On 3/24/2012 1:16 AM, Linda Walsh wrote:

Is this redhat's version of the kernel only? Or does it apply to other linux kernels and other distros?

Any idea what linux kernel versions might cause this?

(from main dovecot webpage news)

Thu Mar 22 14:38:53 EET 2012

Red Hat/CentOS users: A recent kernel update https://bugzilla.redhat.com/show_bug.cgi?id=681578 causes Dovecot to start failing after it has reached 1000 child processes. To fix this, downgrade your kernel until Red Hat releases a fixed kernel.

It appears to be a Red Hat centric regression. They added a patch to fix one thing and broke other things, Dovecot, in the process, because the Red Hat programmer made an incorrect assumption about what real world applications were doing, apparently without investigating such first. Note that one won't see this problem on their REHL/CentOS system if they never hit 1000 child processes.

And as Timo states in the bug report it's *possible* Postfix could suffer the same problem as it uses the same pipe/epoll system. However nobody runs 1000 Postfix smtp[d]s. Few, if any, run over 200. The ones that do usually don't know how to properly tune Postfix, and they use a high smtp[d] daemon count to compensate for suboptimal configuration elsewhere in the system.

A properly setup Postfix server can handle 200-300 msgs/second with the default 100 smtp[d] processes. 1000 smtp[d]s would suggest a message rate 10x that, or 2000-3000 msgs/second. The server plus disk subsystem required to queue that kind of message rate would be impressive, and expensive, for a mail server. This same message rate can typically be achieved by a much less expensive scale out farm.

If anyone on the planet is running a properly tuned 1000 process Postfix server, I'd love to read about it.

-- Stan