On Wed, Aug 22, 2007, rahul wrote:
> [Jyri Virkki:]
> | > and it has a link to the Squid Feature poll that has been setup.
> | > http://vrthra.googlepages.com/squidpoll.html
> | 
> | I'm surprised some of these (like logging) must be chosen at
> | compile-time, that's quite limiting. 
> 
> If the logging is disabled during compile time, It is removed completely
> from the Request Path. (Introducing no overhead at all). While even if
> the logging is enabled during compile time, The administrator can
> disable it from the squid.conf file. (With a --possibly small -- overhead
> for checking the squid.conf option)

Ah, squid performance discussions. :)

Logging itself isn't high overhead. The fact squid-2.6 and Squid-3 right now
log using stdio in the same single-threaded process is the problem if
administrators -wanted- to keep logs.

I've written some proof of concept code (s26_logfile_daemon branch at
squid.sf.net) which breaks out logfile writing to an external process.

If you want to know where the CPU is hiding then grab a copy of polygraph
(http://www.web-polygraph.org/), configure a client/server, setup a Squid
proxy server, run the Polymix-4 test suite and do some Squid profiling.
My s26_logfile_daemon branch removes stdio logfile writes as a bottleneck
(around a couple hundred requests a second with stdio; up to a few
thousand a second..) but the Squid bottlenecks are elsewhere.. :)

(don't forget to compile squid-2.6 with --enable-dev-poll either.)





adrian


Reply via email to