pml/ob1: be more selective when using rdma capable btls

hjelmn · hjelmn · commit 4079eec9749e · 2016-08-09T20:54:42.000-06:00
This commit updates the btl selection logic for the RDMA and RDMA
pipeline protocols to use a btl iff: 1) the btl is also used for eager
messages (high exclusivity), or 2) no other RDMA btl is available on
an endpoint and the pml_ob1_use_all_rdma MCA variable is true. This
fixes a performance regression with shared memory when an RDMA capable
network is available.

Signed-off-by: Nathan Hjelm &lt;hjelmn@lanl.gov&gt;
diff --git a/ompi/mca/pml/ob1/pml_ob1.h b/ompi/mca/pml/ob1/pml_ob1.h
@@ -12,7 +12,7 @@
  *                         All rights reserved.
  * Copyright (c) 2010      Oracle and/or its affiliates.  All rights reserved
  * Copyright (c) 2011      Sandia National Laboratories. All rights reserved.
- * Copyright (c) 2012-2015 Los Alamos National Security, LLC. All rights
+ * Copyright (c) 2012-2016 Los Alamos National Security, LLC. All rights
  *                         reserved.
  * Copyright (c) 2015      Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
@@ -61,6 +61,7 @@ struct mca_pml_ob1_t {
     int max_rdma_per_request;
     int max_send_per_range;
     bool leave_pinned;
+    bool use_all_rdma;
     int leave_pinned_pipeline;
 
     /* lock queue access */
diff --git a/ompi/mca/pml/ob1/pml_ob1_component.c b/ompi/mca/pml/ob1/pml_ob1_component.c
@@ -198,6 +198,12 @@ static int mca_pml_ob1_component_register(void)
 
     mca_pml_ob1_param_register_uint("unexpected_limit", 128, &mca_pml_ob1.unexpected_limit);
 
+    mca_pml_ob1.use_all_rdma = false;
+    (void) mca_base_component_var_register(&mca_pml_ob1_component.pmlm_version, "use_all_rdma",
+                                           "Use all available RDMA btls for the RDMA and RDMA pipeline protocols "
+                                           "(default: false)", MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
+                                           OPAL_INFO_LVL_5, MCA_BASE_VAR_SCOPE_GROUP, &mca_pml_ob1.use_all_rdma);
+
     mca_pml_ob1.allocator_name = "bucket";
     (void) mca_base_component_var_register(&mca_pml_ob1_component.pmlm_version, "allocator",
                                            "Name of allocator component for unexpected messages",
diff --git a/ompi/mca/pml/ob1/pml_ob1_rdma.c b/ompi/mca/pml/ob1/pml_ob1_rdma.c
@@ -10,7 +10,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
- * Copyright (c) 2014-2015 Los Alamos National Security, LLC. All rights
+ * Copyright (c) 2014-2016 Los Alamos National Security, LLC. All rights
  *                         reserved.
  * $COPYRIGHT$
  *
@@ -42,6 +42,7 @@ size_t mca_pml_ob1_rdma_btls(
     mca_pml_ob1_com_btl_t* rdma_btls)
 {
     int num_btls = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_rdma);
+    int num_eager_btls = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_eager);
     double weight_total = 0;
     int num_btls_used = 0;
 
@@ -57,6 +58,22 @@ size_t mca_pml_ob1_rdma_btls(
                     (bml_endpoint->btl_rdma_index + n) % num_btls);
         mca_btl_base_registration_handle_t *reg_handle = NULL;
         mca_btl_base_module_t *btl = bml_btl->btl;
+        /* NTH: go ahead and use an rdma btl if is the only one */
+        bool ignore = !mca_pml_ob1.use_all_rdma;
+
+        /* do not use rdma btls that are not in the eager list. this is necessary to avoid using
+         * btls that exist on the endpoint only to support RMA. */
+        for (int i = 0 ; i < num_eager_btls && ignore ; ++i) {
+            mca_bml_base_btl_t *eager_btl = mca_bml_base_btl_array_get_index (&bml_endpoint->btl_eager, i);
+            if (eager_btl->btl_endpoint == bml_btl->btl_endpoint) {
+                ignore = false;
+                break;
+            }
+        }
+
+        if (ignore) {
+            continue;
+        }
 
         if (btl->btl_register_mem) {
             /* do not use the RDMA protocol with this btl if 1) leave pinned is disabled,
@@ -95,22 +112,66 @@ size_t mca_pml_ob1_rdma_btls(
     return num_btls_used;
 }
 
+size_t mca_pml_ob1_rdma_pipeline_btls_count (mca_bml_base_endpoint_t* bml_endpoint)
+{
+    int num_btls = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_rdma);
+    int num_eager_btls = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_eager);
+    int rdma_count = 0;
+
+    for(int i = 0; i < num_btls && i < mca_pml_ob1.max_rdma_per_request; ++i) {
+        mca_bml_base_btl_t *bml_btl = mca_bml_base_btl_array_get_next(&bml_endpoint->btl_rdma);
+        /* NTH: go ahead and use an rdma btl if is the only one */
+        bool ignore = !mca_pml_ob1.use_all_rdma;
+
+        for (int i = 0 ; i < num_eager_btls && ignore ; ++i) {
+            mca_bml_base_btl_t *eager_btl = mca_bml_base_btl_array_get_index (&bml_endpoint->btl_eager, i);
+            if (eager_btl->btl_endpoint == bml_btl->btl_endpoint) {
+                ignore = false;
+                break;
+            }
+        }
+
+        if (!ignore) {
+            ++rdma_count;
+        }
+    }
+
+    return rdma_count;
+}
+
 size_t mca_pml_ob1_rdma_pipeline_btls( mca_bml_base_endpoint_t* bml_endpoint,
                                        size_t size,
                                        mca_pml_ob1_com_btl_t* rdma_btls )
 {
-    int i, num_btls = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_rdma);
+    int num_btls = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_rdma);
+    int num_eager_btls = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_eager);
     double weight_total = 0;
+    int rdma_count = 0;
+
+    for(int i = 0; i < num_btls && i < mca_pml_ob1.max_rdma_per_request; i++) {
+        mca_bml_base_btl_t *bml_btl = mca_bml_base_btl_array_get_next(&bml_endpoint->btl_rdma);
+        /* NTH: go ahead and use an rdma btl if is the only one */
+        bool ignore = !mca_pml_ob1.use_all_rdma;
+
+        for (int i = 0 ; i < num_eager_btls && ignore ; ++i) {
+            mca_bml_base_btl_t *eager_btl = mca_bml_base_btl_array_get_index (&bml_endpoint->btl_eager, i);
+            if (eager_btl->btl_endpoint == bml_btl->btl_endpoint) {
+                ignore = false;
+                break;
+            }
+        }
+
+        if (ignore) {
+            continue;
+        }
 
-    for(i = 0; i < num_btls && i < mca_pml_ob1.max_rdma_per_request; i++) {
-        rdma_btls[i].bml_btl =
-            mca_bml_base_btl_array_get_next(&bml_endpoint->btl_rdma);
-        rdma_btls[i].btl_reg = NULL;
+        rdma_btls[rdma_count].bml_btl = bml_btl;
+        rdma_btls[rdma_count++].btl_reg = NULL;
 
-        weight_total += rdma_btls[i].bml_btl->btl_weight;
+        weight_total += bml_btl->btl_weight;
     }
 
-    mca_pml_ob1_calc_weighted_length(rdma_btls, i, size, weight_total);
+    mca_pml_ob1_calc_weighted_length (rdma_btls, rdma_count, size, weight_total);
 
-    return i;
+    return rdma_count;
 }
diff --git a/ompi/mca/pml/ob1/pml_ob1_rdma.h b/ompi/mca/pml/ob1/pml_ob1_rdma.h
@@ -1,3 +1,4 @@
+/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
 /*
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
@@ -9,6 +10,8 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2016      Los Alamos National Security, LLC. All rights
+ *                         reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -37,5 +40,8 @@ size_t mca_pml_ob1_rdma_btls(struct mca_bml_base_endpoint_t* endpoint,
  * bandwidth */
 size_t mca_pml_ob1_rdma_pipeline_btls(struct mca_bml_base_endpoint_t* endpoint,
                 size_t size, mca_pml_ob1_com_btl_t* rdma_btls);
+
+size_t mca_pml_ob1_rdma_pipeline_btls_count (mca_bml_base_endpoint_t* bml_endpoint);
+
 #endif
 
diff --git a/ompi/mca/pml/ob1/pml_ob1_recvreq.c b/ompi/mca/pml/ob1/pml_ob1_recvreq.c
@@ -263,7 +263,7 @@ static int mca_pml_ob1_recv_request_ack(
     /* by default copy everything */
     recvreq->req_send_offset = bytes_received;
     if(hdr->hdr_msg_length > bytes_received) {
-        size_t rdma_num = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_rdma);
+        size_t rdma_num = mca_pml_ob1_rdma_pipeline_btls_count (bml_endpoint);
         /*
          * lookup request buffer to determine if memory is already
          * registered.
