Re: [squid-users] Squid loosing connectivity for 30 seconds

On 23/11/2011 8:32 p.m., Elie Merhej wrote:
> Hi,
>
> I am currently facing a problem that I wasn't able to find a solution
> for in the mailing list or on the internet,
> My squid is dying for 30 seconds every one hour at the same exact
> time, squid process will still be running,
> I lose my wccp connectivity, the cache peers detect the squid as a
> dead sibling, and the squid cannot server any requests
> The network connectivity of the sever is not affected (a ping to the
> squid's ip doesn't timeout)
>
> The problem doesn't start immediately when the squid is installed on
> the server (The server is dedicated as a squid)
> It starts when the cache directories starts to fill up,
> I have started my setup with 10 cache directors, the squid will start
> having the problem when the cache directories are above 50% filled
> when i change the number of cache directory (9,8,...) the squid works
> for a while then the same problem
> cache_dir aufs /cache1/squid 90000 140 256
> cache_dir aufs /cache2/squid 90000 140 256
> cache_dir aufs /cache3/squid 90000 140 256
> cache_dir aufs /cache4/squid 90000 140 256
> cache_dir aufs /cache5/squid 90000 140 256
> cache_dir aufs /cache6/squid 90000 140 256
> cache_dir aufs /cache7/squid 90000 140 256
> cache_dir aufs /cache8/squid 90000 140 256
> cache_dir aufs /cache9/squid 90000 140 256
> cache_dir aufs /cache10/squid 80000 140 256
>
> I have 1 terabyte of storage
> Finally I created two cache dircetories (One on each HDD) but the
> problem persisted

You have 2 HDD? but, but, you have 10 cache_dir.
  We repeatedly say "one cache_dir per disk" or similar. In particular
one cache_dir per physical drive spindle (for "disks" made up of
multiple physical spindles) wherever possible with physical
drives/spindles mounting separately to ensure the pairing. Squid
performs a very unusual pattern of disk I/O which stress them down to
the hardware controller level and make this kind of detail critical for
anything like good speed. Avoiding cache_dir object limitations by
adding more UFS-based dirs to one disk does not improve the situation.

That is a problem which will be affecting your Squid all the time
though, possibly making the source of the pause worse.

 From teh description I believe it is garbage collection on the cache
directories. The pauses can be visible when garbage collecting any
caches over a few dozen GB. The squid default "swap_high" and "swap_low"
values are "5" apart, with at minimum being a value of 0 apart. These
are whole % points of the total cache size, being erased from disk in a
somewhat random-access style across the cache area. I did mention
uncommon disk I/O patterns, right?

To be sure what it is, you can use the "strace" tool to the squid worker
process (the second PID in current stable Squids) and see what is
running. But given the hourly regularity and past experience with others
on similar cache sizes, I'm almost certain its the garbage collection.

Amos
Received on Wed Nov 23 2011 - 09:32:03 MST