[Iccrg] Re: [tmrg] Fwd: An aggression metric?

[Lachlan Andrew]

Greetings,

Many new TPC variants are sensitive to things other than the average
loss rate.  For example, CUBIC depends heavily on the burstiness of
the losses, while LEDBAT and CTCP also depend on delay.

It is difficult to find a fundamental "congestion p(t)" that reflects
the notion of congestion seen by all flows.  Steven Low and Kevin Tang
presented some theoretical work on this in INFOCOM 2005.

The notion "congestion = average loss rate" is already too entrenched
IMHO.  It would be great to be able to measure aggression, but a
simple scalar (or even a scalar function of scalar congestion measure)
may not add much insight.

$0.02,
Lachlan

On 14 July 2011 20:22, Bob Briscoe <bob.briscoe at bt.com> wrote:
> Michael,
>
> Congestion-rate fulfils the role of measuring the aggression of a flow,
> doesn't it? And it is in objective units of b/s, avoiding being relative to
> TCP which is what you're looking for.
>
> Congestion-bit-rate is the bit-rate at which congestion events (losses or
> ECN marks) are generated or experienced, for a flow or any other aggregate.
> It measures the response of bit-rate to congestion, which is another way of
> saying aggression.
>
> In a fluid model, congestion-rate, v(t), is the instantaneous product of
> bit-rate x(t) and congestion p(t):
> v(t) = p(t)x(t)
>
> At first I thought this wouldn't be right as a measure of aggression,
> because for TCP it depends on congestion, so I was trying to normalise it to
> be independent of congestion. However, I realised dependence of aggression
> on congestion is a feature of TCP: as congestion rises, aggression rises.
> Also, aggression depends on RTT for TCP.
>
> Taking the simplest TCP macro-model:
>        x = K/(R.sqrt(p))
> where K is a constant, and R is RTT. Then the congestion-rate,
>        v = K.sqrt(p)/R
>
> For a proportionally fair congestion control that converges to a rate
> independent of RTT:
>        x = wp
> where w is the weight (similar role to TCP's constant). Then
> congestion-rate,
>        v = w
> That is, the weight is the aggression of this flow.



-- 
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