[geos-devel] Extent-limited graph building ?

[Sandro Santilli]

On Fri, Sep 19, 2014 at 06:05:09PM +0200, Sandro Santilli wrote:
> On Fri, Sep 19, 2014 at 05:54:10PM +0200, Sandro Santilli wrote:
> 
> > So, to recap, we go from ~15 seconds to ~2.8 seconds with my
> > specific case (polygon with ~250k holes, intersected with small
> > box).
> > 
> > Some nice debugging output:
> > 
> >   Copied 467 nodes out of 10711 for geom 0
> >   Copied 0 nodes out of 1 for geom 1
> >   (computeSelfNodes) Self edges reduced from 10711 to 470
> >   (computeSelfNodes) Self edges reduced from 1 to 1
> >   Self edges reduced from 10711 to 470
> >   Other edges reduced from 1 to 1
> > 
> > It looks like the majority of the time is still spent after
> > the last output line. Now profiling that one.
> 
> Interesting, the profiler shows the same number of calls to 
> index::strtree::STRtree::STRIntersectsOp::intersects(void const*, void const*)
> with and without the envelope restriction optimization.
> 
> Calls to that function represent the 32% of total time with optimization
> activated, and 18% without.
> 
> We're talking about 124380300 calls, all coming from 338673 calls to
> index::strtree::AbstractSTRtree::query, in turn all coming from the
> FastNodingValidator. This means 30% of those ~3 seconds are spent
> validating the output, doesn't it ? Is that step still needed ?
> That's one of the steps that a "loose" (visualization-oriented)
> function would not spend time on.

Removing the noding validator step results in current trunk
taking ~13 seconds instead of ~15 and the envelope-filtering version
taking ~0.9 seconds instead of ~3 seconds

The testsuite shows no failure with the noding validator off (both GEOS
one and PostGIS one). Are there no known cases where that's needed ?
Sounds like a relatively big cost to keep.

Coming down from 15 seconds to < 1 sounds good :)

--strk;

 ()  ASCII ribbon campaign  --  Keep it simple !
 /\  http://strk.keybit.net/rants/ascii_mails.txt