Report on `views with many icons' profiling

[Tomeu Vizoso]

This is really awesome, congrats.

I would like to know how much time takes every switch (including the
redraw), is that 130ms and 170ms? Looks like it should be more to me.

Also, would like to see as well a top-down analysis, which are the top
3-5 high level operations that take most CPU? Are they executed more
often than what would be strictly needed? I see in kcachegrind that
_views_switch_...() was called 2481 times, this means that your script
switched that many times?

Can you please try to assess the impact to the user of any slowness we
may have in there? Which are in your opinion the next steps we should
take?

Thanks,

Tomeu

[Sorry about the late replies]

On Thu, Jul 17, 2008 at 7:43 PM, riccardo <riccardo.lucchese at gmail.com> wrote:
> Hi,
>
> Problem: slow switching between views with many icons
>
> Test-case: the test consist of switching between the favorites view and
> the list view. Test were ran once with the ring layout in the favorite
> view and once with the freeform layout; the xo had 25 activities
> installed all checked as `favorite'. The action of switching was
> automated with a timer with period 130ms when the ring layout was
> selected and 170ms in the case of the freeform layout (as the minimum
> values permitting complete redraw of the views).
>
> Note that there is a noticeable delay when switching to the favorites
> views when the selected layout is freeform.
>
>
> --- RING layout ---
> The following tab. and fig. show cpu time usage of the processes
> taking more cpu time while running the test:
>
> (tot% us+sy) - (partial% us+sy) : cmdline
>             - 63.6             : python /usr/bin/sugar-shell
>  91.2        - 27.5             : /usr/bin/X :0 -fp built-ins...
>  99.5        - 8.2              : picker -t30
>
> http://dev.laptop.org/~rlucchese/views/favorites_ring-list/picker.stats.svg
> (http://dev.laptop.org/~rlucchese/views/favorites_ring-list/picker.stats )
>
> They were obtained by running:
> $ picker -t30
> $ grapher -c3
>
> --- FREEFORM layout ---
> (tot% us+sy) - (partial% us+sy) : cmdline
>             - 82.              : python /usr/bin/sugar-shell
>  91.6        - 9.5              : /usr/bin/X :0 -fp built-ins...
>  99.4        - 7.7              : picker -t30
>
> http://dev.laptop.org/~rlucchese/views/favorites_freeform-list/picker.stats.svg
> (http://dev.laptop.org/~rlucchese/views/favorites_freeform-list/picker.stats )
>
>  ! sugar-shell is taking 20% more cpu time than in the ring layout case.
>
>
>
> cProfile statistics (KCacheGrind format) for sugar-shell:
> --- RING layout ---
> http://dev.laptop.org/~rlucchese/views/favorites_ring-list/cProfile-shell
>
> Ordering by function's self-time:
>  %      func name
> 35.6   : cairo.Context.paint
> 3.9    : gtk.Container.add
> 2.     : sugar.graphics.palette.do_paint_below_children
> 1.9    : __setitem__ sugar.util
> ----------------------------------------------
> 57%
>
> Well, this isn't unexpected. But it's interesting when looking at
> sysprof results (below).
>
> --- FREEFORM layout ---
> http://dev.laptop.org/~rlucchese/views/favorites_freeform-list/cProfile-shell
>
> Ordering by function's self-time:
>  %      func name
> 21.6   : _add_weight in sugar/shell/view/home/grid.py
> 21.5   : _remove_weight in sugar/shell/view/home/grid.py
> 10.6   : cairo.Context.paint
> 8.1    : __setitem__ sugar.util
> 5.7    : _compute_weight in sugar/shell/view/home/grid.py
> ----------------------------------------------
> 57.5%
>
>  ! Box2D would perform better ;)
>
>
> Sysprof results. Well, in sysprof there are many nested levels, so it is
> much more clear to just look at it.
>
> --- RING layout ---
> http://dev.laptop.org/~rlucchese/views/favorites_ring-list/sysprof.data
>
> - most of self-time is spent in the kernel and in X/X-modules.
>
> - time spent in the kernel is due to python and X, respectively 60%-40%.
>
> - time spent `in X' goes mostly to the geode driver, and then, to Xorg
> itself and the libexa module.
>
>
> --- FREEFORM layout ---
> http://dev.laptop.org/~rlucchese/views/favorites_freeform-list/sysprof.data
>
> Notes for the ring layout are valid also here.
>
> There are two (new) entries in this case and they are taking more time
> than the X geode module: python's numpy/core/multiarray.so and
> numpy/core/umath.so. This is in relation with the algorithm used in the
> freeform layout to avoid icons collisions.
>
>
> thanks,
> riccardo
>
> _______________________________________________
> Devel mailing list
> Devel at lists.laptop.org
> http://lists.laptop.org/listinfo/devel
>