Jim Arnott <[EMAIL PROTECTED]> writes on 5 August 1999 at 15:53:53 -0500
> IMHO it seems that there is somthing strange going on when while doing the
> inject test the CPU is 70% idle. (cat test is 0% idle). There seems
> to be a bottleneck somewhere.
That's fairly normal; queue disk bandwidth is generally the bottleneck
on massive inject tests.
A lot of strategies for increasing bandwidth for massing mailings
amount to various ways to get around the queue.
One way is to decide that you don't care if a few messages are lost or
duplicated if you have a crash during the injection phase. In that
case you can hack qmail to not sync after queueing a message -- makes
a *huge* difference, but of course your queue isn't safe from a
crash.
Holding the queue in a battery-backed ramdisk of suitable size is a
neat compromise -- very fast, and fairly stable.
Another way is to make an initial attempt to deliver the message
directly, before queueing it. Since some fairly large proportion of
messages go through on the first try (depends a lot on your address
mix, but 50% to 95%), queueing only the ones that fail of initial
delivery can also be a big win.
--
David Dyer-Bennet ***NOTE ADDRESS CHANGES*** [EMAIL PROTECTED]
http://dd-b.lighthunters.net/ (photos) Minicon: http://www.mnstf.org/minicon
http://www.dd-b.net/dd-b (sf) http://ouroboros.demesne.com/ Ouroboros Bookworms
Join the 20th century before it's too late!
