Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merge Implementation of the mpio driver with selection I/O. #3360

Merged
merged 4 commits into from
Aug 8, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 32 additions & 12 deletions src/H5Dchunk.c
Original file line number Diff line number Diff line change
Expand Up @@ -2731,11 +2731,20 @@ H5D__chunk_read(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)
}
} /* end if */
else if (!skip_missing_chunks) {
/* Set up nonexistent dataset info for (fill value) read from nonexistent chunk */
nonexistent_dset_info.layout_io_info.contig_piece_info = chunk_info;
nonexistent_dset_info.file_space = chunk_info->fspace;
nonexistent_dset_info.mem_space = chunk_info->mspace;
nonexistent_dset_info.nelmts = chunk_info->piece_points;

/* Set request_nelmts. This is not normally set by the upper layers because selection I/O
* usually does not use strip mining (H5D__scatgath_write), and instead allocates buffers
* large enough for the entire I/O. Set request_nelmts to be large enough for all selected
* elements in this chunk because it must be at least that large */
nonexistent_dset_info.type_info.request_nelmts = nonexistent_dset_info.nelmts;

/* Perform the actual read operation from the nonexistent chunk
*/
nonexistent_dset_info.file_space = chunk_info->fspace;
nonexistent_dset_info.mem_space = chunk_info->mspace;
nonexistent_dset_info.nelmts = chunk_info->piece_points;
if ((dset_info->io_ops.single_read)(&nonexistent_io_info, &nonexistent_dset_info) < 0)
HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked read failed");
} /* end if */
Expand Down Expand Up @@ -2866,9 +2875,10 @@ H5D__chunk_read(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)

/* Perform the actual read operation */
assert(chk_io_info->count == 1);
chk_io_info->dsets_info[0].file_space = chunk_info->fspace;
chk_io_info->dsets_info[0].mem_space = chunk_info->mspace;
chk_io_info->dsets_info[0].nelmts = chunk_info->piece_points;
chk_io_info->dsets_info[0].layout_io_info.contig_piece_info = chunk_info;
chk_io_info->dsets_info[0].file_space = chunk_info->fspace;
chk_io_info->dsets_info[0].mem_space = chunk_info->mspace;
chk_io_info->dsets_info[0].nelmts = chunk_info->piece_points;
if ((dset_info->io_ops.single_read)(chk_io_info, &chk_io_info->dsets_info[0]) < 0)
HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked read failed");

Expand Down Expand Up @@ -3055,10 +3065,19 @@ H5D__chunk_write(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)
/* Set up the storage buffer information for this chunk */
cpt_store.compact.buf = chunk;

/* Set up compact dataset info for write to cached chunk */
cpt_dset_info.layout_io_info.contig_piece_info = chunk_info;
cpt_dset_info.file_space = chunk_info->fspace;
cpt_dset_info.mem_space = chunk_info->mspace;
cpt_dset_info.nelmts = chunk_info->piece_points;

/* Set request_nelmts. This is not normally set by the upper layers because selection I/O
* usually does not use strip mining (H5D__scatgath_write), and instead allocates buffers
* large enough for the entire I/O. Set request_nelmts to be large enough for all selected
* elements in this chunk because it must be at least that large */
cpt_dset_info.type_info.request_nelmts = cpt_dset_info.nelmts;

/* Perform the actual write operation */
cpt_dset_info.file_space = chunk_info->fspace;
cpt_dset_info.mem_space = chunk_info->mspace;
cpt_dset_info.nelmts = chunk_info->piece_points;
if ((dset_info->io_ops.single_write)(&cpt_io_info, &cpt_dset_info) < 0)
HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked write failed");

Expand Down Expand Up @@ -3253,9 +3272,10 @@ H5D__chunk_write(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)

/* Perform the actual write operation */
assert(chk_io_info->count == 1);
chk_io_info->dsets_info[0].file_space = chunk_info->fspace;
chk_io_info->dsets_info[0].mem_space = chunk_info->mspace;
chk_io_info->dsets_info[0].nelmts = chunk_info->piece_points;
chk_io_info->dsets_info[0].layout_io_info.contig_piece_info = chunk_info;
chk_io_info->dsets_info[0].file_space = chunk_info->fspace;
chk_io_info->dsets_info[0].mem_space = chunk_info->mspace;
chk_io_info->dsets_info[0].nelmts = chunk_info->piece_points;
if ((dset_info->io_ops.single_write)(chk_io_info, &chk_io_info->dsets_info[0]) < 0)
HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunked write failed");

Expand Down
5 changes: 3 additions & 2 deletions src/H5Dcompact.c
Original file line number Diff line number Diff line change
Expand Up @@ -233,8 +233,9 @@ H5D__compact_io_init(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)
{
FUNC_ENTER_PACKAGE_NOERR

dinfo->store->compact.buf = dinfo->dset->shared->layout.storage.u.compact.buf;
dinfo->store->compact.dirty = &dinfo->dset->shared->layout.storage.u.compact.dirty;
dinfo->store->compact.buf = dinfo->dset->shared->layout.storage.u.compact.buf;
dinfo->store->compact.dirty = &dinfo->dset->shared->layout.storage.u.compact.dirty;
dinfo->layout_io_info.contig_piece_info = NULL;

/* Disable selection I/O */
io_info->use_select_io = H5D_SELECTION_IO_MODE_OFF;
Expand Down
3 changes: 3 additions & 0 deletions src/H5Defl.c
Original file line number Diff line number Diff line change
Expand Up @@ -198,6 +198,9 @@ H5D__efl_io_init(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)

H5MM_memcpy(&dinfo->store->efl, &(dinfo->dset->shared->dcpl_cache.efl), sizeof(H5O_efl_t));

/* No "pieces" selected */
dinfo->layout_io_info.contig_piece_info = NULL;

/* Disable selection I/O */
io_info->use_select_io = H5D_SELECTION_IO_MODE_OFF;
io_info->no_selection_io_cause |= H5D_SEL_IO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET;
Expand Down
30 changes: 16 additions & 14 deletions src/H5Dio.c
Original file line number Diff line number Diff line change
Expand Up @@ -415,8 +415,14 @@ H5D__read(size_t count, H5D_dset_io_info_t *dset_info)
HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode");

/* Only report the collective I/O mode if we're actually performing collective I/O */
if (xfer_mode == H5FD_MPIO_COLLECTIVE)
if (xfer_mode == H5FD_MPIO_COLLECTIVE) {
H5CX_set_mpio_actual_io_mode(io_info.actual_io_mode);

/* If we did selection I/O, report that we used "link chunk" mode, since that's the most
* analogous to what selection I/O does */
if (io_info.use_select_io == H5D_SELECTION_IO_MODE_ON)
H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_LINK_CHUNK);
}
}
#endif /* H5_HAVE_PARALLEL */
}
Expand Down Expand Up @@ -1131,20 +1137,16 @@ H5D__typeinfo_init_phase2(H5D_io_info_t *io_info)
assert(io_info);

/* If selection I/O mode is default (auto), enable it here if the VFD supports it (it will be turned off
* later if something else conflicts), otherwise disable it. If we're using the MPIO VFD, the automatic
* selection will happen in H5D__mpio_opt_possible() inside H5D__ioinfo_adjust(). */
#ifdef H5_HAVE_PARALLEL
if (!io_info->using_mpi_vfd)
#endif /* H5_HAVE_PARALLEL */
if (io_info->use_select_io == H5D_SELECTION_IO_MODE_DEFAULT) {
if (H5F_has_vector_select_io(io_info->dsets_info[0].dset->oloc.file,
io_info->op_type == H5D_IO_OP_WRITE))
io_info->use_select_io = H5D_SELECTION_IO_MODE_ON;
else {
io_info->use_select_io = H5D_SELECTION_IO_MODE_OFF;
io_info->no_selection_io_cause |= H5D_SEL_IO_DEFAULT_OFF;
}
* later if something else conflicts), otherwise disable it */
if (io_info->use_select_io == H5D_SELECTION_IO_MODE_DEFAULT) {
if (H5F_has_vector_select_io(io_info->dsets_info[0].dset->oloc.file,
io_info->op_type == H5D_IO_OP_WRITE))
io_info->use_select_io = H5D_SELECTION_IO_MODE_ON;
else {
io_info->use_select_io = H5D_SELECTION_IO_MODE_OFF;
io_info->no_selection_io_cause |= H5D_SEL_IO_DEFAULT_OFF;
}
}

/* If we're doing type conversion and we might be doing selection I/O, check if the buffers are large
* enough to handle the whole I/O */
Expand Down
22 changes: 3 additions & 19 deletions src/H5Dmpio.c
Original file line number Diff line number Diff line change
Expand Up @@ -622,6 +622,9 @@ H5D__mpio_opt_possible(H5D_io_info_t *io_info)
if (!H5FD_mpi_opt_types_g)
local_cause[0] |= H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED;

/* Decision on whether to use selection I/O should have been made by now */
assert(io_info->use_select_io != H5D_SELECTION_IO_MODE_DEFAULT);

/* Datatype conversions and transformations are allowed with selection I/O. If the selection I/O mode
* is auto (default), disable collective for now and re-enable later if we can */
if (io_info->use_select_io != H5D_SELECTION_IO_MODE_ON) {
Expand Down Expand Up @@ -731,25 +734,6 @@ H5D__mpio_opt_possible(H5D_io_info_t *io_info)
HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code)
} /* end else */

/* If the selection I/O mode is default (auto), decide here whether it should be on or off */
if (io_info->use_select_io == H5D_SELECTION_IO_MODE_DEFAULT) {
/* If the only reason(s) we've disabled collective are type conversions and/or transforms, enable
* selection I/O and re-enable collective I/O since it's supported by selection I/O */
if (global_cause[0] && !(global_cause[0] & ~((unsigned)H5D_MPIO_DATATYPE_CONVERSION |
(unsigned)H5D_MPIO_DATA_TRANSFORMS))) {
assert(!(local_cause[0] &
~((unsigned)H5D_MPIO_DATATYPE_CONVERSION | (unsigned)H5D_MPIO_DATA_TRANSFORMS)));
local_cause[0] = 0;
global_cause[0] = 0;
io_info->use_select_io = H5D_SELECTION_IO_MODE_ON;
}
else {
/* Otherwise, there's currently no benefit to selection I/O, so leave it off */
io_info->use_select_io = H5D_SELECTION_IO_MODE_OFF;
io_info->no_selection_io_cause |= H5D_SEL_IO_DEFAULT_OFF;
}
}

/* Set the local & global values of no-collective-cause in the API context */
H5CX_set_mpio_local_no_coll_cause(local_cause[0]);
H5CX_set_mpio_global_no_coll_cause(global_cause[0]);
Expand Down
Loading