I/O Subcommittee Minutes: 21-23 January 1997

Bill Nitzberg (nitzberg@nas.nasa.gov)
Tue, 04 Feb 1997 09:40:53 -0800

Succinct I/O Subcommittee Minutes: 21-23 January 1997

Major Items
-----------

* An official first vote was taken on most of the I/O chapter sections
(see Votes below).

* All non-blocking collective operations were removed (IOPEN, ICLOSE,
IREAD_ALL, IWRITE_ALL, etc.)

Official Votes
--------------

10.2.1 MPI_OPEN: passed - 15/5/5

Failed: file locking proposal
Passed (16/1/2 - straw vote): MPI_UNIQUE_OPEN
Failed (9/11/4): arguements (filename, amode, info) significant only at root
Friendly amendment: "filename" must refer to the same file on all processes
Passed (11/9/4): "filename" must be textually identical

10.2.2 MPI_IOPEN: failed - 3/18/6

10.2.3 MPI_CLOSE: passed - 24/0/2

10.2.4 MPI_ICLOSE: failed - 2/22/2

10.2.5 MPI_DELETE: passed - 12/5/9

Passed (13/5/9): add "info" arguement
Friendly amendment: errors are handled by MPI_COMM_WORLD
Friendly amendment: name changed to MPI_FILE_DELETE

10.2.6-8 MPI_RESIZE/MPI_PREALLOCATE/MPI_FILE_GET_SIZE: passed - 13/0/5

Failed (2/10/8): make MPI_RESIZE independent
Failed (3/13/5): make MPI_PREALLOCATE independent
Failed (1/5/13): change MPI_RESIZE to truncate only
Friendly amendment: name changed to MPI_FILE_SET_SIZE

10.2.9 Querying File Parameters

Passed (6/5/9): change MPI_FILE_GET_COMM name to MPI_FILE_GET_GROUP
Passed (11/4/5): MPI_FILE_GET_GROUP
Passed (13/0/7): MPI_FILE_GET_AMODE

10.2.10 Info

Passed (14/1/6): MPI_FILE_SET_INFO functionality
Passed (13/2/6): MPI_FILE_GET_INFO functionality

10.3 File Views: passed - 18/0/2

Friendly amendment: erroneous to SET_VIEW with outstanding requests active

10.4.2 Data Access w/Explicit Offsets: passed - 16/0/5

10.4.3 Data Access w/Individual File Pointers: passed - 19/0/2

Friendly amendment: Rather than have MPI_FILE_GET_POSITION return both
offset and byte_offset, add a separate routine MPI_FILE_CONVERT_DISP_X
to convert from offset to byte_offset

10.4.4 Data Access w/Shared File Pointers: passed - 13/4/3

10.5 File Interoperability

Passed (16/0/5): NATIVE & INTERNAL (w/o FILE_TYPE_GET_SIZE)
Passed (13/4/4): CANONICAL functionality

10.8 Error Handling Proposal

Passed (17/0/4 - straw vote) changing I/O error handling:
1. add a new I/O error handler to communicators
2. ERR_NOT_SAME becomes a "regular" error
3. update "7.6 Adding New Error Classes" to indicate
which error handler to use

Discussion Items
----------------

a. PROPOSAL: MPI_Preallocate & associated changes in MPI_Resize

Accepted with minor modifications.

b. PROPOSAL: "Interleaved Collective Operations" semantics

Modified---semantics in original MPI 1.1 were seen as adaquate.

c. PROPOSAL: Error handlers for MPI_File

No proposal forthcoming.

d. PROPOSAL: "Miscellaneous Clarifications" in Semantics section

Accepted.

e. Edited Filetype Constructor section; minor changes

All accepted.

f. PROPOSAL: File locking via amode

Added MPI_UNIQUE_OPEN amode arguement to allow optimizations.
File locking functionality was deemed too filesystem specific,
and the I/O subcommittee has explicitly decided to leave
filesystem specifics out of the standard.

g. Assorted small changes marked by January change bars.

All accepted.

h. MISSING: Canonical Data Representation

Canonical representation proposed by Leslie Hart based
on CANONICAL32 and CANONICAL64 access modes. Awaits further
refinement.

i. MISSING: Portable type constructors for holes, e.g. is it
possible to portably specify a hole of 3 integers?

No action.

j. MISSING: Proposal for mandatory info, or a statement that
implementations cannot require particular info.

No action.

k. MISSING: Statement of fairness semantics

No action.

l. PROPOSAL: Make MPI_File_get_type_size/extent consistent
with MPI-1 MPI_Type_size/extent. [John May]

MPI_FILE_GET_TYPE_SIZE was deleted.

m. PROPOSAL: Directory manipulation routines (e.g. chdir) [Kim Yates]

Directory manipulation routines were deemed too filesystem specific,
and the I/O subcommittee has explicitly decided to leave
filesystem specifics out of the standard.

n. PROPOSAL: "Sequential I/O" [Nick Maclaren]

Minor modifications to the shared file pointer routines should
be able to support "sequential I/O". A proposal is forthcoming.

Additional Items
----------------

o. Remove nonblocking collective data access routines

Passed (11/1/3 - straw vote)

- bill