[OSM-dev] Nominatim's reverse geocoding statistics

[Sarah Hoffmann]

Hi,

On Wed, May 02, 2012 at 09:56:07AM +0200, Charles DESNEUF wrote:
> We are trying to build our own reverse geocoding service, based on
> Nominatim and we have an issue with it. Nominatim does work but when
> running a benchmark on the reverse geocoding page we are unable to get more
> than 13 or 14 pages per second.
> CPU isn't loaded, Memory is fine and IO are goods...
> I would like to know if some of you tried to benchmark his installation and
> know how many request the service can handle ?

poldi, OSM's Nominatim instance[1], currently tops somewhere around
220 request/s[2], although that includes about 5% search requests which
are much more expensive than calls to reverse. Also keep in mind
that it has the update process running in the background.

datendelphin has done some stress testing on a Hetzer EX-4S[3]
and reported that it manages around 120 requests/s (without any updates
running).

Both machines run a full planet with postgresql configured as described
in the wiki.

> And if you have an idea about what parameters (Php, Postgres, Apache ...)
> can we play on to improve the results I'll be glad to ear it.

Switching off keep-alive in Apache did help quite a bit. 
Also, you might want to switch the IO scheduler to deadline.

For postgresql, it might be worth to rerun VACUUM ANALYZE.

Sarah

[1] http://wiki.openstreetmap.org/wiki/Servers/poldi
[2] http://munin.openstreetmap.org/openstreetmap/poldi.openstreetmap/index.html
[3] http://www.hetzner.de/en/hosting/produkte_rootserver/ex4s