summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorjhendersonHDF <jhenderson@hdfgroup.org>2023-09-01 12:20:05 (GMT)
committerGitHub <noreply@github.com>2023-09-01 12:20:05 (GMT)
commitfdeac5309eceeed5159dcbc60de4161f144593ef (patch)
treeef03b76215d539dbf4fb1249910b66327886d821
parent26c8d0c32ed7f3db00ff8d997e5606e07666e846 (diff)
downloadhdf5-fdeac5309eceeed5159dcbc60de4161f144593ef.zip
hdf5-fdeac5309eceeed5159dcbc60de4161f144593ef.tar.gz
hdf5-fdeac5309eceeed5159dcbc60de4161f144593ef.tar.bz2
Fix serial to parallel chunked dataset file space allocation bug (#3394) (#3467)
-rw-r--r--release_docs/RELEASE.txt11
-rw-r--r--src/H5Dint.c35
-rw-r--r--testpar/t_chunk_alloc.c277
-rw-r--r--testpar/testphdf5.c2
-rw-r--r--testpar/testphdf5.h1
5 files changed, 315 insertions, 11 deletions
diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt
index 9b8946d..1059a46 100644
--- a/release_docs/RELEASE.txt
+++ b/release_docs/RELEASE.txt
@@ -109,6 +109,17 @@ Bug Fixes since HDF5-1.14.2 release
===================================
Library
-------
+ - Fixed a file space allocation bug in the parallel library for chunked
+ datasets
+
+ With the addition of support for incremental file space allocation for
+ chunked datasets with filters applied to them that are created/accessed
+ in parallel, a bug was introduced to the library's parallel file space
+ allocation code. This could cause file space to not be allocated correctly
+ for datasets without filters applied to them that are created with serial
+ file access and later opened with parallel file access. In turn, this could
+ cause parallel writes to those datasets to place incorrect data in the file.
+
- Fixed an assertion failure in Parallel HDF5 when a file can't be created
due to an invalid library version bounds setting
diff --git a/src/H5Dint.c b/src/H5Dint.c
index e6a709e..a547a72 100644
--- a/src/H5Dint.c
+++ b/src/H5Dint.c
@@ -1664,12 +1664,13 @@ done:
static herr_t
H5D__open_oid(H5D_t *dataset, hid_t dapl_id)
{
- H5P_genplist_t *plist; /* Property list */
- H5O_fill_t *fill_prop; /* Pointer to dataset's fill value info */
- unsigned alloc_time_state; /* Allocation time state */
- htri_t msg_exists; /* Whether a particular type of message exists */
- hbool_t layout_init = FALSE; /* Flag to indicate that chunk information was initialized */
- herr_t ret_value = SUCCEED; /* Return value */
+ H5P_genplist_t *plist; /* Property list */
+ H5O_fill_t *fill_prop; /* Pointer to dataset's fill value info */
+ unsigned alloc_time_state; /* Allocation time state */
+ htri_t msg_exists; /* Whether a particular type of message exists */
+ hbool_t layout_init = FALSE; /* Flag to indicate that chunk information was initialized */
+ hbool_t must_init_storage = FALSE;
+ herr_t ret_value = SUCCEED; /* Return value */
FUNC_ENTER_PACKAGE_TAG(dataset->oloc.addr)
@@ -1812,12 +1813,24 @@ H5D__open_oid(H5D_t *dataset, hid_t dapl_id)
* Make sure all storage is properly initialized.
* This is important only for parallel I/O where the space must
* be fully allocated before I/O can happen.
+ *
+ * Storage will be initialized here if either the VFD being used
+ * has set the H5FD_FEAT_ALLOCATE_EARLY flag to indicate that it
+ * wishes to force early space allocation OR a parallel VFD is
+ * being used and the dataset in question doesn't have any filters
+ * applied to it. If filters are applied to the dataset, collective
+ * I/O will be required when writing to the dataset, so we don't
+ * need to initialize storage here, as the collective I/O process
+ * will coordinate that.
*/
- if ((H5F_INTENT(dataset->oloc.file) & H5F_ACC_RDWR) &&
- !(*dataset->shared->layout.ops->is_space_alloc)(&dataset->shared->layout.storage) &&
- H5F_HAS_FEATURE(dataset->oloc.file, H5FD_FEAT_ALLOCATE_EARLY))
- if (H5D__alloc_storage(dataset, H5D_ALLOC_OPEN, FALSE, NULL) < 0)
- HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize file storage");
+ must_init_storage = (H5F_INTENT(dataset->oloc.file) & H5F_ACC_RDWR) &&
+ !(*dataset->shared->layout.ops->is_space_alloc)(&dataset->shared->layout.storage);
+ must_init_storage = must_init_storage && (H5F_HAS_FEATURE(dataset->oloc.file, H5FD_FEAT_ALLOCATE_EARLY) ||
+ (H5F_HAS_FEATURE(dataset->oloc.file, H5FD_FEAT_HAS_MPI) &&
+ dataset->shared->dcpl_cache.pline.nused == 0));
+
+ if (must_init_storage && (H5D__alloc_storage(dataset, H5D_ALLOC_OPEN, FALSE, NULL) < 0))
+ HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize file storage");
done:
if (ret_value < 0) {
diff --git a/testpar/t_chunk_alloc.c b/testpar/t_chunk_alloc.c
index 6f862c0..0ffe695 100644
--- a/testpar/t_chunk_alloc.c
+++ b/testpar/t_chunk_alloc.c
@@ -480,3 +480,280 @@ test_chunk_alloc(void)
/* reopen dataset in parallel, read and verify the data */
verify_data(filename, CHUNK_FACTOR, all, CLOSE, &file_id, &dataset);
}
+
+/*
+ * A test to verify the following:
+ *
+ * - That the library forces allocation of all space in the file
+ * for a chunked dataset opened with parallel file access when
+ * that dataset:
+ *
+ * - was created with serial file access
+ * - was created with the default incremental file space
+ * allocation time
+ * - has no filters applied to it
+ *
+ * In this case, the library has to ensure that all the file
+ * space for the dataset is allocated so that the MPI processes
+ * can write to chunks independently of each other and still have
+ * a consistent view of the file.
+ *
+ * - That the library DOES NOT force allocation of all space in
+ * the file for a chunked dataset opened with parallel file access
+ * when that dataset:
+ *
+ * - was created with serial file access
+ * - was created with the default incremental file space
+ * allocation time
+ * - has filters applied to it
+ *
+ * In this case, writes to the dataset are required to be collective,
+ * so file space can be allocated incrementally in a coordinated
+ * fashion.
+ */
+void
+test_chunk_alloc_incr_ser_to_par(void)
+{
+ H5D_space_status_t space_status;
+ const char *filename;
+ hsize_t dset_dims[1];
+ hsize_t start[1];
+ hsize_t stride[1];
+ hsize_t count[1];
+ hsize_t block[1];
+ hsize_t alloc_size;
+ size_t nchunks;
+ herr_t ret;
+ hid_t fid = H5I_INVALID_HID;
+ hid_t fapl_id = H5I_INVALID_HID;
+ hid_t dset_id = H5I_INVALID_HID;
+ hid_t fspace_id = H5I_INVALID_HID;
+ hid_t dxpl_id = H5I_INVALID_HID;
+ int *data = NULL;
+ int *correct_data = NULL;
+ int *read_data = NULL;
+
+ MPI_Comm_size(MPI_COMM_WORLD, &mpi_size);
+ MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
+
+ filename = (const char *)GetTestParameters();
+ if (MAINPROCESS && VERBOSE_MED)
+ printf("Chunked dataset incremental file space allocation serial to parallel test on file %s\n",
+ filename);
+
+ nchunks = (size_t)(CHUNK_FACTOR * mpi_size);
+ dset_dims[0] = (hsize_t)(nchunks * CHUNK_SIZE);
+
+ if (mpi_rank == 0) {
+ hsize_t chunk_dims[1] = {CHUNK_SIZE};
+ hid_t space_id = H5I_INVALID_HID;
+ hid_t dcpl_id = H5I_INVALID_HID;
+
+ fid = H5Fcreate(filename, H5F_ACC_TRUNC, H5P_DEFAULT, H5P_DEFAULT);
+ VRFY((fid >= 0), "H5Fcreate");
+
+ dcpl_id = H5Pcreate(H5P_DATASET_CREATE);
+ VRFY((dcpl_id >= 0), "H5Pcreate");
+
+ ret = H5Pset_chunk(dcpl_id, 1, chunk_dims);
+ VRFY((ret == SUCCEED), "H5Pset_chunk");
+
+ ret = H5Pset_alloc_time(dcpl_id, H5D_ALLOC_TIME_INCR);
+ VRFY((ret == SUCCEED), "H5Pset_alloc_time");
+
+ space_id = H5Screate_simple(1, dset_dims, NULL);
+ VRFY((space_id >= 0), "H5Screate_simple");
+
+ /* Create a chunked dataset without a filter applied to it */
+ dset_id =
+ H5Dcreate2(fid, "dset_no_filter", H5T_NATIVE_INT, space_id, H5P_DEFAULT, dcpl_id, H5P_DEFAULT);
+ VRFY((dset_id >= 0), "H5Dcreate2");
+
+ ret = H5Dclose(dset_id);
+ VRFY((ret == SUCCEED), "H5Dclose");
+
+ /* Create a chunked dataset with a filter applied to it */
+ ret = H5Pset_shuffle(dcpl_id);
+ VRFY((ret == SUCCEED), "H5Pset_shuffle");
+
+ dset_id = H5Dcreate2(fid, "dset_filter", H5T_NATIVE_INT, space_id, H5P_DEFAULT, dcpl_id, H5P_DEFAULT);
+ VRFY((dset_id >= 0), "H5Dcreate2");
+
+ ret = H5Dclose(dset_id);
+ VRFY((ret == SUCCEED), "H5Dclose");
+ ret = H5Pclose(dcpl_id);
+ VRFY((ret == SUCCEED), "H5Pclose");
+ ret = H5Sclose(space_id);
+ VRFY((ret == SUCCEED), "H5Sclose");
+ ret = H5Fclose(fid);
+ VRFY((ret == SUCCEED), "H5Fclose");
+ }
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ fapl_id = H5Pcreate(H5P_FILE_ACCESS);
+ VRFY((fapl_id >= 0), "H5Pcreate");
+
+ ret = H5Pset_fapl_mpio(fapl_id, MPI_COMM_WORLD, MPI_INFO_NULL);
+ VRFY((ret == SUCCEED), "H5Pset_fapl_mpio");
+
+ fid = H5Fopen(filename, H5F_ACC_RDWR, fapl_id);
+ VRFY((fid >= 0), "H5Fopen");
+
+ data = malloc((dset_dims[0] / (hsize_t)mpi_size) * sizeof(int));
+ VRFY(data, "malloc");
+ read_data = malloc(dset_dims[0] * sizeof(int));
+ VRFY(read_data, "malloc");
+ correct_data = malloc(dset_dims[0] * sizeof(int));
+ VRFY(correct_data, "malloc");
+
+ /*
+ * Check the file space allocation status/size and dataset
+ * data before and after writing to the dataset without a
+ * filter
+ */
+ dset_id = H5Dopen2(fid, "dset_no_filter", H5P_DEFAULT);
+ VRFY((dset_id >= 0), "H5Dopen2");
+
+ ret = H5Dget_space_status(dset_id, &space_status);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ VRFY((space_status == H5D_SPACE_STATUS_ALLOCATED), "file space allocation status verification succeeded");
+
+ alloc_size = H5Dget_storage_size(dset_id);
+ VRFY(((dset_dims[0] * sizeof(int)) == alloc_size), "file space allocation size verification succeeded");
+
+ memset(read_data, 255, dset_dims[0] * sizeof(int));
+ memset(correct_data, 0, dset_dims[0] * sizeof(int));
+
+ ret = H5Dread(dset_id, H5T_NATIVE_INT, H5S_BLOCK, H5S_ALL, H5P_DEFAULT, read_data);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ VRFY((0 == memcmp(read_data, correct_data, dset_dims[0] * sizeof(int))), "data verification succeeded");
+
+ fspace_id = H5Dget_space(dset_id);
+ VRFY((ret == SUCCEED), "H5Dget_space");
+
+ start[0] = ((hsize_t)mpi_rank * (dset_dims[0] / (hsize_t)mpi_size));
+ stride[0] = 1;
+ count[0] = (dset_dims[0] / (hsize_t)mpi_size);
+ block[0] = 1;
+
+ ret = H5Sselect_hyperslab(fspace_id, H5S_SELECT_SET, start, stride, count, block);
+ VRFY((ret == SUCCEED), "H5Sselect_hyperslab");
+
+ memset(data, 255, (dset_dims[0] / (hsize_t)mpi_size) * sizeof(int));
+
+ ret = H5Dwrite(dset_id, H5T_NATIVE_INT, H5S_BLOCK, fspace_id, H5P_DEFAULT, data);
+ VRFY((ret == SUCCEED), "H5Dwrite");
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ ret = H5Dget_space_status(dset_id, &space_status);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ VRFY((space_status == H5D_SPACE_STATUS_ALLOCATED), "file space allocation status verification succeeded");
+
+ alloc_size = H5Dget_storage_size(dset_id);
+ VRFY(((dset_dims[0] * sizeof(int)) == alloc_size), "file space allocation size verification succeeded");
+
+ memset(read_data, 0, dset_dims[0] * sizeof(int));
+ memset(correct_data, 255, dset_dims[0] * sizeof(int));
+
+ ret = H5Dread(dset_id, H5T_NATIVE_INT, H5S_BLOCK, H5S_ALL, H5P_DEFAULT, read_data);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ VRFY((0 == memcmp(read_data, correct_data, dset_dims[0] * sizeof(int))), "data verification succeeded");
+
+ ret = H5Sclose(fspace_id);
+ VRFY((ret == SUCCEED), "H5Sclose");
+ ret = H5Dclose(dset_id);
+ VRFY((ret == SUCCEED), "H5Dclose");
+
+ /*
+ * Check the file space allocation status/size and dataset
+ * data before and after writing to the dataset with a
+ * filter
+ */
+ dset_id = H5Dopen2(fid, "dset_filter", H5P_DEFAULT);
+ VRFY((dset_id >= 0), "H5Dopen2");
+
+ ret = H5Dget_space_status(dset_id, &space_status);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ VRFY((space_status == H5D_SPACE_STATUS_NOT_ALLOCATED),
+ "file space allocation status verification succeeded");
+
+ alloc_size = H5Dget_storage_size(dset_id);
+ VRFY((0 == alloc_size), "file space allocation size verification succeeded");
+
+ memset(read_data, 255, dset_dims[0] * sizeof(int));
+ memset(correct_data, 0, dset_dims[0] * sizeof(int));
+
+ ret = H5Dread(dset_id, H5T_NATIVE_INT, H5S_BLOCK, H5S_ALL, H5P_DEFAULT, read_data);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ VRFY((0 == memcmp(read_data, correct_data, dset_dims[0] * sizeof(int))), "data verification succeeded");
+
+ fspace_id = H5Dget_space(dset_id);
+ VRFY((ret == SUCCEED), "H5Dget_space");
+
+ start[0] = ((hsize_t)mpi_rank * (dset_dims[0] / (hsize_t)mpi_size));
+ stride[0] = 1;
+ count[0] = (dset_dims[0] / (hsize_t)mpi_size);
+ block[0] = 1;
+
+ ret = H5Sselect_hyperslab(fspace_id, H5S_SELECT_SET, start, stride, count, block);
+ VRFY((ret == SUCCEED), "H5Sselect_hyperslab");
+
+ memset(data, 255, (dset_dims[0] / (hsize_t)mpi_size) * sizeof(int));
+
+ dxpl_id = H5Pcreate(H5P_DATASET_XFER);
+ VRFY((dxpl_id >= 0), "H5Pcreate");
+
+ ret = H5Pset_dxpl_mpio(dxpl_id, H5FD_MPIO_COLLECTIVE);
+ VRFY((ret == SUCCEED), "H5Pset_dxpl_mpio");
+
+ ret = H5Dwrite(dset_id, H5T_NATIVE_INT, H5S_BLOCK, fspace_id, dxpl_id, data);
+ VRFY((ret == SUCCEED), "H5Dwrite");
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ ret = H5Dget_space_status(dset_id, &space_status);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ VRFY((space_status == H5D_SPACE_STATUS_ALLOCATED), "file space allocation status verification succeeded");
+
+ alloc_size = H5Dget_storage_size(dset_id);
+ VRFY(((dset_dims[0] * sizeof(int)) == alloc_size), "file space allocation size verification succeeded");
+
+ memset(read_data, 0, dset_dims[0] * sizeof(int));
+ memset(correct_data, 255, dset_dims[0] * sizeof(int));
+
+ ret = H5Dread(dset_id, H5T_NATIVE_INT, H5S_BLOCK, H5S_ALL, H5P_DEFAULT, read_data);
+ VRFY((ret == SUCCEED), "H5Dread");
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ VRFY((0 == memcmp(read_data, correct_data, dset_dims[0] * sizeof(int))), "data verification succeeded");
+
+ ret = H5Pclose(dxpl_id);
+ VRFY((ret == SUCCEED), "H5Pclose");
+ ret = H5Sclose(fspace_id);
+ VRFY((ret == SUCCEED), "H5Sclose");
+ ret = H5Dclose(dset_id);
+ VRFY((ret == SUCCEED), "H5Dclose");
+
+ free(correct_data);
+ free(read_data);
+ free(data);
+
+ H5Pclose(fapl_id);
+ H5Fclose(fid);
+}
diff --git a/testpar/testphdf5.c b/testpar/testphdf5.c
index 1cca1fd..430b8a6 100644
--- a/testpar/testphdf5.c
+++ b/testpar/testphdf5.c
@@ -376,6 +376,8 @@ main(int argc, char **argv)
AddTest("selnone", none_selection_chunk, NULL, "chunked dataset with none-selection", PARATESTFILE);
AddTest("calloc", test_chunk_alloc, NULL, "parallel extend Chunked allocation on serial file",
PARATESTFILE);
+ AddTest("chkallocser2par", test_chunk_alloc_incr_ser_to_par, NULL,
+ "chunk allocation from serial to parallel file access", PARATESTFILE);
AddTest("fltread", test_filter_read, NULL, "parallel read of dataset written serially with filters",
PARATESTFILE);
diff --git a/testpar/testphdf5.h b/testpar/testphdf5.h
index 820fdc3..631ad44 100644
--- a/testpar/testphdf5.h
+++ b/testpar/testphdf5.h
@@ -256,6 +256,7 @@ void none_selection_chunk(void);
void actual_io_mode_tests(void);
void no_collective_cause_tests(void);
void test_chunk_alloc(void);
+void test_chunk_alloc_incr_ser_to_par(void);
void test_filter_read(void);
void compact_dataset(void);
void null_dataset(void);