*** NOTE ***
Please read and comment on the current draft of the MPI-2 document
this week, as this will be the last chance to make any corrections
(other than to correct formatting errors) before the SC '96 release!
************
During the October meeting, an official reading of the entire I/O
chapter was held before the entire forum. It is expected that
first votes on chapter will be held at the January meeting.
Discussion Items
----------------
1. "End of File" Discussion (10.1.6, p231/18-30)
Section will be removed; appropriate text will be moved
to MPI_SEEK and MPI_SEEK_SHARED.
2. New section "Progress" (10.1.7, p231/31 - p232/16)
Section approved by the subcommittee, and modified
during the general reading to eliminate the example code
(p231/43 - p232/2).
3. New section "Logical vs. Physical File Layout" (10.1.8, p232/17-33)
Section approved.
4. "File Interoperability" section modified (10.1.9, p233/26 - p244/48)
Section approved: 11/0/3.
Voted on default interoperability mode:
No default Failed: 3/9/2
NATIVE Failed: 5/7/1
INTERNAL Passed: 9/4/1
CANONICAL Passed: 7/4/2
Beauty context between INTERNAL and CANONICAL was ambiguous:
INTERNAL 6
CANONICAL 6
Abstain 3
5. Proposals for file permission (10.2.2, p236/27 - p237/25)
PROPOSAL 1: rejected: 1/10/2
PROPOSAL 2: rejected: 1/6/7
6. Proposal to remove MPI_IOPEN (10.2.3) and MPI_ICLOSE (10.2.7)
Vote was not conclusive: 5/4/4
These routines will remain for the SC '96 release, and
we will re-address them at the next forum meeting.
7. Proposal to split MPI_GET_FILE_PARAMS into separate routines
(10.2.5, p239/1-35)
Approved.
8. Semantics of MPI_CLOSE better specified (10.2.6, p240/1-12)
Approved.
9. MPI_VIEW Discussion (10.2.8, p241/27)
Proposal to change "disp" from absolute byte displacement
to a count of etypes defined by the previous view.
Rejected: 0/13/3
10. New section "File Consistency" (10.2.10, p242/34 - p245/3)
Section approved: 15/0/3
Changed MPI_SET_ATOMICITY to take a boolean flag
rather than ATOMIC/NONATOMIC constants.
Changed the behavior of outstanding requests after
an MPI_FILE_SYNC to be completely implementation dependent.
11. Specified behavior of outstanding requests after:
MPI_CLOSE (10.2.6, p240/8-10)
MPI_ICLOSE (10.2.7, p240/27-30)
MPI_VIEW (10.2.8, p241/44-46)
MPI_SET_ATOMICITY (10.2.10, p243/15-23)
MPI_FILE_SYNC (10.2.10, p244/44-48)
MPI_DELETE (10.2.12, p248/36-37)
MPI_RESIZE (10.2.13, p249/30-36)
Approved.
Changed some behavior to be completely implementation dependent.
12. Added new routines and info keys to "File Info" section
MPI_SET_FILE_INFO (10.2.11, p245/25-48)
MPI_GET_FILE_INFO (10.2.11, p246/1-27)
file_perm (10.2.11, p246/39-42)
chunked* (10.2.11, p246/44 - p247/10)
Approved.
Straw vote on "file_perm": Passed: 9/1/4
13. Semantics of MPI_DELETE better specified (10.2.12, p248/33-48)
Approved.
14. MPI_IWRITE_EXPLICIT Discussion (10.3.4, p254/13-16)
Accessing a buffer used in a nonblocking operation before
the operation has completed is erroneous for point-to-point,
but allowed for one-sided. Which semantic should I/O choose?
A proposal to make it identical to one-sided was approved: 9/3/4
15. Remove nonblocking routines with file pointers
MPI_IREAD* & MPI_IWRITE* (10.4.3, p257/5-48 & 10.4.4, p258/1-48)
MPI_I*_SHARED (10.5.1 p262/8- 263/13)
MPI_I*_SHARED_ORDERED (10.5.2, p265/20 - p267/8)
Vote was not conclusive: 4/3/7
These routines will remain for the SC '96 release, and
we will re-address them at the next forum meeting.
16. MPI_SEEK Discussion (10.4.5, p259/26-27)
Is "offset" signed? If not, how do you seek backwards?
Offset is signed.
17. "Filetype Constructors" section modified (10.6, p268/11 - p278/31)
Constructors take "size and rank" rather than a communicator,
and all constructors are independent.
A scatter/gather example has been added (p269/35 - p270/48).
Approved.
18. New "I/O Error Classes" section (10.7, p278/32 - p279/5)
Approved.
19. Proposal for "I/O Service Interception" (10.8, p279/6 - p280/48)
New proposal will be forthcoming.
Additional Items
----------------
MPI_CLOSE will be changed to set the file handle to MPI_FILE_NULL.
"Status" values will be explicitly be defined (all fields except
for "error" and "count" are undefined).
Type matching will be defined so that MPI_BYTE legally matches
other types for data access.
It was decided that we should leave the definition of an "MPI
environment" as implementation dependent for the purposes of
file interoperability (MPI_DATA_INTERNAL).
Attendees List
--------------
Bill Nitzberg, NASA Ames
Parkson Wong, NASA Ames
Steve Landherr, HP-Convex
Rajeev Thakur, Argonne
Leslie Hart, NOAA/FSLL
Steve Kubica
Jean Pierre Prost, IBM
Pang Chen, Sandia
Elsie Pierce, LLNL
Kim Yates, LLNL
Tyce McLarty, LANL
Nick Nevin, OSC
Manuel Ujaldon, Univ. of Maryland
Hans-Christian Hoppe
Albert Cheng, NCSA
Lyndon Clarke, Edinburgh
Harish Nag
Richard Frost, SDSC