An Investigation into Transport Protocols and Data Transport ...

9.1. Transfer Tests Across Dedicated Private Wide Area Networks 196
observed slight increase in the goodput in Figure 9.1.
Conversely, the decrease factor of H-TCP is programmed such that cwnd ←
β × cwnd and therefore, the calculation upon loss actually results in a lower
cwnd and goodput. The effect of this is most notable on the MB-NG network.
The observed higher goodput achieved with Standard TCP on the DataTAG
link is attributed to the method by which packets are dropped by the receiver
and the limited duration of the test. As the counting of packets also include
those under slow start, a large initial value of cwnd occurs due to the exponential
increase of slow start. Therefore, with lower loss rates, a high average
for the goodput is achieved due to the goodput bias as a result of slow start.
This effect is less apparent for the New-TCP algorithms due to the fast(er)
increases in α which ensures a fair number of congestion epochs and therefore
stable operation of congestion avoidance. A solution to this bias would be
to run the tests of Standard TCP for a prolonged period of time. However,
due to to practical reasons this was not feasible.
On both networks, as lower loss rates result in goodputs reaching 1Gb/sec,
the New-TCP flows are no longer limited by the dynamics of cwnd, but by
physical limitations of the NIC which causes saturation of goodput at the
physical line rate.
9.1.3 Results: CBR Background Traffic with 1 New-
TCP Flow
Figure 9.3 shows the goodput achieved by a single TCP flow competing
against various loads of CBR traffic along both MB-NG and DataTAG. Under
MB-NG, all of the algorithms achieve roughly the same throughput; with H-
TCP actually being able to get a little more bandwidth than all of the other