Chapter 20

p. 277

The effects also depend on the relative synchronisation of the 200ms (and 500ms) timers on the two machines.

p. 285

RFC 2581 recommends that slow start should also be used if no data has been sent for more than one RTO, otherwise re-activating an idle connection can let it flood the network.

p. 285

These algorithms look relatively simple, but in fact RFC 2525 (section 2.3) states that cwnd is occasionally not set: when the second SYN of the three-way hand-shake does not specify an MSS (hence the MSS defaults to 536), some implementations do not set cwnd to the (implied) MSS.

p. 289

The product for an ATM-155 link across the Atlantic, again assuming 60ms RTT, would be

$$\underbrace{\frac{155\rm M}{8}}_{\mbox{bytes/sec}} \times\un... ...imes \underbrace{\frac{6}{100}}_{\mbox{60ms}}\approx1{\rm MB}.$$

This assumes ``single-user'' use: unlikely across the fat pipe during the day. However, a high-bandwidth load would require these sorts of buffer sizes to be effective.

There are also problems with wireless networks: RFC 2757 notes that `` A 3rd Generation wireless service offering 2 Mbps with 200-millisecond latency requires a 50 KB buffer.''

p. 289

There is a good animation of the effect of the window size to be found at http://cable-dsl.home.att.net/rwinanim.htm. The whole site is worth looking at.

This formula assumes no errors: if there is a packet loss rate of $p$, then the ``Reno'' TCP implementation is also limited to maximum sending rate of

$$\frac1{ RTT*\sqrt{2p/3} + RTO*(3\sqrt{3p/8})*p*(1 + 32p^2)}$$

segments per second ([9] and RFC 3155).