Large scale transparent proxying

Has anyone out there done any large scale transparent proxying?

I'm running squid-2.2.STABLE4 under NetBSD 1.4.1 doing transparent
proxying without caching, and with about 200 concurrent users (I don't
have a figure for number of hits per second) the load gets really high
(85% CPU usage or thereabouts) on a Pentium II 450 with 256Mb of RAM.
A lot of that load seems to be system CPU time dealing with the
required NAT translation.

It seems to me that load shouldn't really be high...what it's doing is
really pretty simple, especially given that it's not caching.

Has anyone out there run any large transparent proxies? We need to
scale this up for many more users and want to work out how we go about
doing it. I'm happy to choose the OS and the platform to suit the
task.

--
Peter Yandell
Chief Technology Officer
Freeonline.com.au Pty Ltd
peter_yandell@freeonline.com.au