An Investigation into Transport Protocols and Data Transport ...

5.2. TCP Tuning & Performance Improvements 92
work congestion on the reverse path as more acks are put into the network.
However, as quick acks are designed primarily to reduce the time required for
slow start to complete, the number of quick acks is calculated as n =
rwnd
2×MSS
where rwnd is the reciever advertised window and MSS is the maximum segment
size of the connection. This value, n, corresponds to the number of
packets that the TCP connection should take to exit slow start.
Appropriate Byte Counting (ABC) is a technique to eliminate the slow
growth of cwnd with delayed ack receivers and is outlined in [All03]. It gives
details of how TCP senders should behave when updating their congestion
windows through byte counting rather than the more common packet counting.
This has the benefit of improving TCP responsiveness without increasing
the amount of ack traffic (relative to the throughput) upon the reverse path.
Figure 5.6 shows the effect of running an ABC enabled sender when transferring
data to a normal delayed ack receiver. It can be seen that the dynamics
of using ABC is similar to that of a TCP sender sending data to a
receiver with delayed acks disabled.
It can also be observed that the queueing is apparent as the goodput
remains constant as the cwnd continues increasing until finally the buffer
overflows and fast retransmit and fast recovery takes place.
A potential consequence of implementing ABC is an increase in the burstiness
of the TCP traffic as two back-to-back data packets are sent for every
delayed ack [All99]. In order to reduce the effects of multiple packet bursts,
especially during slow start, a limit L, is introduced that limits the number
of full sized segments that the cwnd will advance by for each delayed or
stretched ack [All03]. The value of L= 2 means that a maximum of 2 new
segments are sent for every one delayed ack received.