>
> One technical correction on TCP slow-start. Although all of the
> theory says the window doubles each RTT, it actually only gets
> multiplied by a factor of 1.5 each RTT. This is "footnote" is often
> passed over when talking about slowstart performance. The reason the
> factor is 1.5 instead of 2 is because the delayed ACK algorithm only
> sends out one ACK for each two packets received. Slowstart increases
> the window by one MSS per ACK, resulting in a factor of 1.5 increase
> with each RTT.
>
> --Jamshid
>
For completeness, it's also often overlooked that the
delayed ACK algorithm has a similar effect on window
growth during congestion avoidance. It's not as dramatic
as in slow-start, but still significant within a long-delay
environment.
Here, the window grows by (1/window) for each ACK received
(RFC 2001 eliminates the additional fraction). If an ACK
were sent for every segment, the window would grow by
up to one segment per RTT. However, since the delayed
ACK algorithm will try and send an ACK for every second
full-sized segment received, the typical growth rate
will be more like .5 segment/RTT.
For a geosynchronous satellite hop, this means I get
to grow my window by about 1 segment/second during
congestion avoidance. Ugh.
Eric
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:29 EST