[e2e] TCP and MAC layer

[Joe Touch]

Harald Alvestrand wrote:

> to make the opposing view:
> 
> in the 95% case, there is a strong resembleance between my current 
> connectivity and the connectivity of 1 minute ago.
> Core route changes generally switch between same-speed paths (and when 
> they don't, the fallback path is so congested as to be useless anyway); 
> intra-technology roaming often will not produce drastic capacity/latency 
> changes; inter-technology roaming doesn't happen that often (on the 
> order of minutes to hours, not seconds, even if we can make it work 
> seamlessly, which we mostly can't today).
> 
> Guessing "right" for initial RTTE and window size 95% of the time is 
> potentially considerably better than what you get with a fixed initial 
> guess.
> 
> Just being heretical....
> 
>                  Harald


Given the fact that you'd hear about being wrong in 1-2 round trips, it 
might not be a terrible strategy. Lots of other things in TCP can burst 
for a round trip anyway, so if your window is based on statistics of 
previous connections, you're not doing more harm than they.

If the scheme keeps failing, then you could again use statistics to 
scale back, e.g., using 1/4 of the past window, etc. This is the 'aging' 
of the state information that could be kept, e.g., by RFC2140-style 
sharing or more intricate systems like the CM.

Joe