[e2e] Agility of RTO Estimates, stability, vulneratibilites

[Detlef Bosau]

O.k., so welcome to the "cage aux folles" :-) *SCNR*

And perhaps, I´m a real fool, because much of that what you say is in 
fact new to me.

David P. Reed wrote:
> The most fundamental problem with RTO estimates in the Internet is that 
> the most significant sources of measured variation (queueing delay, for 
> example) are variables that are being used as signalling channels 
> between multiple independent goal-seeking processes at multiple levels.

Let me put an in example in my own words to see whether I understood 
your remark correctly. And please be patient because I´m a beginner here.

Some example for what you have said could be the congestion signaling 
between a congested router queue and a TCP source. Is this correct?

> 
> Note that the load distribution cannot be characterized by a stable a 
> priori description, because load is itself responsive at all timescales 
> to behavior of humans (users, app designers, cable plant investors, 
> pricing specialists, arbitrage experts, criminal hackers, terrorists, 
> network installers, e-commerce sites, etc.)
> 

O.k.. If we leave out Al Quaida for the moment (apparently, I can´t 
because Al Quaida´s target has moved and appears to be now Good 
Ol´Europe...) the question is: What are the invariants at least for the 
lifetime of a TCP connection?

E.g.:

- Does a path change during that lifetime?
- How often do we encounter multipath routing?


> So you are fooling yourself if you start with a simple a priori model, 
> even if that model passes so-called "peer review" (also called a folie a 
> deux - mutually reinforcing hallucinations about reality) and becomes 
> the common problem statement for a generation of graduate students doing 
> network theory.  In my era, the theorists all assumed that Poisson 
> arrival processes were sufficient.   These days, "heavy tails" are 
> assumed to be correct.   Beware - there's much truth and value, but also 
> a deep and profound lie, in such assertions and conventional wisdoms.

As I said, I´m an absolute beginner here. But when you simply look at 
the assumptions made for the definition of a Poisson process, it´s 
really heavy stuff. And I sometimes wonder, where the justification for 
those assumptions come from. You can extend this to markov chains etc.
One of the first lessons, I´ve learned from textbooks about stochastic 
processes is, that markovian processes are really nice - however reality 
is not quite as nice, it typically is not markovian....;-)

However, I´m not quite sure whether we really need an explicit knowledge 
of a latency distribution to establish a confidence interval.

Of course, it´s comfortable if a given statistic is always N(0,1) 
distributed. If you want a symmetric 0.9975 confidence interval, you 
simply can look it up in a table. And if your distribution is 
N(\mu,\sigma), o.k. that´s more difficult, left to the reader ;-)

However, from what I´ve seen in mearurements from real networks, I´m 
afraid network latencies won´t do us the favour to obey that simple 
distributions.

And what seems realy nasty to me is, that the often used "asymptotic" 
distributions will perhaps not hold here, because although long term 
flows are perhaps comfortable to deal with, they are rare in reality. 
I´ve read statistics and observations that claim that 95 % of all TCP 
flow consist of less then 20 packets in there lifetime.

> 
> Those of you who understand the profound difference between Bayesian and 
> Classical statistical inference will understand ...

O.k., admittedly I do not... So, I have to learn about it.

-- 
Detlef Bosau
Galileistrasse 30
70565 Stuttgart
Mail: detlef.bosau at web.de
Web: http://www.detlef-bosau.de
Mobile: +49 172 681 9937