HDFFV-10625 -- Implemented a process-0 read and then broadcast for collective read of full (HS_ALL), contiguous, atomic datasets by all the processes in the file communicator.

4 files changed, 526 insertions, 75 deletions

diff --git a/testpar/t_bigio.c b/testpar/t_bigio.c
index fdd3488..1d882b8 100644
--- a/testpar/t_bigio.c
+++ b/testpar/t_bigio.c
@@ -671,7 +671,7 @@ dataset_big_write(void)
     /* create a memory dataspace independently */
     mem_dataspace = H5Screate_simple (RANK, dims, NULL);
     VRFY((mem_dataspace >= 0), "");
-    if(!mpi_rank == 0) {
+    if(mpi_rank != 0) {
         ret = H5Sselect_none(mem_dataspace);
         VRFY((ret >= 0), "H5Sset_none succeeded");
     }
@@ -980,7 +980,7 @@ dataset_big_read(void)
     /* create a memory dataspace independently */
     mem_dataspace = H5Screate_simple (RANK, dims, NULL);
     VRFY((mem_dataspace >= 0), "");
-    if(!mpi_rank == 0) {
+    if(mpi_rank != 0) {
         ret = H5Sselect_none(mem_dataspace);
         VRFY((ret >= 0), "H5Sset_none succeeded");
     }
diff --git a/testpar/t_dset.c b/testpar/t_dset.c
index 281d027..35501d8 100644
--- a/testpar/t_dset.c
+++ b/testpar/t_dset.c
@@ -2649,7 +2649,7 @@ compress_readAll(void)
 
             /* Try reading the data */
             ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, xfer_plist, data_read);
-            VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded");
+            VRFY((ret >= 0), "H5Dread succeeded");
 
             /* Verify data read */
             for(u=0; u<dim; u++)
diff --git a/testpar/t_mdset.c b/testpar/t_mdset.c
index 5d989bb..16eb13c 100644
--- a/testpar/t_mdset.c
+++ b/testpar/t_mdset.c
@@ -12,6 +12,7 @@
  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
 
 #include "testphdf5.h"
+#include "H5Dprivate.h"
 
 #define DIM  2
 #define SIZE 32
@@ -311,13 +312,27 @@ void compact_dataset(void)
      VRFY((ret>= 0),"set independent IO collectively succeeded");
     }
 
-
     dataset = H5Dopen2(iof, dname, H5P_DEFAULT);
     VRFY((dataset >= 0), "H5Dopen2 succeeded");
 
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+    hbool_t prop_value;
+    prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF;
+    ret = H5Pinsert2(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, H5D_XFER_COLL_RANK0_BCAST_SIZE, &prop_value,
+                     NULL, NULL, NULL, NULL, NULL, NULL);
+    VRFY((ret >= 0), "H5Pinsert2() succeeded");
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
     ret = H5Dread(dataset, H5T_NATIVE_DOUBLE, H5S_ALL, H5S_ALL, dxpl, inme);
     VRFY((ret >= 0), "H5Dread succeeded");
 
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+    prop_value = FALSE;
+    ret = H5Pget(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value);
+    VRFY((ret >= 0), "H5Pget succeeded");
+    VRFY((prop_value == FALSE && dxfer_coll_type == DXFER_COLLECTIVE_IO),"rank 0 Bcast optimization was performed for a compact dataset");
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
     /* Verify data value */
     for(i = 0; i < size; i++)
         for(j = 0; j < size; j++)
@@ -603,8 +618,8 @@ void dataset_fillvalue(void)
     hsize_t     req_count[4] = {1, 6, 7, 8};
     hsize_t     dset_size;      /* Dataset size */
     int *rdata, *wdata;         /* Buffers for data to read and write */
-    int *twdata, *trdata;        /* Temporary pointer into buffer */
-    int acc, i, j, k, l;        /* Local index variables */
+    int *twdata, *trdata;       /* Temporary pointer into buffer */
+    int acc, i, j, k, l, ii;    /* Local index variables */
     herr_t ret;                 /* Generic return value */
     const char *filename;
 
@@ -645,27 +660,60 @@ void dataset_fillvalue(void)
     /*
      * Read dataset before any data is written.
      */
-    /* set entire read buffer with the constant 2 */
-    HDmemset(rdata,2,(size_t)(dset_size*sizeof(int)));
-    /* Independently read the entire dataset back */
-    ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, H5P_DEFAULT, rdata);
-    VRFY((ret >= 0), "H5Dread succeeded");
 
-    /* Verify all data read are the fill value 0 */
-    trdata = rdata;
-    err_num = 0;
-    for(i = 0; i < (int)dset_dims[0]; i++)
+    /* Create DXPL for I/O */
+    dxpl = H5Pcreate(H5P_DATASET_XFER);
+    VRFY((dxpl >= 0), "H5Pcreate succeeded");
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      hbool_t prop_value;
+      prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF;
+      ret = H5Pinsert2(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, H5D_XFER_COLL_RANK0_BCAST_SIZE, &prop_value,
+                   NULL, NULL, NULL, NULL, NULL, NULL);
+      VRFY((ret >= 0),"testing property list inserted succeeded");
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+    for(ii = 0; ii < 2; ii++) {
+
+      if(ii == 0)
+        ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_INDEPENDENT);
+      else
+        ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_COLLECTIVE);
+      VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded");
+
+      /* set entire read buffer with the constant 2 */
+      HDmemset(rdata,2,(size_t)(dset_size*sizeof(int)));
+
+      /* Read the entire dataset back */
+      ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, dxpl, rdata);
+      VRFY((ret >= 0), "H5Dread succeeded");
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      prop_value = FALSE;
+      ret = H5Pget(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value);
+      VRFY((ret >= 0), "testing property list get succeeded");
+      if(ii == 0)
+        VRFY((prop_value == FALSE), "correctly handled rank 0 Bcast");
+      else
+        VRFY((prop_value == TRUE), "correctly handled rank 0 Bcast");
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* Verify all data read are the fill value 0 */
+      trdata = rdata;
+      err_num = 0;
+      for(i = 0; i < (int)dset_dims[0]; i++)
         for(j = 0; j < (int)dset_dims[1]; j++)
-            for(k = 0; k < (int)dset_dims[2]; k++)
-                for(l = 0; l < (int)dset_dims[3]; l++, twdata++, trdata++)
-		    if(*trdata != 0)
-			if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED)
-			    printf("Dataset Verify failed at [%d][%d][%d][%d]: expect 0, got %d\n", i, j, k, l, *trdata);
-    if(err_num > MAX_ERR_REPORT && !VERBOSE_MED)
+          for(k = 0; k < (int)dset_dims[2]; k++)
+            for(l = 0; l < (int)dset_dims[3]; l++, twdata++, trdata++)
+              if(*trdata != 0)
+                if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED)
+                  printf("Dataset Verify failed at [%d][%d][%d][%d]: expect 0, got %d\n", i, j, k, l, *trdata);
+      if(err_num > MAX_ERR_REPORT && !VERBOSE_MED)
         printf("[more errors ...]\n");
-    if(err_num){
+      if(err_num) {
         printf("%d errors found in check_value\n", err_num);
-	nerrors++;
+        nerrors++;
+      }
     }
 
     /* Barrier to ensure all processes have completed the above test. */
@@ -681,10 +729,6 @@ void dataset_fillvalue(void)
     ret = H5Sselect_hyperslab(memspace, H5S_SELECT_SET, req_start, NULL, req_count, NULL);
     VRFY((ret >= 0), "H5Sselect_hyperslab succeeded on memory dataspace");
 
-    /* Create DXPL for collective I/O */
-    dxpl = H5Pcreate(H5P_DATASET_XFER);
-    VRFY((dxpl >= 0), "H5Pcreate succeeded");
-
     ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_COLLECTIVE);
     VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded");
     if(dxfer_coll_type == DXFER_INDEPENDENT_IO) {
@@ -711,37 +755,64 @@ void dataset_fillvalue(void)
     /*
      * Read dataset after partial write.
      */
-    /* set entire read buffer with the constant 2 */
-    HDmemset(rdata,2,(size_t)(dset_size*sizeof(int)));
-    /* Independently read the entire dataset back */
-    ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, H5P_DEFAULT, rdata);
-    VRFY((ret >= 0), "H5Dread succeeded");
 
-    /* Verify correct data read */
-    twdata=wdata;
-    trdata=rdata;
-    err_num=0;
-    for(i=0; i<(int)dset_dims[0]; i++)
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF;
+      ret = H5Pset(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value);
+      VRFY((ret >= 0), " H5Pset succeeded");
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+    for(ii = 0; ii < 2; ii++) {
+
+      if(ii == 0)
+        ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_INDEPENDENT);
+      else
+        ret = H5Pset_dxpl_mpio(dxpl, H5FD_MPIO_COLLECTIVE);
+      VRFY((ret >= 0), "H5Pset_dxpl_mpio succeeded");
+
+      /* set entire read buffer with the constant 2 */
+      HDmemset(rdata,2,(size_t)(dset_size*sizeof(int)));
+
+      /* Read the entire dataset back */
+      ret = H5Dread(dataset, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, dxpl, rdata);
+      VRFY((ret >= 0), "H5Dread succeeded");
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      prop_value = FALSE;
+      ret = H5Pget(dxpl, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value);
+      VRFY((ret >= 0), "testing property list get succeeded");
+      if(ii == 0)
+        VRFY((prop_value == FALSE), "correctly handled rank 0 Bcast");
+      else
+        VRFY((prop_value == TRUE), "correctly handled rank 0 Bcast");
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* Verify correct data read */
+      twdata=wdata;
+      trdata=rdata;
+      err_num=0;
+      for(i=0; i<(int)dset_dims[0]; i++)
         for(j=0; j<(int)dset_dims[1]; j++)
-            for(k=0; k<(int)dset_dims[2]; k++)
-                for(l=0; l<(int)dset_dims[3]; l++, twdata++, trdata++)
-                    if(i<mpi_size) {
-                        if(*twdata != *trdata )
-                            if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED)
-                                printf("Dataset Verify failed at [%d][%d][%d][%d]: expect %d, got %d\n", i,j,k,l, *twdata, *trdata);
-                    } /* end if */
-                    else {
-                        if(*trdata != 0)
-                            if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED)
-                                printf("Dataset Verify failed at [%d][%d][%d][%d]: expect 0, got %d\n", i,j,k,l, *trdata);
-                    } /* end else */
-    if(err_num > MAX_ERR_REPORT && !VERBOSE_MED)
+          for(k=0; k<(int)dset_dims[2]; k++)
+            for(l=0; l<(int)dset_dims[3]; l++, twdata++, trdata++)
+              if(i<mpi_size) {
+                if(*twdata != *trdata )
+                  if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED)
+                    printf("Dataset Verify failed at [%d][%d][%d][%d]: expect %d, got %d\n", i,j,k,l, *twdata, *trdata);
+              } /* end if */
+              else {
+                if(*trdata != 0)
+                  if(err_num++ < MAX_ERR_REPORT || VERBOSE_MED)
+                    printf("Dataset Verify failed at [%d][%d][%d][%d]: expect 0, got %d\n", i,j,k,l, *trdata);
+              } /* end else */
+      if(err_num > MAX_ERR_REPORT && !VERBOSE_MED)
         printf("[more errors ...]\n");
-    if(err_num){
+      if(err_num){
         printf("%d errors found in check_value\n", err_num);
-	nerrors++;
+        nerrors++;
+      }
     }
-
+    
     /* Close all file objects */
     ret = H5Dclose(dataset);
     VRFY((ret >= 0), "H5Dclose succeeded");
@@ -856,7 +927,7 @@ void collective_group_write(void)
         if(!((m+1) % 10)) {
             printf("created %d groups\n", m+1);
             MPI_Barrier(MPI_COMM_WORLD);
-	}
+        }
 #endif /* BARRIER_CHECKS */
     }
 
diff --git a/testpar/t_pread.c b/testpar/t_pread.c
index 0905d44..74feeb6 100644
--- a/testpar/t_pread.c
+++ b/testpar/t_pread.c
@@ -17,6 +17,7 @@
  */
 
 #include "testpar.h"
+#include "H5Dprivate.h"
 
 /* The collection of files is included below to aid
  * an external "cleanup" process if required.
@@ -34,6 +35,8 @@ const char *FILENAMES[NFILENAME + 1]={"reloc_t_pread_data_file",
 
 #define COUNT 1000
 
+#define LIMIT_NPROC 6
+
 hbool_t pass = true;
 static const char *random_hdf5_text =
 "Now is the time for all first-time-users of HDF5 to read their \
@@ -46,7 +49,7 @@ completely foolproof is to underestimate the ingenuity of complete\n\
 fools.\n";
 
 static int generate_test_file(MPI_Comm comm, int mpi_rank, int group);
-static int test_parallel_read(MPI_Comm comm, int mpi_rank, int group);
+static int test_parallel_read(MPI_Comm comm, int mpi_rank, int mpi_size, int group);
 
 static char *test_argv0 = NULL;
 
@@ -108,6 +111,9 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id )
     hid_t fapl_id   = -1;
     hid_t dxpl_id   = -1;
     hid_t dset_id   = -1;
+    hid_t dset_id_ch = -1;
+    hid_t dcpl_id = H5P_DEFAULT;
+    hsize_t chunk[1];
     float nextValue;
     float *data_slice = NULL;
 
@@ -272,6 +278,55 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id )
         }
     }
 
+
+    /* create a chunked dataset */
+    chunk[0] = COUNT/8;
+    
+    if ( pass ) {
+      if ( (dcpl_id = H5Pcreate (H5P_DATASET_CREATE)) < 0 ) {
+        pass = false;
+          failure_mssg = "H5Pcreate() failed.\n";
+      }
+    }
+
+    if ( pass ) {
+      if ( (H5Pset_chunk (dcpl_id, 1, chunk) ) < 0 ) {
+        pass = false;
+        failure_mssg = "H5Pset_chunk() failed.\n";
+      }
+    }
+    
+    if ( pass ) {
+      
+      if ( (dset_id_ch = H5Dcreate2(file_id, "dataset0_chunked", H5T_NATIVE_FLOAT,
+                                    filespace, H5P_DEFAULT, dcpl_id,
+                                    H5P_DEFAULT)) < 0 ) {
+        pass = false;
+        failure_mssg = "H5Dcreate2() failed.\n";
+      }
+    }
+
+    if ( pass ) {
+      if ( (H5Dwrite(dset_id_ch, H5T_NATIVE_FLOAT, memspace,
+                     filespace, dxpl_id, data_slice)) < 0 ) {
+        pass = false;
+        failure_mssg = "H5Dwrite() failed.\n";
+      }
+    }
+    if ( pass || (dcpl_id != -1)) {
+      if ( H5Pclose(dcpl_id) < 0 ) {
+        pass = false;
+        failure_mssg = "H5Pclose(dcpl_id) failed.\n";
+      }
+    }
+    
+    if ( pass || (dset_id_ch != -1)) {
+      if ( H5Dclose(dset_id_ch) < 0 ) {
+        pass = false;
+        failure_mssg = "H5Dclose(dset_id_ch) failed.\n";
+      }
+    }
+
     /* close file, etc. */
     if ( pass || (dset_id != -1)) {
         if ( H5Dclose(dset_id) < 0 ) {
@@ -413,7 +468,7 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id )
  * Function:    test_parallel_read
  *
  * Purpose:     This actually tests the superblock optimization
- *              and covers the two primary cases we're interested in.
+ *              and covers the three primary cases we're interested in.
  *              1). That HDF5 files can be opened in parallel by
  *                  the rank 0 process and that the superblock
  *                  offset is correctly broadcast to the other
@@ -423,6 +478,10 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id )
  *                  subgroups of MPI_COMM_WORLD and that each
  *                  subgroup operates as described in (1) to
  *                  collectively read the data.
+ *              3). Testing proc0-read-and-MPI_Bcast using
+ *                  sub-communicators, and reading into
+ *                  a memory space that is different from the
+ *                  file space, and chunked datasets.
  *
  *              The global MPI rank is used for reading and
  *              writing data for process specific data in the
@@ -444,7 +503,7 @@ generate_test_file( MPI_Comm comm, int mpi_rank, int group_id )
  *-------------------------------------------------------------------------
  */
 static int
-test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id)
+test_parallel_read(MPI_Comm comm, int mpi_rank, int mpi_size, int group_id)
 {
     const char *failure_mssg;
     const char *fcn_name = "test_parallel_read()";
@@ -457,8 +516,13 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id)
     hid_t fapl_id   = -1;
     hid_t file_id   = -1;
     hid_t dset_id   = -1;
+    hid_t dset_id_ch = -1;
+    hid_t dxpl_id   = H5P_DEFAULT;
     hid_t memspace  = -1;
     hid_t filespace = -1;
+    hid_t filetype  = -1;
+    size_t filetype_size;
+    hssize_t dset_size;
     hsize_t i;
     hsize_t offset;
     hsize_t count = COUNT;
@@ -552,6 +616,14 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id)
         }
     }
 
+    /* open the chunked data set */
+    if ( pass ) {
+      if ( (dset_id_ch = H5Dopen2(file_id, "dataset0_chunked", H5P_DEFAULT)) < 0 ) {
+        pass = FALSE;
+        failure_mssg = "H5Dopen2() failed\n";
+      }
+    }
+
     /* setup memspace */
     if ( pass ) {
         dims[0] = count;
@@ -606,14 +678,6 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id)
         }
     }
 
-    /* close file, etc. */
-    if ( pass || (dset_id != -1) ) {
-        if ( H5Dclose(dset_id) < 0 ) {
-            pass = false;
-            failure_mssg = "H5Dclose(dset_id) failed.\n";
-        }
-    }
-
     if ( pass || (memspace != -1) ) {
         if ( H5Sclose(memspace) < 0 ) {
             pass = false;
@@ -628,6 +692,330 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id)
         }
     }
 
+    /* free data_slice if it has been allocated */
+    if ( data_slice != NULL ) {
+        HDfree(data_slice);
+        data_slice = NULL;
+    }
+
+    /* 
+     * Test reading proc0-read-and-bcast with sub-communicators 
+     */
+
+    /* Don't test with more than LIMIT_NPROC processes to avoid memory issues */
+
+    if( group_size <= LIMIT_NPROC ) {
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      hbool_t prop_value;
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      if ( (filespace = H5Dget_space(dset_id )) < 0 ) {
+        pass = FALSE;
+        failure_mssg = "H5Dget_space failed.\n";
+      }
+
+      if ( (dset_size = H5Sget_simple_extent_npoints(filespace)) < 0 ) {
+        pass = FALSE;
+        failure_mssg = "H5Sget_simple_extent_npoints failed.\n";
+      }
+
+      if ( (filetype = H5Dget_type(dset_id)) < 0 ) {
+        pass = FALSE;
+        failure_mssg = "H5Dget_type failed.\n";
+      }
+
+      if ( (filetype_size = H5Tget_size(filetype)) == 0 ) {
+        pass = FALSE;
+        failure_mssg = "H5Tget_size failed.\n";
+      }
+
+      if ( H5Tclose(filetype) < 0 ) {
+        pass = FALSE;
+        failure_mssg = "H5Tclose failed.\n";
+      };
+
+      if ( (data_slice = (float *)HDmalloc((size_t)dset_size*filetype_size)) == NULL ) {
+        pass = FALSE;
+        failure_mssg = "malloc of data_slice failed.\n";
+      }
+
+      if ( pass ) {
+        if ( (dxpl_id = H5Pcreate(H5P_DATASET_XFER)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Pcreate(H5P_DATASET_XFER) failed.\n";
+        }
+      }
+
+      if ( pass ) {
+        if ( (H5Pset_dxpl_mpio(dxpl_id, H5FD_MPIO_COLLECTIVE)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Pset_dxpl_mpio() failed.\n";
+        }
+      }
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      if ( pass ) {
+        prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF;
+        if(H5Pinsert2(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, H5D_XFER_COLL_RANK0_BCAST_SIZE, &prop_value,
+                   NULL, NULL, NULL, NULL, NULL, NULL) < 0) {
+            pass = FALSE;
+            failure_mssg = "H5Pinsert2() failed\n";
+        }
+      }
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* read H5S_ALL section */
+      if ( pass ) {
+        if ( (H5Dread(dset_id, H5T_NATIVE_FLOAT, H5S_ALL,
+                      H5S_ALL, dxpl_id, data_slice)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Dread() failed\n";
+        }
+      }
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      if ( pass ) {
+        prop_value = FALSE;
+        if(H5Pget(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) {
+            pass = FALSE;
+            failure_mssg = "H5Pget() failed\n";
+        }
+        if (pass) {
+          if(prop_value != TRUE) {
+            pass = FALSE;
+            failure_mssg = "rank 0 Bcast optimization was mistakenly not performed\n";
+          }
+        }
+      }
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* verify the data */
+      if ( pass ) {
+
+        if ( comm == MPI_COMM_WORLD )       /* test 1 */
+          nextValue = 0;
+        else if ( group_id == 0 )           /* test 2 group 0 */
+          nextValue = 0;
+        else                                /* test 2 group 1 */
+          nextValue = (float)((hsize_t)( mpi_size / 2 )*count);
+        
+        i = 0;
+        while ( ( pass ) && ( i < (hsize_t)dset_size ) ) {
+          /* what we really want is data_slice[i] != nextValue --
+           * the following is a circumlocution to shut up the
+           * the compiler.
+           */
+          if ( ( data_slice[i] > nextValue ) ||
+               ( data_slice[i] < nextValue ) ) {
+            pass = FALSE;
+            failure_mssg = "Unexpected dset contents.\n";
+          }
+          nextValue += 1;
+          i++;
+        }
+      }
+
+      /* read H5S_ALL section for the chunked dataset */
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      if ( pass ) {
+        prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF;
+        if(H5Pset(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) {
+            pass = FALSE;
+            failure_mssg = "H5Pset() failed\n";
+        }
+      }
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      for ( i = 0; i < (hsize_t)dset_size; i++) {
+        data_slice[i] = 0;
+      }
+      if ( pass ) {
+        if ( (H5Dread(dset_id_ch, H5T_NATIVE_FLOAT, H5S_ALL,
+                      H5S_ALL, dxpl_id, data_slice)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Dread() failed\n";
+        }
+      }
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      if ( pass ) {
+        prop_value = FALSE;
+        if(H5Pget(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) {
+            pass = FALSE;
+            failure_mssg = "H5Pget() failed\n";
+        }
+        if (pass) {
+          if(prop_value == TRUE) {
+            pass = FALSE;
+            failure_mssg = "rank 0 Bcast optimization was mistakenly performed for chunked dataset\n";
+          }
+        }
+      }
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* verify the data */
+      if ( pass ) {
+
+        if ( comm == MPI_COMM_WORLD )       /* test 1 */
+          nextValue = 0;
+        else if ( group_id == 0 )           /* test 2 group 0 */
+          nextValue = 0;
+        else                                /* test 2 group 1 */
+          nextValue = (float)((hsize_t)( mpi_size / 2 )*count);
+        
+        i = 0;
+        while ( ( pass ) && ( i < (hsize_t)dset_size ) ) {
+          /* what we really want is data_slice[i] != nextValue --
+           * the following is a circumlocution to shut up the
+           * the compiler.
+           */
+          if ( ( data_slice[i] > nextValue ) ||
+               ( data_slice[i] < nextValue ) ) {
+            pass = FALSE;
+            failure_mssg = "Unexpected chunked dset contents.\n";
+          }
+          nextValue += 1;
+          i++;
+        }
+      }
+
+      if ( pass || (filespace != -1) ) {
+        if ( H5Sclose(filespace) < 0 ) {
+          pass = false;
+          failure_mssg = "H5Sclose(filespace) failed.\n";
+        }
+      }
+
+      /* free data_slice if it has been allocated */
+      if ( data_slice != NULL ) {
+        HDfree(data_slice);
+        data_slice = NULL;
+      }
+
+      /* 
+       * Read an H5S_ALL filespace into a hyperslab defined memory space 
+       */
+
+      if ( (data_slice = (float *)HDmalloc((size_t)(dset_size*2)*filetype_size)) == NULL ) {
+        pass = FALSE;
+        failure_mssg = "malloc of data_slice failed.\n";
+      }
+
+      /* setup memspace */
+      if ( pass ) {
+        dims[0] = (hsize_t)dset_size*2;
+        if ( (memspace = H5Screate_simple(1, dims, NULL)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Screate_simple(1, dims, NULL) failed\n";
+        }
+      }
+      if ( pass ) {
+        offset = (hsize_t)dset_size;
+        if ( (H5Sselect_hyperslab(memspace, H5S_SELECT_SET,
+                                  &offset, NULL, &offset, NULL)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Sselect_hyperslab() failed\n";
+        }
+      }
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      if ( pass ) {
+        prop_value = H5D_XFER_COLL_RANK0_BCAST_DEF;
+        if(H5Pset(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) {
+            pass = FALSE;
+            failure_mssg = "H5Pset() failed\n";
+        }
+      }
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* read this processes section of the data */
+      if ( pass ) {
+        if ( (H5Dread(dset_id, H5T_NATIVE_FLOAT, memspace,
+                      H5S_ALL, dxpl_id, data_slice)) < 0 ) {
+          pass = FALSE;
+          failure_mssg = "H5Dread() failed\n";
+        }
+      }
+
+#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
+      if ( pass ) {
+        prop_value = FALSE;
+        if(H5Pget(dxpl_id, H5D_XFER_COLL_RANK0_BCAST_NAME, &prop_value) < 0) {
+            pass = FALSE;
+            failure_mssg = "H5Pget() failed\n";
+        }
+        if (pass) {
+          if(prop_value != TRUE) {
+            pass = FALSE;
+            failure_mssg = "rank 0 Bcast optimization was mistakenly not performed\n";
+          }
+        }
+      }
+#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
+
+      /* verify the data */
+      if ( pass ) {
+
+        if ( comm == MPI_COMM_WORLD )       /* test 1 */
+          nextValue = 0;
+        else if ( group_id == 0 )           /* test 2 group 0 */
+          nextValue = 0;
+        else                                /* test 2 group 1 */
+          nextValue = (float)((hsize_t)(mpi_size / 2)*count);
+
+        i = (hsize_t)dset_size;
+        while ( ( pass ) && ( i < (hsize_t)dset_size ) ) {
+          /* what we really want is data_slice[i] != nextValue --
+           * the following is a circumlocution to shut up the
+           * the compiler.
+           */
+          if ( ( data_slice[i] > nextValue ) ||
+               ( data_slice[i] < nextValue ) ) {
+            pass = FALSE;
+            failure_mssg = "Unexpected dset contents.\n";
+          }
+          nextValue += 1;
+          i++;
+        }
+      }
+      
+      if ( pass || (memspace != -1) ) {
+        if ( H5Sclose(memspace) < 0 ) {
+          pass = false;
+          failure_mssg = "H5Sclose(memspace) failed.\n";
+        }
+      }
+      
+      /* free data_slice if it has been allocated */
+      if ( data_slice != NULL ) {
+        HDfree(data_slice);
+        data_slice = NULL;
+      }
+
+      if ( pass || (dxpl_id != -1) ) {
+        if ( H5Pclose(dxpl_id) < 0 ) {
+          pass = false;
+          failure_mssg = "H5Pclose(dxpl_id) failed.\n";
+        }
+      }
+    }
+
+    /* close file, etc. */
+    if ( pass || (dset_id != -1) ) {
+        if ( H5Dclose(dset_id) < 0 ) {
+            pass = false;
+            failure_mssg = "H5Dclose(dset_id) failed.\n";
+        }
+    }
+
+    if ( pass || (dset_id_ch != -1) ) {
+        if ( H5Dclose(dset_id_ch) < 0 ) {
+            pass = false;
+            failure_mssg = "H5Dclose(dset_id_ch) failed.\n";
+        }
+    }
+
     if ( pass || (file_id != -1) ) {
         if ( H5Fclose(file_id) < 0 ) {
             pass = false;
@@ -668,17 +1056,9 @@ test_parallel_read(MPI_Comm comm, int mpi_rank, int group_id)
             HDfprintf(stdout, "%s: failure_mssg = \"%s\"\n",
                       fcn_name, failure_mssg);
         }
-
         HDremove(reloc_data_filename);
     }
 
-    /* free data_slice if it has been allocated */
-    if ( data_slice != NULL ) {
-        HDfree(data_slice);
-        data_slice = NULL;
-    }
-
-
     return( ! pass );
 
 } /* test_parallel_read() */
@@ -803,7 +1183,7 @@ main( int argc, char **argv)
     }
 
     /* Now read the generated test file (stil using MPI_COMM_WORLD) */
-    nerrs += test_parallel_read( MPI_COMM_WORLD, mpi_rank, which_group);
+    nerrs += test_parallel_read( MPI_COMM_WORLD, mpi_rank, mpi_size, which_group);
 
     if ( nerrs > 0 ) {
         if ( mpi_rank == 0 ) {
@@ -819,7 +1199,7 @@ main( int argc, char **argv)
     }
 
     /* run the 2nd set of tests */
-    nerrs += test_parallel_read(group_comm, mpi_rank, which_group);
+    nerrs += test_parallel_read(group_comm, mpi_rank, mpi_size, which_group);
 
     if ( nerrs > 0 ) {
         if ( mpi_rank == 0 ) {