Our goal in this paper was to quantitatively evaluate a variety of practical mechanisms and policies for realizing performance benefits from ISP multihoming. We focused on the scenario of multihomed enterprises that wish to leverage multiple providers to improve the response time performance for clients who download content from Internet Web servers. Using a real Linux-based route control implementation and an emulated wide-area network testbed, we experimentally evaluated several design alternatives. These included the performance of passive versus active monitoring schemes, sensitivity to various measurement sampling intervals, and techniques to manage performance information for a potentially very large set of target destinations.
Our evaluation shows that both active and passive measurement-based route control schemes offer significant performance benefits in practice, between 15% and 25%, when compared with using the single best-performing ISP provider. Our experiments also show that using historical performance to choose the best ISP link is not necessary - the most current measurement sample gives a good estimate. We showed that the performance penalty from collecting and managing performance data across various destinations is negligible.
Although our evaluation was done using an emulated wide-area network
and actual delay traces, it is valuable to deploy our implementation
in a multihomed site for further experimentation and evaluation. To
this end, we are planning to install our route control proxy device in
a commercial multihomed data center in which we can perform additional
experiments and uncover other wide-area effects.