3.1 Experimental Setup

Figure 5: Experiment setting

Table 1: Compression speed and compression ratio of gzip with 16KB compression buffer

	Compression Speed		Compression Ratio
	Average (MBps)	Std. Dev.(MBps)	Average	Std. Dev.
TXT	0.84569	0.10470	4.0676	1.2047
PS	0.77839	0.18261	3.8816	2.5839
Binaray	0.65497	0.04605	2.0746	0.2169
PDF	0.87562	0.07125	1.1856	0.03035
JPG	0.83335	0.01509	1.0075	0.0099

Experiments are carried out on our local lab testbed. Fig. 5 shows our experiment configuration. There are four host machines, the two slashed squares are the application client and the application server, and the two grid squares are the competing client and the competing server. The two circles are routers, and the link between the routers is the bottleneck link in this simple network; its nominal transmission capacity is 10Mbps. The application client and the application server are executed on two Digital Unix machines. The competing client and the competing server are used to create the competing flow so as to change the available bandwidth on the link. The application uses a single TCP connection to transfer data. While the traffic generated by the competing hosts (also called competing flow) uses UDP packets since it is more accurate to predict the available bandwidth when the background traffic is a UDP flow. A SNMP Collector is deployed to monitor the available bandwidth.

In the experiment, we focus on the comparisons of four pairs of parameters:

1. $predicted\_total\_time$ and $total\_time$
2. $predicted\_compr\_time$ and $compr\_time$
3. $predicted\_send\_time$ and $send\_time$
4. $predicted\_compr\_size$ and $compr\_size$

Here, $predicted\_total\_time$, $predicted\_compr\_time$, $predicted\_send\_time$ and $predicted\_compr\_size$ are the values computed according to formulas (2) - (5). $Total\_time$, $compr\_time$, $send\_time$ and $compr\_size$ are the corresponding measured values in our experiment.