Implementation for mpio driver with selection I/O. #3058
Conversation
This is a draft for preliminary review.
vchoi-hdfgroup
requested review from
lrknox,
derobins,
byrnHDF,
fortnern,
jhendersonHDF,
qkoziol,
bmribler,
glennsong09,
mattjala and
brtnfld
as code owners
vchoi-hdfgroup
added
the
WIP
vchoi-hdfgroup
added
the
Component - C Library
derobins
added
Merge - To 1.14
Priority - 1. High 🔼
src/H5FDint.c
Outdated
| @@ -1555,7 +1562,7 @@ H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, ui | |||
| io_len = MIN(file_len[file_seq_i], mem_len[mem_seq_i]); | |||
|
|
|||
| /* Check if we're using vector I/O */ | |||
| if (use_vector) { | |||
| if (!skip_vector_cb && use_vector) { | |||
There was a problem hiding this comment.
Should use skip_vector_cb to determine the value of use_vector on line 1460, then there will be no need to keep checking it
src/H5FDint.c
Outdated
| *------------------------------------------------------------------------- | ||
| */ | ||
| herr_t | ||
| H5FD_selection_build_types(hbool_t io_op_write, uint32_t num_pieces, H5S_t **file_spaces, H5S_t **mem_spaces, |
There was a problem hiding this comment.
Having this function in a shared place is a little strange. I was originally thinking that the code in H5FDmpio.c and H5Dmpio.c shouldn't be shared. If it's working though it might be worth it. I'll continue thinking about a way to make this function more self-contained/generic/maintainable
There was a problem hiding this comment.
After further thought, I think we should move this function to H5FDmpio.c, and return this functionality to being integrated into link_piece_collective_io (with the frees moved to before the I/O as suggested in my email). I think it'd also be a good idea to move the check for num_pieces out of the function (and into the read/write callbacks), and remove the size_i and mpi_buds_base parameters (calculate them in the read/write_callbacks).
There was a problem hiding this comment.
I will move this function to H5FDmpio.c as indicated.
And I will restore link_piece_collective_io() back to its original state in H5Dmpio.c (with the frees moved to before the IO), is that correct?
There was a problem hiding this comment.
Changes made as indicated for H5FD_selection_build_types().
However, I still need to pass num_pieces to the routine because it is needed to malloc the arrays of mtype, ftype and displacement etc. I can calculate mpi_bufs_base in the read/write callbacks, but I still need to pass the calculated info to H5FD_selection_build_types() for calculating displacements of the pieces.
This is in response to one of the PR review comments.
…into mpio_develop
….c when selection I/O is ON by default in the library. a) Create internal functions H5FD_read/write_vector_from_selection() and H5FD_read/write_from_selection() that correspond to the 4 new APIs H5FDread/write_vector_from_selection() and H5FDread/write_from_selection(). --Call the internal functions H5FD_read/write_vector_from_selection in H5FD__mpio_read/write_selection() for independent xfer_mode.
…into mpio_develop
when selection I/O is ON by default in the library. --Undo the base address addition in internal routines before passing down to the mpio driver
…/t_dsets.c (via testpar/testphdf5.c) when selection I/O is ON by default in the library: --need to check for fixed bufs when trying to figure out mpi_bufs_base in H5FD__selection_build_types().
…into mpio_develop
This is the fix for a failure from test_link_chunk_io_sort_chunk_issue() test in testpar/t_coll_md.c (via testpar/testphdf5.c) when selection I/O is ON by default in the library: --need to check for fixed bufs when trying to figure out mpi_bufs_base in H5FD__selection_build_types().
…into mpio_develop
system() is only used in the iopipe test and the things it calls (which are POSIX-y) are protected by an ifdef.
|
I am going to create a new PR based on the latest "develop" branch and incorporate the mpio driver changes to there. I think it's easier that way instead of resolving all the conflicts/CI failures etc. Then I will close this PR. |
This also addresses the review comments of the previous PR HDFGroup#3058 which will be closed.
|
See PR #3222 |
This is a draft for preliminary review.