Matus UHLAR - fantomas wrote:
Any pointers would be appreciated (I'm new to caching), thank you.
On 28.06.06 10:53, Chris Robertson wrote:
(pretty nice general recommendations, but...)
Thanks, I try. :o)
In no particular order:
*Stay away from regex ACLs.
*Throw as much RAM as you can afford in to the caching servers.
*Use lots of disks.
* Don't RAID your cache disks.
don't raid0/raid5 your cache disks. squid has more cache-disks functionality
that's better (faster and more secure) than raid0, and raid5 KILLS
I/O performance.
And RAID1 is not optimal (I hesitate to say "bad") for writes.
Besides, cached data is very transient, and doesn't need the security
that RAID1 provides. Hence, don't bother with RAID for cache disks... :o)
*Stay away from the ufs store type (aufs seems to be the more stable
than diskd, diskd might be a bit quicker).
since when is diskd faster than aufs?
Recent anecdotal "evidence":
http://www.squid-cache.org/mail-archive/squid-users/200606/0594.html
Hence, the "might be a bit quicker". YMMV and all that.
*Keep your cache_mem setting fairly low (it's only used for objects
fetched from the network), let your system's disk caching work to your
advantage.
however you may set up higher cache_mem if you have enough of memory and
objects are often fetched from servers.
As you wish. The way I look at it is that memory dedicated as cache_mem
is ONLY usable for objects that are fetched from the net. A low
cache_mem does mean that objects will be "flushed to disk" with greater
frequency, but those writes will likely be cached in memory just the
same and the memory is available for other uses (such as... Uh.
Hmmm.). Six of one, half a dozen of an other. *shrug* I've done no
performance testing one way or the other, so take all my advise with the
requisite grain of salt.
Chris
