ob1: as per developer discussion disable rdma retries. the failure path currently suffers from live-lock

This commit was SVN r26575.

ompi/mca/pml/ob1/pml_ob1_component.c

@@ -112,8 +112,14 @@ static int mca_pml_ob1_component_open(void)
         mca_pml_ob1_param_register_int("send_pipeline_depth", 3);
     mca_pml_ob1.recv_pipeline_depth =
         mca_pml_ob1_param_register_int("recv_pipeline_depth", 4);
-    mca_pml_ob1.rdma_retries_limit =
-        mca_pml_ob1_param_register_int("rdma_retries_limit", 5);
+
+    /* NTH: we can get into a live-lock situation in the RDMA failure path so disable
+       RDMA retries for now. Falling back to send may suck but it is better than
+       hanging */
+    mca_pml_ob1.rdma_retries_limit = 0;
+/*     mca_pml_ob1.rdma_retries_limit = */
+/*         mca_pml_ob1_param_register_int("rdma_retries_limit", 5); */
+
     mca_pml_ob1.max_rdma_per_request =
         mca_pml_ob1_param_register_int("max_rdma_per_request", 4);
     mca_pml_ob1.max_send_per_range =