UCP: Fix multi ppn perf estimation (#10937)

Author: brminich

src/ucp/proto/proto_common.c

@@ -392,10 +392,10 @@ ucp_proto_common_get_lane_perf(const ucp_proto_common_init_params_t *params,
     tl_perf->send_pre_overhead  = perf_attr.send_pre_overhead + params->overhead;
     tl_perf->send_post_overhead = perf_attr.send_post_overhead;
     tl_perf->recv_overhead      = perf_attr.recv_overhead + params->overhead;
-    tl_perf->bandwidth          = ucp_tl_iface_bandwidth(context,
-                                                         &perf_attr.bandwidth);
-    tl_perf->path_ratio         = ucp_tl_iface_bandwidth(context,
-                                                         &perf_attr.path_bandwidth) /
+    tl_perf->bandwidth          = ucp_proto_common_iface_bandwidth(
+                                      context, &perf_attr.bandwidth);
+    tl_perf->path_ratio         = ucp_proto_common_iface_bandwidth(
+                                      context, &perf_attr.path_bandwidth) /
                                   tl_perf->bandwidth;
     tl_perf->latency            = ucp_tl_iface_latency(context,
                                                        &perf_attr.latency) +

src/ucp/proto/proto_common.inl

@@ -429,4 +429,12 @@ ucp_proto_common_bandwidth_equal(double bw1, double bw2)
     return fabs(bw1 - bw2) <= UCP_PROTO_PERF_EPSILON;
 }
 
+static UCS_F_ALWAYS_INLINE double
+ucp_proto_common_iface_bandwidth(ucp_context_h context,
+                                 const uct_ppn_bandwidth_t *bandwidth)
+{
+    return bandwidth->dedicated +
+           (bandwidth->shared / ucs_min(context->config.est_num_ppn, 8));
+}
+
 #endif

src/ucp/proto/proto_init.c

@@ -11,6 +11,7 @@
 #include "proto_init.h"
 #include "proto_debug.h"
 #include "proto_select.inl"
+#include "proto_common.inl"
 
 #include <ucp/core/ucp_ep.inl>
 #include <ucs/datastruct/array.h>
@@ -387,7 +388,7 @@ ucp_proto_init_add_buffer_copy_time(ucp_worker_h worker, const char *title,
     perf_factors[buffer_copy_factor_id].c +=
             ucp_tl_iface_latency(context, &perf_attr.latency);
     perf_factors[buffer_copy_factor_id].m +=
-            1.0 / ucp_tl_iface_bandwidth(context, &perf_attr.bandwidth);
+            1.0 / ucp_proto_common_iface_bandwidth(context, &perf_attr.bandwidth);
 
     if ((memtype_op == UCT_EP_OP_GET_SHORT) ||
         (memtype_op == UCT_EP_OP_GET_ZCOPY)) {

src/uct/ib/base/ib_iface.c

@@ -50,7 +50,7 @@
  * XDR single path ratio for RDMA_READ operations.
  * 4 QPs are needed to achieve full bandwidth with PCIe Gen6.
  */
-#define UCT_IB_XDR_READ_PATH_RATIO 0.25
+#define UCT_IB_XDR_READ_PATH_RATIO 0.4
 
 static UCS_CONFIG_DEFINE_ARRAY(path_bits_spec,
                                sizeof(ucs_range_spec_t),