> Keith Owens wrote
> > What about fudging the TCP window size in ACK packets we send back? If
> > you want to throttle a sender, set a couple of variables in our sk_buff
>
> RFC1122 says you are not suppose to shrink your window on people - and
> indeed it does break some crappy stacks.
Agreed, but I meant to close the window (move left edge right) rather than
shrink the window (move right edge left). The idea is to pretend that the
receiving application is overloaded (data still in the TCP receive buffer)
so the sender slows down.
Having though about it some more, forget the minimum window size, let the
window close and use a timer to reopen it later. If the window does not
open far enough this algorithm will interact badly with the silly window
syndrome code.
Another possibility is to delay telling the receiving application that it
has data to read. Instead of fudging ACK data, let the data sit in the
receive buffer until a timer has expired then pass it to the application.
No need to play with IP packets, just slow down the receiving program.
This technique will only work when the traffic is for a single machine, it
cannot be applied to routers without affecting RTT.