Merge pull request #986 from fortnern/parallel_selection_io

Add support for independent parallel I/O with selection I/O

9 files changed, 167 insertions, 62 deletions

diff --git a/src/H5.c b/src/H5.c
index cbd240a..0c7c8c1 100644
--- a/src/H5.c
+++ b/src/H5.c
@@ -147,6 +147,7 @@ done:
 herr_t
 H5_init_library(void)
 {
+    char *env_use_select_io = NULL;
     herr_t ret_value = SUCCEED;
 
     /* Set the 'library initialized' flag as early as possible, to avoid
@@ -278,6 +279,15 @@ H5_init_library(void)
     if (H5VL_init_phase2() < 0)
         HGOTO_ERROR(H5E_FUNC, H5E_CANTINIT, FAIL, "unable to initialize vol interface")
 
+    /* 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"))
+        H5_use_selection_io_g = TRUE;
+
     /* Debugging? */
     H5__debug_mask("-all");
     H5__debug_mask(HDgetenv("HDF5_DEBUG"));
diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c
index 5447233..dcc3baa 100644
--- a/src/H5Dchunk.c
+++ b/src/H5Dchunk.c
@@ -254,7 +254,7 @@ 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(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
+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,
@@ -1057,7 +1057,7 @@ H5D__chunk_is_data_cached(const H5D_shared_t *shared_dset)
  *-------------------------------------------------------------------------
  */
 static herr_t
-H5D__chunk_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
+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)
 {
     const H5D_t *dataset = io_info->dset;       /* Local pointer to dataset info */
@@ -1065,6 +1065,7 @@ H5D__chunk_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_inf
     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        */
 
@@ -1120,6 +1121,11 @@ H5D__chunk_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t *type_inf
     if (H5D__chunk_io_init_selections(io_info, type_info, fm) < 0)
         HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to create file and memory chunk selections")
 
+    /* Check if we're performing selection I/O and save the result */
+    if ((use_selection_io = H5D__chunk_may_use_select_io(io_info)) < 0)
+        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
+    io_info->use_select_io = (hbool_t)use_selection_io;
+
 done:
     /* Reset the global dataspace info */
     fm->file_space = NULL;
@@ -2535,7 +2541,6 @@ 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 */
-    htri_t        use_selection_io = FALSE;    /* Whether to use selection I/O */
     herr_t        ret_value        = SUCCEED;  /*return value        */
 
     FUNC_ENTER_STATIC
@@ -2567,12 +2572,8 @@ H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_
             skip_missing_chunks = TRUE;
     }
 
-    /* Check if we're performing selection I/O */
-    if ((use_selection_io = H5D__chunk_may_use_select_io(io_info)) < 0)
-        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
-
     /* Different blocks depending on whether we're using selection I/O */
-    if (use_selection_io) {
+    if (io_info->use_select_io) {
         size_t num_chunks;
         size_t element_sizes[2] = {type_info->dst_type_size, 0};
         void * bufs[2]          = {io_info->u.rbuf, NULL};
@@ -2815,7 +2816,6 @@ 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 */
-    htri_t        use_selection_io = FALSE;    /* Whether to use selection I/O */
     herr_t        ret_value        = SUCCEED;  /* Return value        */
 
     FUNC_ENTER_STATIC
@@ -2843,12 +2843,8 @@ H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize
     /* Initialize temporary compact storage info */
     cpt_store.compact.dirty = &cpt_dirty;
 
-    /* Check if we're performing selection I/O */
-    if ((use_selection_io = H5D__chunk_may_use_select_io(io_info)) < 0)
-        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
-
     /* Different blocks depending on whether we're using selection I/O */
-    if (use_selection_io) {
+    if (io_info->use_select_io) {
         size_t      num_chunks;
         size_t      element_sizes[2] = {type_info->dst_type_size, 0};
         const void *bufs[2]          = {io_info->u.wbuf, NULL};
diff --git a/src/H5Dcompact.c b/src/H5Dcompact.c
index fe41298..9a6d4b2 100644
--- a/src/H5Dcompact.c
+++ b/src/H5Dcompact.c
@@ -54,7 +54,7 @@
 /* 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(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
+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,
@@ -227,7 +227,7 @@ H5D__compact_is_space_alloc(const H5O_storage_t H5_ATTR_UNUSED *storage)
  *-------------------------------------------------------------------------
  */
 static herr_t
-H5D__compact_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
+H5D__compact_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
                      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)
 {
diff --git a/src/H5Dcontig.c b/src/H5Dcontig.c
index 3b104b8..d2a84da 100644
--- a/src/H5Dcontig.c
+++ b/src/H5Dcontig.c
@@ -91,7 +91,7 @@ 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(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
+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,
@@ -551,16 +551,25 @@ H5D__contig_is_data_cached(const H5D_shared_t *shared_dset)
  *-------------------------------------------------------------------------
  */
 static herr_t
-H5D__contig_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
+H5D__contig_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
                     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)
 {
-    FUNC_ENTER_STATIC_NOERR
+    htri_t use_selection_io = FALSE; /* Whether to use selection I/O */
+    htri_t ret_value = SUCCEED;      /* Return value */
+
+    FUNC_ENTER_STATIC
 
     io_info->store->contig.dset_addr = io_info->dset->shared->layout.storage.u.contig.addr;
     io_info->store->contig.dset_size = io_info->dset->shared->layout.storage.u.contig.size;
 
-    FUNC_LEAVE_NOAPI(SUCCEED)
+    /* Check if we're performing selection I/O */
+    if ((use_selection_io = H5D__contig_may_use_select_io(io_info, H5D_IO_OP_READ)) < 0)
+        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
+    io_info->use_select_io = (hbool_t)use_selection_io;
+
+done:
+    FUNC_LEAVE_NOAPI(ret_value)
 } /* end H5D__contig_io_init() */
 
 /*-------------------------------------------------------------------------
@@ -632,7 +641,6 @@ 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)
 {
-    htri_t use_selection_io = FALSE; /* Whether to use selection I/O */
     herr_t ret_value = SUCCEED;      /* Return value */
 
     FUNC_ENTER_PACKAGE
@@ -644,11 +652,7 @@ H5D__contig_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize
     HDassert(mem_space);
     HDassert(file_space);
 
-    /* Check if we're performing selection I/O */
-    if ((use_selection_io = H5D__contig_may_use_select_io(io_info, H5D_IO_OP_READ)) < 0)
-        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
-
-    if (use_selection_io) {
+    if (io_info->use_select_io) {
         size_t dst_type_size = type_info->dst_type_size;
 
         /* Issue selection I/O call (we can skip the page buffer because we've
@@ -684,7 +688,6 @@ 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)
 {
-    htri_t use_selection_io = FALSE; /* Whether to use selection I/O */
     herr_t ret_value = SUCCEED;      /* Return value */
 
     FUNC_ENTER_PACKAGE
@@ -696,11 +699,7 @@ H5D__contig_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsiz
     HDassert(mem_space);
     HDassert(file_space);
 
-    /* Check if we're performing selection I/O */
-    if ((use_selection_io = H5D__contig_may_use_select_io(io_info, H5D_IO_OP_WRITE)) < 0)
-        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
-
-    if (use_selection_io) {
+    if (io_info->use_select_io) {
         size_t dst_type_size = type_info->dst_type_size;
 
         /* Issue selection I/O call (we can skip the page buffer because we've
diff --git a/src/H5Defl.c b/src/H5Defl.c
index 85c9dba..fea70b0 100644
--- a/src/H5Defl.c
+++ b/src/H5Defl.c
@@ -61,7 +61,7 @@ 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(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, hsize_t nelmts,
+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,
@@ -197,7 +197,7 @@ H5D__efl_is_space_alloc(const H5O_storage_t H5_ATTR_UNUSED *storage)
  *-------------------------------------------------------------------------
  */
 static herr_t
-H5D__efl_io_init(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
+H5D__efl_io_init(H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
                  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)
 {
diff --git a/src/H5Dio.c b/src/H5Dio.c
index 03400cf..cc5f5bb 100644
--- a/src/H5Dio.c
+++ b/src/H5Dio.c
@@ -622,6 +622,10 @@ H5D__ioinfo_init(H5D_t *dset, const H5D_type_info_t *type_info, H5D_storage_t *s
         io_info->io_ops.single_write = H5D__scatgath_write;
     } /* end else */
 
+    /* Start with selection I/O off, layout callback will turn it on if
+     * appropriate */
+    io_info->use_select_io = FALSE;
+
 #ifdef H5_HAVE_PARALLEL
     /* Determine if the file was opened with an MPI VFD */
     io_info->using_mpi_vfd = H5F_HAS_FEATURE(dset->oloc.file, H5FD_FEAT_HAS_MPI);
@@ -841,8 +845,11 @@ 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 */
-    if (io_info->using_mpi_vfd) {
+    /* 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*/) {
         H5FD_mpio_xfer_t xfer_mode; /* Parallel transfer for this request */
         htri_t           opt;       /* Flag whether a selection is optimizable */
 
diff --git a/src/H5Dpkg.h b/src/H5Dpkg.h
index de04ac9..f9014a7 100644
--- a/src/H5Dpkg.h
+++ b/src/H5Dpkg.h
@@ -121,7 +121,7 @@ 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)(const struct H5D_io_info_t *io_info,
+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);
@@ -223,6 +223,7 @@ typedef struct H5D_io_info_t {
     H5D_layout_ops_t layout_ops; /* Dataset layout I/O operation function pointers */
     H5D_io_ops_t     io_ops;     /* I/O operation function pointers */
     H5D_io_op_type_t op_type;
+    hbool_t          use_select_io; /* Whether to use selection I/O */
     union {
         void *      rbuf; /* Pointer to buffer for read */
         const void *wbuf; /* Pointer to buffer to write */
diff --git a/src/H5FDint.c b/src/H5FDint.c
index bdb7c68..f558b59 100644
--- a/src/H5FDint.c
+++ b/src/H5FDint.c
@@ -884,15 +884,15 @@ H5FD__read_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, uin
                         void *tmp_ptr;
 
                         /* Reallocate arrays */
-                        if (NULL == (tmp_ptr = H5MM_realloc(addrs, vec_arr_nalloc * 2)))
+                        if (NULL == (tmp_ptr = H5MM_realloc(addrs, vec_arr_nalloc * sizeof(*addrs) * 2)))
                             HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                                         "memory reallocation failed for address list")
                         addrs = tmp_ptr;
-                        if (NULL == (tmp_ptr = H5MM_realloc(sizes, vec_arr_nalloc * 2)))
+                        if (NULL == (tmp_ptr = H5MM_realloc(sizes, vec_arr_nalloc * sizeof(*sizes) * 2)))
                             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 * 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;
@@ -1503,15 +1503,15 @@ H5FD__write_selection_translate(H5FD_t *file, H5FD_mem_t type, hid_t dxpl_id, ui
                         void *tmp_ptr;
 
                         /* Reallocate arrays */
-                        if (NULL == (tmp_ptr = H5MM_realloc(addrs, vec_arr_nalloc * 2)))
+                        if (NULL == (tmp_ptr = H5MM_realloc(addrs, vec_arr_nalloc * sizeof(*addrs) * 2)))
                             HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                                         "memory reallocation failed for address list")
                         addrs = tmp_ptr;
-                        if (NULL == (tmp_ptr = H5MM_realloc(sizes, vec_arr_nalloc * 2)))
+                        if (NULL == (tmp_ptr = H5MM_realloc(sizes, vec_arr_nalloc * sizeof(*sizes) * 2)))
                             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 * 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;
diff --git a/src/H5FDmpio.c b/src/H5FDmpio.c
index 1db700a..9a99b45 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
@@ -1592,6 +1592,17 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
     H5FD_mpio_xfer_t           xfer_mode;     /* I/O transfer mode */
     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;
+#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;
+#endif
     herr_t                     ret_value = SUCCEED;
 
     FUNC_ENTER_STATIC
@@ -1615,19 +1626,6 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
     HDassert((count == 0) || (sizes[0] != 0));
     HDassert((count == 0) || (types[0] != H5FD_MEM_NOLIST));
 
-    /* sort the vector I/O request into increasing address order if required
-     *
-     * If the vector is already sorted, the base addresses of types, addrs, sizes,
-     * and bufs will be returned in s_types, s_addrs, s_sizes, and s_bufs respectively.
-     *
-     * If the vector was not already sorted, new, sorted versions of types, addrs, sizes, and bufs
-     * 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)
-        HGOTO_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL, "can't sort vector I/O request")
-
     /* Get the transfer mode from the API context
      *
      * This flag is set to H5FD_MPIO_COLLECTIVE if the API call is
@@ -1643,6 +1641,19 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
 
         if (count > 0) { /* create MPI derived types describing the vector write */
 
+            /* sort the vector I/O request into increasing address order if required
+             *
+             * If the vector is already sorted, the base addresses of types, addrs, sizes,
+             * and bufs will be returned in s_types, s_addrs, s_sizes, and s_bufs respectively.
+             *
+             * If the vector was not already sorted, new, sorted versions of types, addrs, sizes, and bufs
+             * 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)
+                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)))) ||
                 (NULL == (mpi_displacments = (MPI_Aint *)HDmalloc((size_t)count * sizeof(MPI_Aint)))) ||
                 (NULL == (mpi_bufs = (MPI_Aint *)HDmalloc((size_t)count * sizeof(MPI_Aint))))) {
@@ -1810,9 +1821,60 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
         if (MPI_SUCCESS != (mpi_code = MPI_File_set_view(file->f, (MPI_Offset)0, MPI_BYTE, MPI_BYTE,
                                                          H5FD_mpi_native_g, file->info)))
             HMPI_GOTO_ERROR(FAIL, "MPI_File_set_view failed", mpi_code)
+
+        /* 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)))
+#else
+        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)
+
+        /* Get the type's size */
+#if MPI_VERSION >= 3
+        if (MPI_SUCCESS != (mpi_code = MPI_Type_size_x(buf_type, &type_size)))
+#else
+        if (MPI_SUCCESS != (mpi_code = MPI_Type_size(buf_type, &type_size)))
+#endif
+            HMPI_GOTO_ERROR(FAIL, "MPI_Type_size failed", mpi_code)
+
+        /* Compute the actual number of bytes requested */
+        io_size = type_size * size_i;
+
+        /* Check for read failure */
+        if (bytes_read < 0 || bytes_read > io_size)
+            HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "file read failed")
+
+        /* Check for incomplete read */
+        n = io_size - bytes_read;
+        if (n > 0) {
+            i = (int)count - 1;
+
+            /* Iterate over sorted array in reverse, filling in zeroes to
+             * sections of the buffers that were not read to */
+            do {
+                HDassert(i >= 0);
+
+#if MPI_VERSION >= 3
+                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]);
+                bytes_read = (int)s_sizes[i] - io_size;
+#endif
+                HDassert(bytes_read >= 0);
+
+                HDmemset((char *)bufs[i] + bytes_read, 0, (size_t)io_size);
+
+                n -= io_size;
+                i--;
+            } while (n > 0);
+        }
     }
     else if (count > 0) {
 
+        haddr_t max_addr = HADDR_MAX;
+
         /* The read is part of an independent operation. As a result,
          * we can't use MPI_File_set_view() (since it it a collective operation),
          * and thus there is no point in setting up an MPI derived type, as
@@ -1832,7 +1894,7 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
 
         for (i = 0; i < (int)count; i++) {
 
-            if (H5FD_mpi_haddr_to_MPIOff(s_addrs[i], &mpi_off) < 0)
+            if (H5FD_mpi_haddr_to_MPIOff(addrs[i], &mpi_off) < 0)
 
                 HGOTO_ERROR(H5E_INTERNAL, H5E_BADRANGE, FAIL, "can't convert from haddr to MPI off")
 
@@ -1845,16 +1907,46 @@ H5FD__mpio_read_vector(H5FD_t *_file, hid_t H5_ATTR_UNUSED dxpl_id, uint32_t cou
                 }
                 else {
 
-                    size = s_sizes[i];
+                    size = sizes[i];
                 }
             }
 
             size_i = (int)size; /* todo: fix potential for overflow */
 
-            if (MPI_SUCCESS !=
-                (mpi_code = MPI_File_read_at(file->f, mpi_off, s_bufs[i], size_i, MPI_BYTE, &mpi_stat)))
+            /* Check if we acutally need to do I/O */
+            if (addrs[i] < max_addr) {
+                /* 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)
+                    HMPI_GOTO_ERROR(FAIL, "MPI_File_read_at failed", mpi_code)
+
+                /* 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
+                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)
+
+                /* Check for read failure */
+                if (bytes_read < 0 || bytes_read > size_i)
+                    HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "file read failed")
+
+                /*
+                 * If we didn't read the entire I/O, fill in zeroes beyond end of
+                 * the physical MPI file and don't issue any more reads at higher
+                 * addresses.
+                 */
+                if ((n = (size_i - bytes_read)) > 0) {
+                    HDmemset((char *)bufs[i] + bytes_read, 0, (size_t)n);
+                    max_addr = addrs[i] + (haddr_t)bytes_read;
+                }
+            }
+            else {
+                /* Read is past the max address, fill in zeroes */
+                HDmemset((char *)bufs[i], 0, size);
+            }
         }
     }