4.3 Experimental Results

Figure 14: Experiment about improved model

Table 2: Estimate the Available Bandwidth

competing BW (Mbps)	64K data transfer time (s)	estimated available BW (Mbps)
0	0.115	4.5
6	0.155	3.4

Fig. 14(a) and (b) show the experimental results without and with UDP competing flow, respectively. The dashed line is computed using formula (6). In order to calculate the theoretical value for $T$, we need to know $t_c$, $t_b$, and $t_s$. The former two can be measured directly in the experiment. $t_s$ is chosen manually from the trace data. We think the measured network transmission time as $t_s$ when $t_c$ is small enough, because the computation part will be completely covered by network transmission. This method of computing $t_s$ is not 100% accurate. The error, we think, leads to the difference between measured value (triangle points) and predicted value (dashed line) in Fig. 14(a). Comparing Fig. 14 with Fig. 13, we can see that the experimental result follows our execution model very well, which we believe confirms the improved model in Fig. 12.

Fig. 14 also provides a way to estimate the real data transfer time $t_s$ (i.e., the turning point between segment (1) and (2) in Fig. 13), which can be further used for available bandwidth computation. For example, fitting Fig. 14 into Fig. 13, we can get the data as in Table 2.

Unfortunately, the estimated values does not make much sense. We think it is due to our assumption that socket send() ($t_s$) starts sending data at the same time as that of buffer copying ($t_b$), which may not be true in reality. The real data transfer time should be smaller than $t_s$, but the exact difference is difficult to measure.