Re: Squid & Advfs disks

On Fri, 14 Apr 2000, Jens-S. Voeckler wrote:

> With Solaris, I ran into one of those UFS fragmentation walls a few times,
> until I limited cache_dir to 80% of the kbytes reported by "df -k".

I've got my three cache_dirs hovering around 95% of the kbytes reported by
"df -k" with fsck reporting 0.2% to 0.5% fragmentation; this is on a box
running Squid/2.2.STABLE5 using async I/O over Solaris 7 UFS filesystems.
The key for me was to optimise for space not time; if optimised for time,
fragmentation quickly rose to the 7-10% vicinity and UFS would start reporting
"No space left on device" and Squid became useless. The only fix was to blow
away the entire cache (even lowering the cache watermarks didn't help), retune
the filesystem for space then start again.

Having said that, I have a second box (not yet upgraded/cleaned-up) running
Squid/2.2.STABLE5 under Solaris 2.5.1 (also with async I/O and UFS
filesystems) but across six cache_dirs and still primarily tuned for time;
"df -k" is reporting 84-89% usage whilst fsck reports 5.1% to 10.7%
fragmentation yet everything's still humming along happily. The difference
seems to be that one of the updated patches for Solaris 2.5.1 includes an
automatic switch between time/space filesystem optimisation (and also being
across six disks) so the system dynamically "copes".

(By "upgraded/cleaned-up", amongst changes that box will move to Solaris 7,
three larger capacity disks instead of six smaller disks and will use space
optimisation on the UFS filesystem to reduce fragmentation to the levels
currently enjoyed by the first box above.)

> VxFS is said to do things better, but I haven't got my hand on any VxFS
> (yet). So far, it was cheaper to leave part of the disk unused (so that
> fragmentation and sequential allocation algorithms can work) than to buy
> VxFS.

Ack.. also, optimising for space instead of time, depending on your system,
doesn't really impose that much of a performance hit (certainly hasn't
here) which has made it a zero-cost "upgrade" to fix the fragmentation
problem (thanks to Chris Tilbury for this tip in the FAQ!).

Cheers..

dave
Received on Fri Apr 14 2000 - 17:45:44 MDT