On Wed, 9 Apr 2003, Nico Erfurth wrote:
> Hmmm, does disabling fsync do such a difference?
extremely. i've done some tests waays back, on ext2fs at that time i
think, some alphaserver hardware (relatively low-spec in every way),
and turning off fsync meant 3 deliveries per second vs 35 deliveries
per second.
> And, it's always the question HOW you send the mails ;)
that was bsmtp, swept up from a local file, fed into exim, and all
deliveries were local as well. no stuff like eximspool/final
destination being on different disks or whatnot; the only peformance
"optimization" (beside turning fsync off) was split_spool_directory.
of course one has to know what risks he's taking with no fsyncs.
being lazy to reply to some other mail in the thread, the original
poster said he was using xfs. i __strongly__ __strongly__ __strongly__
advise using external log. preferably on the starting place of some
completely independent (from the other places involved in
queueing/delivery/anything) drive. this one can be an ide drive, that
shouldn't be a problem (as long as it's fast enough of course, around
ata100), since (at least on linux xfs) the max log size
is 65k blocks, which at 2k blocksize is not much (size-wise); and a
2k block size for mail seems adequeate, partly because "i think so"
:)), partly because the sgi xfs docs say "for news spools, use 2k bs",
and i found much similarity between a news spool and a maildir spool
in (usage) characteristics. 2k bs is for maildir, of course; but imho
anything else would be suicide for high mail loads.
reading the xfs administrator's manual (some sanity filters applied,
as it refers to irix xfs) is highly recommended.
