[e2e] TCP over mobile networks: Are there any problems left?

[Detlef Bosau]

Michael Savoric wrote:
> Hi,
> you can also look into my dissertation
> http://edocs.tu-berlin.de/diss/2004/savoric_michael.htm
> 

Hm. I just looked at the summary (you TKN guys always write so big books 
:-) IIRC, Morten Schlager´s PhD thesis has more than 250 pages as well 
;-) ) of your thesis.

Question: Where exactly do you see the problem for mobile networks here?

In fact, I have to think about sharing network information (I find the 
acronym "NIS" for Network Information Sharing somewhat unfortunate here 
because we should learn to use terms and acronyms with unique, 
unambigous semantics instead of introducing "polymorphism" in the world 
of three letter words...): Which amount of information shall be shared?
How much status information must be kept on intermediate systems?
Is the whole thing worth the effort? NB: To the best of my knowledge, 
the vast majority of e.g. TC flows are short termed flows, i.e. less 
then 20 packets. Must we keep information of this flows? O.k., let´s use 
statistical information then. Is this feasible? We often talk about 
"self similarity" - which means, _extremely_ drastically spoken, that 
useful statistical information does not exist, because what ever 
statistics you chose you cannot expect to have consistent (in the 
statistical sense) values.

With respect to mobile networks, I think the question is, whether there 
exists adverse interactions between the mechanisms of a mobile network 
and e2e congestion control and e2e pacing.

If these interactions do not exist, nothing is to be done.

If there exist any interactions - and I´m still not fully convinced of 
this, but during the last week or so my attitude started to change when 
I started to have a closer look at the scheduling mechanisms used in 3G 
networks, the reasons for the use of scheduling and the consequences for 
upper layers, the solution to overcome them are quite simple: We would
use splitting or spoofing techniques to hide the wireless network from 
the wirebound internet.

(As you see, I only consider the wirebound internet with some mobile or 
cellular "leafs". I don´t consider TCP in MANETs.)

However, before we consider solutions, perhaps compare different ones, 
we first should identify the problem. And until recently, will say until 
before a week or so, I was about to consider e.g. spurious timeouts an 
urban legend. Because I never saw a compelling discussion which 
convinced me of their existence.

I know, there is a huge amount of papers on this matter and a huge 
amount of PhD theses. Particularly on spurious timeouts. Most of them 
start with or contain the allegation: "Spurious timeouts do exist." - 
Period.

There is no proof, there is no consideration what could be the reason 
for SRTO, there are hundreds of pages with a solution looking for a 
perhaps non existent problem.

Perhaps you could help me on this one with the "hiccup" tool that was 
hostet on the pages of your institue for years: What was the motivation 
for this? Delay spikes? What´s the origin of delay spikes?

And this is not the question: "We observed them". First of all: 
irreproducible experiments are always interesting. But rarely helpful. 
Second: Wrong implementations exist. So, an observation is not 
necessarily a structural or scientific problem. It might indicate a 
buggy implementation as well.

I started with the consideration of mobile networks about 6 years ago. 
And the more I think about them, the more I see that the vast majority 
of literature I read about them in a computer science context is not 
very convincing. There is a lot of alleged problems and solutions to them.

But I still miss the "hard scientific stuff".

Spoken more drastically: I don´t know, whether mobile / cellular 
networks are a research topic for computer scientists who deal with TCP.
I don´t know whether there is any particularity in them, which cannot be 
dealt with with well proven and existing methods.

At the moment, I have a look at scheduling issues on layer 2 and whether 
they affect upper layers. Perhaps, there is some interaction, perhaps 
there is none. Thierry Klein writes, he has observed one.

We´ll see ;-)

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