On Thu, 20 Jan 2000 21:04:11 -0800 , Mark Delany writes:
> Are these inbound or outbound transactions. Inbound and the concommitant
> local delivery is usually a lot harder on a system than outbound.
Another issue is multiple deliveries -- if you are
doing header rewriting in the standard, stock qmail
way, you are doing two deliveries per message. To
say this kills performance is an understatement.
> You will need to have separate queues that are load-balanced in
> some way. There are also NVRAM disks to consider as potential
> queue disks with awesome performance, but I've not seen those
> used on qmail.
Solid-state disks are prohibitively expensive -- a
decent sized one costs about as much as a decent
sized house[1]. In other words, if you have enough
money for an SSD, there are generally better things
you can do with it, like run multiple servers, and
use round-robin MXing to do load-balancing and
failover.
If you're trying to keep server counts down, it
might be better to run multiple queues on one machine,
each on its own spindle. You can basically make
multiple qmail installs, modifying conf-qmail to
control where they are installed. Apply Bruce
Guenter's QMAILQUEUE patch to one of them, and use
the smtp daemon from that installation. Wrap the
smtpd with a program that chooses a random qmail
installation, and sets the QMAILQUEUE environment
variable to that installation's qmail-queue.
That will alleviate the qmail-send bottleneck almost
as effectively as having multiple servers.
--
Chris Mikkelson | "Unfortunately, simplicity is a complicated mess
[EMAIL PROTECTED] | of a concept." --Taner Edis
[1] In St. Paul, MN, anyway.....
