[imp] Re: Scalable Webmail HOWTO and LDAP doubts

[Robin McMillon]

On Thu, 11 Dec 2003, Eric Rostetter wrote:

> Quoting Sn!per <sniper at home.net.my>:
>
> > Quoting Eric Rostetter <eric.rostetter at physics.utexas.edu>:
> >
> > > If you have 50K users, then that is a large install.  I would suggest
> > > multiple machines with a load balancer (which is what my university does
> > for
> > > supporting >50K students).
> >
> > can you briefly describe the hardware and software requirement of your load
> > balancer and any particular make/model that *highly* recommended.
> >
> > tia
>
> I'm not invloved on the campus level, so I can't speak authoritatively.

I thought you might have been talking about us.  :)

Well, to shed some light, here's what the campus-wide Webmail uses:

First off, I would not recommend our load balancing setup for an
installation of the size you're talking about.

> But I believe they use Cisco LocalDirector load balancers there.  The

Webmail specifically uses a php script on our front-end servers to do load
balancing.  It works pretty well although it does put an additional load
on the servers instead of off-loading it onto a specialty piece of
hardware.

> mail spool is stored on NetApp appliances.  They have a bunch of Sun machines
> to do the e-mail services (pop, imap, smtp, etc) which are behind the
> load balancers.  I think then also they have another bunch of Sun machines
> which do the web servers also I think behind the load balancers.

What we actually have running Webmail currently are one Sun 280R and three
Sun 220s running the Horde/IMP/Turba code with the Zend Accelerator.
We're using a Sun V120 to run a dedicated MySQL server, and mail is
handled by several completely separate entities running on various makes
and models of hardware (we let people log into any mail server that will
have them).  The web machines are separate from Webmail.

We only have about 30K actual users (many of them faculty and staff that
do not count in the 50K students), but every year that number grows as a
new crop of students trained on Webmail come into the university.

During most non-vacation weeks we have somewhere between 20K and 60K
logins a day.  We also have mini-peaks of 5-20 minutes when classes
change.  When we run low on anything, it's always CPU.  In order to do
most of your traffic with TLS sessions or offer additional Horde modules
to the majority of your user base, you'll definitely need more CPU.  One
other thing you'll probably notice is that the larger people's filter
lists get, the more resources you need to do a login (since most people
have them applied then).  While this might not be noticeable at a smaller
site, it really makes a difference when you're doing >1 login per second
on a single machine!

 > I don't
> know how exactly they do the authentication/prefs, since it is linked into
> our campus wide authentication system via custom code, but I bet it is
> ldap.

The custom code doesn't do LDAP in the current production system.  We
actually wrote our own special driver files, etc.  (Thanks Horde guys for
writing the architecture so it was easy to do that, btw!)

I hope this helps with your planning process.  :)

Robin McMillon
r.mcmillon at its.utexas.edu