An improved DNS server selection algorithm for faster lookups

The success of the Internet owes immensely to the hierarchical architecture of Domain Name System (DNS) that maps domain names to IP addresses. DNS servers that implement this mapping and provide lookup services rely heavily on caching of IP addresses for enhancing the lookup performance. It turns out that the caching of IP addresses of other DNS servers is more beneficial than the caching of IP addresses of end hosts. Since several domains have more than one DNS server, this results in the following important issue that arises frequently: given the cached entries, how do we choose the optimal DNS server for forwarding a query when there are more than one DNS servers handling that particular domain? We refer to this as the DNS server selection problem. In this paper, we investigate the use of auto-regression models for estimating the server response times in the DNS server selection algorithm. As compared to existing implementations such as the one used in BIND, our approach is more resilient and adaptable to fluctuations in network delays because all the parameters are determined based on observed response times, instead of being hard-coded. Experiments with real-query logs indicate that our techniques can significantly enhance the BIND implementation and improve the lookup times by up to 40%.