The tests that Wietse Venema did for Postfix (sorry no ref here - I got
the information from his slides at the Maastrict conference last
November) showed that the MTA spool disk traffic was the biggest
performance limitation on all the mailers he examined. Postfix
attempts to address this slightly, but even so the number of disk ops
(and most of them are directory modifications which necessitate a sync
write to disk) are a major limitation.
Fast wide SCSI would help, however the limitations are more likely to
be related to waggling disk heads around. Spreading the load over more
spindles will help drastically - ie RAID 0. Other RAID levels may help
to a a greater or lesser degree (RAID 5 does normally give a
performance hit on writes). Having a disk subsystem fronted by a
persistent cache would give a very strong performance boost.
A logging filesystem such as VxFS, especially in the sloppy modes -
although you risk data loss on a crash - could give a significant boost.
I certainly would not be at all happy at using a non-persistant file
system for storing queue data. NVRAM might be a possibility if you had
that - in which case you might want to use another exim instance for
soaking up stuff that isn't delivering first time so that the
(presumably) expensive very fast storage is used for transit messages
only.
Nigel.
--
[ Nigel Metheringham Nigel.Metheringham@??? ]
[ Phone: +44 1423 850000 Fax +44 1423 858866 ]
--
*** Exim information can be found at
http://www.exim.org/ ***
