[e2e] Is a non-TCP solution dead?

[David P. Reed]

At 12:10 AM 4/1/2003 -0500, davide+e2e at cs.cmu.edu wrote:
>Or, put another way, if one end-system sends 20 related packets
>"in close succession" to another end-system 5 miles away, isn't it
>pretty likely that most of those 20 packets will traverse the same
>sequence of nodes?

Depends on the architecture of the wireless system.   From a theoretical 
viewpoint, focusing the traffic along a narrow path ignores much of the 
information carrying capacity of the space surrounding the 
nodes.   Spreading the information more widely in space has many 
advantages.   In a real RF environment, you want to take advantage of 
multipath, for example, rather than let multipath degrade the whole 
system.   This is done in a limited way with a RAKE receiver today, and 
BLAST-type systems do so tomorrow, and some of the work I'm doing will 
extend that even further.   The key to all of this are emerging "cognitive 
radios" which can be reconfigured very rapidly (per symbol perhaps in the 
limit).



>Don't CPU requirements practically
>limit you to a medium-sized number of simultaneously synchronized
>partners?

I see no logical connection between CPU requirements and the number of 
sensible associations a node actively maintains (synchronized).
How many web pages do I have the options to fetch within 500 
milliseconds?   Not that I *will* fetch, but how many have a non-zero a 
priori probability that I will want to and be able to fetch them?   This 
measure of "associations" leads me to a number on the order of at least one 
million.  The limit seems to be related to the amount of memory that is 
contained in a particular "speed of light" diameter around me.

Perhaps your thinking involved a different notion of 
"association".   That's because you have limited your thinking to a 
particular architectural use structure.   The telephony folks of the AIN 
world in the 1970's could never imagine the WWW, except if it were 
concentrated on a single server.   Which is why they thought it appropriate 
to design networks where the calls were individually represented in the 
forwarding nodes of the network, rather than either source routing 
or  hierarchical routing (which made forwarding stateless).  [Networking 
folks still think the WWW is "horribly inefficient".   Efficiency trades 
off against flexibility, and valuing flexibility is hard for some people].

>Do we have any experience with architectures
>where each of the 20 packets traverses a *different* sequence of
>nodes?  Would there be much hope of the transmitting end-system
>pacing packets along those 20 independent paths in a reasonable
>way?

In an RF system, congestion control eventually needs to become more 
generalized - dynamic capacity management.   Because depending on the flows 
the capacity of the system will probably vary.   For example, if all 
traffic needs to go to a single "exit point" the total achievable wireless 
capacity will likely be smaller than if the traffic source/destination 
points are more dispersed.

Using the word "paths" to describe wireless traffic seems to me to be a 
conceptual bug or blinder that prevents thinking clearly about wireless 
networks.   It's the same sort of bug that led the telephone network people 
to misunderstand packet networks, because all they thought that mattered 
were long-hold-time, point-to-point, isochronous flows.