In November 2005, PostgreSQL 8.1 was released. It included a number of internal architectural changes, some of which aimed to improve how fast the database would run on a multiprocessor system with many active clients. The result was a major improvement in the ability of the database to scale upwards to handle a heavy load. Benchmarks on modern hardware really highlight just how far that version leapfrogged earlier ones. You can find an excellent performance comparison of versions 8.0 through 8.4 from György Vilmos at http://suckit.blog.hu/2009/09/29/postgresql_history. This shows exactly how dramatic these improvements have been.
This test gives a transactions per second (TPS) figure that measures the total system speed, and you can run it in either a read-only mode or one that includes writes. The read-only performance improved by over four times from 8.0 to 8.1 and more than doubled again by 8.3:
|
Version |
Peak read-only TPS |
# of clients at peak |
|
8.0.21 |
1256 |
4 |
|
8.1.17 |
5620 |
14 |
|
8.2.13 |
8109 |
18 |
|
8.3.7 |
13984 |
22 |
|
8.4.1 |
13546 |
22 |
The rise in the number of clients at the peak load gives us an idea of how well the database internals handle access to shared resources. The area 8.1 in particular included a significant upgrade. Performance improved similarly on the write side, with almost an 8 times gain between 8.0 and 8.3:
|
Version |
Peak write TPS |
# of clients at peak |
|
8.0.21 |
361 |
2 |
|
8.1.17 |
873 |
10 |
|
8.2.13 |
1358 |
14 |
|
8.3.7 |
2795 |
18 |
|
8.4.1 |
2713 |
12 |
The small decrease in performance from 8.3 to 8.4 in both these tests is due to some subtle retuning of the database to improve its worst-case performance. More statistics are collected in 8.4 to improve complicated queries, at the expense of slightly slowing the sort of trivial ones tested here.
These improvements have been confirmed by other benchmarking results, albeit normally not covering such a wide range of versions. It's easy to see that any conclusion about PostgreSQL performance reached before late 2005, when 8.1 shipped, is completely out of date at this point. The speed improvement in 2008's 8.3 release was an additional large leap. Versions before 8.3 are not representative of the current performance and there are other reasons to prefer using that one or a later one too.
