[e2e] TCP, retransmissions, and ISPs with byte caps

[David P. Reed]

At 12:40 AM 6/19/2002 +1000, grenville armitage wrote:

>David P. Reed wrote:
>         [..]
> > Note that packet losses also cause the TCP to send more slowly, which
> > reduces the cost to the user in either case:
>
>I wonder how true that would be for the "think, think, point,
>click, <generate burst of TCP traffic>, repeat" crowd?
>
>The slow down due to retransmissions is probably swamped
>by the duration of each "think, think" phase, consequently
>I suspect there'd be little noticable reduction in
>the user's attempted consumption of network content.

The user does have to wait for a response before the "repeat", and that 
response gets pretty long when retransmissions are interfering with 
slowstart on that flurry of new TCP connections.  I'd be interested in 
experimental results, but it seems that the "HTTP lusers" observe a very 
nonlinear impact on response time as a function of packet drop 
rate.  (compared with those who spend most time sitting in front of big 
FTP's with stopwatches doing performance measurements like us real network 
users ;-) ).