[squid-users] Performance-problems on reverse-proxy squid

Dear all,

I'm running a squid-proxy (squid 2.5.stable7-1) in reverse-proxy mode in
front of two webservers. Squid does equal loadbalancing across the
servers, and answers requests for static pages/images/... itself.
Because of the site-content squid is able to service about 80%-85% of
the requests itself. Statistics report about 500 requests/second hitting
squid, with an output to the internet of about 20mbit/s during peak-times.

The squid-machine has an Intel Pentium IV 3,2GHz with two Intel
Pro/1000-networkadapters (e1000-driver) on-board (one for outbound, one
for internal network to the webservers). The system is running Fedora
Core 3 with latest 2.6.x-kernel from Fedora.
There are 2GB RAM available, with 1024MB allowed for squid, about 500MB
of mem are used for filesystem-buffers, about 200MB used for kernel-buffer.
Access to the harddisks is about 15 i/o writes per second, and a bit
less read-i/o each.

My problem is that during peak-hours the machine runs with very high
cpu-utilisation. About 65% show up as being used by "system".
Unfortunately yet I wasn't able to isolate the bottleneck. I also used
InterruptThrottleRate of the network-adapters to limit their interrupts
and increased the Tx/RxDescriptors - without any change to cpu-utilisation.

Does somebody have an idea how I could "debug" the cpu-utilisation of
"system", and how to lower it? Friends told me to watch out for possible
buffers that could reduce the number of transfers from
kernel-to-userspace or so - but I didn't find much.

Feedback/help is _very_ much appreciated.

Yours sincerely,
 Stefan Neufeind
Received on Sat Mar 12 2005 - 14:50:51 MST