Re: Disk full errors with Squid2.3S1 and Solaris

On Wed, 16 Feb 2000, Mark Bizzell wrote:

> I have recently upgraded our campus proxy servers from squid2.2S5 to 2.3S1.
> Since the upgrade we've had a lot of trouble with the disks filling up -
> regardless as to the Cache_dir size.

The usual Solaris approach is to run "/usr/bin/df -g" and "/usr/sbin/fsck
-n" on the cache disks - for example:

  /usr/bin/df -g /opt/squid/cache/dir1
  /usr/sbin/fsck -n /dev/rdsk/c1t2d0s7

In the first case, you're looking for free space/inodes - in the second,
fragmentation level (I can't recall if the same numbers are available in fsck
output). You should not only have free space/inodes, but fragmentation
shouldn't be high enough to impair the creation of new files.

I think from your previous post you've already noted that you have free space,
inodes and fragmentation so something else must be amiss (are you using file
system tools to measure these, or relying on Squid's own measurements?).

> Squid configuration
> ./configure --enable-icmp --enable-async-io --enable-snmp
> --enable-cache-digests

Have you built a version sans async I/O, "just in case"?

> I have tried the valid combinations of
> newfs -m 2 -b [8092|4096] -f [1024|512] -o space <rdev>
> and also reduced the size of the cache_dir from 1800 down to 1500

My cache disks (Solaris 2.5.1 and 7, running Squid 2.2.STABLE5 with async I/O)
are run up to 96% capacity (both) and retain under 1% fragmentation (Sol7
box); as you've noted above, "-o space" is critical when you don't have
enough spindles. With six spindles I can use "-o time" and Solaris 2.5.1
automatically varies between space/time optimisation across the disks, though
fragmentation does rise to about 7% (interestingly, Solaris 7 doesn't appear
to possess this same functionality, letting fragmentation rise to 17% before
Solaris craps out complaining the disks are full).

Under Solaris 7, I built the latest cache disks using:

  newfs -v -i 8192 -m 2 -o space -r 7200 <rdev>

Altering the bytes/inode ratio I hoped to reduce the number of inodes (most of
which were unused anyway) and increase the amount of data space. At 96% data
capacity, inode usage is about 54% (YMMV).

> We have 2 squid proxys in a sibling relationship. These communicate with 2
> parent squid proxy's. Could the transfer of the cache_digests be filling up
> the disks ?

Shouldn't (certainly isn't on our 2.2.STABLE5 boxes that are using
predominantly cache digests with most of ICP turned off) - cache digests
should be in memory, I'm not aware of them finding their way into the
persistent disk cache. Even if they did, using the file system tools above
should tell you the *real* disk usage, regardless of contents.

Cheers..

dave
Received on Thu Feb 17 2000 - 02:57:38 MST