integrate allgather_reduce

wenduwan · Aug 28, 2023 · f0bf7ee · f0bf7ee
1 parent f8e36c7
commit f0bf7ee
Showing 1 changed file with 35 additions and 9 deletions.
diff --git a/ompi/mca/coll/tuned/coll_tuned_decision_fixed.c b/ompi/mca/coll/tuned/coll_tuned_decision_fixed.c
@@ -60,6 +60,7 @@ ompi_coll_tuned_allreduce_intra_dec_fixed(const void *sbuf, void *rbuf, int coun
 {
     size_t dsize, total_dsize;
     int communicator_size, alg;
+    static int prev_alg = 0;
     communicator_size = ompi_comm_size(comm);
     OPAL_OUTPUT((ompi_coll_tuned_stream, "ompi_coll_tuned_allreduce_intra_dec_fixed"));
 
@@ -72,26 +73,41 @@ ompi_coll_tuned_allreduce_intra_dec_fixed(const void *sbuf, void *rbuf, int coun
      *  {3, "recursive_doubling"},
      *  {4, "ring"},
      *  {5, "segmented_ring"},
-     *  {6, "rabenseifner"
+     *  {6, "rabenseifner"},
+     *  {7, "allgather_reduce"}.
      *
      * Currently, ring, segmented ring, and rabenseifner do not support
      * non-commutative operations.
      */
     if( !ompi_op_is_commute(op) ) {
         if (communicator_size < 4) {
-            if (total_dsize < 131072) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 131072) {
                 alg = 3;
             } else {
                 alg = 1;
             }
         } else if (communicator_size < 8) {
-            alg = 3;
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else {
+                alg = 3;
+            }
         } else if (communicator_size < 16) {
-            if (total_dsize < 1048576) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 1048576) {
                 alg = 3;
             } else {
                 alg = 2;
             }
+        } else if (communicator_size < 64) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else {
+                alg = 3;
+            }
         } else if (communicator_size < 128) {
             alg = 3;
         } else if (communicator_size < 256) {
@@ -119,7 +135,9 @@ ompi_coll_tuned_allreduce_intra_dec_fixed(const void *sbuf, void *rbuf, int coun
         }
     } else {
         if (communicator_size < 4) {
-            if (total_dsize < 8) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 8) {
                 alg = 4;
             } else if (total_dsize < 4096) {
                 alg = 3;
@@ -135,29 +153,37 @@ ompi_coll_tuned_allreduce_intra_dec_fixed(const void *sbuf, void *rbuf, int coun
                 alg = 6;
             }
         } else if (communicator_size < 8) {
-            if (total_dsize < 16) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 16) {
                 alg = 4;
             } else if (total_dsize < 8192) {
                 alg = 3;
             } else {
                 alg = 6;
             }
         } else if (communicator_size < 16) {
-            if (total_dsize < 8192) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 8192) {
                 alg = 3;
             } else {
                 alg = 6;
             }
         } else if (communicator_size < 32) {
-            if (total_dsize < 64) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 64) {
                 alg = 5;
             } else if (total_dsize < 4096) {
                 alg = 3;
             } else {
                 alg = 6;
             }
         } else if (communicator_size < 64) {
-            if (total_dsize < 128) {
+            if (OMPI_COMM_IS_DISJOINT(comm) && total_dsize < 8192) {
+                alg = 7;
+            } else if (total_dsize < 128) {
                 alg = 5;
             } else {
                 alg = 6;