Hi Brian,
This is all true in general. But look at original problem posted by
Angelo. As far as I understood the original problem, MySQL is the
backend for DNS. In DNS, most of requests are very similar short queries
that can be effectively cached. That's why Angelo thought about memcached.
This is very cheap and easy to try MySQL tuning. Results may be or may
be not impressive, but there is no risk.
Peter Zaitsev in his mysqlperformanceblog published a test year ago that
compared multiple cache implementations, including memcached and MySQL
query cache.
http://www.mysqlperformanceblog.com/2006/08/09/cache-performance-comparison/
This is pretty interesting.
P.S. There was a typo in my post, normal value for key_buffer_size is
512M, not K.
Brian Aker wrote:
Hi!
On Oct 31, 2007, at 3:57 AM, Alexander Zaitsev wrote:
MySQL itself has a bunch of optimization options that effectively
turn disk reads into memory reads . In your case I would guess that
following tuning would help:
1. Increase MySQL query cache size
2. increase key_buffer_size (normal value is 512K).
This really isn't the right mailing list for this, but the above is
not quite right. The query cache is purged pretty constantly with
Innodb (and its very accurate except in the most recent versions).
For MyISAM it is better but I've rarely seen the query cache make any
difference outside of benchmarks.
MyISAM relies on the OS'es ability to cache files. With Innodb it is a
bit better, but on current versions of MySQL the buffer pool, which
you need to really scale, is not really well threaded. This has been
fixed and there are additional fixes in the 5.1 tree around the
autoincrement issues that come up with multi-concurrency.
MyISAM is also faster than InnoDB, if your application logic permits
non-transactional DB. MEMORY engine is probably not a good idea since
you can loose all your data.
Not really. MyISAM is at the mercy of the file system cache. It can
sometimes be faster, but a well tuned Innodb is normally faster for
reads (though not writes...). Write concurrency performance is very
dependent on tuning and the application design. If all you need is a
fast writer use Archive :)
If the user has no need for durability, then memory for many
operations is quite fast. You can do backstore with triggers to
persistent engines if you need it.
Cheers,
-Brian
--
_______________________________________________________
Brian "Krow" Aker, brian at tangent.org
Seattle, Washington
http://krow.net/ <-- Me
http://tangent.org/ <-- Software
http://exploitseattle.com/ <-- Fun
_______________________________________________________
You can't grep a dead tree.
--
Alexander Zaitsev
Engineering Manager
AMG Lab Sàrl
