Title: ECO-DNS: Expected Consistency Optimization for DNS

Speaker: Steve Matsumoto (Carnegie Mellon University)

Date: May 22nd, (Wed) 17:00~18:30

Location: 우정정보통신관 (Woojung Building), 208 lecture room, Korea University

Abstract:
The flexibility of the current Domain Name System (DNS) has been stretched to its limits to accommodate delivery networks (CDNs) and Dynamic DNS (DDNS). While the DNS has proven capable of extending to a variety of purposes, an important research question is whether DNS can be enhanced to provide improved properties.

An initial observation is that popularity of names and their associated name resolutions follows a power-law distribution, which suggests that highly popular entries could be handled differently from the heavy-tailed low popularity entries. Moreover, the current DNS architecture typically operates with only two levels, with DNS resolvers directly querying authoritative DNS servers, and Time-To-Live (TTL) offering the only mechanism currently used to enforce cache eviction.

To improve DNS, we observe that by taking popularity of entries into account we can achieve higher consistency at lower overhead and enable multi-level resolver hierarchies. In particular, we propose a new DNS consistency metric, Expected Aggregate Inconsistency (EAI), which incorporates a record's popularity and update frequency, to enable the tradeoff among several key properties: consistency, bandwidth overhead, latency, and server load. We design ECO-DNS, a lightweight system that leverages the information captured by EAI to dynamically select the optimal timeout parameter for a DNS record. Furthermore, ECO-DNS guarantees an upper bound on aggregate inconsistency in multi-level cache hierarchies, thus making them viable. We show that the deployment of ECO-DNS can simultaneously achieve better consistency, lower average query latency and lower overhead.

Bio:
CV can be accessed at http://stevematsumoto.net/cv