Re: [squid-users] cache_dir file systems

From: Henrik Nordstrom <henrik@dont-contact.us>
Date: Fri, 08 Sep 2006 02:45:48 +0200

tor 2006-09-07 klockan 12:22 -0700 skrev Dan Thomson:
> I'm sure this has been discussed before, but I'm curious about what
> people think are the "best" file systems to use for your cache dirs.

reiserfs has been a strong over the years choice, and I think it still
is. But haven't done any recent benchmarks and a lot change when a few
years pass by..

ext2 has been generally fastest on the block thanks to it's simplicity,
but it's drawback in system recoviery time is quite obvious compared to
the more modern journaled filesystems. (hours of fsck isn't that welcome
after a power failure or kernel panic)

ext3 is reasonable, but in past tests I have found it's response time to
be quite uneven under load. But it might just be a matter of tuning it
proper (hasn't done much of that kind..).

> I've read that ReiserFS and XFS are good choices... is there an
> optimal request rate/request size to take into account? Any other hard
> drive tweaking that have yielded favourable results?

* mounting noatime is always a good thing.

* for reiserfs notail is also needed I think. But I am not entirely sure
why.. (gut feeling is that it should actually be the reverse, but past
benchmarks said otherwise..)

* More of them (drives).. More spindles means more seeks/s.

* And making sure there is plenty of memory. Probably the most important
of all..

* And looking into COSS to reduce the seek load on the drives is
starting to become a viable path thanks to our new core team member
Steven Wilton who have done a great job polishing up COSS eleminating a
lot of rough edges.

Regards
Henrik

Received on Thu Sep 07 2006 - 18:45:54 MDT

This archive was generated by hypermail pre-2.1.9 : Sun Oct 01 2006 - 12:00:03 MDT