coll/tuned: Change the bcast default collective algorithm selection

The default algorithm selections were out of date and not performing well. After gathering data using the ompi-collectives-tuning package, new default algorithm decisions are selected for bcast.

Signed-off-by: Jessie Yang <jiaxiyan@amazon.com>

ompi/mca/coll/tuned/coll_tuned.h

@@ -136,6 +136,7 @@ int ompi_coll_tuned_barrier_intra_check_forced_init (coll_tuned_force_algorithm_
 
 /* Bcast */
 int ompi_coll_tuned_bcast_intra_dec_fixed(BCAST_ARGS);
+int ompi_coll_tuned_bcast_intra_disjoint_dec_fixed(BCAST_ARGS);
 int ompi_coll_tuned_bcast_intra_dec_dynamic(BCAST_ARGS);
 int ompi_coll_tuned_bcast_intra_do_this(BCAST_ARGS, int algorithm, int faninout, int segsize);
 int ompi_coll_tuned_bcast_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);

ompi/mca/coll/tuned/coll_tuned_decision_fixed.c

@@ -651,6 +651,78 @@ int ompi_coll_tuned_bcast_intra_dec_fixed(void *buff, int count,
                                                 alg, 0, 0);
 }
 
+
+/*
+ *	bcast_intra_dec for inter node communicators
+ *
+ *	Function:	- selects broadcast algorithm to use
+ *	Accepts:	- same arguments as MPI_Bcast()
+ *	Returns:	- MPI_SUCCESS or error code (passed from the bcast implementation)
+ */
+int ompi_coll_tuned_bcast_intra_disjoint_dec_fixed(void *buff, int count,
+                                               struct ompi_datatype_t *datatype, int root,
+                                               struct ompi_communicator_t *comm,
+                                               mca_coll_base_module_t *module) {
+    size_t total_dsize, dsize;
+    int communicator_size, alg;
+	communicator_size = ompi_comm_size(comm);
+
+    ompi_datatype_type_size(datatype, &dsize);
+    total_dsize = dsize * (unsigned long)count;
+
+    OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_bcast_intra_disjoint_dec_fixed"
+                 " root %d rank %d com_size %d",
+                 root, ompi_comm_rank(comm), communicator_size));
+
+    /** Algorithms:
+     *  {1, "basic_linear"},
+     *  {2, "chain"},
+     *  {3, "pipeline"},
+     *  {4, "split_binary_tree"},
+     *  {5, "binary_tree"},
+     *  {6, "binomial"},
+     *  {7, "knomial"},
+     *  {8, "scatter_allgather"},
+     *  {9, "scatter_allgather_ring"},
+     */
+    if (communicator_size < 4) {
+        alg = 1;
+    } else if (communicator_size < 8) {
+        if (total_dsize < 1048576) {
+            alg = 1;
+        } else {
+            alg = 5;
+        }
+    } else if (communicator_size < 16) {
+        if (total_dsize < 1048576) {
+            alg = 1;
+        } else {
+            alg = 5;
+        }
+    } else if (communicator_size < 32) {
+        if (total_dsize < 262144) {
+            alg = 1;
+        } else if (total_dsize < 1048576) {
+            alg = 7;
+        } else {
+            alg = 5;
+        }
+    } else {
+        if (total_dsize < 65536) {
+            alg = 1;
+        } else if (total_dsize < 1048576) {
+            alg = 7;
+        } else {
+            alg = 5;
+        }
+    }
+
+    return ompi_coll_tuned_bcast_intra_do_this (buff, count, datatype, root,
+                                                comm, module,
+                                                alg, 0, 0);
+}
+
+
 /*
  *	reduce_intra_dec
  *

ompi/mca/coll/tuned/coll_tuned_module.c

@@ -93,15 +93,21 @@ ompi_coll_tuned_comm_query(struct ompi_communicator_t *comm, int *priority)
     /* By default stick with the fixed version of the tuned collectives. Later on,
      * when the module get enabled, set the correct version based on the availability
      * of the dynamic rules.
+     * For some collectives, we distinguish between disjoint communicatiors to make
+     * decision specific for inter node communication.
      */
+    if (OMPI_COMM_IS_DISJOINT(comm)) {
+        tuned_module->super.coll_bcast  = ompi_coll_tuned_bcast_intra_disjoint_dec_fixed;
+    } else {
+        tuned_module->super.coll_bcast  = ompi_coll_tuned_bcast_intra_dec_fixed;
+    }
     tuned_module->super.coll_allgather  = ompi_coll_tuned_allgather_intra_dec_fixed;
     tuned_module->super.coll_allgatherv = ompi_coll_tuned_allgatherv_intra_dec_fixed;
     tuned_module->super.coll_allreduce  = ompi_coll_tuned_allreduce_intra_dec_fixed;
     tuned_module->super.coll_alltoall   = ompi_coll_tuned_alltoall_intra_dec_fixed;
     tuned_module->super.coll_alltoallv  = ompi_coll_tuned_alltoallv_intra_dec_fixed;
     tuned_module->super.coll_alltoallw  = NULL;
     tuned_module->super.coll_barrier    = ompi_coll_tuned_barrier_intra_dec_fixed;
-    tuned_module->super.coll_bcast      = ompi_coll_tuned_bcast_intra_dec_fixed;
     tuned_module->super.coll_exscan     = NULL;
     tuned_module->super.coll_gather     = ompi_coll_tuned_gather_intra_dec_fixed;
     tuned_module->super.coll_gatherv    = NULL;