[e2e] Are we doing sliding window in the Internet?
David P. Reed
dpreed at reed.com
Mon Jan 7 06:46:10 PST 2008
99% of the effect of the original TCP congestion control "innovation"
was due to a simple and robust principle: when a packet is lost, cut the
traffic being sent *drastically*. This converted a tendency of the
community to turn packet losses into a positive feedback loop of
retransmission into a shared understanding that "positive feedback loops
That was a community-scale learning that was useful.
I would argue that the idea that TCP is now "near perfect" as a standard
that *must be enforced lest the Internet enter congestion collapse* has
almost no real-world evidence.
In fact, very little of the traffic on the network today comes in the
form of end-to-end unbounded-rate-demanding streams. TCP's congestion
control (and the definition of "TCP compatible" fairness, too!) is
"tested" only for such a trivial case. No One At All tests TCP
congestion control in a world of HTTP, VoIP, streaming video from
YouTube, Skype, etc. The "gurus" of the IETF that study protocol
*documents* have intuitions that are seriously flawed, if only because
they don't actually have any data about the traffic ANYWHERE in the damn
network. (kc claffy would die to have such data).
So it seems that a bunch of self-elected high-priests of IETF have taken
it upon themselves to claim God-like, experiment-free knowledge of the
Internet as it really is, and to use that knowledge to claim that TCP is
the Gold Standard that must be obeyed lest the network finally do what
Bob Metcalfe erroneously said it would do and ate whatever piece of
clothing he ate as a result.
Do science and real engineering, please. Yeah, the AIMD algorithm in
TCP saved the Internet's ass. But it is not Holy Writ.
More information about the end2end-interest