Re: Squid 2.0 PATCH 2 and cache dir's filling up.

On Mon, 26 Oct 1998, Thilo Manske wrote:

> In message <Pine.LNX.3.96.981026102346.4528D-100000@water.cam.uk.internal>
> John Sloan <johns@uk.uu.net> wrote:
>
> > Needless to say shrinking the cache_dir to 21Kb is not ideal. And
> > indeed after stopping and restarting the cache, it threw away most of the
> > entries in the swaplog file without succeeding in deleteing most of the
> > contents of the disk.
> >
> > I now have a 99% full disk, and a cache with just over 2000 objects in it. :(
> Guess #1:
> Did you remove your old cache after upgrading from 1.1? ;-)

Yes.

> Guess #2:
> Remove your swap.logs after stopping and before starting squid - the
> "clean start" feature didn't worked very well for me with squid
> 2.0.PATCHx: After a "clean start" squid2.0 thought it would use much
> less disk space as it actually did. This led to errors on my system
> similar to yours.

Takes hours to do that, and the response while doing it is dog slow. I
threw my cache away using newfs instead.

My problem was cause by noticing that squid was only using 77% of the disk
despite having a cache_dir setting close to the size of the disk (13000
with 14.5G of disc available), and cache_swap_low/high at 90/95. I
concluded that the discrepancy was due to the store_avg_object_size (since
cachemgr was reporting that squid still though there was 13G of disk
available). I reset that from 13Kb to 10Kb (since our av object size was
10.4kb at the time) at later that night the disc filled up with the
consequences I've already described.

> Hope this helps,
> Thilo.

My bug report is that the nature of the failure is both far from graceful,
and that the adjustmets of cache size downwards were happening too
quickly. I think it's clearly ridiculous to resize the estimate of
cache swap on a given disk down from 13Gb to 27Kb. [I see no evidence
that this estimate ever increases either].

This is one area where tuning the size of a cache is less easy than it was
with Squid 1, where you told it the size of the disk and that was it.
Here, I suspect getting store_avg_object_size is critical, and what to set
it to is far from clear.

John
Received on Mon Oct 26 1998 - 10:28:57 MST