> So you have to keep sending it data in order for it not to restart?
No, this is only one of the conditions. The other half is the last time a
segment is received. If the interval between the send time and the last
received time is greater than 1 RTO, and the data to be sent is new data,
restart slow start. Also note that snd_cwnd is reset to 1 MSS at send
time. As long as the server does not send anything during the time the client
is parsing the page, snd_cwnd won't be changed. Then a new request comes
in, the server responds by sending new data. It does the check and since a
segment is just received, the idle time check fails and TCP will not restart
slow start.
> PS - there's debate on whether this is the correct implementation of the
> intent of the slow-start restart anyway. The idea is that if data isn't
> ackd for a while, you _can_ burst into the net at the line rate, and
> you need to avoid that.
Yes, the implementation may not refect what it is intended. But I guess
the question now is what the behaviour of current implementations is.
....
> At best, the window slams inbetween each user request. Which
> is still a loss.
>
> What am I missing?
I guess it is the fact that window is changed only at send time. So the
window will not shut down between requests. This is an implementation
artifact, btw.
K. Poon
kcpoon@eng.sun.com
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:29 EST