05.08.2014 Views

An Investigation into Transport Protocols and Data Transport ...

An Investigation into Transport Protocols and Data Transport ...

An Investigation into Transport Protocols and Data Transport ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

8.2. Test Calibration 156<br />

1200<br />

ScalableTCP 1<br />

ScalableTCP 2<br />

1200<br />

ScalableTCP 1<br />

ScalableTCP 2<br />

1000<br />

1000<br />

cwnd (packets)<br />

800<br />

600<br />

400<br />

cwnd (packets)<br />

800<br />

600<br />

400<br />

200<br />

200<br />

0<br />

100 200 300 400 500 600<br />

Time (seconds)<br />

0<br />

100 200 300 400 500 600<br />

Time (seconds)<br />

(a) Queuesize 10% BDP<br />

(b) Queuesize 80% BDP<br />

Figure 8.7: cwnd trace of ScalableTCP at various queue-size provisions under<br />

symmetric network conditions (100Mbit/sec, 82ms RTT, queue-size as shown).<br />

It is worth noting that even though Equation 5.6 specifies that 100% utilisation<br />

of St<strong>and</strong>ard TCP should only occur at 100% BDP queue provision,<br />

the calculations assume that only one flow exists in the system or that all<br />

flows are perfectly synchronised. As such, the experimental results demonstrate<br />

the lack of synchronisation between competing St<strong>and</strong>ard TCP flows<br />

which increases bottleneck utilisation.<br />

Fairness between two flows of the same New-TCP algorithm across the<br />

range of queue provisions is shown in Figure 8.6(b). It shows the difficulty<br />

in being able to achieve consistent fairness results from the New-TCP algorithms,<br />

with relatively large errors shown. However, the results do show<br />

that all of the algorithms, with the exception of ScalableTCP, are reasonably<br />

fair across the difficult queue-size provisions. The problems of achieving fairness<br />

with ScalableTCP is shown in Figure 8.7 where the long convergence<br />

times required for steady state are clear which (combined with the method<br />

of measuring fairness) contribute to low fairness.<br />

As all algorithms share the same SACK code base, the overhead gives an<br />

indication of aggressiveness of the underlying congestion control algorithm

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!