[e2e] Agility of RTO Estimates, stability, vulneratibilites

[David P. Reed]

Sireen Habib Malik wrote:

>
> What "network" should one consider when estimating RTO's (as given in 
> the topic)?
>
>
A discursive answer - from the abstract to the concrete - follows:

That depends on why you are doing the measurement - in particular, what 
you intend to use the measurement for...   a corollary of my rant is 
that measurements acquire their meaning partly from their context of use.

So, if you want to measure RTO as part of an adaptive control algorithm, 
the relevant evaluation of measurement approach is the one that is most 
suited for the purposes of that algorithm.  (e.g. tuning a tight control 
loop or deciding when to plan for more capacity are quite different 
contexts).

Of course the algorithm is probably correct based on an argument 
expressed in terms of assumptions about what the measurement measures.   
The logic is essentially meta-circular.  We won't go into that 
recursion, but instead assume that one or more non-trivial fixed points 
exist.

In less abstract terms, there are a large class of candidates to be 
plugged in to the slot in your algorithm called "RTO measurement".   A 
subset of those measurements give outputs that make your algorithm work 
well.   One supposes that candidate measurement algorithms that will 
work give answer sequences in the near neighborhood of an ideal 
"correct" answer which is an idealization that may not even be well 
defined (since the actual RTO exists only when a bit experiences that 
actual RTO value, the "correct" RTO is an extension of actual real RTOs 
over a domain of time instants where its value is not even of interest - 
does a network have an RTO when no packet is actually being sent?).

Consider for example a control problem involving noisy and missing data 
(such as using RTO measurements to control congestion).  It's been shown 
in some cases, for some control algorithms, that the control system can 
still work quite well with very "large" errors, whereas trying to 
correct the errors by some strategy that smooths or delays the arrival 
of measurements actually results in far worse control.

So "accuracy" need not be something that is calculated by comparison of 
the results of the measurement numerically as if there is some 
well-ordered invariant frame of reference.   Confidence intervals are 
usually defined in terms of numerical quantity, not in terms of effect 
on a larger system.

On the other hand if the use of the RTO measurement is getting a paper 
published, accuracy is best calculated as whatever it takes to get peer 
reviewers to nominate your paper for publication.   One hopes that peer 
reviewers are quite familiar with the normal needs for which such 
measurements are done.   But for new fields and for "mature fields" 
where theory has gone a separate way from practice, peers may be just as 
limited as the author in terms of their perspective.   This is the 
"danger" I refer to.