[Iccrg] RE: [tcpm] ECN feedback discussion

[John Leslie]

Ingemar Johansson S <ingemar.s.johansson at ericsson.com> wrote:
> 
> Must admit that I a bit lost here. 

   We probably need to agree on better terminology.

> The idea is that 
> 1) the sender sets ect0

   My understanding is the sender _always_ sets ect0.

> 2a) a congested router sets ce if it implements low density marking
> 2b) alternatively a router that implements high density marking sets
> packets to ect1 or ce. 

   We'd best agree on "low density" and "high density" or some other names.

   I suggest defining:
- "low density" to mean ECN-marking at a high threshhold with the intent
  that the sender should reduce sending rate by half in response to a
  single mark;
- "high density" to mean ECN-marking packets more often than that,
  intending a lesser reaction to any single ECN mark.

   The intent, IMHO, is that high-density marking will change ect0 to
ect1, while low-density markng will change either to ce.

   Clearly, a pure tail-drop router will drop packets regardless, and
any ECN router will drop packets if it runs out of buffer space.

   The original proposal called for a "high-density" (non-legacy) router
to change ect0 to ect1, or ect1 to ce, at a low threshhold. This seems
reasonable, but not critical -- research might show some other behavior
to be superior. (I don't give a lot of thought to detecting different
marking rates for two routers along the path.)

   From a control-theory viewpoint, exponential growth is bad, while
exponential decay is good. Halving the rate, while rather coarse, is
correct if the appropriate target is zero. Control theory talks of
"critical damping" where there's a continuous "correction" towards
a target value, and the "correction" tends to zero as the target is
approached.

   In congestion-control, we suffer from not knowing the "target".
Neither zero nor infinity is a good target. Research using high-density
ECN marking should help us refine the "target" to more reasonable
values, IMHO.

   From there, we can define reactions which start to resemble critical
damping. We are likely to find, with sufficiently "high-density" marking,
that the ideal response will sometimes be _increasing_ the sending rate,
but increasing more slowly than we have been increasing it.

--
John Leslie <john at jlc.net>