[e2e] interaction between TCP and rate limiting in routers

[JFellows at coppermountain.com]

I'm a long time lurker on this list, but would like to now raise an issue
that I suspect is old hat to many of the members of this list.

I'm searching for the current best practice for implementing per flow or per
subscriber rate limits in edge routers.  This is becoming a common request
from consumer oriented ISPs in both the cable and DSL markets.  I've done
the Google search and read a half dozen papers that touch on this subject.
Many of the published papers refer to the classic 'sawtooth' behaviour of
TCP in this configuration.   Some suggest that, in order to avoid multiple
consecutive discards, the token bucket that implements the rate limiter
should have a depth at least equal to the RTT*configured rate.

Are there any non-invasive (not involving on-the-fly modification of TCP
headers) implementations of rate limiters that allow TCP to run fairly
smoothly at the configured rate?

Jonathan