Server capacity estimates?

[Michael Chaney]

On Wed, May 29, 2002 at 04:44:14PM -0300, David Lancaster wrote:
> Hey all,
>
> Are there any sources of information about what server capacity is needed
> to handle various loads from mailscanner and spam assassin?
>
> We're currently looking at handling 3500-4500 (incoming only)mail per
> hour, using mailscanner and SA, with virus checks disabled, and rbl
> enabled in mailscanner.

F-Prot can be had for $300/year, and will add minimal overhead
(especially next to SpamAssassin).  I highly recommend adding it, it's
worth the small investment.

> The server will act as a gateway, passing the
> messages off to another server for actual delivery.
> What sort of capability would a single server (Intel/Linux) need?  IO and
> CPU seem to be the bottlenecks on my test server, would a single 1.4 Ghz
> P3 and a software-mirrored 15K Ultra160 disk be capable of handling this
> throughput? Would dual CPUs increase performance to any marked degree?

My tests (not real-world by any means) show an Athlon 500 or 600 (one of
the two, I don't remember) topping out at under 1000 messages per hour,
sometimes as low as 800 (this while being fed a diet of pure spam to
test the effectiveness of SA).  SA is a large contributor to that, and I
believe that DNS lookups are contributory to that as well, meaning more
CPU won't necessarily help.  It is possible with MailScanner to limit
SA's time, but I've found that it needs to be 20 seconds to catch the
most spam.

My feeling is that preprocessing the DNS lookups would possibly make a
major difference in the time, and I might put that together at some
point as a completely separate piece that works along-side MailScanner.
I have code that I wrote a few years ago that does a massive number of
DNS lookups, well over 100K/hour, by doing 100 or more simultaneously
and asynchronously.  Letting that do the open relay, reverse lookups,
and other blacklists, forcing them to be in the nameserver's cache when
SA asks for the same thing, could potentially speed things up.

Anyway, I'm interested to know what machine is capable of the throughput
that you describe above, too, when you're finished.  My feeling is that
the 1.4GHz P3 might not be quite enough.

Michael
--
Michael Darrin Chaney
mdchaney at michaelchaney.com
http://www.michaelchaney.com/