On Mon, 23 Jul 2007, Patrick - South Valley Internet wrote:
> Christopher - what would you consider "high number of accounts"? Would
> >10,000 count as a lot?
Not really.
The reason for running a hashed structure is to keep the number of
subdirectories (mailboxes) per directory at a sane level, say < 2000 per.
Modern filesystems handle things fairly well, but walk 36, walk 36, walk
1580 (three level tree) is still faster than walk 15,000 (or in my case
walk 98,000ish).
It also makes user operation in the tree alot less painful (think tab
completion in a directory of 10,000).
The reason we run an indirect first level hash via symlinks is to allow
the spreading of storage and load across multiple mounts (in our case
multiple netapps). In our current use a-j use and k-9 are roughly equal.
Even with a single server this would not be a bad idea, as it can be used
to spread across multiple storage volumes while still keeping a unified
tree. We can hald delivery/pickup to 1/36th of the spool just by removing
a symlink, move the sub-tree between volumes, and add the new
symlink.
> The two Postfix servers are Dell dual cpu Opterons with 2gb RAM. The
> NFS is a single Opteron 170 w/2gb RAM and an LSI Logic MegaRAID SATA 6
> port card.
For classic ISP service (20MB or smaller mail quotas and POP service) one
of those duals would be way more than enough for Postfix/Courier service
(If these are also MXes running anti-virus/SPAM I can see it requiring
both).
For more featureful ISP or Corp service it depends. The more IMAP use and
the large the mailboxes, the greater the load.
User wise I have about 80% doing POP only with quotas < 20MB, 15% running
IMAP directly or via webmail with larger boxes, and the remaining 5%
running IMAP with quotas of 2GB (average mailbox size ~ 800 megs and
40,000 messages). Currently I can lose 6 out of 12 systems at peak times
with no impact to the customers. (Systems are all dual CPU w/ 4GB)
Planned increases in quotas and IMAP usage will eat a few more systems,
but the longterm goal is to be able to have up to 4 systems offline
without customer impact (double failure + 2 out for maint).
One thing though, our MX systems that perform all of the SPAM/virus
scanning and blacklisting are totally seperate. All of the anit-SPAM
processing is currently burning 8 machines. These systems do nothing but
relay in/out. We burn more CPU doing anti-virus and SPAM scoring than we
do servicing all per user filters (based off of the scoring on the MXes)
and servicing POP/IMAP.
During high load periods we see about 2,000 SMTP incoming SMTP sessions at
a time, with all 8 MX systems running both CPUs at 100%. < 5% of the CPU
use is postfix.
|
