Am 27.07.2011 08:18, schrieb Kay Drangmeister:
there have been quite some performance tuning measures
(1) number of render processes has been reduced from 8/6 to 4
(2) Kolossos modified expire.rb to render low zoom tiles with low
(3) indexes have been added to the DB for geometry,hstore and osm-id
Is there a good way that we can monitor the results? Especially
(1) should be carefully tracked. I can see no significant changes
in IO throughput
This decision has been made to try if offloading the database
result in less render timeouts.
and not even in postgres connections
The load and CPU usage has been decreased a bit. My guess would
be that more processes would result in a better CPU utilization
(and thus faster overall rendering).
To monitor this we need two figures: (a) average tile rendering
time (per process) and (b) tiles rendered per second (by all
processes). Can we set up munin to track it?
I don't think tirex allows
capturing the tile throughput on a
per-process base, I guess it would need to be modified to allow that.
The whole tirex block has disappeared from the statistics.
Munin is not
listing the plugins anymore:
osm@ptolemy:~$ telnet localhost 4949
telnet: connect to address ::1: Connection refused
Connected to localhost.
Escape character is '^]'.
# munin node at ptolemy.esi.toolserver.org
apache_accesses apache_processes apache_volume cpu df if_e1000g0
io_busy_sd io_bytes_sd io_ops_sd iostat load mod_tile_fresh
mod_tile_response mod_tile_zoom netstat ntp_kernel_err
ntp_kernel_pll_freq ntp_kernel_pll_off ntp_offset ntp_states
postfix_mailqueue postfix_mailstats postfix_mailvolume postgres_bgwriter
postgres_cache_osm_mapnik postgres_checkpoints postgres_connections_db
postgres_tuples_osm_mapnik postgres_users postgres_xlog processes
replication_delay2 uptime users
This seems like a munin misconfiguration. Sometimes only munin-node
needs to be restarted.
And another question: earlier, two slots have been
prio 1 queue requests (i.e. missing tiles). Is there a reserve
available currently? Otherwise one would have to wait in that
I just reduced the max. number of render processed by two. The
configuration now looks like this:
osm@ptolemy:~$ less tirex/etc/tirex/tirex.conf
# Buckets for different priorities.
bucket name=missing minprio=1 maxproc=6 maxload=20
bucket name=dirty minprio=2 maxproc=4 maxload=8
bucket name=bulk minprio=10 maxproc=3 maxload=6
bucket name=background minprio=20 maxproc=3 maxload=4