tuned: use tree instead of bruck at scale

The switch from tree to bruck between 512 and 1023 processes leads
to unexpected latency changes in benchmarks of other collectives.
We should be consistent here. There is no good reason for why bruck
would perform better in that range but not beyond.


Signed-off-by: Joseph Schuchart <[email protected]>

Author: devreal

ompi/mca/coll/tuned/coll_tuned_decision_fixed.c

@@ -490,14 +490,8 @@ int ompi_coll_tuned_barrier_intra_dec_fixed(struct ompi_communicator_t *comm,
         alg = 3;
     } else if (communicator_size < 256) {
         alg = 4;
-    } else if (communicator_size < 512) {
-        alg = 6;
-    } else if (communicator_size < 1024) {
-        alg = 4;
-    } else if (communicator_size < 4096) {
-        alg = 6;
     } else {
-        alg = 4;
+        alg = 6;
     }
 
     return ompi_coll_tuned_barrier_intra_do_this (comm, module,