Slow rendering (openstreetmap-carto)

Hello,

I just set up a planet DB on my server, following the switch2osm tutorial (https://switch2osm.org/manually-building-a-tile-server-18-04-lts). I'm using the default openstreetmap-carto style. Tiles rendering is way too slow to be usable: I plan to pre-render up to zoom 12, and generating a single metatile at zoom 13 takes 20 seconds. It still takes 6 seconds at zoom 16.

Here's what I've done so far to speed things up:

• put my DB on my server's SSD
• execute the indexes.sql script (https://github.com/gravitystorm/openstreetmap-carto/blob/master/indexes.sql )
• fine tune my PostgreSQL config and set my cpufreq-set to performance as recommended here: https://wiki.openstreetmap.org/wiki/User:Species/PostGIS_Tuning

What more should I do? I added below more information about my setup. Interestingly rendering durations with render_list seems fine (see below).

Here's my server's config:

• Intel Core i7-8700 Hexa-Core (12 threads)
• 64 GB of RAM
• 2 TB SSD (for the DB)
• 2 * 6 TB HDD (set up in RAID 1, for the OS & the tile cache)

I imported the planet.pbf file using this command:

time osm2pgsql -d gis --create --slim
-G --hstore -C 48000 --number-processes 6 \
     --tag-transform-script ~/src/map-styles/openstreetmap-carto/openstreetmap-carto.lua \
     --style ~/src/map-styles/openstreetmap-carto/openstreetmap-carto.style \
    planet-latest.osm.pbf

It took 36 hours to complete, creating a DB of around 900 GB.

Here's what I added to my postgresql.conf:

synchronous_commit = off
fsync = off
random_page_cost = 1.1
wal_buffers = 16MB
effective_cache_size = 48GB
maintenance_work_mem = 4GB
shared_buffers = 8MB
work_mem = 128MB
autovacuum = on
effective_io_concurrency = 200
max_connections = 300
checkpoint_completion_target = 0.7
default_statistics_target = 100
min_wal_size = 1GB
max_wal_size = 2GB
max_worker_processes = 12
max_parallel_workers_per_gather = 6
max_parallel_workers = 12

I also changed my cpufreq-set with the following command:

sudo cpufreq-set -g performance

Here are some rendering durations I got with render_list. They seem fine to me:

• zooms 0 -> 4: 13mn
• zooms 5 -> 6: 15mn
• zooms 7: 13mn

What more can I do? Are we supposed to create other indexes than the ones listed in the indexes.sql script?

I was expecting rendering performance to be something like 1-5 secs on that server (depending on the zoom level). Is that realistic?