Yet another Quadruple DNS?
nanog at ics-il.net
Sun Apr 1 17:05:15 UTC 2018
Unless you want optimum CDN performance, then your recursive servers belong pushed back in your network until there are no more diverse upstream\peer paths to choose from.
Yes, I know there's an extension to DNS to help remove this need, but until that's universally supported, you can't abandon the old way.
Intelligent Computing Solutions
----- Original Message -----
From: "Stephen Satchell" <list at satchell.net>
To: nanog at nanog.org
Sent: Sunday, April 1, 2018 11:22:10 AM
Subject: Re: Yet another Quadruple DNS?
On 04/01/2018 08:18 AM, Matt Hoppes wrote:
> Why not just implement recursive cache severs on end user routers?
> Why does an end user CPE need to query one or two specific DNS
Recursive lookups take bandwidth and wall time. The closer you can get
your recursive DNS server to the core of the internet, the faster the
lookups. This is particularly true of mobile and satellite.
Implementing large caches in that close-to-the-core DNS server can add
another benefit: lookups to common and popular endpoints, such as
Google, would require far less real time to resolve. As the traffic
tides change, the cache would change with it, so flash-in-the-pan sites
would be served from cache, and forgotten in time when said sites drift
back to obscurity.
(I wonder if the Internet Systems Consortium has considered adding a
cache-hit counter, or even a very coarse heat map (say, four 16-bit
counters cycled every five minutes), to DNS entries, to figure out the
best ones to drop? It would increase the complexity of BIND, but the
benefit for a BIND server serving a largish customer population could be
significant. If I were younger, I might try to model such a change.
Sort by hits, then by time-to-die. Drop the oldest 250 or so entries
when the cache is full. That way, the oldest least-used cache entries
ISPs provide to their customers DNS addresses to servers that,
allegedly, are closer to the core than the customers are. (Pipe dream,
I know; which is why so many ISPs have decided to specify 18.104.22.168 and
22.214.171.124, because Google is closer to the core than the ISP is.)
I've not personally measured the number of times a look-up could be
satisfied from a cache in a production environment; it's been 15 years
since I worked in such a job. It would be interesting to see if someone
has taken the time to gather those statistics and published them.
The main reason for *not* implementing recursion exclusively in CPE is
that a recursive lookup is a complex operation, and I have my doubts if
BIND or equivalent could be maintained properly in, say, a wireless
access point (router) -- how would you update a hints table?
More information about the NANOG