TCP Self-clocking
Allen B. Downey
Olin College
                                                                                
Although Jacobson described TCP self-clocking in 1988, subsequent
models of TCP have ignored it.  As a result, a lot of what we think we
know about TCP turns out to be wrong.  If there is enough buffer space
at the bottleneck link to support self-clocking, the following
statements are all false:
                                                                                
1) During TCP slow start, the congestion window grows exponentially.
 
2) In steady state, the congestion window grows linearly between
   drops.
 
3) When the congestion window exceeds the bandwidth-delay product,
   packets get dropped.
 
4) In steady state, TCP uses additive-increase, multiplicative-
   decrease congestion avoidance.
 
5) In steady state, throughput is proportional to 1/sqrt(p),
   where p is the drop rate.
                                                                                
6) When connections share a bottleneck, they get throughput
   proportional to 1/rtt, where rtt is the round trip time.
                                                                                
7) After slow start, TCP is slow to discover available capacity.
                                                                                
I will present a simple network model that allows us to correct
these statements.  I will also present measurements that show
that self-clocking is common in the Internet, and that when it
occurs, the predictions of the model are correct.