[squid-users] Performance Question

Hello all,

After reading all the docs I could find about squid performance tuning a
question remains.
As far as the 5min-values of squid can tell, the line is at 50% load
during the day and mostly idle at night.
I wonder: Is my squid not fast enough, are there other bottlenecks or is
simply the demand not higher.

I know I could benchmark the squid, but it is quite difficult to
benchmark the whole system including ADSL uplink.
I would like to have a second opinion to my squid values:

my setup:
line is 4MBit/384kbit ADSL
parent proxy is OpenBSD with pppoe/squid 2.5.x (memory cache only)
directly connected to the line.
child proxy is Debian squid 2.4.6 which is used by the clients, all
non-local requests are forwarded to parent (no-delay) as long as the
parent is alive.
Both systems are writing an access.log: On the child I have to analyse
the logs (sarg) on the parent the log is for troubleshooting only.
The internal squid is equipped with squidguard, 15 redirectors with
bypass on.

Throughput: 200...250kbyte/s
Requests/s: 35...40
Hit Ratios - Request: 40%, Byte: 15%
Hit Service Time: 3ms
Miss Service Time: 250ms
The miss service time is the same on the parent as on the child, the
proxy-chain/firewalls/network itself doesn't seem to add noticeable delays.

IO Num of reads: 100...150/s
FD Usage: around 300
Number of Clients: about 1000 per day
The cache_dir ist at 32GB (2 disks). Traffic is about 8...10GB a day:

cache_replacement_policy lru
memory_replacement_policy heap GDSF
cache_dir ufs /var/spool/squid/store1 16384 16 256
cache_dir ufs /var/spool/squid/store2 16384 16 256

cache_mem: 386MB, process size is at 850MB, with 2GB RAM total.
Server is mostly used for squid, apache and some monitoring.
Load average ist at 0.5...0.8 (single xeon, HT enabled), load sometimes
at 1.0...1.2 when monitoring probes are forked.

I tried using diskd, but I had sometimes a very high load and bad
response times when squid started purging the cache.
Otherwise diskd did not improve throughput, that is why I switched back
to ufs. I never trusted aufs enough...
On the other hand, diskd can't improve the situation if there is not
more demand.

I would appreciate any opinion about those data.

Thanks in advance.

Hendrik Voigtländer
Received on Wed May 19 2004 - 15:52:35 MDT