Statistics used in perftest_unittest.py and perftest_integrationtest.py are bogus

[Ryosuke Niwa]

Right now, run-perf-tests simply reads statistics off of test output text for non-page-loading tests. While this is desirable in the long term where we can have all statistics functions implemented only in JavaScript, it blocks our work to use multiple instances of DRT/WTR to smooth out between-run variance in the bug 97510.

Now that Dromaeo and other perf. tests all report results from each iteration, we can reliably compute statistics in python code instead.

[Radar WebKit Bug Importer]

<rdar://problem/12955987>

[Ryosuke Niwa]

Created attachment 181282 [details]
Patch

[Ryosuke Niwa]

Let me change the scope of this bug. Instead of refactoring the unit & integration tests and modifying perftest.py, we can concentrate on refactoring unit & integration tests.

[Ryosuke Niwa]

Created attachment 181283 [details]
Patch

[Ryosuke Niwa]

Committed r138810: <http://trac.webkit.org/changeset/138810>

[Csaba Osztrogonác]

(In reply to comment #5)
> Committed r138810: <http://trac.webkit.org/changeset/138810>

It broke a webkitpy tests:

   File "/ramdisk/qt-linux-release/build/Tools/Scripts/webkitpy/performance_tests/perftestsrunner_integrationtest.py", line 306, in test_run_memory_test
      self.assertEqual(results['Parser/memory-test'], MemoryTestData.results)
  AssertionError: {u'min': 1080.0, u'max': 1120.0, u'median': 1101.0, u'values': [1080.0, 1120.0, 1095.0, 1101.0, 1104.0], u'stdev': 14.508599999999999, u'avg': 1100.0, u'unit': u'ms'} != {'min': 1080, 'max': 1120, 'median': 1101, 'values': [1080, 1120, 1095, 1101, 1104], 'stdev': 14.508609999999999, 'avg': 1100, 'unit': 'ms'}

Could you fix it, please?

[Ryosuke Niwa]

(In reply to comment #6)
> (In reply to comment #5)
> > Committed r138810: <http://trac.webkit.org/changeset/138810>
> 
> It broke a webkitpy tests:
> 
>    File "/ramdisk/qt-linux-release/build/Tools/Scripts/webkitpy/performance_tests/perftestsrunner_integrationtest.py", line 306, in test_run_memory_test
>       self.assertEqual(results['Parser/memory-test'], MemoryTestData.results)
>   AssertionError: {u'min': 1080.0, u'max': 1120.0, u'median': 1101.0, u'values': [1080.0, 1120.0, 1095.0, 1101.0, 1104.0], u'stdev': 14.508599999999999, u'avg': 1100.0, u'unit': u'ms'} != {'min': 1080, 'max': 1120, 'median': 1101, 'values': [1080, 1120, 1095, 1101, 1104], 'stdev': 14.508609999999999, 'avg': 1100, 'unit': 'ms'}
> 
> Could you fix it, please?

Huh, do you have a very old version of python? It appears to me that there's some significant rounding error there.

[Csaba Osztrogonác]

I have Python 2.6.6 (Debian Squeeze), but it fails on Qt, GTK and Chromium 
bots too. But it passes for me with Python 2.7.3 (Ubuntu 12.04).

I don't think if the proper fix is updating python on several bots ...

[Ryosuke Niwa]

(In reply to comment #8)
> I have Python 2.6.6 (Debian Squeeze), but it fails on Qt, GTK and Chromium 
> bots too. But it passes for me with Python 2.7.3 (Ubuntu 12.04).

I mean... it's really bad that standard deviation computation has such a large computation error. It's correct for only 6 decimal points...

[Ryosuke Niwa]

(In reply to comment #9)
> (In reply to comment #8)
> > I have Python 2.6.6 (Debian Squeeze), but it fails on Qt, GTK and Chromium 
> > bots too. But it passes for me with Python 2.7.3 (Ubuntu 12.04).
> 
> I mean... it's really bad that standard deviation computation has such a large computation error. It's correct for only 6 decimal points...

Ugh... I mean 6 significant figures.

[Ryosuke Niwa]

It's mind blowing that people think python is great for scientific computation when its numerical accuracy is much worse than that of JavaScript.

[Ryosuke Niwa]

Attempted a fix in http://trac.webkit.org/changeset/138965.