Summary

  • RFC 6298 lowered TCP's recommended initial retransmission timeout from three seconds to one, based on historical observations that most measured RTTs were below one second.
  • If a sub-three-second timer expires while waiting for a SYN acknowledgment, the standard requires returning to three seconds before data transmission begins.

A timer before evidence

TCP normally learns a retransmission timeout from observed round-trip times. At the opening of a connection, however, no such observation exists. If the first SYN or its reply is lost, the sender cannot consult a path-specific estimate; it must act on an initial value chosen in advance.

That value controls a real trade-off. A timer set too high makes a genuine opening loss unnecessarily expensive. A timer set too low retransmits on a path that is merely slow, injecting a duplicate SYN and disturbing the connection before it has measured anything. The initial RTO is therefore not just a performance constant. It is a policy for acting under missing evidence.

RFC 2988 had used three seconds. RFC 6298 changed the recommended initial setting to one second, while allowing an implementation to remain more conservative with any value greater than one second. The distinction matters: one second was a SHOULD, not an unconditional MUST.

What the measurements supported

Appendix A did not argue from speed as an abstraction. It recorded the evidence available at the time. A large-scale analysis found that more than 97.5 percent of observed connection RTTs were below one second. It also observed handshake retransmission at roughly two percent, making faster recovery relevant to a non-trivial group of connections.

The same evidence preserved the adverse case. Roughly 2.5 percent of the observed connections had RTTs above one second. On such a long-RTT path, a one-second starting timer could cause an unnecessary SYN retransmission during establishment. Later trace analysis found the estimated spurious-retransmission share below 0.1 percent in most datasets, but about 1.1 percent in a Dartmouth wireless trace.

The benefit estimates were likewise bounded to those traces. Among connections that retransmitted a SYN, the share expected to gain at least ten percent ranged from 43 to 87 percent; the share expected to gain at least fifty percent ranged from 17 to 73 percent. These were inputs to a 2011 standards decision, not measurements of today's Internet.

The retreat built into the advance

RFC 6298 paired the shorter starting point with two restraints. First, timeout still triggered exponential backoff: the host had to double its RTO. Second, if a timer below three seconds expired while the sender awaited acknowledgment of a SYN, the RTO had to be reinitialized to three seconds when the handshake completed and data transmission began.

That reset contained the inference the sender was allowed to make. A SYN timeout could mean loss, but it could also reveal that one second was too short for this path. The standard accepted the extra SYN and, under RFC 5681, an initial congestion window limited to one segment. It then prevented the unmeasured one-second assumption from governing the data phase unchanged.

The fallback did not make every long path easy. RFC 6298 explicitly noted that a path with an RTT above three seconds would still encounter difficulty. What it did was bound the new risk introduced by lowering the default.

Sources