io/ompio: add a new grouping option avoiding communication

the new grouping option simple+ performs all calculations used
for the aggregator selection as if the default file view would be used,
thus avoiding communication in file_set_view all together. This mode
is useful for applications that do not set a file view, but use
explicit offset operations on the default file view.

Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>

ompi/mca/common/ompio/common_ompio_file_view.c

@@ -30,7 +30,7 @@
 #include "ompi/mca/fcoll/base/base.h"
 #include "ompi/mca/topo/topo.h"
 
-static OMPI_MPI_OFFSET_TYPE get_contiguous_chunk_size (mca_io_ompio_file_t *);
+static OMPI_MPI_OFFSET_TYPE get_contiguous_chunk_size (mca_io_ompio_file_t *, int flag);
 static int datatype_duplicate (ompi_datatype_t *oldtype, ompi_datatype_t **newtype );
 static int datatype_duplicate  (ompi_datatype_t *oldtype, ompi_datatype_t **newtype )
 {
@@ -139,7 +139,13 @@ int mca_common_ompio_set_view (mca_io_ompio_file_t *fh,
     // in orig_file type, No need to set args on this one.
     ompi_datatype_duplicate (newfiletype, &fh->f_filetype);
 
-    fh->f_cc_size = get_contiguous_chunk_size (fh);
+
+    if( SIMPLE_PLUS == mca_io_ompio_grouping_option ) {
+        fh->f_cc_size = get_contiguous_chunk_size (fh, 1);
+    }
+    else {
+        fh->f_cc_size = get_contiguous_chunk_size (fh, 0);
+    }
 
     if (opal_datatype_is_contiguous_memory_layout(&etype->super,1)) {
         if (opal_datatype_is_contiguous_memory_layout(&filetype->super,1) &&
@@ -166,7 +172,7 @@ int mca_common_ompio_set_view (mca_io_ompio_file_t *fh,
        }
     }
 
-    if ( SIMPLE != mca_io_ompio_grouping_option ) {
+    if ( SIMPLE != mca_io_ompio_grouping_option || SIMPLE_PLUS != mca_io_ompio_grouping_option ) {
 
         ret = mca_io_ompio_fview_based_grouping(fh,
                                                 &num_groups,
@@ -179,6 +185,7 @@ int mca_common_ompio_set_view (mca_io_ompio_file_t *fh,
     else {
         int done=0;
         int ndims;
+        
         if ( fh->f_comm->c_flags & OMPI_COMM_CART ){
             ret = fh->f_comm->c_topo->topo.cart.cartdim_get( fh->f_comm, &ndims);
             if ( OMPI_SUCCESS != ret ){
@@ -253,7 +260,7 @@ exit:
     return ret;
 }
 
-OMPI_MPI_OFFSET_TYPE get_contiguous_chunk_size (mca_io_ompio_file_t *fh)
+OMPI_MPI_OFFSET_TYPE get_contiguous_chunk_size (mca_io_ompio_file_t *fh, int flag)
 {
     int uniform = 0;
     OMPI_MPI_OFFSET_TYPE avg[3] = {0,0,0};
@@ -268,6 +275,10 @@ OMPI_MPI_OFFSET_TYPE get_contiguous_chunk_size (mca_io_ompio_file_t *fh)
     ** 2. each section in the file view has exactly the same size
     */
 
+    if ( flag  ) {
+        global_avg[0] = MCA_IO_DEFAULT_FILE_VIEW_SIZE;
+    }
+    else {
         for (i=0 ; i<(int)fh->f_iov_count ; i++) {
             avg[0] += fh->f_decoded_iov[i].iov_len;
             if (i && 0 == uniform) {
@@ -322,6 +333,8 @@ OMPI_MPI_OFFSET_TYPE get_contiguous_chunk_size (mca_io_ompio_file_t *fh)
             fh->f_flags |= OMPIO_UNIFORM_FVIEW;
         }
 #endif
+    }
+
     return global_avg[0];
 }
 

ompi/mca/io/ompio/io_ompio.h

@@ -111,7 +111,7 @@ OMPI_DECLSPEC extern int mca_io_ompio_coll_timing_info;
 #define OPTIMIZE_GROUPING               4
 #define SIMPLE                          5
 #define NO_REFINEMENT                   6
-
+#define SIMPLE_PLUS                     7
 
 #define OMPIO_UNIFORM_DIST_THRESHOLD     0.5
 #define OMPIO_CONTG_THRESHOLD        1048576

ompi/mca/io/ompio/io_ompio_aggregators.c

@@ -498,7 +498,8 @@ int mca_io_ompio_set_aggregator_props (struct mca_io_ompio_file_t *fh,
 
     if (-1 == num_aggregators) {
         if ( SIMPLE        == mca_io_ompio_grouping_option ||
-            NO_REFINEMENT == mca_io_ompio_grouping_option ) {
+             NO_REFINEMENT == mca_io_ompio_grouping_option ||
+             SIMPLE_PLUS   == mca_io_ompio_grouping_option ) {
             fh->f_aggregator_index = 0;
             fh->f_final_num_aggrs  = fh->f_init_num_aggrs;
             fh->f_procs_per_group  = fh->f_init_procs_per_group;

ompi/mca/io/ompio/io_ompio_component.c

@@ -212,7 +212,7 @@ static int register_component(void)
                                            "Option for grouping of processes in the aggregator selection "
                                            "1: Data volume based grouping 2: maximizing group size uniformity 3: maximimze "
                                            "data contiguity 4: hybrid optimization  5: simple (default) "
-                                           "6: skip refinement step",
+                                           "6: skip refinement step 7: simple+: grouping based on default file view",
                                            MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
                                            OPAL_INFO_LVL_9,
                                            MCA_BASE_VAR_SCOPE_READONLY,