[e2e] Policing TCP flows

[Dennis Ferguson]

Greg,

> here's a way to think of it:  shaping means delaying; delaying means 
> increasing the RTT.
> 
> (here's another way to think of it:  shaping means slowing down the rate at 
> which bits are injected into the pipe; slowing down means decreasing the 
> bandwidth.)
> 
> it's been a while since i've looked at this.  i think my feeling at the time 
> was that one needed to decide on a "reasonable" window size, and be prepared 
> to hold on to that many packets in the shaping device.  2 packets doesn't seem 
> reasonable, but i don't remember how many packets i buffered in my 
> experiments.  it could have been as many as 64.

This I understand.  What I'm trying to point out, however, is that the
effect that delay has on the TCP dynamics you were concerned with depends
primarily on the relative increase over the uncongested RTT that the queuing
delay represents.

For example, if you do your experiment on a LAN path that has an uncongested
RTT of 500 microseconds, a 64 packet queue may represent an increase over
the uncongested RTT by a factor of 30, but the same 64 packet queue on a
path with a 70 millisecond uncongested RTT may represent an increase over
the uncongested RTT by a factor of of 1.2.  A factor of 30 increase in the
RTT has a significant effect on the behaviour of TCP, but a factor of 1.2
may not matter much at all.

I think we agree that it is possible to design a policer which closely
emulates the queuing discipline of a real (shaped) queue with the exception
of the queuing delay that a shaped queue would introduce, and if this is
true the question of whether a policer might be a useful approximation to
a shaped queue reduces to an evaluation of how much that delay matters.
I believe your result, that the delay makes a significant difference to
loss rates if not goodput, is exactly right for even well-designed policers
on very low (uncongested) delay paths.  All I'm really pointing out is that
if the experiments were done on very low RTT paths then you were exercising
an extreme corner case whose results might not closely match what you'd see
with traffic more closely representing a wide area Internet mix where
uncongested RTTs tend to be much longer on average.

This should not be taken to imply that I'm a fan of policers, by the way.
I'm only trying to match your result with my understanding of what factors
should matter in what circumstances.

Dennis Ferguson