[14990] in North American Network Operators' Group
Re: backbone routers' priority settings for ICMP & UDP
daemon@ATHENA.MIT.EDU (Vern Paxson)
Wed Feb 4 16:52:07 1998
To: Alex Bligh <amb@gxn.net>
Cc: nanog@merit.edu
In-reply-to: Your message of Wed, 04 Feb 1998 19:57:40 PST.
Date: Wed, 04 Feb 1998 13:47:40 PST
From: Vern Paxson <vern@ee.lbl.gov>
> As is (reasonably) well known, TCP has its own congestion control built
> in to an extent.
This should be put much stronger - TCP's congestion control is *the*
mechanism that keeps the network stable today.
> Or is Source Quench really broken by design?
Pretty much. It's flawed in at least two important ways: (1) it adds
traffic to the network during times of congestion; and (2) if a source
quench packet is dropped, then the intended target fails to receive the
notification to slow down.
Both of these lead to the wrong sort of feedback cycle when the network
is heavily congested - a cycle that can make congestion self-sustaining,
i.e., congestion collapse.
Vern
(There may be a niche for source quench, though, in environments with
high-latency links, where the quicker feedback than that obtainable from
RED+ECN might be a sufficiently big win.)
