[e2e] Which RTT to use ?

[Woojune Kim]

I just ran into the following question : Normally we say tcp needs a bw * RTT amount of buffering to perform optimally. This is also related to the receiver / sender windows. 
 
Now the question is, what should the RTT be ? Should it be the RTT measured for tcp packets and the segments that contain their ACK (which is the way the TCP RTT algorithm measures the RTT) or should it be RTT of a packet sent along the same path (eg. the packet's actual delay IF the receiving node had replied ASAP) ?
 
Doesn't the TCP RTT algorithm also end up including delayed acks, nagle algorithm effects, timer granularities (old UNIX implementations having 500ms ticks) etc. ? 
 
thanks
Woojune Kim