Employing History to Estimate Performance

Figure 9: Impact of history: The performance achieved by relying on historical samples to varying degrees. These results are for the passive measurement-based strategy with a sampling interval of 30s.

Figure 9 plots the performance of the passive measurement scheme for three different values of the parameter $\alpha$. These correspond to assigning 80%, 50% and 20% weight to the current measurement sample and the remaining weight to the past samples. Although we only show results for a sampling interval of 30s, the performance from other interval sizes are similar. The figure also plots the performance when no history is employed ($\alpha = 0$) and the performance from using ISP 3 alone. Notice that the performance from employing history is uniformly inferior in all situations, relative to employing no history. In fact, historical samples only serve to bring performance close to that from using the single best provider. These results show that the best way to estimate provider performance is to just use the current performance sample as an estimate of near-term performance.