[webkit-dev] Instrumentation framework for WebKit

[James Robinson]

It's difficult to understand the performance of a large web app nowadays -
think GMail scale and beyond, in part because a lot of things happen in the
UI thread outside of javascript.  For example layout, selector matching and
painting can all happen in a deferred way, take a lot of time and yet not
show up in any javascript-based profiling or metrics.  Even within
javascript it's not always obvious why things happen when they do - for
example a setTimeout() might get blocked behind an XHR readystatechange
handler or a layout causing the timer code to run much later than expected.


To help with this I propose adding a number of instrumentation points within
WebCore designed to bracket the more significant pieces of logic that can
block the UI thread.  Then a client could register itself for the
instrumentation points and aggregate statistics or build reports of what's
going on while a web app is running.


This would not be another generic profiler per se but more targeted at
things useful to a web developer.  As a nice side effect the instrumentation
can have a significantly smaller observer effect than full-scale profiler.


I have a basic implementation of the framework and plumbing for WebKit/mac
at https://bugs.webkit.org/show_bug.cgi?id=25503 with a simple aggregator
for the DumpRenderTree tool.  The output looks something like this:


$ ./WebKitTools/Scripts/run-webkit-tests fast/dynamic/ --dump


Running tests from /usr/local/home/jamesr/WebKit/LayoutTests

Testing 57 test cases.

fast/dynamic .........................................................

1.92s total testing time

html_write time = 38.554668ms count = 230

recalc_style time = 3.255128ms count = 248

paint time = 58.205367ms count = 70

layout time = 117.240668ms count = 240


all 57 test cases succeeded

Any thoughts or feedback?


Thanks,

- James
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-dev/attachments/20090430/2b6c598c/attachment.html>