Re: GTK2 performance and Dillo

[Owen Taylor <otaylor redhat com>]

On Thu, 2003-10-16 at 14:09, Robin Rowe wrote:
> Hi. Dillo is a very small fast web browser (400kb executable). Stephan Lewis
> posted some interesting performances tests yesterday comparing GTK1 and GTK2
> running Dillo.
>
> Can anyone illuminate further on the size and performance differences
> between GTK1 and GTK2?

I think these observations are basically irrelevant to most GTK+-2.x
applications. GTK+-2.x *is* bigger and slower (this is an almost
inevitable consequence of having more features and rendering nicer),
but it's not megs bigger and minutes slower.

> > Test 1: Base load times
> > dillo1
> >     Time to load+render: ~1s
> >     Time to quit: instantaneous
> >     Size (from top): 7MB
> > dillo2
> >     Time to load+render: ~3s
> >     Time to quit: 0.5s (estimated, "felt" slower than dillo1)
> >     Size: 14MB

This is almost a plausible difference between GTK+-1.2 and GTK+-2.0
for a large program; though I suspect that reading the common FAQ's
about how to read top sizes would show that the memory difference in
actual memory used is much smaller than described above.

Small GTK+-2.0 programs load in a few tenths of a second and malloc
maybe ~600k of memory on X11. (Probably considerably less on
Windows since a large amount that RAM is being used for font lists and
fonts.)

> > Test2: Loading a 1.8MB html file, styled text, bulletted lists, tables
> >        of text
> > dillo1
> >     Time to load: 3s
> >         extra time to finish rendering: 6s
> >     Time to quit: ~2s
> >     Size: 56MB
> > dillo2
> >     Time to load: 1m20s real according to the time command, more like
> >                   2m30s due to swapping
> >         extra time to finish rendering: 10s
> >     Time to close: 1m20s
> >     Size: 220MB

This is a dillo bug and has nothing to do with GTK+. (Or alternatively
dillo is misusing GTK+ badly; if it's creating a GtkLabel per paragraph
or something you might get this kind of change. But the 56meg "before"
is pretty bad too.)

> > A couple of behavioural notes:
> > 1) Progressive rendering wasn't useful in dillo2, you get an initial
> >    draw (which is largely useless), then dillo2 blocks, then the final
> >    draw.

dillo2 issue; the expose models *are* different between GTK+-1.2 and
GTK+-2.0. This needs to be taken into account if you are doing something
like progressive rendering.

Regards,
                                                Owen