central amavis machine for in and outbound

[Patrick Ben Koetter]

* Tobias Hachmer <lists at kokelnet.de>:
> Am 17.11.2011 19:55, schrieb Patrick Ben Koetter:
> >>Well, at the moment there are 30000~ incoming mails per month,
> >
> >30.000 msg overall, including spam, or filtered messages?
> 
> Yes, overall. Got this values from mailgraph...

2 servers for in- and 2 for outbound mail is FAR too much!

Look at this breakdown:

30.000÷30 = 1.000 msg/day
1.000÷24  = 41,666666667 msg/hour
42÷60     = 0,7 msg/minute
0,7÷60    = 0,011666667 msg/sec

Ordinary hardware (!) with cheap HDs can do about 20 msg/sec without any
problem. If you go and buy a good RAID controller incl. a BBU and add
reasonably fast harddiscs you can easily take this up to 60 msg/sec. Better
hardware gives you 100 msg/sec and more.

Ordinary hardware:
20×60        = 1.200 msg/min
1.200×60     = 72.000 msg/hour
72.000×24    = 1.728.000 msg/day
1.728.000×30 = 51.840.000 msg/month

That's 1.728 times much more performance than you need right now.

My example isn't all correct, because there will be other applications that
consume ressources and spam filtering is expensive (bayes tokens etc.), but
the tendency is clear: You don't need 4 servers for performance. You might
need two servers for redundancy.

> >Content filtering is CPU expensive. If you centralize the filter you will
> >have to buy hardware that can filter for more than one server at the same
> >time AND can deal with peak load. You will end up buying extra powerful
> >CPUs (including board and RAM).
> >
> >Regular servers bring 8 core CPUs, of whom most idle most of the time on a
> >mail server because mail is I/O bound. There's plenty of ressources left to
> >run an additional content filter.
> 
> Yeah, you're right, doesn't thought of that, yet.
> 
> >Sharing Ressources
> >If you run a centralized content filter you don't have to worry about
> >knowledge distribution. All knowledge required (e.g. bayes DB, quarantine)
> >resides on one server.
> >
> >I don't consider this a problem either. Accessing bayes and quarantine over
> >a network is easy - both can use SQL as storage. All amavis tools are
> >storage agnostic. You configure the backend in amavisd and tool
> >communication takes place with amavisd, which deals with backend
> >communication.
> 
> Is it a problem when multiple amavisd instances want to write to bayes db
> ore awl db at the same time?

I doubt that 0,011666667 msg/sec impose a problem for a shared database.

p at rick


-- 
All technical questions asked privately will be automatically answered on the
list and archived for public access unless privacy is explicitely required and
justified.

saslfinger (debugging SMTP AUTH):
<http://postfix.state-of-mind.de/patrick.koetter/saslfinger/>