Measuring a round trip
A sender measures elapsed time between transmission and corresponding acknowledgment. Both directions and the feedback mechanism contribute, not just physical propagation between machines.
Samples vary with queuing, scheduling and acknowledgment timing. One low or high value does not characterize every later request or the time required for an application operation.
Estimating RTT and choosing timeouts
TCP’s RFC 6298 algorithm maintains both a smoothed estimate and variation before deriving a retransmission timeout. The newest sample is not used as an unqualified deadline.
Too short a timeout causes spurious retransmission of data still in flight; too long a timeout delays recovery. The estimator balances responsiveness with uncertainty in the observations.
References: [1]
Ambiguity from retransmission and direction
After retransmission, an acknowledgment may not reveal which transmission it answers. Karn’s algorithm avoids ambiguous samples; TCP timestamps can disambiguate under the applicable rules.
A round trip combines both directions, so halving it is not a precise one-way measurement. Explaining page speed additionally requires identifying naming, connection, application processing and transfer stages.