Merge pull request #1036 from fortnern/parallel_selection_io

Implement parallel collective support for selection I/O.

14 files changed, 224 insertions, 151 deletions

diff --git a/src/H5.c b/src/H5.c
index 0c7c8c1..f97fe31 100644
--- a/src/H5.c
+++ b/src/H5.c
@@ -147,8 +147,8 @@ done:
 herr_t
 H5_init_library(void)
 {
-    char *env_use_select_io = NULL;
-    herr_t ret_value = SUCCEED;
+    char * env_use_select_io = NULL;
+    herr_t ret_value         = SUCCEED;
 
     /* Set the 'library initialized' flag as early as possible, to avoid
      * possible re-entrancy.
@@ -281,11 +281,10 @@ H5_init_library(void)
 
     /* Check for HDF5_USE_SELECTION_IO env variable */
     env_use_select_io = HDgetenv("HDF5_USE_SELECTION_IO");
-    if (NULL != env_use_select_io && HDstrcmp(env_use_select_io, "")
-        && HDstrcmp(env_use_select_io, "0") && HDstrcmp(env_use_select_io, "no")
-        && HDstrcmp(env_use_select_io, "No") && HDstrcmp(env_use_select_io, "NO")
-        && HDstrcmp(env_use_select_io, "false") && HDstrcmp(env_use_select_io, "False")
-        && HDstrcmp(env_use_select_io, "FALSE"))
+    if (NULL != env_use_select_io && HDstrcmp(env_use_select_io, "") && HDstrcmp(env_use_select_io, "0") &&
+        HDstrcmp(env_use_select_io, "no") && HDstrcmp(env_use_select_io, "No") &&
+        HDstrcmp(env_use_select_io, "NO") && HDstrcmp(env_use_select_io, "false") &&
+        HDstrcmp(env_use_select_io, "False") && HDstrcmp(env_use_select_io, "FALSE"))
         H5_use_selection_io_g = TRUE;
 
     /* Debugging? */
diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c
index dcc3baa..8de023d 100644
--- a/src/H5Dchunk.c
+++ b/src/H5Dchunk.c
@@ -254,9 +254,8 @@ typedef struct H5D_chunk_coll_info_t {
 /* Chunked layout operation callbacks */
 static herr_t H5D__chunk_construct(H5F_t *f, H5D_t *dset);
 static herr_t H5D__chunk_init(H5F_t *f, const H5D_t *dset, hid_t dapl_id);
-static herr_t H5D__chunk_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
-                                 hsize_t nelmts, const H5S_t *file_space, const H5S_t *mem_space,
-                                 H5D_chunk_map_t *fm);
+static herr_t H5D__chunk_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
+                                 const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t *fm);
 static herr_t H5D__chunk_io_init_selections(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
                                             H5D_chunk_map_t *fm);
 static herr_t H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
@@ -1064,10 +1063,10 @@ H5D__chunk_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsi
     hssize_t     old_offset[H5O_LAYOUT_NDIMS];  /* Old selection offset */
     htri_t       file_space_normalized = FALSE; /* File dataspace was normalized */
     unsigned     f_ndims;                       /* The number of dimensions of the file's dataspace */
-    int          sm_ndims;            /* The number of dimensions of the memory buffer's dataspace (signed) */
-    htri_t       use_selection_io = FALSE;      /* Whether to use selection I/O */
-    unsigned     u;                   /* Local index variable */
-    herr_t       ret_value = SUCCEED; /* Return value        */
+    int          sm_ndims; /* The number of dimensions of the memory buffer's dataspace (signed) */
+    htri_t       use_selection_io = FALSE; /* Whether to use selection I/O */
+    unsigned     u;                        /* Local index variable */
+    herr_t       ret_value = SUCCEED;      /* Return value        */
 
     FUNC_ENTER_STATIC
 
@@ -2541,7 +2540,7 @@ H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_
     H5S_t *       chunk_file_spaces_static[8]; /* Static buffer for chunk_file_spaces */
     haddr_t *     chunk_addrs = NULL;          /* Array of chunk addresses */
     haddr_t       chunk_addrs_static[8];       /* Static buffer for chunk_addrs */
-    herr_t        ret_value        = SUCCEED;  /*return value        */
+    herr_t        ret_value = SUCCEED;         /*return value        */
 
     FUNC_ENTER_STATIC
 
@@ -2650,10 +2649,10 @@ H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_
         /* Issue selection I/O call (we can skip the page buffer because we've
          * already verified it won't be used, and the metadata accumulator
          * because this is raw data) */
-        if (num_chunks > 0 &&
-            H5F_shared_select_read(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, (uint32_t)num_chunks,
-                                   (const H5S_t * const *)chunk_mem_spaces, (const H5S_t * const *)chunk_file_spaces,
-                                   chunk_addrs, element_sizes, bufs) < 0)
+        if (H5F_shared_select_read(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, (uint32_t)num_chunks,
+                                   (const H5S_t *const *)chunk_mem_spaces,
+                                   (const H5S_t *const *)chunk_file_spaces, chunk_addrs, element_sizes,
+                                   bufs) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunk selection read failed")
 
         /* Clean up memory */
@@ -2816,7 +2815,7 @@ H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize
     H5S_t *       chunk_file_spaces_static[8]; /* Static buffer for chunk_file_spaces */
     haddr_t *     chunk_addrs = NULL;          /* Array of chunk addresses */
     haddr_t       chunk_addrs_static[8];       /* Static buffer for chunk_addrs */
-    herr_t        ret_value        = SUCCEED;  /* Return value        */
+    herr_t        ret_value = SUCCEED;         /* Return value        */
 
     FUNC_ENTER_STATIC
 
@@ -2985,10 +2984,10 @@ H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize
         /* Issue selection I/O call (we can skip the page buffer because we've
          * already verified it won't be used, and the metadata accumulator
          * because this is raw data) */
-        if (num_chunks > 0 && H5F_shared_select_write(
-                                  H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, (uint32_t)num_chunks,
-                                  (const H5S_t * const *)chunk_mem_spaces, (const H5S_t * const *)chunk_file_spaces,
-                                  chunk_addrs, element_sizes, bufs) < 0)
+        if (H5F_shared_select_write(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, (uint32_t)num_chunks,
+                                    (const H5S_t *const *)chunk_mem_spaces,
+                                    (const H5S_t *const *)chunk_file_spaces, chunk_addrs, element_sizes,
+                                    bufs) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "chunk selection read failed")
 
         /* Clean up memory */
diff --git a/src/H5Dcompact.c b/src/H5Dcompact.c
index 9a6d4b2..f68a93a 100644
--- a/src/H5Dcompact.c
+++ b/src/H5Dcompact.c
@@ -54,9 +54,8 @@
 /* Layout operation callbacks */
 static herr_t  H5D__compact_construct(H5F_t *f, H5D_t *dset);
 static hbool_t H5D__compact_is_space_alloc(const H5O_storage_t *storage);
-static herr_t  H5D__compact_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
-                                    hsize_t nelmts, const H5S_t *file_space, const H5S_t *mem_space,
-                                    H5D_chunk_map_t *cm);
+static herr_t  H5D__compact_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
+                                    const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t *cm);
 static ssize_t H5D__compact_readvv(const H5D_io_info_t *io_info, size_t dset_max_nseq, size_t *dset_curr_seq,
                                    size_t dset_size_arr[], hsize_t dset_offset_arr[], size_t mem_max_nseq,
                                    size_t *mem_curr_seq, size_t mem_size_arr[], hsize_t mem_offset_arr[]);
diff --git a/src/H5Dcontig.c b/src/H5Dcontig.c
index d2a84da..d7ebbee 100644
--- a/src/H5Dcontig.c
+++ b/src/H5Dcontig.c
@@ -91,9 +91,8 @@ typedef struct H5D_contig_writevv_ud_t {
 /* Layout operation callbacks */
 static herr_t  H5D__contig_construct(H5F_t *f, H5D_t *dset);
 static herr_t  H5D__contig_init(H5F_t *f, const H5D_t *dset, hid_t dapl_id);
-static herr_t  H5D__contig_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
-                                   hsize_t nelmts, const H5S_t *file_space, const H5S_t *mem_space,
-                                   H5D_chunk_map_t *cm);
+static herr_t  H5D__contig_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
+                                   const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t *cm);
 static ssize_t H5D__contig_readvv(const H5D_io_info_t *io_info, size_t dset_max_nseq, size_t *dset_curr_seq,
                                   size_t dset_len_arr[], hsize_t dset_offset_arr[], size_t mem_max_nseq,
                                   size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[]);
@@ -555,8 +554,8 @@ H5D__contig_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED
                     hsize_t H5_ATTR_UNUSED nelmts, const H5S_t H5_ATTR_UNUSED *file_space,
                     const H5S_t H5_ATTR_UNUSED *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *cm)
 {
-    htri_t use_selection_io = FALSE; /* Whether to use selection I/O */
-    htri_t ret_value = SUCCEED;      /* Return value */
+    htri_t use_selection_io = FALSE;   /* Whether to use selection I/O */
+    htri_t ret_value        = SUCCEED; /* Return value */
 
     FUNC_ENTER_STATIC
 
@@ -601,10 +600,10 @@ H5D__contig_may_use_select_io(const H5D_io_info_t *io_info, H5D_io_op_type_t op_
     /* Don't use selection I/O if it's globally disabled, if there is a type
      * conversion, or if it's not a contiguous dataset, or if the sieve buffer
      * exists (write) or is dirty (read) */
-    if (!H5_use_selection_io_g || io_info->io_ops.single_read != H5D__select_read
-        || io_info->layout_ops.readvv != H5D__contig_readvv
-        || (op_type == H5D_IO_OP_READ && io_info->dset->shared->cache.contig.sieve_dirty)
-        || (op_type == H5D_IO_OP_WRITE && io_info->dset->shared->cache.contig.sieve_buf))
+    if (!H5_use_selection_io_g || io_info->io_ops.single_read != H5D__select_read ||
+        io_info->layout_ops.readvv != H5D__contig_readvv ||
+        (op_type == H5D_IO_OP_READ && io_info->dset->shared->cache.contig.sieve_dirty) ||
+        (op_type == H5D_IO_OP_WRITE && io_info->dset->shared->cache.contig.sieve_buf))
         ret_value = FALSE;
     else {
         htri_t page_buf_enabled;
@@ -641,7 +640,7 @@ herr_t
 H5D__contig_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
                  const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *fm)
 {
-    herr_t ret_value = SUCCEED;      /* Return value */
+    herr_t ret_value = SUCCEED; /* Return value */
 
     FUNC_ENTER_PACKAGE
 
@@ -657,8 +656,8 @@ H5D__contig_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize
 
         /* Issue selection I/O call (we can skip the page buffer because we've
          * already verified it won't be used, and the metadata accumulator
-         * because this is raw data) Only call funciton if nelmts > 0. */
-        if (nelmts > 0 && H5F_shared_select_read(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, 1,
+         * because this is raw data) */
+        if (H5F_shared_select_read(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, nelmts > 0 ? 1 : 0,
                                    &mem_space, &file_space, &(io_info->store->contig.dset_addr),
                                    &dst_type_size, &(io_info->u.rbuf)) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "contiguous selection read failed")
@@ -666,7 +665,7 @@ H5D__contig_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize
     else
         /* Read data through legacy (non-selection I/O) pathway */
         if ((io_info->io_ops.single_read)(io_info, type_info, nelmts, file_space, mem_space) < 0)
-            HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "contiguous read failed")
+        HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "contiguous read failed")
 
 done:
     FUNC_LEAVE_NOAPI(ret_value)
@@ -688,7 +687,7 @@ herr_t
 H5D__contig_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
                   const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t H5_ATTR_UNUSED *fm)
 {
-    herr_t ret_value = SUCCEED;      /* Return value */
+    herr_t ret_value = SUCCEED; /* Return value */
 
     FUNC_ENTER_PACKAGE
 
@@ -704,16 +703,16 @@ H5D__contig_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsiz
 
         /* Issue selection I/O call (we can skip the page buffer because we've
          * already verified it won't be used, and the metadata accumulator
-         * because this is raw data). Only call funciton if nelmts > 0. */
-        if (nelmts > 0 && H5F_shared_select_write(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, 1,
-                                   &mem_space, &file_space, &(io_info->store->contig.dset_addr),
-                                   &dst_type_size, &(io_info->u.wbuf)) < 0)
+         * because this is raw data) */
+        if (H5F_shared_select_write(H5F_SHARED(io_info->dset->oloc.file), H5FD_MEM_DRAW, nelmts > 0 ? 1 : 0,
+                                    &mem_space, &file_space, &(io_info->store->contig.dset_addr),
+                                    &dst_type_size, &(io_info->u.wbuf)) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "contiguous selection write failed")
     } /* end if */
     else
         /* Write data through legacy (non-selection I/O) pathway */
         if ((io_info->io_ops.single_write)(io_info, type_info, nelmts, file_space, mem_space) < 0)
-            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "contiguous write failed")
+        HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "contiguous write failed")
 
 done:
     FUNC_LEAVE_NOAPI(ret_value)
@@ -1649,4 +1648,3 @@ done:
 
     FUNC_LEAVE_NOAPI(ret_value)
 } /* end H5D__contig_copy() */
-
diff --git a/src/H5Defl.c b/src/H5Defl.c
index fea70b0..ce6d481 100644
--- a/src/H5Defl.c
+++ b/src/H5Defl.c
@@ -60,9 +60,9 @@ typedef struct H5D_efl_writevv_ud_t {
 /********************/
 
 /* Layout operation callbacks */
-static herr_t H5D__efl_construct(H5F_t *f, H5D_t *dset);
-static herr_t H5D__efl_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
-                               const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t *cm);
+static herr_t  H5D__efl_construct(H5F_t *f, H5D_t *dset);
+static herr_t  H5D__efl_io_init(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
+                                const H5S_t *file_space, const H5S_t *mem_space, H5D_chunk_map_t *cm);
 static ssize_t H5D__efl_readvv(const H5D_io_info_t *io_info, size_t dset_max_nseq, size_t *dset_curr_seq,
                                size_t dset_len_arr[], hsize_t dset_offset_arr[], size_t mem_max_nseq,
                                size_t *mem_curr_seq, size_t mem_len_arr[], hsize_t mem_offset_arr[]);
diff --git a/src/H5Dio.c b/src/H5Dio.c
index cc5f5bb..c7f8a3e 100644
--- a/src/H5Dio.c
+++ b/src/H5Dio.c
@@ -845,11 +845,8 @@ H5D__ioinfo_adjust(H5D_io_info_t *io_info, const H5D_t *dset, const H5S_t *file_
         H5CX_set_mpio_actual_io_mode(H5D_MPIO_NO_COLLECTIVE);
     } /* end if */
 
-    /* Make any parallel I/O adjustments.  Do not use collective code path if
-     * we're using selection I/O - in this case the file driver will handle it.
-     */
-     /* Check for selection/vector support in file driver? -NAF */
-    if (io_info->using_mpi_vfd /*&& !H5_use_selection_io_g*/) {
+    /* Make any parallel I/O adjustments */
+    if (io_info->using_mpi_vfd) {
         H5FD_mpio_xfer_t xfer_mode; /* Parallel transfer for this request */
         htri_t           opt;       /* Flag whether a selection is optimizable */
 
@@ -867,12 +864,17 @@ H5D__ioinfo_adjust(H5D_io_info_t *io_info, const H5D_t *dset, const H5S_t *file_
 
         /* Check if we can use the optimized parallel I/O routines */
         if (opt == TRUE) {
-            /* Override the I/O op pointers to the MPI-specific routines */
-            io_info->io_ops.multi_read   = dset->shared->layout.ops->par_read;
-            io_info->io_ops.multi_write  = dset->shared->layout.ops->par_write;
-            io_info->io_ops.single_read  = H5D__mpio_select_read;
-            io_info->io_ops.single_write = H5D__mpio_select_write;
-        } /* end if */
+            /* Override the I/O op pointers to the MPI-specific routines, unless
+             * selection I/O is to be used - in this case the file driver will
+             * handle collective I/O */
+            /* Check for selection/vector support in file driver? -NAF */
+            if (!io_info->use_select_io) {
+                io_info->io_ops.multi_read   = dset->shared->layout.ops->par_read;
+                io_info->io_ops.multi_write  = dset->shared->layout.ops->par_write;
+                io_info->io_ops.single_read  = H5D__mpio_select_read;
+                io_info->io_ops.single_write = H5D__mpio_select_write;
+            } /* end if */
+        }     /* end if */
         else {
             /* Check if there are any filters in the pipeline. If there are,
              * we cannot break to independent I/O if this is a write operation;
diff --git a/src/H5Dpkg.h b/src/H5Dpkg.h
index f9014a7..22712ee 100644
--- a/src/H5Dpkg.h
+++ b/src/H5Dpkg.h
@@ -121,9 +121,8 @@ typedef herr_t (*H5D_layout_construct_func_t)(H5F_t *f, H5D_t *dset);
 typedef herr_t (*H5D_layout_init_func_t)(H5F_t *f, const H5D_t *dset, hid_t dapl_id);
 typedef hbool_t (*H5D_layout_is_space_alloc_func_t)(const H5O_storage_t *storage);
 typedef hbool_t (*H5D_layout_is_data_cached_func_t)(const H5D_shared_t *shared_dset);
-typedef herr_t (*H5D_layout_io_init_func_t)(struct H5D_io_info_t *io_info,
-                                            const H5D_type_info_t *type_info, hsize_t nelmts,
-                                            const H5S_t *file_space, const H5S_t *mem_space,
+typedef herr_t (*H5D_layout_io_init_func_t)(struct H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
+                                            hsize_t nelmts, const H5S_t *file_space, const H5S_t *mem_space,
                                             struct H5D_chunk_map_t *cm);
 typedef herr_t (*H5D_layout_read_func_t)(struct H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
                                          hsize_t nelmts, const H5S_t *file_space, const H5S_t *mem_space,
diff --git a/src/H5FDint.c b/src/H5FDint.c
index f558b59..b23b918 100644
--- a/src/H5FDint.c
+++ b/src/H5FDint.c
@@ -712,7 +712,7 @@ done:
  */
 static herr_t
 H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count,
-                               const H5S_t * const *mem_spaces, const H5S_t * const *file_spaces,
+                               const H5S_t *const *mem_spaces, const H5S_t *const *file_spaces,
                                haddr_t offsets[], size_t element_sizes[], void *bufs[] /* out */)
 {
     hbool_t        extend_sizes = FALSE;
@@ -751,7 +751,6 @@ H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uin
     /* Sanity checks */
     HDassert(file);
     HDassert(file->cls);
-    HDassert(count > 0);
     HDassert(vec_arr_nalloc == sizeof(sizes_static) / sizeof(sizes_static[0]));
     HDassert(vec_arr_nalloc == sizeof(vec_bufs_static) / sizeof(vec_bufs_static[0]));
     HDassert(mem_spaces);
@@ -823,9 +822,9 @@ H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uin
         /* Initialize values so sequence lists are retrieved on the first
          * iteration */
         file_seq_i = H5FD_SEQ_LIST_LEN;
-        mem_seq_i = H5FD_SEQ_LIST_LEN;
-        file_nseq = 0;
-        mem_nseq = 0;
+        mem_seq_i  = H5FD_SEQ_LIST_LEN;
+        file_nseq  = 0;
+        mem_nseq   = 0;
 
         /* Loop until all elements are processed */
         while (file_seq_i < file_nseq || nelmts > 0) {
@@ -892,7 +891,8 @@ H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uin
                             HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                                         "memory reallocation failed for size list")
                         sizes = tmp_ptr;
-                        if (NULL == (tmp_ptr = H5MM_realloc(vec_bufs, vec_arr_nalloc * sizeof(*vec_bufs) * 2)))
+                        if (NULL ==
+                            (tmp_ptr = H5MM_realloc(vec_bufs, vec_arr_nalloc * sizeof(*vec_bufs) * 2)))
                             HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                                         "memory reallocation failed for buffer list")
                         vec_bufs = tmp_ptr;
@@ -1010,8 +1010,8 @@ done:
  *-------------------------------------------------------------------------
  */
 herr_t
-H5FD_read_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, const H5S_t * const *mem_spaces,
-                    const H5S_t * const *file_spaces, haddr_t offsets[], size_t element_sizes[],
+H5FD_read_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, const H5S_t *const *mem_spaces,
+                    const H5S_t *const *file_spaces, haddr_t offsets[], size_t element_sizes[],
                     void *bufs[] /* out */)
 {
     hbool_t  offsets_cooked = FALSE;
@@ -1284,8 +1284,9 @@ H5FD_read_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem_
         }
 
         /* Translate to vector or scalar I/O */
-        if (H5FD__read_selection_translate(file, type, dxpl_id, count, (const H5S_t * const *)mem_spaces,
-                                           (const H5S_t * const *)file_spaces, offsets, element_sizes, bufs) < 0)
+        if (H5FD__read_selection_translate(file, type, dxpl_id, count, (const H5S_t *const *)mem_spaces,
+                                           (const H5S_t *const *)file_spaces, offsets, element_sizes,
+                                           bufs) < 0)
             HGOTO_ERROR(H5E_VFL, H5E_READERROR, FAIL, "translation to vector or scalar read failed")
     }
 
@@ -1331,7 +1332,7 @@ done:
  */
 static herr_t
 H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uint32_t count,
-                                const H5S_t * const *mem_spaces, const H5S_t * const *file_spaces,
+                                const H5S_t *const *mem_spaces, const H5S_t *const *file_spaces,
                                 haddr_t offsets[], size_t element_sizes[], const void *bufs[])
 {
     hbool_t        extend_sizes = FALSE;
@@ -1370,7 +1371,6 @@ H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, ui
     /* Sanity checks */
     HDassert(file);
     HDassert(file->cls);
-    HDassert(count > 0);
     HDassert(vec_arr_nalloc == sizeof(sizes_static) / sizeof(sizes_static[0]));
     HDassert(vec_arr_nalloc == sizeof(vec_bufs_static) / sizeof(vec_bufs_static[0]));
     HDassert(mem_spaces);
@@ -1442,9 +1442,9 @@ H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, ui
         /* Initialize values so sequence lists are retrieved on the first
          * iteration */
         file_seq_i = H5FD_SEQ_LIST_LEN;
-        mem_seq_i = H5FD_SEQ_LIST_LEN;
-        file_nseq = 0;
-        mem_nseq = 0;
+        mem_seq_i  = H5FD_SEQ_LIST_LEN;
+        file_nseq  = 0;
+        mem_nseq   = 0;
 
         /* Loop until all elements are processed */
         while (file_seq_i < file_nseq || nelmts > 0) {
@@ -1511,7 +1511,8 @@ H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, ui
                             HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                                         "memory reallocation failed for size list")
                         sizes = tmp_ptr;
-                        if (NULL == (tmp_ptr = H5MM_realloc(vec_bufs, vec_arr_nalloc * sizeof(*vec_bufs) * 2)))
+                        if (NULL ==
+                            (tmp_ptr = H5MM_realloc(vec_bufs, vec_arr_nalloc * sizeof(*vec_bufs) * 2)))
                             HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                                         "memory reallocation failed for buffer list")
                         vec_bufs = tmp_ptr;
@@ -1627,8 +1628,8 @@ done:
  *-------------------------------------------------------------------------
  */
 herr_t
-H5FD_write_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, const H5S_t * const *mem_spaces,
-                     const H5S_t * const *file_spaces, haddr_t offsets[], size_t element_sizes[],
+H5FD_write_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, const H5S_t *const *mem_spaces,
+                     const H5S_t *const *file_spaces, haddr_t offsets[], size_t element_sizes[],
                      const void *bufs[])
 {
     hbool_t  offsets_cooked = FALSE;
@@ -1886,8 +1887,9 @@ H5FD_write_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem
         }
 
         /* Translate to vector or scalar I/O */
-        if (H5FD__write_selection_translate(file, type, dxpl_id, count, (const H5S_t * const *)mem_spaces,
-                                            (const H5S_t * const *)file_spaces, offsets, element_sizes, bufs) < 0)
+        if (H5FD__write_selection_translate(file, type, dxpl_id, count, (const H5S_t *const *)mem_spaces,
+                                            (const H5S_t *const *)file_spaces, offsets, element_sizes,
+                                            bufs) < 0)
             HGOTO_ERROR(H5E_VFL, H5E_WRITEERROR, FAIL, "translation to vector or scalar write failed")
     }
 
diff --git a/src/H5FDmpio.c b/src/H5FDmpio.c
index 9a99b45..707715d 100644
--- a/src/H5FDmpio.c
+++ b/src/H5FDmpio.c
@@ -1281,7 +1281,7 @@ H5FD__mpio_read(H5FD_t *_file, H5FD_mem_t H5_ATTR_UNUSED type, hid_t H5_ATTR_UNU
 #endif
             HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", 0)
 
-    /* Get the type's size */
+            /* Get the type's size */
 #if MPI_VERSION >= 3
     if (MPI_SUCCESS != (mpi_code = MPI_Type_size_x(buf_type, &type_size)))
 #else
@@ -1570,9 +1570,10 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
     hbool_t                    vector_was_sorted = TRUE;
     hbool_t                    fixed_size        = FALSE;
     size_t                     size;
-    H5FD_mem_t *               s_types           = NULL;
-    haddr_t *                  s_addrs           = NULL;
-    size_t *                   s_sizes           = NULL;
+    H5FD_mem_t *               s_types = NULL;
+    haddr_t *                  s_addrs = NULL;
+    size_t *                   s_sizes = NULL;
+    size_t                     s_size;
     void **                    s_bufs            = NULL;
     int *                      mpi_block_lengths = NULL;
     char                       unused            = 0; /* Unused, except for non-NULL pointer value */
@@ -1593,17 +1594,18 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
     H5FD_mpio_collective_opt_t coll_opt_mode; /* whether we are doing collective or independent I/O */
     int                        size_i;
 #if MPI_VERSION >= 3
-    MPI_Count                  bytes_read = 0; /* Number of bytes read in */
-    MPI_Count                  type_size;      /* MPI datatype used for I/O's size */
-    MPI_Count                  io_size;        /* Actual number of bytes requested */
-    MPI_Count                  n;
+    MPI_Count bytes_read = 0; /* Number of bytes read in */
+    MPI_Count type_size;      /* MPI datatype used for I/O's size */
+    MPI_Count io_size;        /* Actual number of bytes requested */
+    MPI_Count n;
 #else
-    int                        bytes_read = 0; /* Number of bytes read in */
-    int                        type_size;      /* MPI datatype used for I/O's size */
-    int                        io_size;        /* Actual number of bytes requested */
-    int                        n;
+    int bytes_read = 0; /* Number of bytes read in */
+    int type_size;      /* MPI datatype used for I/O's size */
+    int io_size;        /* Actual number of bytes requested */
+    int n;
 #endif
-    herr_t                     ret_value = SUCCEED;
+    hbool_t rank0_bcast = FALSE; /* If read-with-rank0-and-bcast flag was used */
+    herr_t  ret_value   = SUCCEED;
 
     FUNC_ENTER_STATIC
 
@@ -1639,7 +1641,23 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
 
     if (xfer_mode == H5FD_MPIO_COLLECTIVE) {
 
-        if (count > 0) { /* create MPI derived types describing the vector write */
+        if (count == 1) {
+            /* Single block.  Just use a series of MPI_BYTEs for the file view.
+             */
+            size_i        = (int)sizes[0];
+            buf_type      = MPI_BYTE;
+            file_type     = MPI_BYTE;
+            mpi_bufs_base = bufs[0];
+
+            /* Setup s_sizes (needed for incomplete read filling code) */
+            vector_was_sorted = TRUE;
+            s_sizes           = sizes;
+
+            /* some numeric conversions */
+            if (H5FD_mpi_haddr_to_MPIOff(addrs[0], &mpi_off) < 0)
+                HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI offset")
+        }
+        else if (count > 0) { /* create MPI derived types describing the vector write */
 
             /* sort the vector I/O request into increasing address order if required
              *
@@ -1650,8 +1668,8 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
              * are allocated, populated, and returned in s_types, s_addrs, s_sizes, and s_bufs respectively.
              * In this case, this function must free the memory allocated for the sorted vectors.
              */
-            if (H5FD_sort_vector_io_req(&vector_was_sorted, count, types, addrs, sizes, bufs, &s_types, &s_addrs,
-                                        &s_sizes, &s_bufs) < 0)
+            if (H5FD_sort_vector_io_req(&vector_was_sorted, count, types, addrs, sizes, bufs, &s_types,
+                                        &s_addrs, &s_sizes, &s_bufs) < 0)
                 HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "can't sort vector I/O request")
 
             if ((NULL == (mpi_block_lengths = (int *)HDmalloc((size_t)count * sizeof(int)))) ||
@@ -1750,6 +1768,10 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
             if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(&file_type)))
 
                 HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit for file_type failed", mpi_code)
+
+            /* some numeric conversions */
+            if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, &mpi_off) < 0)
+                HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0")
         }
         else {
 
@@ -1763,14 +1785,14 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
 
             /* MPI count to read */
             size_i = 0;
+
+            /* some numeric conversions */
+            if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, &mpi_off) < 0)
+                HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0")
         }
 
         /* Portably initialize MPI status variable */
-        HDmemset(&mpi_stat, 0, sizeof(MPI_Status));
-
-        /* some numeric conversions */
-        if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, &mpi_off) < 0)
-            HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0")
+        HDmemset(&mpi_stat, 0, sizeof(mpi_stat));
 
 #ifdef H5FDmpio_DEBUG
         if (H5FD_mpio_Debug[(int)'r'])
@@ -1782,6 +1804,10 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
                                                          H5FD_mpi_native_g, file->info)))
             HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
 
+        /* Reset mpi_off to 0 since the view now starts at the data offset */
+        if (H5FD_mpi_haddr_to_MPIOff((haddr_t)0, &mpi_off) < 0)
+            HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't set MPI off to 0")
+
         /* Get the collective_opt property to check whether the application wants to do IO individually.
          */
         if (H5CX_get_mpio_coll_opt(&coll_opt_mode) < 0)
@@ -1793,15 +1819,30 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
         if (H5FD_mpio_Debug[(int)'r'])
             HDfprintf(stdout, "%s: using MPIO collective mode\n", FUNC);
 #endif
-
         if (coll_opt_mode == H5FD_MPIO_COLLECTIVE_IO) {
 #ifdef H5FDmpio_DEBUG
             if (H5FD_mpio_Debug[(int)'r'])
                 HDfprintf(stdout, "%s: doing MPI collective IO\n", FUNC);
 #endif
+            /* Check whether we should read from rank 0 and broadcast to other ranks */
+            if (H5CX_get_mpio_rank0_bcast()) {
+#ifdef H5FDmpio_DEBUG
+                if (H5FD_mpio_Debug[(int)'r'])
+                    HDfprintf(stdout, "%s: doing read-rank0-and-MPI_Bcast\n", FUNC);
+#endif
+                /* Indicate path we've taken */
+                rank0_bcast = TRUE;
 
-            if (MPI_SUCCESS != (mpi_code = MPI_File_read_at_all(file->f, mpi_off, mpi_bufs_base, size_i,
-                                                                buf_type, &mpi_stat)))
+                /* Read on rank 0 Bcast to other ranks */
+                if (file->mpi_rank == 0)
+                    if (MPI_SUCCESS != (mpi_code = MPI_File_read_at(file->f, mpi_off, mpi_bufs_base, size_i,
+                                                                    buf_type, &mpi_stat)))
+                        HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at_all failed", mpi_code)
+                if (MPI_SUCCESS != (mpi_code = MPI_Bcast(mpi_bufs_base, size_i, buf_type, 0, file->comm)))
+                    HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code)
+            } /* end if */
+            else if (MPI_SUCCESS != (mpi_code = MPI_File_read_at_all(file->f, mpi_off, mpi_bufs_base, size_i,
+                                                                     buf_type, &mpi_stat)))
                 HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at_all failed", mpi_code)
         } /* end if */
         else if (size_i > 0) {
@@ -1822,15 +1863,34 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
                                                          H5FD_mpi_native_g, file->info)))
             HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
 
-        /* How many bytes were actually read? */
+        /* Only retrieve bytes read if this rank _actually_ participated in I/O */
+        if (!rank0_bcast || (rank0_bcast && file->mpi_rank == 0)) {
+            /* How many bytes were actually read? */
 #if MPI_VERSION >= 3
-        if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read)))
+            if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, buf_type, &bytes_read)))
 #else
-        if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read)))
+            if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&mpi_stat, MPI_BYTE, &bytes_read)))
 #endif
-            HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code)
+                HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code)
+        } /* end if */
 
-        /* Get the type's size */
+        /* If the rank0-bcast feature was used, broadcast the # of bytes read to
+         * other ranks, which didn't perform any I/O.
+         */
+        /* NOTE: This could be optimized further to be combined with the broadcast
+         *          of the data.  (QAK - 2019/1/2)
+         *       Or have rank 0 clear the unread parts of the buffer prior to
+         *          the bcast.  (NAF - 2021/9/15)
+         */
+        if (rank0_bcast)
+#if MPI_VERSION >= 3
+            if (MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_COUNT, 0, file->comm))
+#else
+            if (MPI_SUCCESS != MPI_Bcast(&bytes_read, 1, MPI_INT, 0, file->comm))
+#endif
+                HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", 0)
+
+                /* Get the type's size */
 #if MPI_VERSION >= 3
         if (MPI_SUCCESS != (mpi_code = MPI_Type_size_x(buf_type, &type_size)))
 #else
@@ -1856,10 +1916,10 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
                 HDassert(i >= 0);
 
 #if MPI_VERSION >= 3
-                io_size = MIN(n, (MPI_Count)s_sizes[i]);
+                io_size    = MIN(n, (MPI_Count)s_sizes[i]);
                 bytes_read = (MPI_Count)s_sizes[i] - io_size;
 #else
-                io_size = MIN(n, (int)s_sizes[i]);
+                io_size    = MIN(n, (int)s_sizes[i]);
                 bytes_read = (int)s_sizes[i] - io_size;
 #endif
                 HDassert(bytes_read >= 0);
@@ -1915,13 +1975,16 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
 
             /* Check if we acutally need to do I/O */
             if (addrs[i] < max_addr) {
+                /* Portably initialize MPI status variable */
+                HDmemset(&mpi_stat, 0, sizeof(mpi_stat));
+
                 /* Issue read */
                 if (MPI_SUCCESS !=
                     (mpi_code = MPI_File_read_at(file->f, mpi_off, bufs[i], size_i, MPI_BYTE, &mpi_stat)))
 
                     HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code)
 
-                /* How many bytes were actually read? */
+                    /* How many bytes were actually read? */
 #if MPI_VERSION >= 3
                 if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&mpi_stat, MPI_BYTE, &bytes_read)))
 #else
diff --git a/src/H5FDprivate.h b/src/H5FDprivate.h
index 03b1e9e..60784f4 100644
--- a/src/H5FDprivate.h
+++ b/src/H5FDprivate.h
@@ -129,12 +129,12 @@ H5_DLL herr_t  H5FD_read_vector(H5FD_t *file, uint32_t count, H5FD_mem_t types[]
                                 size_t sizes[], void *bufs[] /* out */);
 H5_DLL herr_t  H5FD_write_vector(H5FD_t *file, uint32_t count, H5FD_mem_t types[], haddr_t addrs[],
                                  size_t sizes[], const void *bufs[] /* out */);
-H5_DLL herr_t  H5FD_read_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, const H5S_t * const *mem_spaces,
-                                   const H5S_t * const *file_spaces, haddr_t offsets[], size_t element_sizes[],
-                                   void *bufs[] /* out */);
-H5_DLL herr_t  H5FD_write_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count, const H5S_t * const *mem_spaces,
-                                    const H5S_t * const *file_spaces, haddr_t offsets[], size_t element_sizes[],
-                                    const void *bufs[]);
+H5_DLL herr_t  H5FD_read_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count,
+                                   const H5S_t *const *mem_spaces, const H5S_t *const *file_spaces,
+                                   haddr_t offsets[], size_t element_sizes[], void *bufs[] /* out */);
+H5_DLL herr_t  H5FD_write_selection(H5FD_t *file, H5FD_mem_t type, uint32_t count,
+                                    const H5S_t *const *mem_spaces, const H5S_t *const *file_spaces,
+                                    haddr_t offsets[], size_t element_sizes[], const void *bufs[]);
 H5_DLL herr_t  H5FD_read_selection_id(H5FD_t *file, H5FD_mem_t type, uint32_t count, hid_t mem_space_ids[],
                                       hid_t file_space_ids[], haddr_t offsets[], size_t element_sizes[],
                                       void *bufs[] /* out */);
diff --git a/src/H5Fio.c b/src/H5Fio.c
index 9982dbc..123fe40 100644
--- a/src/H5Fio.c
+++ b/src/H5Fio.c
@@ -254,8 +254,8 @@ done:
  *-------------------------------------------------------------------------
  */
 herr_t
-H5F_shared_select_read(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, const H5S_t * const *mem_spaces,
-                       const H5S_t * const *file_spaces, haddr_t offsets[], size_t element_sizes[],
+H5F_shared_select_read(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, const H5S_t *const *mem_spaces,
+                       const H5S_t *const *file_spaces, haddr_t offsets[], size_t element_sizes[],
                        void *bufs[] /* out */)
 {
     H5FD_mem_t map_type;            /* Mapped memory type */
@@ -300,8 +300,9 @@ done:
  *-------------------------------------------------------------------------
  */
 herr_t
-H5F_shared_select_write(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, const H5S_t * const *mem_spaces,
-                        const H5S_t * const *file_spaces, haddr_t offsets[], size_t element_sizes[], const void *bufs[])
+H5F_shared_select_write(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count, const H5S_t *const *mem_spaces,
+                        const H5S_t *const *file_spaces, haddr_t offsets[], size_t element_sizes[],
+                        const void *bufs[])
 {
     H5FD_mem_t map_type;            /* Mapped memory type */
     herr_t     ret_value = SUCCEED; /* Return value */
diff --git a/src/H5Fprivate.h b/src/H5Fprivate.h
index cd9682d..2779a2b 100644
--- a/src/H5Fprivate.h
+++ b/src/H5Fprivate.h
@@ -924,11 +924,13 @@ H5_DLL herr_t H5F_block_write(H5F_t *f, H5FD_mem_t type, haddr_t addr, size_t si
 
 /* Functions that operate on selections of elements in the file */
 H5_DLL herr_t H5F_shared_select_read(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count,
-                                     const struct H5S_t * const *mem_spaces, const struct H5S_t * const *file_spaces,
-                                     haddr_t offsets[], size_t element_sizes[], void *bufs[] /* out */);
+                                     const struct H5S_t *const *mem_spaces,
+                                     const struct H5S_t *const *file_spaces, haddr_t offsets[],
+                                     size_t element_sizes[], void *bufs[] /* out */);
 H5_DLL herr_t H5F_shared_select_write(H5F_shared_t *f_sh, H5FD_mem_t type, uint32_t count,
-                                      const struct H5S_t * const *mem_spaces, const struct H5S_t * const *file_spaces,
-                                      haddr_t offsets[], size_t element_sizes[], const void *bufs[]);
+                                      const struct H5S_t *const *mem_spaces,
+                                      const struct H5S_t *const *file_spaces, haddr_t offsets[],
+                                      size_t element_sizes[], const void *bufs[]);
 
 /* Functions that flush or evict */
 H5_DLL herr_t H5F_flush_tagged_metadata(H5F_t *f, haddr_t tag);
diff --git a/testpar/t_coll_chunk.c b/testpar/t_coll_chunk.c
index 651a392..29341d7 100644
--- a/testpar/t_coll_chunk.c
+++ b/testpar/t_coll_chunk.c
@@ -832,7 +832,10 @@ coll_chunktest(const char *filename, int chunk_factor, int select_factor, int ap
     VRFY((status >= 0), "dataset write succeeded");
 
 #ifdef H5_HAVE_INSTRUMENTED_LIBRARY
-    if (facc_type == FACC_MPIO) {
+    /* Only check chunk optimization mode if selection I/O is not being used -
+     * selection I/O bypasses this IO mode decision - it's effectively always
+     * multi chunk currently */
+    if (facc_type == FACC_MPIO && !H5_use_selection_io_g) {
         switch (api_option) {
             case API_LINK_HARD:
                 status = H5Pget(xfer_plist, H5D_XFER_COLL_CHUNK_LINK_HARD_NAME, &prop_value);
diff --git a/testpar/t_dset.c b/testpar/t_dset.c
index 3e4a304..d65a2a9 100644
--- a/testpar/t_dset.c
+++ b/testpar/t_dset.c
@@ -3348,32 +3348,38 @@ actual_io_mode_tests(void)
     MPI_Comm_size(MPI_COMM_WORLD, &mpi_size);
     MPI_Comm_size(MPI_COMM_WORLD, &mpi_rank);
 
-    test_actual_io_mode(TEST_ACTUAL_IO_NO_COLLECTIVE);
+    /* Only run these tests if selection I/O is not being used - selection I/O
+     * bypasses this IO mode decision - it's effectively always multi chunk
+     * currently */
+    if (!H5_use_selection_io_g) {
+        test_actual_io_mode(TEST_ACTUAL_IO_NO_COLLECTIVE);
+
+        /*
+         * Test multi-chunk-io via proc_num threshold
+         */
+        test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_IND);
+        test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_COL);
 
-    /*
-     * Test multi-chunk-io via proc_num threshold
-     */
-    test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_IND);
-    test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_COL);
+        /* The Multi Chunk Mixed test requires atleast three processes. */
+        if (mpi_size > 2)
+            test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_MIX);
+        else
+            HDfprintf(stdout, "Multi Chunk Mixed test requires 3 proceses minimum\n");
 
-    /* The Multi Chunk Mixed test requires atleast three processes. */
-    if (mpi_size > 2)
-        test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_MIX);
-    else
-        HDfprintf(stdout, "Multi Chunk Mixed test requires 3 proceses minimum\n");
+        test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_MIX_DISAGREE);
 
-    test_actual_io_mode(TEST_ACTUAL_IO_MULTI_CHUNK_MIX_DISAGREE);
+        /*
+         * Test multi-chunk-io via setting direct property
+         */
+        test_actual_io_mode(TEST_ACTUAL_IO_DIRECT_MULTI_CHUNK_IND);
+        test_actual_io_mode(TEST_ACTUAL_IO_DIRECT_MULTI_CHUNK_COL);
 
-    /*
-     * Test multi-chunk-io via setting direct property
-     */
-    test_actual_io_mode(TEST_ACTUAL_IO_DIRECT_MULTI_CHUNK_IND);
-    test_actual_io_mode(TEST_ACTUAL_IO_DIRECT_MULTI_CHUNK_COL);
+        test_actual_io_mode(TEST_ACTUAL_IO_LINK_CHUNK);
+        test_actual_io_mode(TEST_ACTUAL_IO_CONTIGUOUS);
 
-    test_actual_io_mode(TEST_ACTUAL_IO_LINK_CHUNK);
-    test_actual_io_mode(TEST_ACTUAL_IO_CONTIGUOUS);
+        test_actual_io_mode(TEST_ACTUAL_IO_RESET);
+    }
 
-    test_actual_io_mode(TEST_ACTUAL_IO_RESET);
     return;
 }