Defining and Running Benchmarks

This guide contains instructions for defining and running a TensorFlow benchmark. These benchmarks store output in TestResults format. If these benchmarks are added to the TensorFlow github repo, we will run them daily with our continuous build and display a graph on our dashboard:

Defining a Benchmark

Defining a TensorFlow benchmark requires extending the tf.test.Benchmark class and calling the self.report_benchmark method. Below, you'll find an example of benchmark code:

import time

import tensorflow as tf

# Define a class that extends from tf.test.Benchmark.
class SampleBenchmark(tf.test.Benchmark):

  # Note: benchmark method name must start with `benchmark`.
  def benchmarkSum(self):
    with tf.Session() as sess:
      x = tf.constant(10)
      y = tf.constant(5)
      result = tf.add(x, y)

      iters = 100
      start_time = time.time()
      for _ in range(iters):
      total_wall_time = time.time() - start_time

      # Call report_benchmark to report a metric value.
          # This value should always be per iteration.

if __name__ == "__main__":

See the full example for SampleBenchmark.

Key points to note in the example above:

  • Benchmark class extends from tf.test.Benchmark.
  • Each benchmark method should start with benchmark prefix.
  • Benchmark method calls report_benchmark to report the metric value.

Running with Python

Use the --benchmarks flag to run the benchmark with Python. A BenchmarkEntries proto will be printed.

python --benchmarks=SampleBenchmark

Setting the flag as --benchmarks=. or --benchmarks=all works as well.

(Please ensure that Tensorflow is installed to successfully import the package in the line import tensorflow as tf. For installation instructions, see Installing TensorFlow. This step is not necessary when running with Bazel.)

Adding a bazel Target

We have a special target called tf_py_logged_benchmark for benchmarks defined under the TensorFlow github repo. tf_py_logged_benchmark should wrap around a regular py_test target. Running a tf_py_logged_benchmark would print a TestResults proto. Defining a tf_py_logged_benchmark also lets us run it with TensorFlow continuous build.

First, define a regular py_test target. See example below:

  name = "sample_benchmark",
  srcs = [""],
  srcs_version = "PY2AND3",
  deps = [

You can run benchmarks in a py_test target by passing the --benchmarks flag. The benchmark should just print out a BenchmarkEntries proto.

bazel test :sample_benchmark --test_arg=--benchmarks=all

Now, add the tf_py_logged_benchmark target (if available). This target would pass in --benchmarks=all to the wrapped py_test target and provide a way to store output for our TensorFlow continuous build. The target tf_py_logged_benchmark should be available in TensorFlow repository.

load("//tensorflow/tools/test:performance.bzl", "tf_py_logged_benchmark")

    name = "sample_logged_benchmark",
    target = "//tensorflow/examples/benchmark:sample_benchmark",

Use the following command to run the benchmark target:

bazel test :sample_logged_benchmark