra space "est )" (removed all spaces before ")" and
after "(" in the whole chapter
26) p51:21-22 - "every process called MPI_Ibarrier"
27) p51:14 - operation -> call
28) pxx:xx - operation -> call! everywhere -> the word operation is
overloaded (MPI Operation and Collective Operation) - do
we care?
#85 - Bronis, Jeff, Rolf, Bill
29) p67:24 - mixed -> blocking calls can be interleaved
30) p67:34 - after -> until
31) p68:33 - dropped "huge"
32) p69:44-48 - made advice weaker
33) p69:45 - should -> could
34) p69:44 - can potentially -> may
35) p69:44 - dropped "huge"
36) p69:44-48 - rephrase advice to implementors: Keith, Jeff, Bronis, Rolf
#86 - Keith, Jeff, Bronis, Rolf
37) p70:11-24 - add picture
38) p70:27 - remove "very"
Best,
Torsten
--
bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler | Postdoctoral Researcher
Open Systems Lab | Indiana University
150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA
Lindley Hall Room 135 | +01 (812) 855-3608
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 197 bytes
Desc: Digital signature
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-coll/attachments/20081229/4efdf7e5/attachment.pgp>
More information about the mpiwg-coll
mailing list