[Maps-l] PostgreSQL tweak on cassini

[Kai Krueger]

On 04/19/2010 05:35 PM, Ævar Arnfjörð Bjarmason wrote:
> Cassini used to only assign 24MB of shared memory to PostgreSQL, I
> changed that to 128MB.

It seems that all of postgresql default memory settings are awfully low 
for such a big database as mapnik's planet.osm. Even 128Mb may still be 
too low. Although I don't have any own experience on this, from talking 
to others, it appears that both the shared_buffers setting as well as 
the maintanance_work_mem (especially during initial import) should be in 
the gigabyte range. Unfortunately non of the documentation on the wiki 
seem to have any good guidelines on setting these values, or example 
configs available in svn.

Does ptolemy need similar tweaks, or has its config been optimized already?

>
> We mostly seemed to be IO-bound on cassini so it looks like this has helped..

The rendering performance varies quite heavily with the zoom level. For 
low zoom tiles it doesn't appear to be uncommon to need in the range of 
tens of seconds to minutes, whereas the the high zoom tiles are 
typically rendered within the 3 second timeout of mod_tile.

It does appear as if cassini is rendering a disproportionate amount of 
low zoom tiles. Most of the low zoom tiles I have checked where renderd 
within the last day or two. How does the expiry work on low zoom tiles?

On the otherhand some of the tiles don't seem to be updating at all. An 
example is 
http://cassini.toolserver.org/tiles/osm-like/de/2/1/1.png/status This 
claims to need rerendering so presumably every time some one views this 
tile, mod_tile will send a rendering request to renderd for it. However, 
the rendering never seems successfully completed and thus doesn't clear 
the "dirty flag". Is it possible that renderd doesn't have permission to 
write to some of the sub directories? Thus rendering the tiles 
successfully, using up resources, but not being able to write the 
results to disk?

It may also be worth updating to the latest version of renderd and 
mod_tile, where I had added some more munin graphs allowing to see the 
proportion of zoom_levels for served and rendered tiles. This may allow 
to see if indeed there is a disproportionate amount of low zoom tiles, 
which could explain the low performance.

Kai

>
> _______________________________________________
> Maps-l mailing list
> Maps-l at lists.wikimedia.org
> https://lists.wikimedia.org/mailman/listinfo/maps-l