Compute runtime && iteration standard deviation.

PiperOrigin-RevId: 731888477
Change-Id: I04c4f2f6cf63097bed88f91ff86e84eb08e15147

Author: liyuying0000

fleetbench/parallel/reporter.py

@@ -92,9 +92,7 @@ def GenerateBenchmarkReport(
 
   # Remove "fleetbench (" prefix and ")" suffix
   df["Benchmark"] = (
-      df["Benchmark"]
-      .astype(str)
-      .str.replace(r"fleetbench \((.*)\)", r"\1", regex=True)
+      df["Benchmark"].astype(str).str.replace(r".* \((.*)\)", r"\1", regex=True)
   )
 
   grouped_results = (
@@ -104,21 +102,33 @@ def GenerateBenchmarkReport(
           Mean_Wall_Time=("WallTimes", "mean"),
           Mean_CPU_Time=("CPUTimes", "mean"),
           Mean_Iterations=("Iterations", "mean"),
+          Wall_Time_std=("WallTimes", "std"),
+          CPU_Time_std=("CPUTimes", "std"),
+          Iterations_std=("Iterations", "std"),
       )
       .round(3)
   )
   grouped_results["Mean_Iterations"] = grouped_results[
       "Mean_Iterations"
   ].astype(int)
 
+  # We only show the following columns in the console report.
+  # More metrics can be founded in the dumped JSON file.
+  selected_columns = [
+      "Count",
+      "Mean_Wall_Time",
+      "Mean_CPU_Time",
+  ]
+
   # Combine perf_counter_df and benchmark run results on the same
   # "benchmark" entry.
   if perf_counter_df is not None:
     grouped_results = pd.merge(
         grouped_results, perf_counter_df, on="Benchmark", how="left"
     )
+    selected_columns.extend(perf_counter_df.columns.tolist())
 
-  print(grouped_results.to_string())
+  print(grouped_results[selected_columns].to_string())
   return grouped_results
 
 
@@ -139,6 +149,9 @@ def SaveBenchmarkResults(output_dir, df: pd.DataFrame) -> None:
           "Mean_Wall_Time": "real_time",
           "Mean_CPU_Time": "cpu_time",
           "Mean_Iterations": "iterations",
+          "Wall_Time_std": "real_time_std",
+          "CPU_Time_std": "cpu_time_std",
+          "Iterations_std": "iterations_std",
       }
   )
   data = df.reset_index().to_dict(orient="records")

fleetbench/parallel/reporter_test.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import json
+import math
 import os
 import shutil
 
@@ -150,6 +151,9 @@ def test_generate_benchmark_report(self):
             "Mean_Wall_Time": 4.5,
             "Mean_CPU_Time": 5.0,
             "Mean_Iterations": 30,
+            "Wall_Time_std": 2.121,
+            "CPU_Time_std": 2.828,
+            "Iterations_std": 28.284,
             "instructions": 130.0,
             "cycles": 3.0,
         },
@@ -159,6 +163,9 @@ def test_generate_benchmark_report(self):
             "Mean_Wall_Time": 4,
             "Mean_CPU_Time": 5.0,
             "Mean_Iterations": 20,
+            "Wall_Time_std": math.nan,
+            "CPU_Time_std": math.nan,
+            "Iterations_std": math.nan,
             "instructions": 200.0,
             "cycles": 2.0,
         },