[Iccrg] Delay as a metric of congestion
John Leslie
john at jlc.net
Sun Feb 21 22:12:58 GMT 2010
Matt Mathis <mathis at psc.edu> wrote:
>
> John and I went around on this for a bit in a private conversation,
> where I argued against his position.
I'm not positive what Matt "argued against," but it certainly wasn't
my "position"...
> However, Wes's comments about the history made me realize that I
> actually very strongly agree with a slightly weaker, conditional
> version of John's position.
>
> Given that:
>
> 1) When delay sensing works it generally provides better, more precise,
> congestion control with less disruption to other applications than loss
> based congestion control.
I did not take that position, but I won't argue against it...
> 2) Delay sensing does not work in all environments and must be considered
> to be an optimization that is conditionally applied in addition to the
> primary loss based congestion control.
I don't even agree with that...
> 3) For the vast majority of Internet users (e.g. home users behind a slow,
> over buffered access link and perhaps wireless users) #1 applies extremely
> strongly.
I think Matt means that if we can determine that the bottleneck is
between the end-user and his/her ISP, delay is a better metric. I would
agree with that.
`
> Therefor: all stacks should include a secondary congestion control
> mechanism that detects when they are causing large queues in the network
> and regulates their congestion window accordingly.
I couldn't agree with that exact wording, but Matt & I may be close
here. Perhaps:
"
" Stacks intending to "play nice" with TCP should implement a delay-based
" mechanism to avoid reaching a congestion level where packets they send
" are dropped instead of delivered. Research into such methods is an
" appropriate field of study of the ICCRG.
> It may be an appropriate future work item of the ICCRG and IETF to
> attempt to standardize such a mechanism.
I think in appropriate _now_ to collect data on how well delay-based
reductions (not necessarily multiplicative reductions) in window size
signal "gzornenplat" and avoid packets being dropped.
("gzornenplat" is a placeholder for what I actually wanted to talk
about: the onset of what I call "congestion", but not so much of it as
to require packets to be dropped.)
> In the short term it would be sufficient to merely agree on a statement
> of principle or intent.
I really don't see how that would help...
(Nor did I intend to talk of anything like this in my proposed
presentation...)
--
John Leslie <john at jlc.net>
More information about the Iccrg
mailing list