[e2e] Congestion control and avoidance - QoS?

[Randall Stewart]

Michael Welzl wrote:
> 
> Folks,
> 
> > The idea of implementing improved congestion control on an
> > end-to-end,
> > cooperative basis in a new protocol (replacing TCP as the
> > basis for HTTP)
> > sounds like a worthy goal.
> 
> ... wouldn't SCTP be a good place to start?
> After all, from what I've read about it, I figure that it may
> indeed replace TCP in the long run.
> 

I don't know if it can ever replace TCP.. supplement it for
certain types of application yes... Is HTTP one of those? it may
be but I am not that versed in the web... someone with a greater
insight into the deep details of HTTP would need to judge if
SCTP would be a better choice than TCP..

> I do not believe that there has been sufficient research to
> qualify for such changes in TCP yet, but on the other hand,
> I really don't think we should stick with AIMD for all ages.
> 

SCTP uses AIMD as well so you will not get away from AIMD without
first going back and getting changes to the basic CC algorithms
defined for SCTP... a tough sell in my mind..


> e.g., see:
> Harrick Vin and Sergey Gorinsky, "Additive Increase Appears
> Inferior", http://www.cs.utexas.edu/users/gorinsky/pubs.html
> There could indeed be better linear controls.
> 
> And there is more than just the stability issue - some argue
> that AIMD satisfies Frank Kelly's desirable "proportional
> fairness" property, which is not true according to Jean-Yves
> Le Boudec in "The Fairness of Additive Increase and
> Multiplicative Decrease", which can be found at
> http://icawww.epfl.ch/ICAPublic/asp/Publication_selection.asp
> 
> The SCTP folks could get rid of things which are only there
> for backwards compatibility - such as the "ECN capable" flag.

Hmm. I take it you mean the "ECN capable" flags at startup of
an assocaition. Yes, these could be eliminated when we attempt
to go to draft standard status. The reason that this did NOT
happen first cut is that at the time we were putting ECN (or
attempting to I should say) into SCTP, ECN was still experimental.
This made it so we had to relegate it to an appendix with "reserved
chunk types". Now that ECN is standards track I would hope we
discard the whole negotiation thing and just make it mandatory...

Time will tell though.


R

-- 
Randall R. Stewart
randall at stewart.chicago.il.us 815-342-5222 (cell phone)