Hey,
I've written an ICP responder, that would help a lot in squid-apache web server cluster balancing. It does delay requests to frontend caches at high loads, sends immediate responses at idle cpu capacity and might be remotely turned on or off at emergencies. Right now due to Linux kernel scheduler granularity it works at 0/10/20/..ms steps, therefore it does not achieve my design goals (milisecond-level delays), but still, it works with superior functionality, and does not need hacked squids installed on apaches.
Source location CVS:extensions/icpagent wm:~midom/extensions/icpagent
Docs: http://wp.wikidev.net/ICP_agent
Cheers, Domas
On Jan 3, 2005, at 8:28 AM, Domas Mituzas wrote:
I've written an ICP responder, that would help a lot in squid-apache web server cluster balancing. It does delay requests to frontend caches at high loads, sends immediate responses at idle cpu capacity and might be remotely turned on or off at emergencies. Right now due to Linux kernel scheduler granularity it works at 0/10/20/..ms steps, therefore it does not achieve my design goals (milisecond-level delays), but still, it works with superior functionality, and does not need hacked squids installed on apaches.
Cool... I tried it out for a few minutes while we were playing with slow things. Haven't really got comprehensive stats; I think the capping of high load is more or less working, but at the low-end load we still have a great deal of unevenness as the minimum network delay results in bad distribution.
Also there was a bit where it kind of freaked out, but this may have been a restart or something.
-- brion vibber (brion @ pobox.com)
wikitech-l@lists.wikimedia.org