squid stats

to fellow squiders:

i few months back i started keeping stats on squid using mrtg, the
cachemgr and calamaris. i have a few questions as to squid's performance.
our main concern is fast content delivery first, bw savings second.

Background info:

Squid-2.3-STABLE2
Enterprise 450 Solaris 7 (latest and greatest patches)
1GB phyical memory
.12 pagefaults-with-physical-IO/ httprequests ratio.
Solaris TCP tweeks. changed the hiwater/lowater marks to 16K rather than
the standard 8K
8000 clients.
4 GB disk cache over 2 scsi disks
60k-100k http reqests/hr
joined the nlanr hierarchy cachedigests enabled with htcp
rtp.us and sv.us are a sibling to us...proxy-only
full T1 connection to the internet
100Mb lan connetion

squid stats using mrtg:

what exactly are the definitions of:

cachehttperrors
cachehttpinkb
cachehttpoutkb
cachenumobjcount
cacheserverrequests
cachememusage
cacheclienthttprequests
cachecpuusage

why is that for 3 months, cachememusage did not have a significant value
but on wednesday it decided to register about 10 Bytes/second? fwiw,
according to top, the squid mem usage was around 300MB, and today
(thursday) it is hovering around 325 MB.

why are there some spikes in cachecpuusage?

refer too:

http://www.newpaltz.edu/Squid-monitor/cachememusage.html
http://www.newpaltz.edu/Squid-monitor

in all the squid-mrtg graphs, are the values/units on the y cooridnates
actually correct? for example, in cacheclienthttprequests

http://www.newpaltz.edu/Squid-monitor/cacheclienthttprequests.html

you will notice that the value on the y-axis is 0-100kBytes/sec, so does
that mean that at peak times, 12pm-6pm that squid is receiving it's
maximal ammount of http requests from the clients? does that mean that i
need another squid cache to balance the load? this question pertains to
all categories on the squid-monitor webpage.

what is the difference between cacheserverrequets and
cacheclienthttprequests?

overall, my real concern is an overloaded web cache. at what point do i
need to split the load to another squid box? how do i know that this
squid box is not saturated with numerous requests? in the near future,
we'd like to set up a cache farm, and transparent proxying using the
router.

also, i can't seem to increase the hit-ratio. according to the cachemgr,
5 minute and 60 min avg's are 'bout 27%-37%. squid-montior/cachehttphits
indicates about 30% Bytes/sec. byte-hit-ratio is usually between 5-19%.
i've increased the size of the largest object to about 32 MB, and then 64
MB and i did not see an increase in byte-hit-ratio..nor an increase in
cachehttphits.

calamaris indicates about 28-30% hit ratio...but a measly 2% are being
fetched from the siblings (rtp.us.ircache.net and sv.us.ircache.net).
it's almost not worth peering i guess?.?.?

are there any suggestions? ideas are always welcome.

General Observations:

last year i ran the proxy on a linux box, same ammount of RAM, same #of
clients, but overall a little less number of requests/hr. it was a dell
poweredge 2500, PII350Mhz, scsi everything, same disk cache config...slim
and trim latest kernel at that time. i noticed that the
pagefaults-with-physical-IO/httprequests ratio was a little higher.
rather than .10-.12 on the solaris machine, on the linux box it was around
.30-.39 at peek times. why would that be?

i would attribute it to solaris/linux differences and differences in
hardware (obviously). the e450 has a 4MB level2 cache, as compared to a
512KB L2 cache on the intels. also the e450 has a 1.6GB/sec backplane to
handle alot of IO. i would assume for performance reasons, the smaller
that ratio, the better off you are b/c the machine has less page faults as
the number of requests increases. hence...i am better off running squid
on solaris rather than linux (cost not being a factor)?.?.? the solaris
machine now is also a mail server, and a socks server too, as well as dns.

anyways...sorry for the rant..but these are just some general observations
questions that i have.

-tia

-andrew

#############################
Andrew Stack
UNIX Systems Administrator
State University at New Paltz
Phone: 914.257.3828

GNU/Linux - The choice of a GNU generation.....

One good solution to solve all your windows problems
is to install some flavor of UNIX. 8-)
#############################
Received on Thu Apr 06 2000 - 11:26:13 MDT