Modified slow-start:

For HSTCP, we have introduced a modified version of slow-start, to
avoid increasing the congestion window by thousands or tens of
thousands of packets in a single round-trip time.  Such a dramatic
increase in the congestion window is hard on the network, and also
is counter-productive for the TCP flow itself.  Thousands or tens
of thousands of packets can easily be dropped from a single window
of data, resulting ultimately in retransmit timeouts, and the TCP
connection continuing in congestion avoidance phase with a very
small congestion window.

Drawing from conversations with Tom Kelly, we have introduced a
modified slow-start to NS, to allow better startups for flows with
very high congestion windows.  A parameter "max_ssthresh_" has been
added to NS to specify the maximum congestion window for the initial
slow-start.  That is, for (cwnd_ < max_ssthresh_) during the initial
slow-start, cwnd_ is increased by one packet for every arriving
ACK.

For (cwnd_ > max_ssthresh_) during the initial slow-start, the
invariant is maintained that the congestion window is increased by
at most max_ssthresh_/2 per RTT.  This is done as follows:

Set K = int(cwnd/(0.5 maxthresh)).  The, for (cwnd_ > max_ssthresh_)
during the initial slow-start, the window is increased by 1/K for
each arriving ACK, instead of by 1 as in the standard slow-start.

To illustrate, with (cwnd_ == max_ssthresh_) the window is increased
by 1/2, with (cwnd_ == 1.5 max_ssthresh_) the window is increased
by 1/3, and so on.

With this method it takes: 
 log(max_ssthresh_) RTTs 
to reach a congestion window of max_ssthresh_, and it takes:
 log(max_ssthresh_) + (cwnd_ - max_ssthresh_)/(max_ssthresh_/2) RTTS
to reach a congestion window of cwnd_.

This is illustrated in: 
"./test-all-tcpHighspeed tcp1A"
and:
"./test-all-tcpHighspeed tcpHighspeed1"
in tcl/lib.