Merge pull request #10726 from rakhmets/topic/cuda-cpy-h2h-bw-1.19

UCT/CUDA/CUDA_COPY: H2H performance estimation is not supported - v1.19.x

Author: brminich

src/ucp/proto/proto_common.c

@@ -319,6 +319,21 @@ static void ucp_proto_common_tl_perf_reset(ucp_proto_common_tl_perf_t *tl_perf)
     tl_perf->max_frag           = SIZE_MAX;
 }
 
+static void ucp_proto_common_perf_attr_set_mem_type(
+        const ucp_proto_common_init_params_t *params,
+        uct_perf_attr_t *perf_attr)
+{
+    const ucp_rkey_config_key_t *rkey_config_key = params->super.rkey_config_key;
+
+    perf_attr->field_mask       |= UCT_PERF_ATTR_FIELD_LOCAL_MEMORY_TYPE;
+    perf_attr->local_memory_type = params->reg_mem_info.type;
+
+    if (rkey_config_key != NULL) {
+        perf_attr->field_mask        |= UCT_PERF_ATTR_FIELD_REMOTE_MEMORY_TYPE;
+        perf_attr->remote_memory_type = rkey_config_key->mem_type;
+    }
+}
+
 ucs_status_t
 ucp_proto_common_get_lane_perf(const ucp_proto_common_init_params_t *params,
                                ucp_lane_index_t lane,
@@ -367,6 +382,7 @@ ucp_proto_common_get_lane_perf(const ucp_proto_common_init_params_t *params,
                            UCT_PERF_ATTR_FIELD_PATH_BANDWIDTH |
                            UCT_PERF_ATTR_FIELD_LATENCY;
     perf_attr.operation  = params->send_op;
+    ucp_proto_common_perf_attr_set_mem_type(params, &perf_attr);
 
     status = ucp_worker_iface_estimate_perf(wiface, &perf_attr);
     if (status != UCS_OK) {
@@ -647,6 +663,22 @@ ucp_proto_common_reg_md_map(const ucp_proto_common_init_params_t *params,
     return reg_md_map;
 }
 
+static int ucp_proto_common_find_lanes_check_mem_type(
+        const ucp_proto_common_init_params_t *params, ucp_lane_index_t lane)
+{
+    uct_perf_attr_t perf_attr = {0};
+    ucp_rsc_index_t rsc_index;
+    ucp_worker_iface_t *wiface;
+
+    ucp_proto_common_perf_attr_set_mem_type(params, &perf_attr);
+
+    rsc_index = ucp_proto_common_get_rsc_index(&params->super, lane);
+    wiface    = ucp_worker_iface(params->super.worker, rsc_index);
+    /* TODO: Use memory reachability UCT API, when available, to check memory
+       type support */
+    return uct_iface_estimate_perf(wiface->iface, &perf_attr) == UCS_OK;
+}
+
 ucp_lane_index_t ucp_proto_common_find_lanes_with_min_frag(
         const ucp_proto_common_init_params_t *params, ucp_lane_type_t lane_type,
         uint64_t tl_cap_flags, ucp_lane_index_t max_lanes,
@@ -688,6 +720,10 @@ ucp_lane_index_t ucp_proto_common_find_lanes_with_min_frag(
             continue;
         }
 
+        if (!ucp_proto_common_find_lanes_check_mem_type(params, lane)) {
+            continue;
+        }
+
         lanes[num_valid_lanes++] = lane;
         if (num_valid_lanes >= max_lanes) {
             break;

src/uct/cuda/cuda_copy/cuda_copy_iface.c

@@ -198,6 +198,14 @@ uct_cuda_copy_estimate_perf(uct_iface_h tl_iface, uct_perf_attr_t *perf_attr)
     const double ss_factor         = zcopy ? 1 : 0.95;
     uct_ppn_bandwidth_t bandwidth  = {};
 
+    if ((src_mem_type == UCS_MEMORY_TYPE_HOST) &&
+        (dst_mem_type == UCS_MEMORY_TYPE_HOST)) {
+        ucs_trace("src_mem_type:%s to dst_mem_type:%s is not supported",
+                  ucs_memory_type_names[src_mem_type],
+                  ucs_memory_type_names[dst_mem_type]);
+        return UCS_ERR_UNSUPPORTED;
+    }
+
     if (uct_perf_attr_has_bandwidth(perf_attr->field_mask)) {
         if (uct_ep_op_is_fetch(op)) {
             ucs_swap(&src_mem_type, &dst_mem_type);

test/gtest/uct/v2/test_uct_query.cc

@@ -79,7 +79,8 @@ UCS_TEST_P(test_uct_query, query_perf)
                             UCT_PERF_ATTR_FIELD_SEND_POST_OVERHEAD |
                             UCT_PERF_ATTR_FIELD_RECV_OVERHEAD |
                             UCT_PERF_ATTR_FIELD_BANDWIDTH;
-    EXPECT_EQ(iface_estimate_perf(&perf_attr), UCS_OK);
+    EXPECT_EQ(iface_estimate_perf(&perf_attr),
+              has_transport("cuda_copy") ? UCS_ERR_UNSUPPORTED : UCS_OK);
 
     perf_attr.remote_memory_type = UCS_MEMORY_TYPE_CUDA;
     perf_attr.operation          = UCT_EP_OP_PUT_SHORT;