[postgis-devel] chunked GeomUnion commit

[Bill Binko]

On Tue, 28 Jun 2005 strk at refractions.net wrote:

> Bill, sorry but there has been many changes recently in ALL
> repositories...
> 
> I will now freeze GeomUnion related works in all of them
> to take allow you running tests.

Actually, if you could just define a tag for each codeset you'd like me to 
use, I'll update with them and run the tests against that code.

> 
> Briefly:
> 	GEOS improvements have been ported from JTS
> 	and customly made for Buffer op. Generally
> 	speaking HEAD branch should be faster then
> 	2.1-branch. 2.1-branch should be *faster* 
> 	then 2.1.2, if you really see other results
> 	we have a problem!
> 

That's actually not 100% what I'm seeing :)  If you look at the results I
posted, the GiST-sorted ordering gets quite a bit slower from 2.1.1 to CVS 
HEAD.  Random sorted ordering is considerably faster.

I have not tested against the 2.1-branch, and will wait until you tell me 
what tags to pull (or you decide to really freeze -- which I don't think 
is necessary)

> 	PostGIS GeomUnion function now uses 'chunked'
> 	buffering of input. Size of chunks is defined
> 	by number of vertex in input and you can
> 	set it using a define at top of lwgeom_geos.c.	
> 	This is both in HEAD and 1.0-branch.
> 	In both cases using buffer(collect(p),0) would
> 	use a single BIG chunk thus giving better results
> 	if available memory is enough.
> 	Buffer(collect(p),0) would work on all postgis
> 	*releases*.
> 

I have not updated my postGIS, so I am using postgis-1.0rc6 for all of 
these tests.  I don't mind testing a new one (please tell me if I need to 
dump/reload), but don't really want to add another axis to measure 
(postgis version X GEOS version X compile switches X Ordering).

> When online you can talk to me on freenode irc network.

What channel?  Is there a #postgis that is fairly stable?


Bill