next up previous contents
Next: Bibliography Up: A DNS-based Client Redirector Previous: Micro-Benchmarks of the AS-path   Contents


Conclusion

This thesis presents the client redirection problem and the requirements that should be met by a good solution. These requirements include: transparency, scalability, maintainability, and efficiency. Among three popular redirection mechanisms (HTTP redirection, TCP handoff, and DNS redirection), we conclude that a DNS-based redirection mechanism meets the four above requirements best.

Our implementation, called NetAirt, proves that with little effort Apache can be made to service DNS packets, no matter which transport-layer protocol is used: UDP or TCP. Moreover, our module can easily be integrated with an external replication-managing component, designed either as a separate Apache module (like Globule, another project of ours), or as a completely independent program.

We have paid attention to the development of a reasonable set of redirection policies: static name-to-address mapping, round-robin replica selection, and the AS-path length policy. We have incorporated them into NetAirt, thus demonstrating its flexibility, and allowing users to switch between them. We have described in detail how the most advanced one, the AS-path length policy, can be implemented. Firstly, BGP routing tables can be exploited to associate an ASN with any IP address, as well as to build a map of the ASes. This map allows to define the AS-path length metric, which is the base of the AS-path length redirection policy.

We have also conveyed two experiments. The first of them investigates the following three issues: the impact of the transport protocol on the Round Trip Time (RTT) of a DNS packet, the overhead introduced by running the round-robin address selection instead of the static name-to-address mapping, and the correspondence between the number of IP addresses returned in a single DNS response and the RTT. The results prove that sending DNS queries over TCP is almost 80% slower than in case of UDP. On the other hand, the overhead introduced by the round-robin policy, and the cost of responding with additional IP addresses turn out to be negligible.

The second experiment evaluates the overhead due to the AS-path length policy. The analysis shows that the majority of Autonomous Systems is located at distance from 3 to 5 from a typical client. We have also measured the time overhead generated by the AS-path length policy when it searches for a closest replica. The results show that exploring the entire map of the Internet that we use for this purpose takes no longer than 3.1 milliseconds, and on average only 0.64 milliseconds. We find this time small compared to the overall RTT for a typical DNS request sent over a wide-area network. Thus, considering the beneficial impact of running the AS-path length policy, we believe that redirecting clients based on network distance calculations can be nearly as efficient as using other, less sophisticated policies.

NetAirt will soon be released for public use. We hope that it can contribute to the development of worldwide-distributed services, and thus help sustaining the continuous growth of the Internet.


next up previous contents
Next: Bibliography Up: A DNS-based Client Redirector Previous: Micro-Benchmarks of the AS-path   Contents
root 2002-08-27