by LOEWENTHAL Simon-2
<http://old.nabble.com/user/UserProfile.jtp?user=747235> Oct 03, 2011;
11:08amHi there:
I have to set-up a few low power SA boxes. Currently I'm used to
using Intel Xeon 2.6Ghz with 16Gb of memory, but these proposed boxes
are small. I won't buy one unless I know it can do the job. I know the
figures below are tiny, but I don't know the Intel Atoms and what they
can really do.
# of active Email addresses (excluding Email aliases) : 80
# of messages (including rejected) appox 3,500 daily
running : Debian/ SA 3.3.1 and spamass-milter (with MTA postfix,
clamav-milter).
hardware:
http://soekris.eu/shop/net6501_en/
1.6 Ghz Intel Atom E660 (1 core, 2 threads)
1024Mb RAM
Transcend mSATA SSD 32Gb MLC
Cheers for any commentary.
This year I decommissioned a box with a similar workload, but also
running Apache and mySQL. It had a Celeron 700 and 256Mb of RAM. It was
a little under-specified - it needed more RAM for peak loads, but as it
was email, it just had to queue for a few seconds. It'd still be running
but for one of the PATA drives - went flaky and replacements not readily
available - and it was ~12 years old so owed me nothing.
For what it's worth, this was running FreeBSD, sendmail, postfix and
dovecot (along with Apache/mySQL - which was pushing it). With just
email and 1Gb of RAM you should be laughing. You could probably allocate
half of it as a ramdisk for good effect. If you're using X.11, all bets
are off!
You don't say what you're doing *with* the mail. If it's going into
mailboxes with an IMAP server, you'll need more disk unless the users
are well trained - or NFS?
Don't be scared of the Atom - I use them all the time. My concern is low
energy, low maintenance (little heat = no fan). The boards I use take
about <1A @ 12V (with 2.5" drives) and can realistically be run off
Solar and a car battery.
Just remember Bayesian filtering is what eats CPU, if it comes to
optimisation. But you're talking about a couple of emails a minute,
which is nothing - I'd estimate they'd take between two and ten seconds
each, even with Bayesian filtering - don't even thing of worrying until
they hit fifteen a minute.
Regards, Frank.
--
--------------
Sent from my Cray XT5