Re: GtkPerf data for GTK+ 2.6 vs GTK+ 2.10 vs Maemo-GTK+ 2.6 (on a Nokia 770)



On Fri, 28 Jul 2006 19:38:35 +0300, "Kalle Vahlman" wrote:
> Absolutely, so that's why I grabbed the gtk-engines 2.6.9 package
> (oldish, yes, but I needed to compile for GTK+ 2.6 too so...) and ran
> the (admittedly limited set of) GtkPerf tests with 8 of them (all or
> at least almost all of them?). Compiled with vanilla 2.6 and 2.10 and
> the maemo-GTK+.

Kalle,

I really appreciate you running these tests. Ever since you did that,
I've been anxious to replicate the experiment.

I've finally gotten around to doing something like that. The GTK+ 2.6
setup I built consists of the following libraries:

glib	2.6.6
pango	1.8.2
gtk+	2.6.10

While the GTK+ 2.10 setup I have is as follows:

atk	1.9.1
glib	2.12.1
cairo	1.2.2
pango	1.12.3
gtk+	2.10.1

> So when you add the three builtin theme runs and one with the sapwood
> engine, you get a total of 28 different theme/GTK+ setups, all of them
> with 14 results plus a total time spent.

Oh, reading your message now I see I must have missed some themes. I
grabbed gtk-engines 2.6.9 and found 8 themes available for both builds
with gtk-theme-switch2. (Somehow Raleigh appeared in my 2.10
directory, but not in 2.6.) But how does one get at the "three builtin
themes"?

> Needless to say, that spells "a lot of data". In fact, it seems to be
> a bit too much to show in one chart in any reasonable way.

Yes, I struggled with this a bit myself. Here's what I came up with
(just tables so far--no fancy charting) but hopefully this is still
usable:

http://cairographics.org/~cworth/gtk2.6_vs_2.10/

The idea there is that you can see the average 2.10 slowdown across
all widgets for any theme, (by reporting times relative to the same
theme in gtk2.6), in each of the three different areas that
gtk-theme-torturer tests, and then you can drill down to more
details. (And if you chop the last page off the URL you can find the
raw results exactly as they came out of the torturer.)

My setup differs in two significant ways from yours:

1) This is on an x86 laptop, (a ThinkPad x31 with a Pentium M 1.7GHz)

2) This is with gtk-theme-torturer instead of gtkperf

I haven't looked closely at the results yet, but I think these will be
pointing to some useful optimizations to be made. It may be that 2.10
actually fares worse than these numbers show as I suspect there's a
fair amount of noise in some of the faster measurements, but I haven't
done a lot of runs to determine that yet. (I did have one previous run
where the averages for a couple of themes were as high as 2.82 and
3.14, but that wasn't as controlled a run---more things were happening
on the machine at that time.) We may just need to increase the number
of iterations from the default in gtk-theme-torturer to reduce that
noise and get reliable results.

I could also integrate gtkperf into this test rig if that would be
useful.

Anyway, I'm all setup now to just push a button and have my machine
re-run all the tests and generate all the tables. So that will be
interesting to do both as gtk-theme-torturer picks up new tests and as
we make improvements to cairo and GTK+ to target specific problems
pointed out here.

And charts like these will hopefully make it easy for people to drill
down with profiling tools, and also write micro-benchmarks for the
cairo performance test suite.

Any feedback on the above would be most welcome.

Oh, if anyone is interested, here is the script I used for running the
tests and generating the reports:

http://gitweb.freedesktop.org/?p=users/cworth/soc-gtk.git;a=blob_plain;h=a69a47aef074a5766429dc8f4bd2de31eb19bbb1;f=gtk-theme-torturer/compare_2.6_and_2.10

-Carl

Attachment: pgpCt9A5lIHto.pgp
Description: PGP signature



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]