[e2e] DNS

[Hari Balakrishnan]

We've been collecting client-side traffic traces of DNS traffic, jointly with 
the driving TCP connection workload, for a couple of years now, and a recent 
paper analyzing these data sets may be of interest to many on this list.  Our 
data spans many months of collection at the border router connecting MIT's LCS 
and AI Lab to the rest of the Internet, as well as a border router at KAIST in 
Korea.

The paper, "DNS Performance and the Effectiveness of Caching," 
by Jaeyeon Jung, Emil Sit, Hari Balakrishnan, and Robert Morris, is at 
http://nms.lcs.mit.edu/papers/dns-imw2001.html

Some of our non-obvious (to us!) findings and conclusions are: 

- About a quarter of all DNS lookups never get an answer. More than 50% of the 
DNS-related packets in the wide-area correspond to such lookups...

- The DNS retransmission protocol appeears to be overly persistent: while most 
successful answers are received in at most 2-3 retransmissions, the lack of an 
answer or response causes a much larger number of retransmissions and a 
corresponding number of DNS packets traverse the wide-area.

- Replacing the A-record TTL's for most (or all) Internet hosts to a value as 
small of 10 minutes is not likely to degrade the scalability of DNS in any 
noticeable way.

- The scalability of DNS has little to do with its hierarchical organization or 
good A-record caching. Most of the DNS name space is a flat, two-level 
structure.  A-record caching does not seem to add much more to the per-host or 
per-application caching done by end clients today.  Rather, the scalability 
derives from the good name space partitioning achieved by the cacheability of 
NS records, which avoid load on the root and top-level name servers.

The paper details the combination of trace analysis and trace-driven cache 
simulation that we used to arrive at these conclusions.