[Profiler] Add optimization advice for reducing the host-to-TPU data transfer time.

PiperOrigin-RevId: 312388184
Change-Id: I2fc8a60af6724467e447026dde7a8d6925ed1357

tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.cc

@@ -752,5 +752,17 @@ std::string GetSummaryNextStep(absl::string_view input_classification,
   return summary_next_step;
 }
 
+double HostToDeviceTransferAsPercentOfInputTime(
+    const InputTimeBreakdown& breakdown) {
+  // Thanks to the scaling trick we did in GenerateHostResult(), we can
+  // estimate the percentage of input-time spent on host-to-device transfer in
+  // the following way.
+  double total_input_time_us =
+      breakdown.demanded_file_read_us() + breakdown.advanced_file_read_us() +
+      breakdown.preprocessing_us() + breakdown.enqueue_us() +
+      breakdown.unclassified_non_enqueue_us();
+  return 100.0 * SafeDivide(breakdown.enqueue_us(), total_input_time_us);
+}
+
 }  // namespace profiler
 }  // namespace tensorflow

tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.h

@@ -31,6 +31,17 @@ limitations under the License.
 namespace tensorflow {
 namespace profiler {
 
+// If the percent of input-time spent on host-to-device transfer is greater than
+// kHostToDeviceTimePercentAsSignificant, we should advise the
+// user to optimize this transfer.
+constexpr double kHostToDeviceTimePercentAsSignificant = 10.0;
+
+// If the percent of input-time spent on host-to-device transfer is greater than
+// kHostToDeviceTimePercentAsDominant, we should ONLY advise the
+// user to optimize this transfer; we won't bother to suggest optimization for
+// tf.data.
+constexpr double kHostToDeviceTimePercentAsDominant = 90.0;
+
 // Computes the summary of step time in milliseconds.
 StepSummary ComputeStepTimeSummaryInMs(
     const ::tensorflow::protobuf::RepeatedPtrField<PerCoreStepInfo>&
@@ -62,6 +73,11 @@ void OutputAnalysis(double output_percent, std::string* output_classification,
 string GetSummaryNextStep(absl::string_view input_classification,
                           const InputTimeBreakdown& breakdown);
 
+// Returns the percentage of the input time that is spent on transferring the
+// data from host to device.
+double HostToDeviceTransferAsPercentOfInputTime(
+    const InputTimeBreakdown& breakdown);
+
 void AddErrorMessages(const OpStats& op_stats,
                       InputPipelineAnalysisResult* result);
 

tensorflow/core/profiler/protobuf/overview_page.proto

@@ -81,6 +81,8 @@ message OverviewPageRecommendation {
   // A statement for input that recommends the next steps for investigating the
   // bottleneck.
   string statement = 2;
+  // A list of tips for tackling input bottleneck.
+  repeated OverviewPageTip input_tips = 11;
   // A statement for output that recommends the next steps for investigating the
   // bottleneck.
   string output_statement = 9;