Merge branch 'master' of https://github.com/grpc/grpc into ALL-the-things

13 files changed, 693 insertions, 265 deletions

diff --git a/tools/doxygen/Doxyfile.core.internal b/tools/doxygen/Doxyfile.core.internal
index 70e332970f..766c20f59b 100644
--- a/tools/doxygen/Doxyfile.core.internal
+++ b/tools/doxygen/Doxyfile.core.internal
@@ -948,6 +948,7 @@ src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_ev_driver.h \
 src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_ev_driver_posix.c \
 src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.c \
 src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h \
+src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_fallback.c \
 src/core/ext/filters/client_channel/resolver/dns/native/README.md \
 src/core/ext/filters/client_channel/resolver/dns/native/dns_resolver.c \
 src/core/ext/filters/client_channel/resolver/fake/fake_resolver.c \
diff --git a/tools/jenkins/run_performance.sh b/tools/jenkins/run_performance.sh
index fc61b3b900..177b076650 100755
--- a/tools/jenkins/run_performance.sh
+++ b/tools/jenkins/run_performance.sh
@@ -23,4 +23,4 @@ BENCHMARKS_TO_RUN="bm_fullstack_unary_ping_pong bm_fullstack_streaming_ping_pong
 cd $(dirname $0)/../..
 
 tools/run_tests/start_port_server.py
-tools/profiling/microbenchmarks/bm_diff.py -d origin/$ghprbTargetBranch -b $BENCHMARKS_TO_RUN
+tools/profiling/microbenchmarks/bm_diff/bm_main.py -d origin/$ghprbTargetBranch -b $BENCHMARKS_TO_RUN
diff --git a/tools/profiling/microbenchmarks/README.md b/tools/profiling/microbenchmarks/README.md
new file mode 100644
index 0000000000..035888ee18
--- /dev/null
+++ b/tools/profiling/microbenchmarks/README.md
@@ -0,0 +1,4 @@
+Microbenchmarks
+====
+
+This directory contains helper scripts for the microbenchmark suites.
diff --git a/tools/profiling/microbenchmarks/bm_diff.py b/tools/profiling/microbenchmarks/bm_diff.py
deleted file mode 100755
index 3de5c663ce..0000000000
--- a/tools/profiling/microbenchmarks/bm_diff.py
+++ /dev/null
@@ -1,244 +0,0 @@
-#!/usr/bin/env python2.7
-# Copyright 2017 gRPC authors.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import sys
-import json
-import bm_json
-import tabulate
-import argparse
-from scipy import stats
-import subprocess
-import multiprocessing
-import collections
-import pipes
-import os
-sys.path.append(os.path.join(os.path.dirname(sys.argv[0]), '..', '..', 'run_tests', 'python_utils'))
-import comment_on_pr
-import jobset
-import itertools
-import speedup
-import random
-import shutil
-import errno
-
-_INTERESTING = (
-  'cpu_time',
-  'real_time',
-  'locks_per_iteration',
-  'allocs_per_iteration',
-  'writes_per_iteration',
-  'atm_cas_per_iteration',
-  'atm_add_per_iteration',
-  'cli_transport_stalls_per_iteration',
-  'cli_stream_stalls_per_iteration',
-  'svr_transport_stalls_per_iteration',
-  'svr_stream_stalls_per_iteration'
-  'nows_per_iteration',
-)
-
-def changed_ratio(n, o):
-  if float(o) <= .0001: o = 0
-  if float(n) <= .0001: n = 0
-  if o == 0 and n == 0: return 0
-  if o == 0: return 100
-  return (float(n)-float(o))/float(o)
-
-def median(ary):
-  ary = sorted(ary)
-  n = len(ary)
-  if n%2 == 0:
-    return (ary[n/2] + ary[n/2+1]) / 2.0
-  else:
-    return ary[n/2]
-
-def min_change(pct):
-  return lambda n, o: abs(changed_ratio(n,o)) > pct/100.0
-
-_AVAILABLE_BENCHMARK_TESTS = ['bm_fullstack_unary_ping_pong',
-                              'bm_fullstack_streaming_ping_pong',
-                              'bm_fullstack_streaming_pump',
-                              'bm_closure',
-                              'bm_cq',
-                              'bm_call_create',
-                              'bm_error',
-                              'bm_chttp2_hpack',
-                              'bm_chttp2_transport',
-                              'bm_pollset',
-                              'bm_metadata',
-                              'bm_fullstack_trickle']
-
-argp = argparse.ArgumentParser(description='Perform diff on microbenchmarks')
-argp.add_argument('-t', '--track',
-                  choices=sorted(_INTERESTING),
-                  nargs='+',
-                  default=sorted(_INTERESTING),
-                  help='Which metrics to track')
-argp.add_argument('-b', '--benchmarks', nargs='+', choices=_AVAILABLE_BENCHMARK_TESTS, default=['bm_cq'])
-argp.add_argument('-d', '--diff_base', type=str)
-argp.add_argument('-r', '--repetitions', type=int, default=1)
-argp.add_argument('-l', '--loops', type=int, default=20)
-argp.add_argument('-j', '--jobs', type=int, default=multiprocessing.cpu_count())
-args = argp.parse_args()
-
-assert args.diff_base
-
-def avg(lst):
-  sum = 0.0
-  n = 0.0
-  for el in lst:
-    sum += el
-    n += 1
-  return sum / n
-
-def make_cmd(cfg):
-  return ['make'] + args.benchmarks + [
-      'CONFIG=%s' % cfg, '-j', '%d' % args.jobs]
-
-def build(dest):
-  shutil.rmtree('bm_diff_%s' % dest, ignore_errors=True)
-  subprocess.check_call(['git', 'submodule', 'update'])
-  try:
-    subprocess.check_call(make_cmd('opt'))
-    subprocess.check_call(make_cmd('counters'))
-  except subprocess.CalledProcessError, e:
-    subprocess.check_call(['make', 'clean'])
-    subprocess.check_call(make_cmd('opt'))
-    subprocess.check_call(make_cmd('counters'))
-  os.rename('bins', 'bm_diff_%s' % dest)
-
-def collect1(bm, cfg, ver, idx):
-  cmd = ['bm_diff_%s/%s/%s' % (ver, cfg, bm),
-         '--benchmark_out=%s.%s.%s.%d.json' % (bm, cfg, ver, idx),
-         '--benchmark_out_format=json',
-         '--benchmark_repetitions=%d' % (args.repetitions)
-         ]
-  return jobset.JobSpec(cmd, shortname='%s %s %s %d/%d' % (bm, cfg, ver, idx+1, args.loops),
-                             verbose_success=True, timeout_seconds=None)
-
-build('new')
-
-where_am_i = subprocess.check_output(['git', 'rev-parse', '--abbrev-ref', 'HEAD']).strip()
-subprocess.check_call(['git', 'checkout', args.diff_base])
-try:
-  build('old')
-finally:
-  subprocess.check_call(['git', 'checkout', where_am_i])
-  subprocess.check_call(['git', 'submodule', 'update'])
-
-jobs = []
-for loop in range(0, args.loops):
-  jobs.extend(x for x in itertools.chain(
-    (collect1(bm, 'opt', 'new', loop) for bm in args.benchmarks),
-    (collect1(bm, 'counters', 'new', loop) for bm in args.benchmarks),
-    (collect1(bm, 'opt', 'old', loop) for bm in args.benchmarks),
-    (collect1(bm, 'counters', 'old', loop) for bm in args.benchmarks),
-  ))
-random.shuffle(jobs, random.SystemRandom().random)
-
-jobset.run(jobs, maxjobs=args.jobs)
-
-class Benchmark:
-
-  def __init__(self):
-    self.samples = {
-      True: collections.defaultdict(list),
-      False: collections.defaultdict(list)
-    }
-    self.final = {}
-
-  def add_sample(self, data, new):
-    for f in args.track:
-      if f in data:
-        self.samples[new][f].append(float(data[f]))
-
-  def process(self):
-    for f in sorted(args.track):
-      new = self.samples[True][f]
-      old = self.samples[False][f]
-      if not new or not old: continue
-      mdn_diff = abs(median(new) - median(old))
-      print '%s: new=%r old=%r mdn_diff=%r' % (f, new, old, mdn_diff)
-      s = speedup.speedup(new, old)
-      if abs(s) > 3 and mdn_diff > 0.5:
-        self.final[f] = '%+d%%' % s
-    return self.final.keys()
-
-  def skip(self):
-    return not self.final
-
-  def row(self, flds):
-    return [self.final[f] if f in self.final else '' for f in flds]
-
-
-def eintr_be_gone(fn):
-  """Run fn until it doesn't stop because of EINTR"""
-  while True:
-    try:
-      return fn()
-    except IOError, e:
-      if e.errno != errno.EINTR:
-        raise
-
-
-def read_json(filename):
-  try:
-    with open(filename) as f: return json.loads(f.read())
-  except ValueError, e:
-    return None
-
-
-def finalize():
-  benchmarks = collections.defaultdict(Benchmark)
-
-  for bm in args.benchmarks:
-    for loop in range(0, args.loops):
-      js_new_ctr = read_json('%s.counters.new.%d.json' % (bm, loop))
-      js_new_opt = read_json('%s.opt.new.%d.json' % (bm, loop))
-      js_old_ctr = read_json('%s.counters.old.%d.json' % (bm, loop))
-      js_old_opt = read_json('%s.opt.old.%d.json' % (bm, loop))
-
-      if js_new_ctr:
-        for row in bm_json.expand_json(js_new_ctr, js_new_opt):
-          print row
-          name = row['cpp_name']
-          if name.endswith('_mean') or name.endswith('_stddev'): continue
-          benchmarks[name].add_sample(row, True)
-      if js_old_ctr:
-        for row in bm_json.expand_json(js_old_ctr, js_old_opt):
-          print row
-          name = row['cpp_name']
-          if name.endswith('_mean') or name.endswith('_stddev'): continue
-          benchmarks[name].add_sample(row, False)
-
-  really_interesting = set()
-  for name, bm in benchmarks.items():
-    print name
-    really_interesting.update(bm.process())
-  fields = [f for f in args.track if f in really_interesting]
-
-  headers = ['Benchmark'] + fields
-  rows = []
-  for name in sorted(benchmarks.keys()):
-    if benchmarks[name].skip(): continue
-    rows.append([name] + benchmarks[name].row(fields))
-  if rows:
-    text = 'Performance differences noted:\n' + tabulate.tabulate(rows, headers=headers, floatfmt='+.2f')
-  else:
-    text = 'No significant performance differences'
-  print text
-  comment_on_pr.comment_on_pr('```\n%s\n```' % text)
-
-
-eintr_be_gone(finalize)
diff --git a/tools/profiling/microbenchmarks/bm_diff/README.md b/tools/profiling/microbenchmarks/bm_diff/README.md
new file mode 100644
index 0000000000..caa4770229
--- /dev/null
+++ b/tools/profiling/microbenchmarks/bm_diff/README.md
@@ -0,0 +1,116 @@
+The bm_diff Family
+====
+
+This family of python scripts can be incredibly useful for fast iteration over
+different performance tweaks. The tools allow you to save performance data from
+a baseline commit, then quickly compare data from your working branch to that
+baseline data to see if you have made any performance wins.
+
+The tools operate with three concrete steps, which can be invoked separately,
+or all together via the driver script, bm_main.py. This readme will describe 
+the typical workflow for these scripts, then it will include sections on the
+details of every script for advanced usage.
+
+## Normal Workflow
+
+Let's say you are working on a performance optimization for grpc_error. You have
+made some significant changes and want to see some data. From your branch, run
+(ensure everything is committed first):
+
+`tools/profiling/microbenchmarks/bm_diff/bm_main.py -b bm_error -l 5 -d master`
+
+This will build the `bm_error` binary on your branch, and then it will checkout 
+master and build it there too. It will then run these benchmarks 5 times each. 
+Lastly it will compute the statistically significant performance differences 
+between the two branches. This should show the nice performance wins your 
+changes have made.
+
+If you have already invoked bm_main with `-d master`, you should instead use 
+`-o` for subsequent runs. This allows the script to skip re-building and 
+re-running the unchanged master branch. For example:
+
+`tools/profiling/microbenchmarks/bm_diff/bm_main.py -b bm_error -l 5 -o`
+
+This will only build and run `bm_error` on your branch. It will then compare
+the output to the saved runs from master.
+
+## Advanced Workflow
+
+If you have a deeper knowledge of these scripts, you can use them to do more
+fine tuned benchmark comparisons. For example, you could build, run, and save
+the benchmark output from two different base branches. Then you could diff both
+of these baselines against your working branch to see how the different metrics
+change. The rest of this doc goes over the details of what each of the
+individual modules accomplishes.
+
+## bm_build.py
+
+This scrips builds the benchmarks. It takes in a name parameter, and will
+store the binaries based on that. Both `opt` and `counter` configurations
+will be used. The `opt` is used to get cpu_time and real_time, and the
+`counters` build is used to track other metrics like allocs, atomic adds,
+etc etc etc.
+
+For example, if you were to invoke (we assume everything is run from the 
+root of the repo):
+
+`tools/profiling/microbenchmarks/bm_diff/bm_build.py -b bm_error -n baseline`
+
+then the microbenchmark binaries will show up under 
+`bm_diff_baseline/{opt,counters}/bm_error`
+
+## bm_run.py
+
+This script runs the benchmarks. It takes a name parameter that must match the
+name that was passed to `bm_build.py`. The script then runs the benchmark
+multiple times (default is 20, can be toggled via the loops parameter). The
+output is saved as `<benchmark name>.<config>.<name>.<loop idx>.json`
+
+For example, if you were to run:
+
+`tools/profiling/microbenchmarks/bm_diff/bm_run.py -b bm_error -b baseline -l 5`
+
+Then an example output file would be `bm_error.opt.baseline.0.json`
+
+## bm_diff.py
+
+This script takes in the output from two benchmark runs, computes the diff
+between them, and prints any significant improvements or regressions. It takes
+in two name parameters, old and new. These must have previously been built and
+run.
+
+For example, assuming you had already built and run a 'baseline' microbenchmark
+from master, and then you also built and ran a 'current' microbenchmark from
+the branch you were working on, you could invoke:
+
+`tools/profiling/microbenchmarks/bm_diff/bm_diff.py -b bm_error -o baseline -n current -l 5`
+
+This would output the percent difference between your branch and master.
+
+## bm_main.py
+
+This is the driver script. It uses the previous three modules and does
+everything for you. You pass in the benchmarks to be run, the number of loops,
+number of CPUs to use, and the commit to compare to. Then the script will:
+* Build the benchmarks at head, then checkout the branch to compare to and
+  build the benchmarks there
+* Run both sets of microbenchmarks
+* Run bm_diff.py to compare the two, outputs the difference.
+
+For example, one might run:
+
+`tools/profiling/microbenchmarks/bm_diff/bm_main.py -b bm_error -l 5 -d master`
+
+This would compare the current branch's error benchmarks to master.
+
+This script is invoked by our infrastructure on every PR to protect against
+regressions and demonstrate performance wins.
+
+However, if you are iterating over different performance tweaks quickly, it is
+unnecessary to build and run the baseline commit every time. That is why we
+provide a different flag in case you are sure that the baseline benchmark has
+already been built and run. In that case use the --old flag to pass in the name
+of the baseline. This will only build and run the current branch. For example:
+
+`tools/profiling/microbenchmarks/bm_diff/bm_main.py -b bm_error -l 5 -o old`
+
diff --git a/tools/profiling/microbenchmarks/bm_diff/bm_build.py b/tools/profiling/microbenchmarks/bm_diff/bm_build.py
new file mode 100755
index 0000000000..650ccdc2b2
--- /dev/null
+++ b/tools/profiling/microbenchmarks/bm_diff/bm_build.py
@@ -0,0 +1,75 @@
+#!/usr/bin/env python2.7
+#
+# Copyright 2017 gRPC authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" Python utility to build opt and counters benchmarks """
+
+import bm_constants
+
+import argparse
+import subprocess
+import multiprocessing
+import os
+import shutil
+
+
+def _args():
+  argp = argparse.ArgumentParser(description='Builds microbenchmarks')
+  argp.add_argument(
+    '-b',
+    '--benchmarks',
+    nargs='+',
+    choices=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    default=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    help='Which benchmarks to build')
+  argp.add_argument(
+    '-j',
+    '--jobs',
+    type=int,
+    default=multiprocessing.cpu_count(),
+    help='How many CPUs to dedicate to this task')
+  argp.add_argument(
+    '-n',
+    '--name',
+    type=str,
+    help='Unique name of this build. To be used as a handle to pass to the other bm* scripts'
+  )
+  args = argp.parse_args()
+  assert args.name
+  return args
+
+
+def _make_cmd(cfg, benchmarks, jobs):
+  return ['make'] + benchmarks + ['CONFIG=%s' % cfg, '-j', '%d' % jobs]
+
+
+def build(name, benchmarks, jobs):
+  shutil.rmtree('bm_diff_%s' % name, ignore_errors=True)
+  subprocess.check_call(['git', 'submodule', 'update'])
+  try:
+    subprocess.check_call(_make_cmd('opt', benchmarks, jobs))
+    subprocess.check_call(_make_cmd('counters', benchmarks, jobs))
+  except subprocess.CalledProcessError, e:
+    subprocess.check_call(['make', 'clean'])
+    subprocess.check_call(_make_cmd('opt', benchmarks, jobs))
+    subprocess.check_call(_make_cmd('counters', benchmarks, jobs))
+  os.rename(
+    'bins',
+    'bm_diff_%s' % name,)
+
+
+if __name__ == '__main__':
+  args = _args()
+  build(args.name, args.benchmarks, args.jobs)
diff --git a/tools/profiling/microbenchmarks/bm_diff/bm_constants.py b/tools/profiling/microbenchmarks/bm_diff/bm_constants.py
new file mode 100644
index 0000000000..2bbd987b20
--- /dev/null
+++ b/tools/profiling/microbenchmarks/bm_diff/bm_constants.py
@@ -0,0 +1,29 @@
+#!/usr/bin/env python2.7
+#
+# Copyright 2017 gRPC authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" Configurable constants for the bm_*.py family """
+
+_AVAILABLE_BENCHMARK_TESTS = [
+  'bm_fullstack_unary_ping_pong', 'bm_fullstack_streaming_ping_pong',
+  'bm_fullstack_streaming_pump', 'bm_closure', 'bm_cq', 'bm_call_create',
+  'bm_error', 'bm_chttp2_hpack', 'bm_chttp2_transport', 'bm_pollset',
+  'bm_metadata', 'bm_fullstack_trickle'
+]
+
+_INTERESTING = ('cpu_time', 'real_time', 'locks_per_iteration',
+        'allocs_per_iteration', 'writes_per_iteration',
+        'atm_cas_per_iteration', 'atm_add_per_iteration',
+        'nows_per_iteration',)
diff --git a/tools/profiling/microbenchmarks/bm_diff/bm_diff.py b/tools/profiling/microbenchmarks/bm_diff/bm_diff.py
new file mode 100755
index 0000000000..881e157ccd
--- /dev/null
+++ b/tools/profiling/microbenchmarks/bm_diff/bm_diff.py
@@ -0,0 +1,201 @@
+#!/usr/bin/env python2.7
+#
+# Copyright 2017 gRPC authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Computes the diff between two bm runs and outputs significant results """
+
+import bm_constants
+import bm_speedup
+
+import sys
+import os
+
+sys.path.append(os.path.join(os.path.dirname(sys.argv[0]), '..'))
+import bm_json
+
+import json
+import tabulate
+import argparse
+import collections
+import subprocess
+
+verbose = False
+
+
+def _median(ary):
+  assert (len(ary))
+  ary = sorted(ary)
+  n = len(ary)
+  if n % 2 == 0:
+    return (ary[(n - 1) / 2] + ary[(n - 1) / 2 + 1]) / 2.0
+  else:
+    return ary[n / 2]
+
+
+def _args():
+  argp = argparse.ArgumentParser(
+    description='Perform diff on microbenchmarks')
+  argp.add_argument(
+    '-t',
+    '--track',
+    choices=sorted(bm_constants._INTERESTING),
+    nargs='+',
+    default=sorted(bm_constants._INTERESTING),
+    help='Which metrics to track')
+  argp.add_argument(
+    '-b',
+    '--benchmarks',
+    nargs='+',
+    choices=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    default=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    help='Which benchmarks to run')
+  argp.add_argument(
+    '-l',
+    '--loops',
+    type=int,
+    default=20,
+    help='Number of times to loops the benchmarks. Must match what was passed to bm_run.py'
+  )
+  argp.add_argument('-n', '--new', type=str, help='New benchmark name')
+  argp.add_argument('-o', '--old', type=str, help='Old benchmark name')
+  argp.add_argument(
+    '-v', '--verbose', type=bool, help='Print details of before/after')
+  args = argp.parse_args()
+  global verbose
+  if args.verbose: verbose = True
+  assert args.new
+  assert args.old
+  return args
+
+
+def _maybe_print(str):
+  if verbose: print str
+
+
+class Benchmark:
+
+  def __init__(self):
+    self.samples = {
+      True: collections.defaultdict(list),
+      False: collections.defaultdict(list)
+    }
+    self.final = {}
+
+  def add_sample(self, track, data, new):
+    for f in track:
+      if f in data:
+        self.samples[new][f].append(float(data[f]))
+
+  def process(self, track, new_name, old_name):
+    for f in sorted(track):
+      new = self.samples[True][f]
+      old = self.samples[False][f]
+      if not new or not old: continue
+      mdn_diff = abs(_median(new) - _median(old))
+      _maybe_print('%s: %s=%r %s=%r mdn_diff=%r' %
+             (f, new_name, new, old_name, old, mdn_diff))
+      s = bm_speedup.speedup(new, old)
+      if abs(s) > 3 and mdn_diff > 0.5:
+        self.final[f] = '%+d%%' % s
+    return self.final.keys()
+
+  def skip(self):
+    return not self.final
+
+  def row(self, flds):
+    return [self.final[f] if f in self.final else '' for f in flds]
+
+
+def _read_json(filename, badjson_files, nonexistant_files):
+  stripped = ".".join(filename.split(".")[:-2])
+  try:
+    with open(filename) as f:
+      return json.loads(f.read())
+  except IOError, e:
+    if stripped in nonexistant_files:
+      nonexistant_files[stripped] += 1
+    else:
+      nonexistant_files[stripped] = 1
+    return None
+  except ValueError, e:
+    if stripped in badjson_files:
+      badjson_files[stripped] += 1
+    else:
+      badjson_files[stripped] = 1
+    return None
+
+
+def diff(bms, loops, track, old, new):
+  benchmarks = collections.defaultdict(Benchmark)
+
+  badjson_files = {}
+  nonexistant_files = {}
+  for bm in bms:
+    for loop in range(0, loops):
+      for line in subprocess.check_output(
+        ['bm_diff_%s/opt/%s' % (old, bm),
+         '--benchmark_list_tests']).splitlines():
+        stripped_line = line.strip().replace("/", "_").replace(
+          "<", "_").replace(">", "_").replace(", ", "_")
+        js_new_ctr = _read_json('%s.%s.counters.%s.%d.json' %
+                    (bm, stripped_line, new, loop),
+                    badjson_files, nonexistant_files)
+        js_new_opt = _read_json('%s.%s.opt.%s.%d.json' %
+                    (bm, stripped_line, new, loop),
+                    badjson_files, nonexistant_files)
+        js_old_ctr = _read_json('%s.%s.counters.%s.%d.json' %
+                    (bm, stripped_line, old, loop),
+                    badjson_files, nonexistant_files)
+        js_old_opt = _read_json('%s.%s.opt.%s.%d.json' %
+                    (bm, stripped_line, old, loop),
+                    badjson_files, nonexistant_files)
+
+        if js_new_ctr:
+          for row in bm_json.expand_json(js_new_ctr, js_new_opt):
+            name = row['cpp_name']
+            if name.endswith('_mean') or name.endswith('_stddev'):
+              continue
+            benchmarks[name].add_sample(track, row, True)
+        if js_old_ctr:
+          for row in bm_json.expand_json(js_old_ctr, js_old_opt):
+            name = row['cpp_name']
+            if name.endswith('_mean') or name.endswith('_stddev'):
+              continue
+            benchmarks[name].add_sample(track, row, False)
+
+  really_interesting = set()
+  for name, bm in benchmarks.items():
+    _maybe_print(name)
+    really_interesting.update(bm.process(track, new, old))
+  fields = [f for f in track if f in really_interesting]
+
+  headers = ['Benchmark'] + fields
+  rows = []
+  for name in sorted(benchmarks.keys()):
+    if benchmarks[name].skip(): continue
+    rows.append([name] + benchmarks[name].row(fields))
+  note = 'Corrupt JSON data (indicates timeout or crash) = %s' % str(
+    badjson_files)
+  note += '\n\nMissing files (new benchmark) = %s' % str(nonexistant_files)
+  if rows:
+    return tabulate.tabulate(rows, headers=headers, floatfmt='+.2f'), note
+  else:
+    return None, note
+
+
+if __name__ == '__main__':
+  args = _args()
+  diff, note = diff(args.benchmarks, args.loops, args.track, args.old,
+            args.new)
+  print('%s\n%s' % (note, diff if diff else "No performance differences"))
diff --git a/tools/profiling/microbenchmarks/bm_diff/bm_main.py b/tools/profiling/microbenchmarks/bm_diff/bm_main.py
new file mode 100755
index 0000000000..0136c6aa57
--- /dev/null
+++ b/tools/profiling/microbenchmarks/bm_diff/bm_main.py
@@ -0,0 +1,135 @@
+#!/usr/bin/env python2.7
+#
+# Copyright 2017 gRPC authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" Runs the entire bm_*.py pipeline, and possible comments on the PR """
+
+import bm_constants
+import bm_build
+import bm_run
+import bm_diff
+
+import sys
+import os
+import argparse
+import multiprocessing
+import subprocess
+
+sys.path.append(
+  os.path.join(
+    os.path.dirname(sys.argv[0]), '..', '..', 'run_tests', 'python_utils'))
+import comment_on_pr
+
+
+def _args():
+  argp = argparse.ArgumentParser(
+    description='Perform diff on microbenchmarks')
+  argp.add_argument(
+    '-t',
+    '--track',
+    choices=sorted(bm_constants._INTERESTING),
+    nargs='+',
+    default=sorted(bm_constants._INTERESTING),
+    help='Which metrics to track')
+  argp.add_argument(
+    '-b',
+    '--benchmarks',
+    nargs='+',
+    choices=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    default=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    help='Which benchmarks to run')
+  argp.add_argument(
+    '-d',
+    '--diff_base',
+    type=str,
+    help='Commit or branch to compare the current one to')
+  argp.add_argument(
+    '-o',
+    '--old',
+    default='old',
+    type=str,
+    help='Name of baseline run to compare to. Ususally just called "old"')
+  argp.add_argument(
+    '-r',
+    '--repetitions',
+    type=int,
+    default=1,
+    help='Number of repetitions to pass to the benchmarks')
+  argp.add_argument(
+    '-l',
+    '--loops',
+    type=int,
+    default=20,
+    help='Number of times to loops the benchmarks. More loops cuts down on noise'
+  )
+  argp.add_argument(
+    '-j',
+    '--jobs',
+    type=int,
+    default=multiprocessing.cpu_count(),
+    help='Number of CPUs to use')
+  args = argp.parse_args()
+  assert args.diff_base or args.old, "One of diff_base or old must be set!"
+  if args.loops < 3:
+    print "WARNING: This run will likely be noisy. Increase loops."
+  return args
+
+
+def eintr_be_gone(fn):
+  """Run fn until it doesn't stop because of EINTR"""
+
+  def inner(*args):
+    while True:
+      try:
+        return fn(*args)
+      except IOError, e:
+        if e.errno != errno.EINTR:
+          raise
+
+  return inner
+
+
+def main(args):
+
+  bm_build.build('new', args.benchmarks, args.jobs)
+
+  old = args.old
+  if args.diff_base:
+    old = 'old'
+    where_am_i = subprocess.check_output(
+      ['git', 'rev-parse', '--abbrev-ref', 'HEAD']).strip()
+    subprocess.check_call(['git', 'checkout', args.diff_base])
+    try:
+      bm_build.build('old', args.benchmarks, args.jobs)
+    finally:
+      subprocess.check_call(['git', 'checkout', where_am_i])
+      subprocess.check_call(['git', 'submodule', 'update'])
+
+  bm_run.run('new', args.benchmarks, args.jobs, args.loops, args.repetitions)
+  bm_run.run(old, args.benchmarks, args.jobs, args.loops, args.repetitions)
+
+  diff, note = bm_diff.diff(args.benchmarks, args.loops, args.track, old,
+                'new')
+  if diff:
+    text = 'Performance differences noted:\n' + diff
+  else:
+    text = 'No significant performance differences'
+  print('%s\n%s' % (note, text))
+  comment_on_pr.comment_on_pr('```\n%s\n\n%s\n```' % (note, text))
+
+
+if __name__ == '__main__':
+  args = _args()
+  main(args)
diff --git a/tools/profiling/microbenchmarks/bm_diff/bm_run.py b/tools/profiling/microbenchmarks/bm_diff/bm_run.py
new file mode 100755
index 0000000000..3457af916b
--- /dev/null
+++ b/tools/profiling/microbenchmarks/bm_diff/bm_run.py
@@ -0,0 +1,113 @@
+#!/usr/bin/env python2.7
+#
+# Copyright 2017 gRPC authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" Python utility to run opt and counters benchmarks and save json output """
+
+import bm_constants
+
+import argparse
+import subprocess
+import multiprocessing
+import random
+import itertools
+import sys
+import os
+
+sys.path.append(
+  os.path.join(
+    os.path.dirname(sys.argv[0]), '..', '..', '..', 'run_tests',
+    'python_utils'))
+import jobset
+
+
+def _args():
+  argp = argparse.ArgumentParser(description='Runs microbenchmarks')
+  argp.add_argument(
+    '-b',
+    '--benchmarks',
+    nargs='+',
+    choices=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    default=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+    help='Benchmarks to run')
+  argp.add_argument(
+    '-j',
+    '--jobs',
+    type=int,
+    default=multiprocessing.cpu_count(),
+    help='Number of CPUs to use')
+  argp.add_argument(
+    '-n',
+    '--name',
+    type=str,
+    help='Unique name of the build to run. Needs to match the handle passed to bm_build.py'
+  )
+  argp.add_argument(
+    '-r',
+    '--repetitions',
+    type=int,
+    default=1,
+    help='Number of repetitions to pass to the benchmarks')
+  argp.add_argument(
+    '-l',
+    '--loops',
+    type=int,
+    default=20,
+    help='Number of times to loops the benchmarks. More loops cuts down on noise'
+  )
+  args = argp.parse_args()
+  assert args.name
+  if args.loops < 3:
+    print "WARNING: This run will likely be noisy. Increase loops to at least 3."
+  return args
+
+
+def _collect_bm_data(bm, cfg, name, reps, idx, loops):
+  jobs_list = []
+  for line in subprocess.check_output(
+    ['bm_diff_%s/%s/%s' % (name, cfg, bm),
+     '--benchmark_list_tests']).splitlines():
+    stripped_line = line.strip().replace("/", "_").replace(
+      "<", "_").replace(">", "_").replace(", ", "_")
+    cmd = [
+      'bm_diff_%s/%s/%s' % (name, cfg, bm), '--benchmark_filter=^%s$' %
+      line, '--benchmark_out=%s.%s.%s.%s.%d.json' %
+      (bm, stripped_line, cfg, name, idx), '--benchmark_out_format=json',
+      '--benchmark_repetitions=%d' % (reps)
+    ]
+    jobs_list.append(
+      jobset.JobSpec(
+        cmd,
+        shortname='%s %s %s %s %d/%d' % (bm, line, cfg, name, idx + 1,
+                         loops),
+        verbose_success=True,
+        timeout_seconds=60 * 2))
+  return jobs_list
+
+
+def run(name, benchmarks, jobs, loops, reps):
+  jobs_list = []
+  for loop in range(0, loops):
+    for bm in benchmarks:
+      jobs_list += _collect_bm_data(bm, 'opt', name, reps, loop, loops)
+      jobs_list += _collect_bm_data(bm, 'counters', name, reps, loop,
+                      loops)
+  random.shuffle(jobs_list, random.SystemRandom().random)
+  jobset.run(jobs_list, maxjobs=jobs)
+
+
+if __name__ == '__main__':
+  args = _args()
+  run(args.name, args.benchmarks, args.jobs, args.loops, args.repetitions)
diff --git a/tools/profiling/microbenchmarks/speedup.py b/tools/profiling/microbenchmarks/bm_diff/bm_speedup.py
index 3f45cc38fb..3d126efa62 100644..100755
--- a/tools/profiling/microbenchmarks/speedup.py
+++ b/tools/profiling/microbenchmarks/bm_diff/bm_speedup.py
@@ -1,3 +1,5 @@
+#!/usr/bin/env python2.7
+#
 # Copyright 2017 gRPC authors.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -17,13 +19,17 @@ import math
 
 _THRESHOLD = 1e-10
 
+
 def scale(a, mul):
-  return [x*mul for x in a]
+  return [x * mul for x in a]
+
 
 def cmp(a, b):
   return stats.ttest_ind(a, b)
 
+
 def speedup(new, old):
+  if (len(set(new))) == 1 and new == old: return 0
   s0, p0 = cmp(new, old)
   if math.isnan(p0): return 0
   if s0 == 0: return 0
@@ -31,7 +37,7 @@ def speedup(new, old):
   if s0 < 0:
     pct = 1
     while pct < 101:
-      sp, pp = cmp(new, scale(old, 1 - pct/100.0))
+      sp, pp = cmp(new, scale(old, 1 - pct / 100.0))
       if sp > 0: break
       if pp > _THRESHOLD: break
       pct += 1
@@ -39,14 +45,15 @@ def speedup(new, old):
   else:
     pct = 1
     while pct < 100000:
-      sp, pp = cmp(new, scale(old, 1 + pct/100.0))
+      sp, pp = cmp(new, scale(old, 1 + pct / 100.0))
       if sp < 0: break
       if pp > _THRESHOLD: break
       pct += 1
     return pct - 1
 
+
 if __name__ == "__main__":
-  new=[66034560.0, 126765693.0, 99074674.0, 98588433.0, 96731372.0, 110179725.0, 103802110.0, 101139800.0, 102357205.0, 99016353.0, 98840824.0, 99585632.0, 98791720.0, 96171521.0, 95327098.0, 95629704.0, 98209772.0, 99779411.0, 100182488.0, 98354192.0, 99644781.0, 98546709.0, 99019176.0, 99543014.0, 99077269.0, 98046601.0, 99319039.0, 98542572.0, 98886614.0, 72560968.0]
-  old=[60423464.0, 71249570.0, 73213089.0, 73200055.0, 72911768.0, 72347798.0, 72494672.0, 72756976.0, 72116565.0, 71541342.0, 73442538.0, 74817383.0, 73007780.0, 72499062.0, 72404945.0, 71843504.0, 73245405.0, 72778304.0, 74004519.0, 73694464.0, 72919931.0, 72955481.0, 71583857.0, 71350467.0, 71836817.0, 70064115.0, 70355345.0, 72516202.0, 71716777.0, 71532266.0]
+  new = [1.0, 1.0, 1.0, 1.0]
+  old = [2.0, 2.0, 2.0, 2.0]
   print speedup(new, old)
   print speedup(old, new)
diff --git a/tools/run_tests/generated/sources_and_headers.json b/tools/run_tests/generated/sources_and_headers.json
index bfcdc6060a..16ff417a27 100644
--- a/tools/run_tests/generated/sources_and_headers.json
+++ b/tools/run_tests/generated/sources_and_headers.json
@@ -8518,7 +8518,8 @@
       "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_ev_driver.h", 
       "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_ev_driver_posix.c", 
       "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.c", 
-      "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h"
+      "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h", 
+      "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_fallback.c"
     ], 
     "third_party": false, 
     "type": "filegroup"
diff --git a/tools/run_tests/run_microbenchmark.py b/tools/run_tests/run_microbenchmark.py
index 3eecd8eff4..312abd59c4 100755
--- a/tools/run_tests/run_microbenchmark.py
+++ b/tools/run_tests/run_microbenchmark.py
@@ -23,18 +23,8 @@ import argparse
 import python_utils.jobset as jobset
 import python_utils.start_port_server as start_port_server
 
-_AVAILABLE_BENCHMARK_TESTS = ['bm_fullstack_unary_ping_pong',
-                              'bm_fullstack_streaming_ping_pong',
-                              'bm_fullstack_streaming_pump',
-                              'bm_closure',
-                              'bm_cq',
-                              'bm_call_create',
-                              'bm_error',
-                              'bm_chttp2_hpack',
-                              'bm_chttp2_transport',
-                              'bm_pollset',
-                              'bm_metadata',
-                              'bm_fullstack_trickle']
+sys.path.append(os.path.join(os.path.dirname(sys.argv[0]), '..', 'profiling', 'microbenchmarks', 'bm_diff'))
+import bm_constants
 
 flamegraph_dir = os.path.join(os.path.expanduser('~'), 'FlameGraph')
 
@@ -199,8 +189,8 @@ argp.add_argument('-c', '--collect',
                   default=sorted(collectors.keys()),
                   help='Which collectors should be run against each benchmark')
 argp.add_argument('-b', '--benchmarks',
-                  choices=_AVAILABLE_BENCHMARK_TESTS,
-                  default=_AVAILABLE_BENCHMARK_TESTS,
+                  choices=bm_constants._AVAILABLE_BENCHMARK_TESTS,
+                  default=bm_constants._AVAILABLE_BENCHMARK_TESTS,
                   nargs='+',
                   type=str,
                   help='Which microbenchmarks should be run')