In the experiment, the application depends on Remos to provide available bandwidth information. Clearly, the accuracy of the network performance information can affect the final prediction accuracy significantly. We first use a simple experiment to get an idea of the accuracy of the Remos information.
Fig. 6 shows the experimental results. In the experiment, we keep monitoring the available bandwidth of the network path between two machines on our testbed, which generally carries no traffic. The measured value is plotted using the solid line in the figure. At around 6 second (competing flow starts in Fig. 6), a 6Mbps UDP flow is added to generate the competing traffic. We can see, with the introduction of competing traffic, Remos starts reporting reduced available bandwidth. After around 4 seconds, the reported value gets stable, with the final value around 3.4Mbps. We think it is reasonable since the link capacity is set as 10Mbps.
The reason that we use UDP traffic instead of TCP traffic as the competing flow is that UDP traffic sender side throughput is not affected by other traffic on the same path. In our experiment, we need different rates of competing traffic, so we can easily control its traffic throughput.
