diff --git a/benchmarks/python/__init__.py b/benchmarks/python/__init__.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/benchmarks/python/__init__.py
diff --git a/benchmarks/python/py_benchmark.py b/benchmarks/python/py_benchmark.py
new file mode 100755
index 0000000..6942d20
--- /dev/null
+++ b/benchmarks/python/py_benchmark.py
@@ -0,0 +1,152 @@
+import sys
+import os
+import timeit
+import math
+import argparse
+import fnmatch
+import json
+
+parser = argparse.ArgumentParser(description="Python protobuf benchmark")
+parser.add_argument("data_files", metavar="dataFile", nargs="+", 
+                    help="testing data files.")
+parser.add_argument("--json", action="store_const", dest="json",
+                    const="yes", default="no",
+                    help="Whether to output json results")
+parser.add_argument("--behavior_prefix", dest="behavior_prefix",
+                    help="The output json format's behavior's name's prefix",
+                    default="")
+# BEGIN CPP GENERATED MESSAGE
+parser.add_argument("--cpp_generated", action="store_const",
+                    dest="cpp_generated", const="yes", default="no",
+                    help="Whether to link generated code library")
+# END CPP GENERATED MESSAGE
+args = parser.parse_args()
+# BEGIN CPP GENERATED MESSAGE
+# CPP generated code must be linked before importing the generated Python code
+# for the descriptor can be found in the pool
+if args.cpp_generated != "no":
+  sys.path.append( os.path.dirname( os.path.dirname( os.path.abspath(__file__) ) ) + "/.libs" )
+  import libbenchmark_messages
+  sys.path.append( os.path.dirname( os.path.dirname( os.path.abspath(__file__) ) ) + "/tmp" )
+# END CPP GENERATED MESSAGE
+
+
+import datasets.google_message1.proto2.benchmark_message1_proto2_pb2 as benchmark_message1_proto2_pb2
+import datasets.google_message1.proto3.benchmark_message1_proto3_pb2 as benchmark_message1_proto3_pb2
+import datasets.google_message2.benchmark_message2_pb2 as benchmark_message2_pb2
+import datasets.google_message3.benchmark_message3_pb2 as benchmark_message3_pb2
+import datasets.google_message4.benchmark_message4_pb2 as benchmark_message4_pb2
+import benchmarks_pb2 as benchmarks_pb2
+
+
+def run_one_test(filename):
+  data = open(filename).read()
+  benchmark_dataset = benchmarks_pb2.BenchmarkDataset()
+  benchmark_dataset.ParseFromString(data)
+  benchmark_util = Benchmark(full_iteration=len(benchmark_dataset.payload),
+                             module="py_benchmark",
+                             setup_method="init")
+  result={}
+  result["filename"] =  filename
+  result["message_name"] =  benchmark_dataset.message_name
+  result["benchmarks"] = {}
+  benchmark_util.set_test_method("parse_from_benchmark")
+  result["benchmarks"][args.behavior_prefix + "_parse_from_benchmark"] = \
+    benchmark_util.run_benchmark(setup_method_args='"%s"' % (filename))
+  benchmark_util.set_test_method("serialize_to_benchmark")
+  result["benchmarks"][args.behavior_prefix + "_serialize_to_benchmark"] = \
+    benchmark_util.run_benchmark(setup_method_args='"%s"' % (filename))
+  return result
+
+
+def init(filename):
+  global benchmark_dataset, message_class, message_list, counter
+  message_list=[]
+  counter = 0
+  data = open(os.path.dirname(sys.argv[0]) + "/../" + filename).read()
+  benchmark_dataset = benchmarks_pb2.BenchmarkDataset()
+  benchmark_dataset.ParseFromString(data)
+
+  if benchmark_dataset.message_name == "benchmarks.proto3.GoogleMessage1":
+    message_class = benchmark_message1_proto3_pb2.GoogleMessage1
+  elif benchmark_dataset.message_name == "benchmarks.proto2.GoogleMessage1":
+    message_class = benchmark_message1_proto2_pb2.GoogleMessage1
+  elif benchmark_dataset.message_name == "benchmarks.proto2.GoogleMessage2":
+    message_class = benchmark_message2_pb2.GoogleMessage2
+  elif benchmark_dataset.message_name == "benchmarks.google_message3.GoogleMessage3":
+    message_class = benchmark_message3_pb2.GoogleMessage3
+  elif benchmark_dataset.message_name == "benchmarks.google_message4.GoogleMessage4":
+    message_class = benchmark_message4_pb2.GoogleMessage4
+  else:
+    raise IOError("Message %s not found!" % (benchmark_dataset.message_name))
+
+  for one_payload in benchmark_dataset.payload:
+    temp = message_class()
+    temp.ParseFromString(one_payload)
+    message_list.append(temp)
+
+
+def parse_from_benchmark():
+  global counter, message_class, benchmark_dataset
+  m = message_class().ParseFromString(benchmark_dataset.payload[counter % len(benchmark_dataset.payload)])
+  counter = counter + 1
+
+
+def serialize_to_benchmark():
+  global counter, message_list, message_class
+  s = message_list[counter % len(benchmark_dataset.payload)].SerializeToString()
+  counter = counter + 1
+
+
+class Benchmark:
+  def __init__(self, module=None, test_method=None,
+               setup_method=None, full_iteration = 1):
+    self.full_iteration = full_iteration
+    self.module = module
+    self.test_method = test_method
+    self.setup_method = setup_method
+
+  def set_test_method(self, test_method):
+    self.test_method = test_method
+
+  def full_setup_code(self, setup_method_args=''):
+    setup_code = ""
+    setup_code += "from %s import %s\n" % (self.module, self.test_method)
+    setup_code += "from %s import %s\n" % (self.module, self.setup_method)
+    setup_code += "%s(%s)\n" % (self.setup_method, setup_method_args)
+    return setup_code
+
+  def dry_run(self, test_method_args='', setup_method_args=''):
+    return timeit.timeit(stmt="%s(%s)" % (self.test_method, test_method_args),
+                         setup=self.full_setup_code(setup_method_args),
+                         number=self.full_iteration);
+
+  def run_benchmark(self, test_method_args='', setup_method_args=''):
+    reps = self.full_iteration;
+    t = self.dry_run(test_method_args, setup_method_args);
+    if t < 3 :
+      reps = int(math.ceil(3 / t)) * self.full_iteration
+    t = timeit.timeit(stmt="%s(%s)" % (self.test_method, test_method_args),
+                      setup=self.full_setup_code(setup_method_args),
+                      number=reps);
+    return 1.0 * t / reps * (10 ** 9)
+  
+
+if __name__ == "__main__":
+  results = []
+  for file in args.data_files:
+    results.append(run_one_test(file))
+  
+  if args.json != "no":
+    print json.dumps(results)
+  else:
+    for result in results:
+      print "Message %s of dataset file %s" % \
+          (result["message_name"], result["filename"])
+      print "Average time for parse_from_benchmark: %.2f ns" % \
+          (result["benchmarks"][ \
+                      args.behavior_prefix + "_parse_from_benchmark"])
+      print "Average time for serialize_to_benchmark: %.2f ns" % \
+          (result["benchmarks"][ \
+                      args.behavior_prefix + "_serialize_to_benchmark"])
+      print ""
diff --git a/benchmarks/python/python_benchmark_messages.cc b/benchmarks/python/python_benchmark_messages.cc
new file mode 100644
index 0000000..ded16fe
--- /dev/null
+++ b/benchmarks/python/python_benchmark_messages.cc
@@ -0,0 +1,29 @@
+#include <Python.h>
+
+#include "benchmarks.pb.h"
+#include "datasets/google_message1/proto2/benchmark_message1_proto2.pb.h"
+#include "datasets/google_message1/proto3/benchmark_message1_proto3.pb.h"
+#include "datasets/google_message2/benchmark_message2.pb.h"
+#include "datasets/google_message3/benchmark_message3.pb.h"
+#include "datasets/google_message4/benchmark_message4.pb.h"
+
+static PyMethodDef python_benchmark_methods[] = {
+    {NULL, NULL, 0, NULL}        /* Sentinel */
+};
+
+
+PyMODINIT_FUNC
+initlibbenchmark_messages() {
+  benchmarks::BenchmarkDataset().descriptor();
+  benchmarks::proto3::GoogleMessage1().descriptor();
+  benchmarks::proto2::GoogleMessage1().descriptor();
+  benchmarks::proto2::GoogleMessage2().descriptor();
+  benchmarks::google_message3::GoogleMessage3().descriptor();
+  benchmarks::google_message4::GoogleMessage4().descriptor();
+
+  PyObject *m;
+
+  m = Py_InitModule("libbenchmark_messages", python_benchmark_methods);
+  if (m == NULL)
+      return;
+}