[Chris: We're discussing your presentation on the Qmail list]
On Sat, 27 Jan 2001, Steve Fulton wrote:
> http://www.users.qwest.net/~presentations/cmikk/
I notice that Chris Mikkelson and the people at qwest.net use multiple
qmail-sends on multiple queues, with a note that qmail-send is the big
bottleneck. I wonder if they are using the big concurrency patch.
Once the big concurrency patch is installed, it is trivial to get a
concurrency of 500. The 500 number is based on Linux's limits--I would
not be surprised if FreeBSD has far bigger limits.
This means we can send out 500 messages at any given instant.
Let us suppose that 99% of the messages can be delivered to the remote
machine in one second, and 1% of the messages take 60 seconds (one minute)
to deliver or time out. [1] With this simplified model, our imaginary
high-traffic mail server can handle a continuous stream of the maximum
concurrency of our qmail-send process multiplied by .625.
Which means a concurrency of 500 can handle about 300 messages a second,
which is well over a million messages a day. In fact, with a concurrency
of 500, a heavily loaded mail server should handle over a million messages
in an hour.
- Sam
[1] 60 seconds is how long qmail-remote will wait before giving up on a
dead host.
