A 20k+ paper

[Anders Berg]

Hi,

I just found this paper about how to scale apache beyond 20k+ connections.

http://www.stdlib.net/~colmmacc/Apachecon-EU2005/scaling-apache-handout.pdf

Fun reading, but it also gives us a heads up on one problem.

Under section 3.4:

3.4 SendFile

If at all possible, use of the sendfile call will almost certainly boost
performance and increase effciency. Unfortunately in our case, Linux's
sendfile implementation does not operate on our hardware (or any that we
can find) without corrupting IPv6 sessions (3). As HEAnet is committed to
offering all production services over IPv6 as well as IPv4, simply turning
off IPv6 was not an option for us.

(3) The bug is actually caused by TCP checksum miscalculations by almost
all network interfaces, and Linux uses checksum offoading during use of
the sendfile call. This was tracked down with the help of Joe Orton.

Well, now we know that.

Anders Berg