Implications and Discussion

The key findings from our evaluation are as follows:

1. The route control schemes we describe can significantly improve the performance of client transfers at a multihomed site, up to 25% in our experiments.

2. We show that relying on historical samples to monitor performance of ISPs (e.g., using EWMA) is not very useful, and sometimes may be detrimental to performance. The most current measurement sample is a very good estimator of near-term performance of an ISP link.

3. Both passive and active measurement-based schemes offer competitive performance, with the latter offering better performance for lighter client workloads. For the generic Web workloads we tested with, both active measurement implementations - SlidingWindow and FrequencyCounts - showed similar performance benefits.

4. The overhead introduced by aggressive performance sampling may slightly reduce the overall performance benefit of route control schemes. A sampling interval on a minutes timescale, e.g., 60s, seems to offer very good performance overall.

5. The overhead from measurements and frequent updates to the NAT table are negligible. Most of the performance penalties arise from the inaccuracies of the measurement and estimation techniques.

Figure 12: DNS responsiveness: This figure shows traffic volume over time just before and after a DNS change. The left graph (a) shows a 2-day period around the end of the event, while (b) focuses on a 2-hour period around the time of the DNS update.