Re: [Exim] large scale exim/mail system

> > Exim is IO limited, so get some 2 gigs RAM disks. They cost about £5k
> > each, but that's about the price of a compac server. However, the
> > preformace of exim increases by a factor or 4-5. Thus you are effectivly
> > saving money.

I tried a RAM disk for the hints databases and under high load, it helps
a lot to reduce mail latency. Once Linux tmpfs becomes stable (ramfs
is _not_ up to the job), it will certainly be an option to consider.

> Can I just remind everyone that has a disc IO bottleneck that the software
> RAID that comes with Linux (version 0.90 as provided by all the major
> distributions on 2.2.x, as well as all 2.4.x kernels) is a simple, cheap way
> of increasing IO bandwidth for both read and write.
>
> Lots of 'small' SCSI 160Mhz discs on RAID 0 _really_ flys (and do so even
> more if you have dual controllers as a lot of the onboard SCSI controllers
> are these days).

Even with a single controller, four striped disks get amazing I/O values.
In my other life, I am a newsmaster and I run multiple RAID0 md devices
on four disks for a newsfeeder, each with a different stripe size, for
different partitions. The load spreads very well and the transactions
per second of the entire system as well as the throughput continues to
amaze me.

Are there any experiences with different stripe sizes for use with Exim,
possibly even which stripe size to use for hints db and which for queue
and maildirs?

While talking about RAM: Exim needs a new feature, actually four new
features. :) TCP window size for receiving/sending, SMTP client/server.
Just recently I used smaller window sizes for connections with a
request/reply behaviour and known small packets to save lots of RAM, which
in turn helped the page cache and thus made the whole system a lot faster.

Michael