[Geocoding] Maximum realistic performance (geocoded addresses per second)

[Sarah Hoffmann]

On Fri, Mar 27, 2015 at 09:29:57AM +0100, Mathias Versichele wrote:
> I see. Judging by the graphs, the server handles around 1000 successful
> results per minute or almost 17 per second. Which is lower than what I am
> seeing right now. Does that mean that I shouldn't expect any more gains? 

Not necessarily. The server is serving up to 8k reverse queries
per minute at the same time. They are cheaper but still take time to process.
Also, you need to add the search queries that yield 0 results. They are listed
extra because in most cases they are more expensive (Nominatim needs to try
more interpretations of the search query).

If you look at it the other way around, a search query takes 250ms on average
when the server is under full load. With 12 real CPUs and given that IO has
very little influence because of the SSDs, I'd expect it can do easily
50 search queries per second. Naturally, that's all just back-of-the-envelope
computation. And it's only a factor of 2. I suspect you can gain more by
making sure your queries are well formed (adding commas between address
parts, for example) and by reducing the database size to only the data
you really need. 

Sarah


> I understand the different factors influencing performance, but I'm more
> trying to get a grasp of the maximum I should be aiming at (on a dedicated
> system).
> 
> On Fri, Mar 27, 2015 at 9:10 AM, Sarah Hoffmann <lonvia at denofr.de> wrote:
> 
> > Hi,
> >
> > On Wed, Mar 25, 2015 at 10:56:30PM +0100, Mathias Versichele wrote:
> > > using batch mode and experimenting with batch-size and the number of
> > > simultaneous batch requests, my geocoding speed seems to level out at
> > > around 25 addresses per second. Is it realistic to be still aiming for a
> > > considerable improvement ? Without going through the details, are there
> > any
> > > kind of benchmarking results available (sort of like the planet imports).
> >
> > I'm not aware of any systematic benchmarking. You can have a look at the
> > server stats for nominatim.osm.org at
> >
> > http://munin.osm.org/openstreetmap/pummelzacken.openstreetmap/index.html#nominatim
> > to get an idea how that server is doing.
> >
> > Note that query speed can be vastly different depending on the hardware you
> > have, what kind of area is imported and what kind of queries you have (even
> > what kind of format your search term has can make a huge difference). So
> > I'm
> > not even sure benchmarks could have more than anecdotic value.
> >
> > Sarah
> >
> 
> 
> 
> -- 
> Tel: +32 (09)/241.56.35
> Mobile: +32 (0)485.16.07.08 <%2B32%20%280%29473.44.59.29>
> www.geointelligence.be
> <http://www.geointelligence.be/home-2/>