[e2e] Some simple, perhaps stupid, idea. was: Re: Agility of RTO Estimates, stability, vulneratibilites

[Detlef Bosau]

Detlef Bosau wrote:
> 
> Lucky me, I found Raj Jains Paper on Divergence of Timeout Algorithms...
> online. And now I´m trying to obtain the paper by Lixia Zhang, Why TCP
> Timers don´t work well. And primarily of course I want to get and to
> read the paper on Edge´s algorithm. I think, that´s what I want and what
> is my next step. Understanding the rationale behind Edge´s algorithm,
> which is used until today AFAIK.
> 

...

> 
> The simple question is: When I hide a network´s last mile, e.g. a mobile
> wireless network, behind some kind of PEP and thus provide the sender
> with a "changed RTT behaviour", which any spoofing or splitting PEP
> does, what is the ideal RTT behaviour? I.e. an RTT behaviour which makes
> Edge´s algorithm work perfect?


Eventually, I got Edge´s paper. Perhaps, we should turn away from 
computers and go for good old books. O.k., I got Edge´s paper as pdf 
File. However: The longer I study literature on TCP, the more I get the 
impression: The older it is, the better it is. This may a stupid 
prejudice. But much of the somewhat older papers are really carefully 
thought through. And perhaps do not follow blindly the "publish or 
perish" principle, which is perhaps modern nowadays.

In this post, I simply want to share a very spontaneous idea, presumably 
it´s stoneaged, but it is quite simple and clear. And it illustrates my 
thoughts.

Edge poses rather weak requirements for the RTT process. E.g., the 
individual stochastic variables T(n) must share a common expectation and 
variance. And there is some requirement concerning the convariance.

This is far from being "i.i.d." or memoryless or Poissonian or something 
like that.

Particularly, there is absolutely no assumption for the T(n) to obey 
some specific distribution funkcion. The rationale for the RTO itself is 
based on Chebyshev´s inequality, thus it is absolutely generic.

However, Edge wants the T(n) to share a common expectation and variance.

Now, when I think of RED strategies, I remember a strategy where there 
are two thresholds a, b, a < b, for a queuelength q. If q < a, packets 
are accepted. If b < q, packets are rejected. If a <= q <= b packets are 
rejected randomly with a probality p which is linear increased from p=0 
if q=a to p=1 if q=b.

Question: Would it make sense to chose a and b that way, that
i) q has a constant expectation and
ii) q has a constand variance
for certain periods of time?

Expection and variance could well be chosen appropriate for the load 
situation.

When I consider a network as a sequence of links and queues (I know..... 
  but I will do this for the moment), the varying part of the RTT is the 
queueing delay, as long as the path for a connection does not change.
So, if every router on the path would try to maintain a constant 
expectation and variance for queue lenghts, the queueing delays would
have a constant expectation and variance.

Therefore, the observed Tn would have a constant expectation and 
variance, at least for small periods of time.

Would it be possible to achieve this by management of the thresholds a 
and b? If so, this could be achieved by each router individually.

As a consequence, at least the requirement for a common expectation and 
variation of the T(n) would be met.

So far. It´s spontaneous, it´s perhaps stupid.

But it shall illustrate my way of thinking, that it may be reasonable to 
perhaps make the network meet a protocol´s requirements instead of 
always make a protocol suitable for a network.

However, I expect that someone has discussed this before, it´s just too 
simple.

Detlef


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