[e2e] Some thoughts on WLAN etc., was: Re: RES: Why Buffering?

[Detlef Bosau]

Hi to all. 

In the last two weeks, I got some first hand experience with WLAN in 
some educational setups, and perhaps it's interesting to share this 
experience.

Perhaps the most astonishing experience for me was to see packet round 
trip times from my PC to a neighbour, with about 5 meters distance, 
about 80 (literally: eighty!) milliseconds.

I well remember that sometimes attendees ridiculed about a talk where 
the speaker talked about RTT about 20 ms.

Of course, this huge RTT value was a rare, extreme value.

However, even a WLAN setup with about 10 people and, for educational 
purposes, in three infrastructure based networks causes extreme 
difficulties in communication. (In Germany, we have 13 WLAN channels 
AFAIK, so actually we can only run three WLANs in parallel in a 
classroom when the cells shall use disjoint frequency ranges.)

Of course, in some WWAN setups we pursue a better frequency setup and a 
better antenna setup - but basically the problems are the same.

1. Obviously, depending on the setup we use, intercell interference may 
cause extreme difficulties in wireless communication.
2. I strongly guess, that this holds true not only for intercell 
interference but for intracell interference, interferences from external 
sources and other problems as well.
3. I don't expect any realistic possibility to forecast causes for 
interference for a "wireless channel".
4. Consequently, models which model a "wireless channel" as some memory 
less, well behaved and easy going entity may be a bit questionable and 
may be restricted for academic purpose, but do not really reflect the 
real world.

Hence, some of my questions are:

- What is a wireless "channel"? What is a wireless "connection"?

I once was asked about the behaviour of some protocol in the presence of 
"short time disconnections". What is a "short time disconnection"? Or, 
the other way round, what is a connection all about in a wireless network?

In my opinion, there simply are no connections in a packet switching 
network. Consequently, they are no disconnections in a packet switching 
networks. (And as always, this rule is strongly confirmed by quite a 
couple of exceptions. Exactly that's why it is a rule.)

I would prefer the question: How likely is a packet to reach the 
receiver without corruption, once it has been sent?

To the best of my knowledge, even adaptation schemes for wireless 
networks which make use of some kind of "pilot signal" only give a rough 
estimation of the "channel's quality". However, even when a channel 
with  some well known AWGN in a lab will lead to a transport block 
corruption rate of, say, 10 %, there is absolutely no guarantee that 
this will hold in the real world.

So, I think we should abandon the idea of a really useful estimation of 
something like a channel's block corruption ratio.
We may have some reasonable guess - but not really more.

Another question is:

- How do we deal with the aim of resource fairness?

In wireline networks, this is simple. In a wireline link, packet service 
times  vary about the same average and are the same for all users. So, 
by Little's Law (and IMHO this is its correct place in this discussion) 
we have, that a flow's share of throughput on a wireline link is 
directly proportional to its share of space / resource consumption.

As an immediate consequence, the TCP congestion algorithm (see congavoid 
paper) which targets at a fair distribution of ressources / spaces in 
the first, leads to a fair distribution of rates as a side effect.

This hardly holds true in a network where Little's law does only apply 
as a long term average and where quite a number of packets are not 
successfully received, i.e a flow's throughput is not necessarily 
proportional to the number of served packets because some packets are 
sent once, others twice or more, and some packets are sent up to n 
times, n depending on the setup, and there is no successful transmission 
at all.

Hence, the algorithm as stated in the congavoid paper will be of little 
use here, the same hold's true for the fairness consideration of 
Kelly's  paper "Charging and rate control for elastic traffic" from 1997.

 From an end to end point of view, I would like to abandon terms like 
"channel" and "connection" in a wireless environment. A connection could 
only be defined as a logical association of a pair of peers.

The term "channel" may reflect a lab setup with some sender placed in 
the one corner of the lab and a receiver in the opposite corner - and 
some well behaved source of noise in between. For the real world, this 
scenario is of limited use.
(Please note: I don't say: of no use. It's limited use - as long as we 
deal with this model in a reasonable way.)

We can well use adaptation mechanisms to choose appropriate line coding 
and channel coding schemes, however there's no way to forecast a flows 
corruption ratio within narrow limits, neither is there a forecast for a 
flow's achievable throughput or goodput.

O.k., that's just my experience from the last week. And I would like to 
here some comments.

Particularly as this view is not reflected in the text books I know so far.

Detlef






-- 
Detlef Bosau		Galileistraße 30	70565 Stuttgart
phone: +49 711 5208031	mobile: +49 172 6819937	skype: detlef.bosau	
ICQ: 566129673		http://detlef.bosau@web.de