Re: your results..

Hi,

Clifton Royston:
>
> Maciej Kozinski writes:
> >
> > here are the results gathered from my squid's log files between
> > 23.07-2.08. This has been done using my own script called "squeezer"
> > (http://www.uck.uni.torun.pl/~maciek/w3cache/). A bunch of
> > interesting results including supremacy of cache digests over ICP
> > (the fastest sibling use CD!) and many more.
> >
> > Enjoy,
> > Maciej
>
> That really is interesting. What size cache disk and cache RAM are you
> using, and how did you set the maximum object size to get these high
> bytes-hit ratios? Did you adjust the default refresh patterns any?

OK, a few words about the setup...
I currently use Squid 2.2.STABLE3 on Sun's SparcStation 20MP (2xSuperSparc
60MHz) with 288MB RAM and 27 GB HDD running Solaris 7 (2.7, 5.7). I have 72
MB RAM set up as cache_mem and 3050 MB of HDD on two disks and 4 different
partitions for the squid (the disks are at the same SCSI controller). The
maximum object size is 16384 KB.
 What is the main factor increasing hit ratio? As Clifton supposed it is
adjusting refresh_patterns in the right way. As you could see in the results I
had included, ca 70% of requests and ca 55% of traffic are the files of the
three types: GIF, JPEG and HTML. I assume, that the pictures change rarely -
they are logos, photos, buttons and the other gadgets hard to work over and
making a style and form. The HTML is a content, which is much easier to
change and it _is_ changed more often. So my refresh patterns do look like:

# CGI
refresh_pattern /cgi-bin/ 0 0% 0
# Local files - for no_cache not working :)))
refresh_pattern \.torun\.pl 0 0% 0
# Pictures and other stable documents
refresh_pattern \.(gif|jpg|jpeg|ps|pdf|dvi)$ 10800 50% 86400
# Various HTML documents
refresh_pattern \.html$ 120 25% 2880
refresh_pattern \.htm$ 120 25% 2880
refresh_pattern \.shtml$ 120 25% 2880
# Anything of unknown type which comes vi HTTP - relaxing it could improve
# hit ratios a lot ;)
refresh_pattern ^http://*/$ 120 25% 2880

 It stops my cache from serving too much stale documents whiche still keeping
good hit and the transfer ratio. Another thing is to make squid working
really fast.
 Another thing is squid fetching whole object the users have requested - it
does not abort fetching when the users abort.

quick_abort_min -1 KB

 Yet another thing is to stop squid to fetch object which are available via
the fast links - I use ICMP pings and SOURCE_FASTEST option compiled in to
prevent this kind of cache behavour ;)

Regards,
Maciej

-- 
           Maciej Kozinski         http://www.uck.uni.torun.pl/~maciek/ 
	   Remember: Un*x _IS_ user friendly... It's just selective about
		     who it's friends are.