Dave Sill wrote:
>
> Daemeon Reiydelle <[EMAIL PROTECTED]> wrote:
> >
> >Since the bottleneck will be your internet pipe, ...
>
> Probably the spool disk, unless you've got multiple boxes flooding the
> pipe.
A couple of pieces. Performance tuning any system consists of finding
which of a series of potential bottlenecks is THE bottleneck limiting
performance at the moment. The original author was very clearly talking
about outbound message delivery. I made an assumption that anyone
looking at the volumes she was discussing probably didn't have a
ten-thousand-dollar-per-month OC-12 network connection. I therefore
limited myself to discussions of less-than 100mbit issues. NONE of the
issues discussed by Dave would affect performance at network speeds
under 50mbit (OC-3). (Even with mirrored drives as long as they are at
least fast SCSI).
Please don't let comments like Dave's make you think there is any pain
in Qmail on Linux, FreeBSD or even Solaris. You do need to have some
reasonable skills in the OS and a reasonable time to learn this new
technology (incidentally, not only do I work with both UNIX and NT, but
I am currently moving a client from an NT Exchange environment to Qmail
on Solaris. I *believe* I have a good feeling for how easy it really is
to make this happen.).
If you have e.g. an OC-12 internet pipe, then the tuning points
mentioned by Dave would need to be considered. The next bottleneck you
are likely to see would indeed be the qmail directory file system, and
yes, you will need to add an accellerated (battery backed cache) disk
subsystem. After you do that, you will need to make some additional
kernel changes (just as you would on Solaris, HP-UX, SGI-Irix, etc.)
(FYI, both OS's in their current kernel builds will support >1024 child
processes with appropriate kernel optimizations).
Going out on a limb, my guess would be you could expect 50-100 1000 byte
messages per second (or better) outbound with an OC-12, hardware assist,
etc. However, there are some FreeBSD/Linux limitations on throughput at
the moment that *might* come into play here. Recent IIS vs. Apache/Linux
testing identified a limitation on threading in the IP stack at high
HTTP volumes in recent releases that don't occur with Solaris X_86. This
limitation on the IP stack applies (or applied) to both FreeBSD and
Linux. I am sure that these limitations will (or may already have as you
read this) go away soon. As Silicon Graphics gets more involved in
Linux, these extreme load point constraints should start to fall even
more rapidly.
Your mileage may differ of course ;{)
>
> >Either FB or Linux will build qmail with ease and works fine.
>
> True, but if one or the other handles >1024 file descriptors so
> concurrencyremote can go over 256, that would be a major plus since
> the hardware involved could potentially drive that many
> qmail-remotes.
>
> -Dave
--
Daemeon Reiydelle
Systems Engineer, Anthropomorphics Inc.
[EMAIL PROTECTED]
