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
