Improving TileDrawer rendering speed
I have been playing with TileDrawer recently. After a first rapid successful test on a throw-away EC2 instance I decided to deploy one on a vm for intranet use.
The installation process is pretty straightforward, just run the script copied from the TileDrawer page as root. FYI make sure you have installed curl and python-cssutils before launching the script.
After the script completed I started browsing the map and noticed that tiles took a looong time to render.
A look at top from the server console showed that postgres was hogging the cpu. Memory was fine with no signs of swapping (it is a 1GB instance which I promptly upgraded to 2, running on server class hardware with Xeon CPU). Even after the memory upgrade tile rendering was so slow that the browser would sometime give up and show a white tile.
I decided to look into it a little further and started by using the technique I have already described in another article. The database cache looked fine though as most relationships were completely cached and database size is a measly 300MB which is no problem at all as the postgres param shared_buffers was already set to 512MB.
I then decided to enable slow queries logging (scroll down to 18.8.2) in postgresql.conf:
I started with a threshold of 100ms, which means pg will log all queries that take more than 100ms to complete. The log quickly filled with lots of queries. Armed with patience I started indexing all the obvious fields like landuse, z_order, amenity, etc.
Eventually I used EXPLAIN on a frequent query that was well above 500ms and to my surprise found out that the planet_osm_polygon table was missing the spatial index on the way geometry. After the index was created no long queries were logged at all and tile rendering was snappy enough to not require cache seeding.
The complete list of all indexes I have created on the database planet_osm (copy and paste):
The installation process is pretty straightforward, just run the script copied from the TileDrawer page as root. FYI make sure you have installed curl and python-cssutils before launching the script.
After the script completed I started browsing the map and noticed that tiles took a looong time to render.
A look at top from the server console showed that postgres was hogging the cpu. Memory was fine with no signs of swapping (it is a 1GB instance which I promptly upgraded to 2, running on server class hardware with Xeon CPU). Even after the memory upgrade tile rendering was so slow that the browser would sometime give up and show a white tile.
I decided to look into it a little further and started by using the technique I have already described in another article. The database cache looked fine though as most relationships were completely cached and database size is a measly 300MB which is no problem at all as the postgres param shared_buffers was already set to 512MB.
I then decided to enable slow queries logging (scroll down to 18.8.2) in postgresql.conf:
log_min_duration_statement = 100
I started with a threshold of 100ms, which means pg will log all queries that take more than 100ms to complete. The log quickly filled with lots of queries. Armed with patience I started indexing all the obvious fields like landuse, z_order, amenity, etc.
Eventually I used EXPLAIN on a frequent query that was well above 500ms and to my surprise found out that the planet_osm_polygon table was missing the spatial index on the way geometry. After the index was created no long queries were logged at all and tile rendering was snappy enough to not require cache seeding.
The complete list of all indexes I have created on the database planet_osm (copy and paste):
create index planet_osm_polygon_landuse_idx on planet_osm_polygon(landuse);
create index planet_osm_polygon_x_idx on planet_osm_polygon(z_order);
create index planet_osm_line_highway_idx on planet_osm_line(highway);
create index planet_osm_polygon_amenity_idx on planet_osm_polygon(amenity);
create index planet_osm_line_waterway_idx on planet_osm_line(waterway);
create index planet_osm_poly_building_idx on planet_osm_polygon(building);
create index planet_osm_poly_wayarea_idx on planet_osm_polygon(way_area);
create index planet_osm_poly_natural_idx on planet_osm_polygon("natural");
create index planet_osm_poly_buildingnn_idx on planet_osm_polygon(building) where (building is not null);
create index planet_osm_poly_way_idx on planet_osm_polygon using gist(way);