On Tue, 27 Apr 2004 Hanno.Wagner@boerse-stuttgart.de wrote:

> I simply recreated the filesystem.
> Unfortunately, I can not just switch over to Linux for the squid (I'd
> love to but our company wants only Solaris-Boxes), therefore I can not
> do much to change the behaviour.

Tuning the filesystem to optimize for space rather than speed quite
effectively avoids the fragmentation issue. See the Solaris UFS
documentation or the Squid FAQ section on Solaris for details.

I think there is also a patch from Sun which improves UFS to more
intelligently use the block fragments, but it probably needs some tuning
to work well with Squid.

Also, configuring squid.conf to use a smaller cache_dir also avoids the
issue, as it is only a problem if the filesystem is filled beyond it's
block capacity.

> Since even Solaris ufs is quite slow with small files, my question is:
> has anyone experience with Veritas Filesystem used for Caching purposes?

No experience of veritas here, but I do remember some guy which seemed to
know the veritas filesystem quite well claiming it is not suitable for
Squid due to it's extents based design. Details a bit fuzzy as it was very
long ago (several years).

But I have had quite positive reports about using Solaris UFS + Disksuite
transaction log (on separate device). If you have a Solaris 8 server then
this is a viable option as Disksuite should be included in your server
package. But this too was quite many years ago.

Also remember to mount the cache partition with noatime. This helps on
any OS (Solaris included) by reducing the amount of inode writes. Squid
does not use the filesystem atime detail so you can safely disable this on
a Squid cache partition. In fact I know of very few tools using the
atime timestamps.