Обсуждение: High concurrency OLTP database performance tuning
Good morning,
I'd like to ask you some advice on pg tuning in a high
concurrency OLTP-like environment.
The application I'm talking about is running on Pg 8.0.1.
Under average users load, iostat and vmstat show that iowait stays
well under 1%. Tables and indexes scan and seek times are also good.
I can be reasonably sure that disk I/O is not the *main* bottleneck
here.
These OLTP transactions are composed each of 50-1000+ small queries, on
single tables or 2/3 joined tables. Write operations are very frequent,
and done concurrently by many users on the same data.
Often there are also queries which involve record lookups like:
SELECT DISTINCT rowid2 FROM table
WHERE rowid1 IN (<long_list_of_numerical_ids>) OR
refrowid1 IN (<long_list_of_numerical_ids>)
These files are structured with rowid fields which link
other external tables, and the links are fairly complex to follow.
SQL queries and indexes have been carefully(?) built and tested,
each with its own "explain analyze".
The problem is that under peak load, when n. of concurrent transactions
raises, there is a sensible performance degradation.
I'm looking for tuning ideas/tests. I plan to concentrate,
in priority order, on:
- postgresql.conf, especially:
effective_cache_size (now 5000)
bgwriter_delay (500)
commit_delay/commit_siblings (default)
- start to use tablespaces for most intensive tables
- analyze the locks situation while queries run
- upgrade to 8.1.n
- convert db partition filesystem to ext2/xfs?
(now ext3+noatime+data=writeback)
- ???
Server specs:
2 x P4 Xeon 2.8 Ghz
4 Gb RAM
LSI Logic SCSI 2x U320 controller
6 disks in raid 1 for os, /var, WAL
14 disks in raid 10 for db on FC connected storage
Current config is now (the rest is like the default):
max_connections = 100
shared_buffers = 8192
work_mem = 8192
maintenance_work_mem = 262144
max_fsm_pages = 200000
max_fsm_relations = 1000
bgwriter_delay = 500
fsync = false
wal_buffers = 256
checkpoint_segments = 32
effective_cache_size = 5000
random_page_cost = 2
Thanks for your ideas...
--
Cosimo
Cosimo, On 8/31/06, Cosimo Streppone <cosimo@streppone.it> wrote: > The problem is that under peak load, when n. of concurrent transactions > raises, there is a sensible performance degradation. Could you give us more information about the performance degradation? Especially cpu load/iostat/vmstat data when the problem occurs can be interesting. -- Guillaume
--On August 31, 2006 5:45:18 PM +0200 Cosimo Streppone <cosimo@streppone.it> wrote: > Good morning, > - postgresql.conf, especially: > effective_cache_size (now 5000) > bgwriter_delay (500) > commit_delay/commit_siblings (default) commit delay and siblings should be turned up, also you'll want to probably increase log_segments, unless you're not getting any warnings about it. also increase shared_buffers. i'd also make sure write caching is on on the RAID arrays as long as they're battery backed caches. > - start to use tablespaces for most intensive tables > - analyze the locks situation while queries run > - upgrade to 8.1.n > - convert db partition filesystem to ext2/xfs? > (now ext3+noatime+data=writeback) > - ??? > > Server specs: > 2 x P4 Xeon 2.8 Ghz > 4 Gb RAM > LSI Logic SCSI 2x U320 controller > 6 disks in raid 1 for os, /var, WAL > 14 disks in raid 10 for db on FC connected storage > > Current config is now (the rest is like the default): > max_connections = 100 > shared_buffers = 8192 > work_mem = 8192 > maintenance_work_mem = 262144 > max_fsm_pages = 200000 > max_fsm_relations = 1000 > bgwriter_delay = 500 > fsync = false > wal_buffers = 256 > checkpoint_segments = 32 > effective_cache_size = 5000 > random_page_cost = 2 > > Thanks for your ideas... > > -- > Cosimo > > > ---------------------------(end of broadcast)--------------------------- > TIP 4: Have you searched our list archives? > > http://archives.postgresql.org > -- "Genius might be described as a supreme capacity for getting its possessors into trouble of all kinds." -- Samuel Butler
On 8/31/06, Cosimo Streppone <cosimo@streppone.it> wrote: > Good morning, > - postgresql.conf, especially: > effective_cache_size (now 5000) > bgwriter_delay (500) > commit_delay/commit_siblings (default) while thse settings may help, don't expect too much. ditto shared buffers. your fsync is false btw. the major gotcha in high transaction volume systems is stats_command_string (leave it off). > - start to use tablespaces for most intensive tables this is an i/o optimization mostly. again, dont expect much. > - analyze the locks situation while queries run > - upgrade to 8.1.n absolutely you want to do this. when I moved my converted isam projects which dont sound too far from your workload, I saw a huge speed increase with 8.1. > - convert db partition filesystem to ext2/xfs? > (now ext3+noatime+data=writeback) > - ??? meh. :-) I think application level improvements are the name of the game here. Make sure your application or middleware is using the parameterized query interface in libpq. Another possible optimiation is to attempt application level caching in conjunction with some server side locking, Since details are light, only general hints are possible :) consider move to opteron or intel woodcrest platform. a single opteron 170 will easily beat your two xeons, and 2x270 will be a whole new world. woodcrests are great as well if you can get them. also, if you are not already on a *nix kernel, get yourself on one. Merlin
On 31-Aug-06, at 11:45 AM, Cosimo Streppone wrote: > Good morning, > > I'd like to ask you some advice on pg tuning in a high > concurrency OLTP-like environment. > The application I'm talking about is running on Pg 8.0.1. > Under average users load, iostat and vmstat show that iowait stays > well under 1%. Tables and indexes scan and seek times are also good. > I can be reasonably sure that disk I/O is not the *main* bottleneck > here. > > These OLTP transactions are composed each of 50-1000+ small > queries, on > single tables or 2/3 joined tables. Write operations are very > frequent, > and done concurrently by many users on the same data. > > Often there are also queries which involve record lookups like: > > SELECT DISTINCT rowid2 FROM table > WHERE rowid1 IN (<long_list_of_numerical_ids>) OR > refrowid1 IN (<long_list_of_numerical_ids>) > > These files are structured with rowid fields which link > other external tables, and the links are fairly complex to follow. > SQL queries and indexes have been carefully(?) built and tested, > each with its own "explain analyze". > > The problem is that under peak load, when n. of concurrent > transactions > raises, there is a sensible performance degradation. > I'm looking for tuning ideas/tests. I plan to concentrate, > in priority order, on: > > - postgresql.conf, especially: > effective_cache_size (now 5000) > bgwriter_delay (500) > commit_delay/commit_siblings (default) > - start to use tablespaces for most intensive tables > - analyze the locks situation while queries run > - upgrade to 8.1.n > - convert db partition filesystem to ext2/xfs? > (now ext3+noatime+data=writeback) > - ??? > > Server specs: > 2 x P4 Xeon 2.8 Ghz > 4 Gb RAM > LSI Logic SCSI 2x U320 controller > 6 disks in raid 1 for os, /var, WAL > 14 disks in raid 10 for db on FC connected storage > > Current config is now (the rest is like the default): > max_connections = 100 > shared_buffers = 8192 way too low, shared buffers should be 50k > work_mem = 8192 > maintenance_work_mem = 262144 > max_fsm_pages = 200000 why ? > max_fsm_relations = 1000 > bgwriter_delay = 500 > fsync = false you will lose data with this! > wal_buffers = 256 > checkpoint_segments = 32 > effective_cache_size = 5000 way too low should be on the order of 300k > random_page_cost = 2 again why ? > > Thanks for your ideas... > > -- > Cosimo > > > ---------------------------(end of > broadcast)--------------------------- > TIP 4: Have you searched our list archives? > > http://archives.postgresql.org >
It will be very important to determine if as performance degrades you are either i/o bound, cpu bound or hindered by some other contention (db locks, context switching, etc). Try turning on statement duration logging for all statments or "slow" statments (like those over 100ms or some arbitrary threshold). Either eyeball or write a script to see which statement(s) are frequently slowest. This can greatly aid in tuning. You say the db is write intensive. In what way, inserts or updates? The former tend to be much cheaper than the latter. If the latter are things being adequately vacuumed? loss of dead tuple space can really hurt performance. If you have lots of concurrent writes, commit_delay/ commit_siblings can help, as can increasing checkpoint_segments further. I see you have fsync off, are you feeling lucky? ;^) If you are i/o bound see what the disks are doing. How fast are they reading/writing? How close are they to their max throughput? Typically I find the disks are nowhere near that due to excessive seeking. If that's the case you can typically only fix it by putting more of the DB in RAM -- buy more RAM, crank up shared_buffers I would say double what you have it, maybe more (much more with 8.1), or by arranging the data better on disk (clustering, denormalizing data, putting tables and indices on different disks, etc). -Casey On Aug 31, 2006, at 8:45 AM, Cosimo Streppone wrote: > Good morning, > > I'd like to ask you some advice on pg tuning in a high > concurrency OLTP-like environment. > The application I'm talking about is running on Pg 8.0.1. > Under average users load, iostat and vmstat show that iowait stays > well under 1%. Tables and indexes scan and seek times are also good. > I can be reasonably sure that disk I/O is not the *main* bottleneck > here. > > These OLTP transactions are composed each of 50-1000+ small > queries, on > single tables or 2/3 joined tables. Write operations are very > frequent, > and done concurrently by many users on the same data. > > Often there are also queries which involve record lookups like: > > SELECT DISTINCT rowid2 FROM table > WHERE rowid1 IN (<long_list_of_numerical_ids>) OR > refrowid1 IN (<long_list_of_numerical_ids>) > > These files are structured with rowid fields which link > other external tables, and the links are fairly complex to follow. > SQL queries and indexes have been carefully(?) built and tested, > each with its own "explain analyze". > > The problem is that under peak load, when n. of concurrent > transactions > raises, there is a sensible performance degradation. > I'm looking for tuning ideas/tests. I plan to concentrate, > in priority order, on: > > - postgresql.conf, especially: > effective_cache_size (now 5000) > bgwriter_delay (500) > commit_delay/commit_siblings (default) > - start to use tablespaces for most intensive tables > - analyze the locks situation while queries run > - upgrade to 8.1.n > - convert db partition filesystem to ext2/xfs? > (now ext3+noatime+data=writeback) > - ??? > > Server specs: > 2 x P4 Xeon 2.8 Ghz > 4 Gb RAM > LSI Logic SCSI 2x U320 controller > 6 disks in raid 1 for os, /var, WAL > 14 disks in raid 10 for db on FC connected storage > > Current config is now (the rest is like the default): > max_connections = 100 > shared_buffers = 8192 > work_mem = 8192 > maintenance_work_mem = 262144 > max_fsm_pages = 200000 > max_fsm_relations = 1000 > bgwriter_delay = 500 > fsync = false > wal_buffers = 256 > checkpoint_segments = 32 > effective_cache_size = 5000 > random_page_cost = 2 > > Thanks for your ideas... > > -- > Cosimo > > > ---------------------------(end of > broadcast)--------------------------- > TIP 4: Have you searched our list archives? > > http://archives.postgresql.org