Speaker
Description
Summary
Traditional wisdom is that connection setup will balloon latency for clients and overwhelm servers. We provide data to show that these assumptions are overblown---our model of end-to-end latency shows TLS to the recursive resolver is only about 21% slower, with UDP to the authoritative server. End-to-end latency is 90% slower with TLS to recursive and TCP to authoritative. Experiments behind these models show that after connection establishment, TCP and TLS latency is equivalent to UDP. Using diverse trace data we show that frequent connection reuse is possible (60--95% for stub and recursive resolvers, although half that for authoritative servers). With conservative timeouts (20 s at authoritative servers and 60 s elsewhere) we show that server memory requirements match current hardware: a large recursive resolver may have 25k active connections consuming about 9 GB of RAM. These results depend on specific design and implementation decisions---query pipelining, out-of-order responses, TLS connection resumption, and plausible timeouts.
We hope to solicit feedback from the OARC community about this work to understand design and operational concerns if T-DNS deployment was widespread. The work in the talk is Liang Zhu, Zi Hu, and John Heidemann (all of USC/ISI), Duane Wessels and Allison Mankin (both of Verisign), and Nikita Somaiya (USC/ISI).
A technical report describing the work is at
http://www.isi.edu/~johnh/PAPERS/Zhu14a.pdf
and the protocol changes are described as
http://datatracker.ietf.org/doc/draft-hzhwm-start-tls-for-dns/
