Our page (http://www.tilllate.com) contains a dozen embedded images and is served through a ipvs load balancer. We have about 20 requests/sec on one cluster member. (Apache2 on Gentoo)

We found a relatively low KeepAlive value of about 5-10 seconds as sweet spot. In this way, for fast clients the page along with the embedded images is loaded quickly and with small overhead. However, the slower connections (DSL, dialup) are dropped after 10 secs and the associated apache/php processes don’t eat up all the memory.

We would not turn off KeepAlive as the page loads noticeabely slower. Connection setup would generate load on the machines.

It got too technical for me sometimes, seeing sa these are dimensions I don’t need to work with often, but some good points were made, and I learned quite a bit :)

It’s difficult to summarize as there was a ton of useful ideas and information. Urban and Bernhard’s experiences are based on running primarily read-only sites - tillate probably has more writes (e.g. user comments) to contend with. One or two random points that stuck in my head;

- Monitoring, monitoring and more monitoring - you must do it. As well as keeping track of traffic, you want to be building up stuff like histories of CPU load and page response times. i.e. invest in building data capture (and visualization) tools

- For http load balancing, tools like Perlbal do the job (i.e. no need to blow money on Cisco)

- Persistent HTTP connections + dial up connections are evil. At the same time persistent HTTP connections can help mitigate latency effects (e.g. DSL clients often suffer relatively high connection setup costs) - a smart reverse proxy can handle this and save Apache from getting many processes locked up with persistent connections

- memcached helps : http://www.danga.com/memcached/

Loads more information than that but you had to be there.

More in Christian’s recap - we found Switzerlands only Ruby on Rails user and press-ganged him into the talking at the next meet up.