[e2e] What should e2e protocols know about lower layers?

[Marcel Waldvogel]

Michael

So here is a (I hope) convincing argument to those that believe TCP congestion 
control should be turned off locally. Again, I'm assuming MTU=MSS=1.5kB, Link 
Rate=10Gbps, RTT=0.2ms, which seem to be reasonable in a LAN environment.

- Slow start fills the available bandwidth in about 3ms, i.e., instantaneously

- Even in congestion avoidance, when N-1 flows suddenly disappear, our lonely 
flow will be back at full link rate in at most 33ms [*]. For a longer RTT of 
2ms [**] OR a link rate of 1Tbps (!), this goes up to 3s to fill a link. 
Anyone complaining that this is too long: Why did you accept sharing this link 
with N-1 other flows in the first place?

I guess until we get significantly above 1Tbps to the desktop, this should 
convincingly dispel the rumors that congestion control is bad on the LAN.

-Marcel

*: 250kb sent per RTT, which equals 166 packets. It thus requires 166 RTT to 
get up to full bandwidth in congestion avoidance mode, resulting in 33ms.

**: This is about 400km of propagation in copper/fiber, so probably much more 
than most would consider local. Queueing delays are negligible by definition, 
otherwise the network would be congested and we *would* need congestion avoidance.