Обсуждение: How to interpret view pg_stat_bgwriter
Hi,
I was reading an article of Gregory Smith http://www.westnet.com/~gsmith/content/postgresql/chkp-bgw-83.htm and tried to do some analysis on our database.
postgres=# select * from pg_stat_bgwriter;
-[ RECORD 1 ]------+------------
checkpoints_timed | 42435
checkpoints_req | 629448
buffers_checkpoint | 1821978480
buffers_clean | 117710078
maxwritten_clean | 23796
buffers_backend | 1284631340
buffers_alloc | 32829025268
postgres=# show checkpoint_segments ;
-[ RECORD 1 ]-------+----
checkpoint_segments | 128
postgres=# show checkpoint_timeout ;
-[ RECORD 1 ]------+------
checkpoint_timeout | 10min
bgwriter_delay bgwriter_lru_maxpages bgwriter_lru_multiplier
postgres=# show bgwriter_delay;
-[ RECORD 1 ]--+------
bgwriter_delay | 100ms
postgres=# show bgwriter_lru_maxpages;
-[ RECORD 1 ]---------+-----
bgwriter_lru_maxpages | 1000
postgres=# show bgwriter_lru_multiplier;
-[ RECORD 1 ]-----------+--
bgwriter_lru_multiplier | 5
based on one snapshot, below are my thoughts after reading the example reading the example Greg used, it might be completely wrong as I’m just starting the learning process of checkpoint mechanism in PG. If anything missing/wrong, appreciate if you can help to point out.
# checkpoints_req is much bigger than checkpoints_timed, suggest that I may increase checkpoint_segments in our system
#maxwritten_clean is high, suggests increase bgwriter_lru_maxpages
# buffers_backend is much smaller than buffers_alloc, suggests increasing bgwriter_lru_maxpages, bgwriter_lru_multiplier, and decreasing bgwriter_delay.
Thanks,
Suya
On 9/16/14 2:42 AM, Huang, Suya wrote: > # checkpoints_req is much bigger than checkpoints_timed, suggest that I > may increase checkpoint_segments in our system That's not strictly necessary, but personally I think both of those settings are too low for you. > #maxwritten_clean is high, suggests increase bgwriter_lru_maxpages Yes. > # buffers_backend is much smaller than buffers_alloc, suggests > increasing bgwriter_lru_maxpages, bgwriter_lru_multiplier, and > decreasing bgwriter_delay. You should rather compare buffers_backend against buffers_checkpoint. Also consider graphing these quantities over time so you can see how different workloads affects the results and what the effects of your tuning are. Also turn on log_checkpoints and look at the sync times. In my experience that dominates all the bgwriter tuning.
-----Original Message-----
From: Peter Eisentraut [mailto:peter_e@gmx.net]
Sent: Thursday, September 18, 2014 6:41 AM
To: Huang, Suya; pgsql-admin@postgresql.org
Subject: Re: [ADMIN] How to interpret view pg_stat_bgwriter
>You should rather compare buffers_backend against buffers_checkpoint.
>Also consider graphing these quantities over time so you can see how different workloads affects the results and what
theeffects of your tuning are.
>Also turn on log_checkpoints and look at the sync times. In my experience that dominates all the bgwriter tuning.
Thank you Peter, comes with more question:
How to compare buffers_backend against buffers_checkpoint?
Below is some statistics of check point activities happened on the day while backend is loading data:
Day Hour Written buffers Write time Sync time Total time
06 107,338 1,475.583s 3.142s 1,478.828s
07 601,672 1,041.482s 348.215s 1,390.153s
08 3,613 361.422s 0.053s 361.535s
09 1,006,704 1,456.898s 14.501s 1,471.545s
10 1,116,463 1,987.896s 6.164s 1,995.11s
Can we tell something about the configuration?
Thanks.
Suay
On 9/19/14 1:54 AM, Huang, Suya wrote: > How to compare buffers_backend against buffers_checkpoint? This is a very general question, but generally you want buffers_backend to be as small as possible. > Below is some statistics of check point activities happened on the day while backend is loading data: > > Day Hour Written buffers Write time Sync time Total time > 06 107,338 1,475.583s 3.142s 1,478.828s > 07 601,672 1,041.482s 348.215s 1,390.153s > 08 3,613 361.422s 0.053s 361.535s > 09 1,006,704 1,456.898s 14.501s 1,471.545s > 10 1,116,463 1,987.896s 6.164s 1,995.11s > > Can we tell something about the configuration? If this is while loading data, there isn't much to say. I'd be worried about the 348s sync time if this were during normal transaction processing load.