imitatio creatio co we łbie piszczy

pgbouncer mini HOWTO + benchmark

pgbouncer is a lightweight connection pooler for PostgreSQL.

I've decided to write this mini howto in order to prove that pgbouncer:
- is easy to install and configure
- is really useful, even in minimal setups (same machine as postgres, 10 clients).

As a prerequisite, we will need:
- postgres up and running (well, that's what pgbouncer was made for).
- standard set of development tools needed to compile C programs (gcc+make+binutils).
I assume you already have these.

My Linux distro is Ubuntu 9.10, kernel/software versions:

filip@srv:~$ uname -r
2.6.27.7-smp
filip@srv:~$ pg_config --version
PostgreSQL 8.4.3
filip@srv:~$ gcc --version
gcc (Ubuntu 4.4.1-4ubuntu9) 4.4.1

Installing pgbouncer

First we need libevent (event notification library). Luckilly, in recent Ubuntu version it's packaged, so let's use it:

filip@srv:~$ apt-cache search libevent
libevent-1.4-2 - An asynchronous event notification library
libevent-core-1.4-2 - An asynchronous event notification library (core)
libevent-dev - Development libraries, header files and docs for libevent
(...)

We need libevent-dev package, to have header and include files needed for pgbouncer compilation.

filip@srv:~$ sudo apt-get install libevent-dev
(...) done.

OK so now we have libevent installed. Next we go for pgbouncer itself.

It's not packaged for Ubuntu at the moment, so we have to compile it from sources. I chose to install to /usr/local - simply because I had no better idea. YMMV.

Download, unpack and configure:

filip@srv:~/src$ wget http://pgfoundry.org/frs/download.php/2608/pgbouncer-1.3.2.tgz
(...) `pgbouncer-1.3.2.tgz' saved [166756/166756]
filip@srv:~/src$ tar xzf pgbouncer-1.3.2.tgz
filip@srv:~/src$ cd pgbouncer-1.3.2/
filip@srv:~/src/pgbouncer-1.3.2$ ./configure --prefix=/usr/local
(...)
configure: creating ./config.status
config.status: creating config.mak
config.status: creating include/config.h

OK it's configured, lets compile and install:

filip@srv:~/src/pgbouncer-1.3.2$ make
(...)
make[1]: Nothing to be done for `all'.
make[1]: Leaving directory `/home/filip/src/pgbouncer-1.3.2/doc'
filip@srv:~/src/pgbouncer-1.3.2$ sudo make install
(...)

Pgbouncer is now installed.

All in one command, for your convenience:

wget http://pgfoundry.org/frs/download.php/2608/pgbouncer-1.3.2.tgz \
&& tar xzf pgbouncer-1.3.2.tgz \
&& cd pgbouncer-1.3.2 \
&& ./configure --prefix=/usr/local \
&& make && sudo make install

Configuring pgbouncer

Now let's create a config file. We do this by copying distributed ini file sample to /etc:

filip@srv:~$ sudo cp /usr/local/share/doc/pgbouncer/pgbouncer.ini /etc/pgbouncer.ini

Next we edit this file and configure needed options.

Most important is pool_mode. I choose transaction pooling mode because it gives most performance boost for normal postgresql usage.

As for connection settings, we have pgbouncer on port 6432, routing connections to port 5432 (my postgres is on same host and standard port).

File paths are adjusted for typical Ubuntu setup. pgbouncer will be run from postgres account, so we just point auth_file to pg_auth.

Here is my pgbouncer.ini file:

[databases]
* = port=5432
[pgbouncer]
logfile = /var/log/postgresql/pgbouncer.log
pidfile = /var/log/postgresql/pgbouncer.pid
listen_addr = *
listen_port = 6432
unix_socket_dir = /var/run/postgresql
auth_type = trust
auth_file = /var/lib/postgresql/8.4/main/global/pg_auth
admin_users = postgres
stats_users = postgres
pool_mode = transaction
server_reset_query = DISCARD ALL;
server_check_query = select 1
server_check_delay = 10
max_client_conn = 1000
default_pool_size = 20
log_connections = 1
log_disconnections = 1
log_pooler_errors = 1

Starting pgbouncer

Now let's start the beast:

filip@srv:~$ sudo su - postgres
postgres@srv:~$ pgbouncer -d /etc/pgbouncer.ini
2010-04-23 18:37:05.969 20068 LOG File descriptor limit: 10000 (H:15000), max_client_conn: 1000, max fds possible: 1010

OK, now let's check if this really works. Connect to admin console (virtual db "pgbouncer"):

filip@srv:~$ psql -Upostgres -p6432 pgbouncer
psql (8.4.3, server 8.0/bouncer)
WARNING: psql version 8.4, server version 8.0.
         Some psql features might not work.
Type "help" for help.

postgres@pgbouncer=#

It works!.

Benchmarking

Now let's do some testing. We use good'ol' pgbench.

Create and initialize test database with scale 10:

filip@srv:~$ createdb bench
filip@srv:~$ /usr/lib/postgresql/8.4/bin/pgbench -i -s 10 bench
(...)
vacuum...done.

Test with 10 clients, separate connection for each transaction, direct connection to postgres, run test for one minute:

(mandela)filip@ratel:~$ /usr/lib/postgresql/8.4/bin/pgbench -c 10 -C -T 60 bench
starting vacuum...end.
transaction type: TPC-B (sort of)
scaling factor: 10
query mode: simple
number of clients: 10
duration: 60 s
number of transactions actually processed: 1528
tps = 25.410170 (including connections establishing)
tps = 53.186072 (excluding connections establishing)

Now the same test, but connecting via pgbouncer:

(mandela)filip@ratel:~$ /usr/lib/postgresql/8.4/bin/pgbench -c 10 -C -T 60 -p 6432 bench
starting vacuum...end.
transaction type: TPC-B (sort of)
scaling factor: 10
query mode: simple
number of clients: 10
duration: 60 s
number of transactions actually processed: 2601
tps = 43.308068 (including connections establishing)
tps = 55.619391 (excluding connections establishing)

Looks promising.

Now both tests repeated several times, just bare results for brevity:

(direct)
tps = 25.583194 (including connections establishing)
tps = 55.247968 (excluding connections establishing)
(pgbouncer)
tps = 51.769025 (including connections establishing)
tps = 73.188059 (excluding connections establishing)
(direct)
tps = 25.857126 (including connections establishing)
tps = 64.090508 (excluding connections establishing)
(pgbouncer)
tps = 61.633963 (including connections establishing)
tps = 87.375610 (excluding connections establishing)
(direct)
tps = 21.134134 (including connections establishing)
tps = 50.005559 (excluding connections establishing)
(pgbouncer)
tps = 50.122482 (including connections establishing)
tps = 74.693641 (excluding connections establishing)
(direct)
tps = 18.925272 (including connections establishing)
tps = 49.249117 (excluding connections establishing)
(pgbouncer)
tps = 63.616117 (including connections establishing)
tps = 94.977040 (excluding connections establishing)
(direct)
tps = 22.444140 (including connections establishing)
tps = 43.382705 (excluding connections establishing)
(pgbouncer)
tps = 68.886017 (including connections establishing)
tps = 102.644402 (excluding connections establishing)
(direct)
tps = 19.979776 (including connections establishing)
tps = 52.215144 (excluding connections establishing)
(pgbouncer)
tps = 57.047613 (including connections establishing)
tps = 85.300031 (excluding connections establishing)

Lets make it more readable, and calculate performance gain (pgbench vs direct):

direct vs pgbouncer pgbench results - non persistent connections

Test #	direct incl connections	direct excl connections	pgbouncer incl connections	pgbouncer excl connections
1	25,41	53,19	43,31	55,62
2	25,58	55,25	51,77	73,19
3	25,86	64,09	61,63	87,38
4	21,13	50,01	50,12	74,69
5	18,93	49,25	63,62	94,98
6	22,44	43,38	68,89	102,64
7	19,98	52,22	57,05	85,3
AVG	22,76	52,48	56,63	81,97
PGBOUNCER GAIN PERCENT			148,78%	56,19%

We can see that - including time consumed by connections handling - pgbouncer gives about 150% speedup compared to raw postgres.

56,19% is also a very interesting result - this difference probably comes from postgres session initialization, but maybe pgbouncer handles connections/disconnections faster than postgresql itself.

These results are very good but of course they are heavily influenced by pgbench "-C" switch (separate connection for each transaction). Let's see what are the results for pgbench without "-C":

direct vs pgbouncer pgbench results - persistent connections

Test #	direct incl conn	direct excl conn	pgbouncer incl conn	pgbouncer excl conn
1	77,64	77,78	55,67	55,7
2	73,16	73,38	79,59	79,6
3	80,91	81,03	67,45	67,46
4	61,8	61,94	78,97	79,02
5	79,45	79,57	80,66	80,7
AVG	74,59	74,74	72,47	72,5
PGBOUNCER GAIN PERCENT			-2,85%	-3,00%

As you see, for persistent connections there is no gain - even a small overhead.

BTW, both test were executed like this:

filip@srv:~$ for n in 1 2 3 4 5 ; do echo "$n (direct)"; /usr/lib/postgresql/8.4/bin/pgbench -c 10 -T 60 bench   2>&1 | grep tps; echo "$n (pgbouncer)"; /usr/lib/postgresql/8.4/bin/pgbench -c 10 -T 60 -p 6432 bench 2>&1 | grep tps; done

Note: this was on a very weak machine - desktop class PC from circa 2006.

Conclusion and disclamer

Using pgbench is definitely a good idea if you have clients connecting many times repeatedly.

These result would probably be very similar using any other decent connection pooler.

I am not an expert on TPC-B and I also do not take responsibility for any damage made to your system and/or database by the code written above.

I do not guarantee that you achieve same results - maybe the whole test was crippled and useless.

As always, please let me know if you see any errors / omissions in the article.