Re: Performance data for Theme capplet

[Michael Meeks <michael ximian com>]

Hi Muktha,

On Mon, 2002-10-28 at 13:38, muktha wrote:
> The report highlights only the top five expensive functions. Please get
> back to us in case you need additional information / data.

	These performance reports (as you may have noticed) are remarkable only
in their uniformity, they all flag lookup_type_node_I,
g_type_instance_is_a etc. as the top bottlenecks.

	Sadly this is pretty useless to people trying to optimise; they're more
interested in "30% of the time is spent in the children of
my_widget_frobnicate" - rather than a precise breakdown of how that time
was spent furkling various types.

	ie. the information is far, far too granular - we need more levels of
stack trace for the most critical spots, and we also need to ignore all
this type checking / underlying library code and cut to the algorithmic
inefficiencies in the application so they can be fixed.

	ie. this code:

for (i = 0; i < 10000; i++)
	a[i] = g_object_new (GTK_TYPE_WIDGET, NULL);
}

	is not going to show at all on your profile, almost certainly the
locking overhead of object instantiation will swamp anything meaningful
and a single stack frame will hide the real problem.

	Do the tools do a better job if you tweak them ? it seems to me that
ultimately to do sensible optimisation you need a more in-depth,
interactive approach, not reflected in the profiles.

	Regards,

		Michael.

-- 
 mmeeks gnu org  <><, Pseudo Engineer, itinerant idiot