[e2e] Re: [Tsvwg] FTCP algorithm
Dohy.Hong at ens.fr
Fri Apr 26 11:21:18 PDT 2002
On Fri, 26 Apr 2002, Lloyd Wood wrote:
> I have this half-assed half-baked untested intuition that complex
> sending strategies like tcp pacing and complex queuing strategies like
> red are guaranteed to interoperate badly - you want such complexity in
> the endhosts or in the routers, but preferably not both, since they'll
> just interfere destructively, and use ridiculous amounts of processing
> power while doing so. (the endhost at least gets feedback from the
> open loop that is the path, so it's better-placed to adapt to the
> conditions imposed by intermediate queues.)
I'm not sure to share your intuition. But I believe that if
your should implement something complex (even if I think ideas
of FTCP or TCP pacing are simple and natural),
this should be first beared at end-user level
(making things scalable, robust, decentralized).
> If I read it right, Part F of Aggarwal's paper seems to bear this out
> by getting red to approximate drop-tail via tweaking thresholds and a
> low marking probability for best results; just as we test
> modifications to TCP against SACK and Reno and with and without
> delayed acks, we also need to test TCP (pacing algorithms) against a
> range of drop-tail and overly complicated queuing scenarios.
Yes, I agree that TCP pacing (or FTCP :) ) algorithms should
be tested in more various scenarios and see its robustness.
In Part F of Aggarwal's paper, I read that 'pacing builds
up a queue ...' So I'm really wandering if their results
are not biased (from mine) by the absence of MPPN control.
What about your simulations? Is this kind of control is
I still believe that a correct TCP pacing (I hope FTCP does it :) )
should be very robust in quite different scenarios to improve
throughput and decrease synchronization effect.
Projet TREC (Théorie des Réseaux Et Communications)
ECOLE NORMALE SUPERIEURE
45 rue d'Ulm, 75230 PARIS CEDEX 05, FRANCE
Tel : 33 (1) 4432 2112, Fax : 33 (1) 4432 2151
Email : Dohy.Hong at ens.fr
More information about the end2end-interest