[Openstreetmap-dev] Re: Ideas to lower server resources

[Lars Aronsson]

Previously, I wrote:

> My conclusions: 1) The Squid is doing a great job!  2) The tile 
> server is quite slow, even at zoom=7 where no streets need to be 
> drawn.  In two samples out of five, tile generation took longer 
> than 10 seconds.

As could be seen from my samples, my measuring method tends to 
phase-lock into a 48 hour cycle, so the non-cached requests all 
come at the same time of day.  One would have to "force reload"  
(shift+reload in modern browsers) the fetches in order to 
investigate the behaviour of the tile generation.  This can be 
done with "wget --no-cache", and I have not done this over any 
longer period of time.

Just to get a quick idea, I went to 
http://www.openstreetmap.org/index.html?lat=59.17171346909417&lon=14.870644&zoom=12 
and found the center tile 
http://tile.openstreetmap.org/cgi-bin/steve/mapserv?map=/usr/lib/cgi-bin/steve/wms.map&service=WMS&WMTVER=1.0.0&REQUEST=map&STYLES=&TRANSPARENT=TRUE&LAYERS=landsat,streets&WIDTH=256&HEIGHT=128&BBOX=14.85341,59.16525936483926,14.897355000000001,59.17652007041727
which does contain some streets.  If the streets are drawn, 
the returned PNG image is 6558 bytes.  If only the background 
satellite image is returned, the PNG image is 4802 bytes.

I fetched this URL with wget --no-cache in a shell script loop 
interleaved with a sleep 15.  I did this at night (European time) 
when server load should be at a minimum.  Response times were 
between 2 and 5 seconds, which really is way too slow, but 
unfortunately very typical.  There was no big difference in 
response times between the 4802 bytes and 6558 bytes images.

The first 40 fetches all returned 4802 bytes.  The next 52 fetches 
returned a remarkable pattern: One image of 6558 bytes followed by 
two images of 4802 bytes.  Then 24 images of 4802 bytes, and one 
of 6558 bytes.  When I increased the sleep to 60 seconds, the 
striped 2+1 pattern came back.

Of 140 fetches, 115 were 4802 bytes (without roads) and only 25 
(or 18%) were 6558 bytes (with roads drawn).

These numbers should supplement trac ticket #134,
http://www.openstreetmap.org/trac/ticket/134



-- 
  Lars Aronsson (lars at aronsson.se)
  Aronsson Datateknik - http://aronsson.se