[postgis-devel] chunked GeomUnion commit

[Bill Binko]

strk,

I have run some more of the test (finally -- sorry about the delay).

I hope I have run some that will be useful to you.. I have not had time to 
run all of them.

I have spreadsheets (in OOCalc and exported XLS formats) available here:

http://www.mapshine.com/Perf2.sxc
http://www.mapshine.com/Perf2.xls

Sorry, I don't have Excel here for a pretty export: I will do so when I 
have more data filled in.

I put some notes on the first page of the spreadsheet.  However, a couple 
of important ones are here:

1) I'm using CVS HEAD for this.  Is that what you wanted?  I had to
rebuild postgis to pick up the libgeos.so.3, but it seems to work fine.

2) I see no difference with the #define set to 1 or 0 for the 
short-circuiting (including no stderr outputs.)

3) GEOS 2.1.1 seems to be considerably faster than CVS HEAD for the GIST
ordered case.  (See the GIST Compare tab) I have not had time to do the
non-ordered and BTree cases for the new code (it simply ramps up to fast
for me to do tonight).

4) Are you sure that buffer(collect(p), 0) is equivalent to GeomUnion(p)?  
If so, it is FAR faster in all respects no matter how the shapes come in, 
etc.  In fact, the randomly ordered  buffer(collect()) method came in 
faster than the GIST of all of the other methods (!).  Does 
Buffer(collect()) leave holes where GeomUnion() does?  Or is it a solid 
outline?

Anyway, I look forward to feedback.

Bill


On Mon, 27 Jun 2005 strk at refractions.net wrote:

> I've committed the chunked implementation of unite_garray
> in the head branch.
> 
> This one is the version that doesn't require dump/reload
> so will still get full geometry set in input, but will
> chunk it before passing to GEOS for processing.
> 
> I don't feel comfortable breaking the scripts interface
> so far. Please make your tests with this setup.
> 
> You can fine-tune chunk size changing the MAXGEOMSPOINTS
> #define around line 267
> 
> --strk;
> 
> _______________________________________________
> postgis-devel mailing list
> postgis-devel at postgis.refractions.net
> http://postgis.refractions.net/mailman/listinfo/postgis-devel
> 
>