[Webkit-unassigned] [Bug 70699] [WK2][GTK] Each test takes roughly half a second to load

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=70699





--- Comment #2 from Zan Dobersek <zandobersek at gmail.com>  2012-01-01 09:30:29 PST ---
So, I've taken some time to further investigate this bug, and I'm glad that time did not go to waste.

I used strace on WebKitTestRunner to see what exactly is going on during that 400-500 milliseconds that are wasted upon each load. The trace showed that the LoadURL message is enqueued and dispatched properly and that those milliseconds are not wasted here as I previously assumed. This meant that the LoadURL message was successfully transmitted to WebProcess and that I should investigate there, and that the WebKit2 message structure is working fine, so there's no problem there.

I attached strace to the WebKitWebProcess and it was soon clear that all the time wasted was due to clearing the application cache. This is not done in the DumpRenderTree for the Gtk port, but is done for the Mac port. In WebKitTestRunner, this is done for every port that utilizes it. The clearing in WebKitTestRunner is actually done in the injected bundle before every test[1]. This call ends up in WebCore, in ApplicationCacheStorage::deleteAllEntries[2], where the sqlite database is emptied and vacuumed.

At the start, my ApplicationCache.db file in ~/.cache/webkitgtk/applications was 28KB in size, so this resulted in about that half a second being wasted on both emptying and vacuuming when the database was populated, but would still take about that half a second just to vacuum an empty database.

The Web Dev edition of the HTML5 spec at whatwg.org gives an appcache of about 3.5MB. At the loading of first test with a populated database, it took over 4 seconds to empty and vacuum it, and then again about half a second when the database was empty.

The solution could vary - we could simply opt out of cleaning the appcache before each test. A better option would be to provide a clean (as in empty, or rather non-existing) cache for testing purposes that we would point the WebProcess at before any tests. That could be done through an env and some work in NRWT. Or perhaps there's a way to optimize SQLite operations. The mac port does not seem to be affected time-wise by emptying out the cache before each test, but the reason may be that it doesn't exist (yet there's no indication that the mac port deletes the cache before running WKTR).


[1] http://trac.webkit.org/browser/trunk/Tools/WebKitTestRunner/InjectedBundle/InjectedBundle.cpp#L211 
[2] http://trac.webkit.org/browser/trunk/Source/WebCore/loader/appcache/ApplicationCacheStorage.cpp#L1519

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.