diff --git a/tensorflow/core/common_runtime/direct_session.cc b/tensorflow/core/common_runtime/direct_session.cc
index 40a1ffc42da..d9c6a6e4c98 100644
--- a/tensorflow/core/common_runtime/direct_session.cc
+++ b/tensorflow/core/common_runtime/direct_session.cc
@@ -463,7 +463,7 @@ Status DirectSession::RunInternal(int64 step_id, const RunOptions& run_options,
                                   CallFrameInterface* call_frame,
                                   ExecutorsAndKeys* executors_and_keys,
                                   RunMetadata* run_metadata) {
-  const uint64 start_time_usecs = Env::Default()->NowMicros();
+  const uint64 start_time_usecs = options_.env->NowMicros();
   string session_id_meta = strings::StrCat("SessionRun #id=", step_id, "#");
   tracing::ScopedActivity activity(session_id_meta);
 
@@ -720,7 +720,7 @@ Status DirectSession::RunInternal(int64 step_id, const RunOptions& run_options,
       exec_and_lib.graph->ToGraphDef(partition_graph_def);
     }
   }
-  metrics::UpdateGraphExecTime(Env::Default()->NowMicros() - start_time_usecs);
+  metrics::UpdateGraphExecTime(options_.env->NowMicros() - start_time_usecs);
 
   return Status::OK();
 }
diff --git a/tensorflow/core/common_runtime/direct_session.h b/tensorflow/core/common_runtime/direct_session.h
index bcac3415440..662044f8ec3 100644
--- a/tensorflow/core/common_runtime/direct_session.h
+++ b/tensorflow/core/common_runtime/direct_session.h
@@ -390,7 +390,7 @@ class DirectSession : public Session {
   std::atomic<int64> edge_name_counter_ = {0};
   std::atomic<int64> handle_name_counter_ = {0};
 
-  // For generating step ids that are unique across this sessions.
+  // For generating step ids that are unique among all sessions.
   static std::atomic_int_fast64_t step_id_counter_;
 
   // Global timeout for all blocking operations in this session.
diff --git a/tensorflow/core/common_runtime/metrics.cc b/tensorflow/core/common_runtime/metrics.cc
index fcdab26d3db..6a8e8c818f8 100644
--- a/tensorflow/core/common_runtime/metrics.cc
+++ b/tensorflow/core/common_runtime/metrics.cc
@@ -15,6 +15,7 @@ limitations under the License.
 
 #include "tensorflow/core/common_runtime/metrics.h"
 #include "tensorflow/core/lib/monitoring/counter.h"
+#include "tensorflow/core/lib/monitoring/sampler.h"
 
 namespace tensorflow {
 namespace metrics {
@@ -29,6 +30,12 @@ auto* graph_run_time_usecs = monitoring::Counter<0>::New(
     "/tensorflow/core/graph_run_time_usecs",
     "The total time spent on executing graphs in microseconds.");
 
+auto* graph_run_time_usecs_histogram = monitoring::Sampler<0>::New(
+    {"/tensorflow/core/graph_run_time_usecs_histogram",
+     "The wall-clock time spent on executing graphs in microseconds."},
+    // Power of 2 with bucket count 20 (> 17 minutes)
+    {monitoring::Buckets::Exponential(1000, 2, 20)});
+
 auto* tf_data_autotune_counter = monitoring::Counter<1>::New(
     "/tensorflow/data/autotune", "tf.data autotuning", "name");
 
@@ -81,6 +88,7 @@ void UpdateGraphExecTime(const uint64 running_time_usecs) {
   if (running_time_usecs > 0) {
     graph_runs->GetCell()->IncrementBy(1);
     graph_run_time_usecs->GetCell()->IncrementBy(running_time_usecs);
+    graph_run_time_usecs_histogram->GetCell()->Add(running_time_usecs);
   }
 }