Dromaeo perf-tests results are wrong

• Component: Tools / Tests
• Status: RESOLVED INVALID

[Kentaro Hara]

Regarding Dromaeo/dom-attr.html, the ./run-perf-test result is as follows:
  RESULT Dromaeo: dom-attr= 6270.733871 ms
  median= 0.0 ms, stdev= 38.4424019328 ms, min= 6203.83270023 ms, max= 6354.54507606 ms
  Finished: 30.665235 s


When I manually change the loop count from 10240 to 1, the ./run-perf-test result becomes:
  RESULT Dromaeo: dom-attr= 21839544.4 ms
  median= 0.0 ms, stdev= 50594.7730808 ms, min= 21685174.0 ms, max= 21970762.0 ms
  Finished: 30.582642 s

As you can see, the "ms" number does not make sense. The number should represent the execution time per iteration and thus should be "smaller is better". Also, "median" should not be "0.0 ms".

[Kentaro Hara]

Created attachment 131576 [details]
Patch

[Hajime Morrita]

Comment on attachment 131576 [details]
Patch

Please consider to keep using "run/s" since it's stable.
I tried ms. But making them resulted unexpected (and typically longer) running time, especially we cannot assume underlying environment.

Also, I guess |runStyle| could be specified externally via URL query string. 
Please do it if possible. Since this Dromaeo source is copied from upstream. If there is any obvious errors there, 
I encourage you to upstream the fix.

[Kentaro Hara]

(In reply to comment #2)
> (From update of attachment 131576 [details])
> Please consider to keep using "run/s" since it's stable.
> I tried ms. But making them resulted unexpected (and typically longer) running time, especially we cannot assume underlying environment.

This is the reason why I removed 'if (runStyle === "runs/s")' part. Now the running time does not change between run/s and ms.

> Also, I guess |runStyle| could be specified externally via URL query string. 
> Please do it if possible.

Let me try it.

> Since this Dromaeo source is copied from upstream.
> If there is any obvious errors there, 
> I encourage you to upstream the fix.

If the patch looks OK, I'll try it. But it will take time, and I would like to land the patch into WebKit first. The current Dromaeo results have been preventing me from evaluating performance optimization patches.

[Kentaro Hara]

Created attachment 131582 [details]
Patch

[Kentaro Hara]

Created attachment 131599 [details]
Patch

[Kentaro Hara]

Created attachment 131600 [details]
Patch

[WebKit Review Bot]

Comment on attachment 131600 [details]
Patch

Clearing flags on attachment: 131600

Committed r110559: <http://trac.webkit.org/changeset/110559>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.

[Benjamin Poulain]

I think this broke most of the Dromaeo tests on WebKit-perf: http://webkit-perf.appspot.com/
They all report a time of 0 now.

[Ryosuke Niwa]

Comment on attachment 131600 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=131600&action=review

> PerformanceTests/Dromaeo/resources/dromaeorunner.js:16
> +                 median += 1000 / item.median;

This makes no sense. The sum of medians isn't the median of all samples.

[Ryosuke Niwa]

What we need to do is to fix https://bugs.webkit.org/show_bug.cgi?id=78303.  We don't want to invent our own score for Dromaeo.