[squid-users] Optimal cache_dir entries for 3 drives i.e. coss and aufs distribution

From: Mark Powell <M.S.Powell@dont-contact.us>
Date: Fri, 20 Apr 2007 15:39:01 +0100 (BST)

Hi,
   Forgive me if this has been answered before. I've been searching the
archives, but cannot find anything that properly answers this.
   I've upgrading our old squid 2.5 servers with something newer. The new
systems have 4GB RAM, hardware RAID (Dell PERC 4/DC controllers with 256MB
write-back cache) and 5x72GB drives. They are using 32bit FreeBSD
6.2-STABLE with the FreeBSD squid port, which is currently 2.6.STABLE12.
I'm new to COSS and not sure of how best to set everything up.
   I have placed the 1st 2 drives into RAID1 for OS and everything except
cache_dirs (/dev/amrd0). That leaves 3 drives for cache usage, which will
not be RAID (/dev/amrd[123]). I'm wondering what is the best configuration
for this. Currently I have 3 cache_dir entries:

cache_dir coss /dev/amrd1 69881 max-size=16384 block-size=8192
cache_dir aufs /2 62000 16 256
cache_dir aufs /3 62000 16 256

Would 2 coss cache_dirs perform better?

cache_dir coss /dev/amrd1 69881 max-size=16384 block-size=8192
cache_dir coss /dev/amrd2 69881 max-size=16384 block-size=8192
cache_dir aufs /3 62000 16 256

If so would a min-size option on the 2nd and a higher max-size be better?
Is this possible?

cache_dir coss /dev/amrd1 69881 max-size=16384 block-size=8192
cache_dir coss /dev/amrd2 69881 max-size=32768 min-size=16384 block-size=8192
cache_dir aufs /3 62000 16 256

If so, does it make sense to configure it that way?
   Just rechecking, AFAICT there is no min-size option for COSS? Therefore
the 2nd configuration with 2 COSS drives would balance objects less than
16K, between the 2 drives using LRU, and any larger objects would get
placed onto the 3rd aufs drive?
   Also, what happens if drive for line 3 fails and I remove it's cache_dir
entry. Does this mean objects over 16KB would never make it onto disk? And
if so, is this something I need worry about?
   Total cache size is around the 180GB mark. That means about 1.8GB of RAM
will be used just storing the in-memory object index? I assume that won't
be a problem with 4GB RAM to play with? Or should I reduce the amount
cached on each drive?
   That's lots of questions :)
   Many thanks for any thoughts.

-- 
Mark Powell - UNIX System Administrator - The University of Salford
Information Services Division, Clifford Whitworth Building,
Salford University, Manchester, M5 4WT, UK.
Tel: +44 161 295 4837  Fax: +44 161 295 5888  www.pgp.com for PGP key
Received on Fri Apr 20 2007 - 08:39:12 MDT

This archive was generated by hypermail pre-2.1.9 : Tue May 01 2007 - 12:00:01 MDT