Re: Thin Client Benchmark Testing

[Dave Richards <drichard largo com>]

All-

     I appreciate waking up to patches to try.

  I put this patch down on Cairo 1.0.2.  I needed to use 1.0.2 because it contains Endian fixes for remote display to our RISC based thin clients.


I cleaned up my chart and added new columns to test with the patch applied.   I understand Federicos concerns about using gtkperf, but it does provide something we can measure instead of just how Evolution responds based on my usage.  I did see maybe a very slight increase in speed in Evolution, but not a major one.   I marked the widgets that seem to gain the most screen from the patch.  Some of them were much faster in the benchmark.  What's also interesting to me is that there was an increase in speed on the HP thin client which *does* contain RENDER.  I didn't expect those numbers to change much.  The NCD column is the most important column to me right now, because that is the one suffering.

On Tue, 2005-11-01 at 22:26 -0600, Billy Biggs wrote:
> Federico Mena Quintero (federico ximian com):
>
> > I don't know yet why that happens.  Interestingly enough, running
> > sysprof shows that *all* the time (94%) goes to the X server.  I
> > wonder why that happens, since, well, aren't we falling back to doing
> > everything client-side?
> > 
> > Your best bet is to see what GTK+ does in the fallback code that makes
> > it so slow.
>
>   Federico rocks.  On IRC he very quickly found the most likely cause of
> the performance regression in the cairo code.
>
>   Attached is a patch to cairo which should make it much more usable,
> and maybe gets close to solving the performance regression.  There is a
> missing optimization in cairo for non-RENDER X servers where it should
> be using XFillRectangle() instead of a more involved fallback.
>
>   Dave, is there any way you could try this patch out?  It was made
> against cairo CVS HEAD but I have verified that it applies to the 1.0
> branch.
>
>   The patch can still be improved, it currently allocates a colour
> through X where it could calculate the pixel value itself for TrueColor
> or DirectColor visuals.
>
>   -Billy