Re: [DNSOP] interop problems with getaddrinfo() address selection

[Edward Lewis <[e-mail]>]

I stumbled across this:

http://www.nanog.org/mtg-0310/pdf/wessels.pdf

Look for the (unnumbered) slide "Distribution of Queries to 
Nameservers" and the ones following it.

At 14:47 -0800 12/4/07, Edward Lewis wrote:
>I knew this came up before:
>
>http://lists.oarci.net/pipermail/dns-operations/2007-September/002021.html
>
>Note that the RFC's text says "implementation" and not "application" 
>or "protocol."  There isn't a standard DNS algorithm for choosing 
>the server to be queried, the algorithms are up to the implementers 
>(as the algorithm isn't an inter-operability issue) of DNS instances.
>
>I looked for a bit but can't find an excellent paper in which a 
>research stood up different name servers with no connectivity to 
>others and then measured the number of (unanswered) queries sent to 
>each of the root servers.  One server implementation sent them all 
>to A, one sent 1/13th to each of the 13 servers, others did other 
>distributions.  This mostly demonstrated the spectrum of strategies.
>
>Note that DNS has reasons not to follow what's in 3484 - namely that 
>DNS servers are supposed to be far flung, finding a local one is not 
>going to be advantageous (as opposed to other application data 
>sources).
>
>At 21:44 +0000 12/4/07, Tony Finch wrote:
>>Rule 9 of RFC 3484 specifies that the IP addresses (v4 and v6) returned by
>>getaddrinfo() should be sorted according to the size of their common
>>prefix with the local host's chosen source IP address. This defeats DNS
>>round robin load balancing which has led to some severe operational
>>problems. DNS round robin needs to be documented in an RFC, and RFC 3484
>>should be updated with rule 9 deleted or substantially modified for
>>compatibility with DNS round robin.
>>
>>Ian Jackson has written a thorough description of the problem from the
>>perspective of a proposal that Debian's version of glibc should have the
>>troublesome sorting behaviour removed:
>>
>>http://lists.debian.org/debian-ctte/2007/11/msg00029.html
>>
>>FreeBSD's getaddrinfo() only implements Rule 9 if the kernel's address
>>selection policy is configured, which is not done by default so DNS round
>>robin is passed through.
>>
>>Tony.
>>--
>>f.a.n.finch  <[e-mail]>  http://dotat.at/
>>FISHER GERMAN BIGHT HUMBER THAMES DOVER WIGHT PORTLAND SOUTHWEST 5 TO 7,
>>INCREASING GALE 8 AT TIMES. MAINLY ROUGH. RAIN THEN SHOWERS. MODERATE OR POOR
>>BECOMING GOOD.
>>
>>_______________________________________________
>>DNSOP mailing list
>>[e-mail]
>>https://www1.ietf.org/mailman/listinfo/dnsop
>
>--
>-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
>Edward Lewis                                                [phone]
>NeuStar
>
>Think glocally.  Act confused.
>
>_______________________________________________
>DNSOP mailing list
>[e-mail]
>https://www1.ietf.org/mailman/listinfo/dnsop

-- 
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Edward Lewis                                                [phone]
NeuStar

Think glocally.  Act confused.

_______________________________________________
DNSOP mailing list
[e-mail]
https://www1.ietf.org/mailman/listinfo/dnsop