Hi, Sorry for send this email twice but it seems it fits the performance group than admin group...
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
