Re: placement of secondary name servers for .uk.

["Brian W. Spolarich" <briansp@walid.com>]

On Tue, 21 Nov 2000, Peter Gradwell wrote:

| I'm also speaking to the people at CAIDA about their research.
| Currently it seems that the US, Asia and Europe require equal amounts
| of "attention". Their work is excellent and highly commended. But you
| knew that already!

  Peter, let us know what you hear from CAIDA.  Their work is indeed
excellent (I used to work with some of these folks at ANS).

| One final question. Randy posted a traceroute to ns0.ja.net. However,
| for .uk. we have ns.uu.net and ns.eu.net, which might be "closer" to
| randy (for this example). Which server would Randy prefer to use? Do
| resolvers take the "closest?". AFAIK, from reading "DNS & BIND" it's
| all rather random. Thoughts? Does this randomness affect the placement
| of the servers?

  I recall BIND having a clever set of algorithms that would enable named
to return records with network latency in mind.  If multiple NS records
were found the caching nameserver would use a nice mechansim based on
network round-trip time and number of times that record was used to
determine which record to send.  The result was that servers that were
closer were queried more often than those that were far away.  A colleage
and I did some modeling of this behaviour a few years ago and wrote a
short whitepaper on it (which I can't find).  As I recall, the behaviour
was satisfactory for reasonable load-balancing as long as the number of
records in the rrset was relatively small (<5).

  In BIND8, bin/named/ns_resp.c seems to contain this code.  BIND9 is
reorganized enough that I couldn't find it quickly.

  Is this documented somewhere?  I've always thought it was cool but
didn't find much in the way of documentation on it in 1997, and haven't
looked since. :-)

  -bws