[Tmrg] Proposal to increase TCP initial CWND

lachlan.andrew at gmail.com (Lachlan Andrew) Sun, 18 July 2010 09:58 UTC

From: lachlan.andrew at gmail.com (Lachlan Andrew)
Date: Sun, 18 Jul 2010 19:58:18 +1000
Subject: [Tmrg] Proposal to increase TCP initial CWND
In-Reply-To: <AANLkTimgBjTOZQDi8wHaum8KomAKo3TJOBUMeGXjEX1B@mail.gmail.com>
References: <AANLkTil937lyUzRvUtdqd2qdl9RN7AZ-Mo_cT-dtmqXz@mail.gmail.com> <AANLkTimgBjTOZQDi8wHaum8KomAKo3TJOBUMeGXjEX1B@mail.gmail.com>
Message-ID: <AANLkTinruuD89W7WLSiQDSruq2omiCr3gLWO6L5Jqsr-@mail.gmail.com>

On 18 July 2010 16:56, Nandita Dukkipati <nanditad at google.com> wrote:
>
>> To start things off, here is one observation:
>> - An increase in retransmissions from 3.94% to 5.79% seems very
>> significant, if that rate reflects the loss rate.
>
> Quoting a particular retransmission rate (3.94% -> 5.79%) out of context is
> extremely misleading at best. We had made a particular note that this
> significant increase is coming from:
> 1) Applications operating with multiple simultaneous TCP connections (such
> as Maps traffic in this example), and not from a *single* flow as mentioned
> above.
> 2) Furthermore, such a high increase was only observed in an experiment
> performed in one of the data-centers serving subnets with poor connectivity.
>
> While, I am all for inferring and modeling the impact of IW=10 on IW=3, but
> it's pointless to have a discussion based on an incorrect premise.

Thanks for your prompt reply, Nandita.

True, I should have given the context, and pointed out that this was
the worst case reported in the experiments.  I certainly wasn't
meaning to say "we should reject the idea because of this particular
statistic", and apologise if it came across that way!  It is up to
TCPM to make  judgements like that, not TMRG.  (Instead of "seems very
significant", please read the more objective "seems much larger than
1.85%"...)

Ideally, I'd like to see a way that the initial rate can scale as
capacities scale up, and from that point of view I would like the TMRG
to help you achieve your overall objective.  The first step is to make
your experiments as compelling as possible.

My intention was to suggest that the relevant statistic is not the
change in the small fraction of packets retransmitted, but the
relative increase in congestion signals.  (My meta-intention was to
start a discussion, and so I mentioned the first point that jumped out
at me.)

Regarding whether it is a "single flow" or not, I'd argue that even if
it is multiple TCP connections, they're all really part of the same
"flow", since they're aiming to carry (different parts of) the same
map.  Terminology aside, my point was that the effect observed when
only a (small?) subset of users adopt the scheme is different from
what we would expect when everybody adopts it.  The TMRG issue is "How
can we predict the latter, given we only observer the former?"  Of
course, we could just simulate toy networks as has been done in the
past, but it would be much more compelling if we could use the
extensive data you have collected.

I hope this removes the incorrect premise, and that this discussion is
one you're happy to be part of...

Cheers,
Lachlan

-- 
Lachlan Andrew? Centre for Advanced Internet Architectures (CAIA)
Swinburne University of Technology, Melbourne, Australia
<http://caia.swin.edu.au/cv/landrew>
Ph +61 3 9214 4837