Re: TCP (and SCTP) sucks on high speed networks

[Mark Allman <mallman@grc.nasa.gov>]

> Once new ACKs start coming back, you'll increase the congestion
> window by a segment size for each ACK.

No.  You want to increase cwnd by something on the order of 1/cwnd
for each ACK that arrives during congestion avoidance.  Increasing
the cwnd by a segment for each ACK results in exponential cwnd
growth (ala slow start) rather than linear growth.  See RFC 2581.

(BTW, If there really are implementations that do what you describe,
I'd be quite interested in an example trace!)

allman


---
http://roland.grc.nasa.gov/~mallman/