Re: djbdns-1.05-epoll + speedup patch

[Pete Ehlke <lists-djbdns@rfc822.net>]

On Mon Mar 05, 2007 at 10:13:16 -0600, Matthew R. Dempsky wrote:
>On Mon, Mar 05, 2007 at 04:47:46PM +0200, Sami Farin wrote:
>> Not that it matters much, but version 20070305T031624Z
>> already compares >= 0.  It also does one useless loop
>> if n == 1 before for-loop unless I check for it.  But dns_sortip
>> is called by dnscache only with n == 64 (16).
>
>I'm refering to the fact that the former code uses the Knuth shuffle
>algorithm to generate a uniform distribution over all possible
>permutations.  The new code does not.
>
The important operational question is not the efficiency or correctness
of randomization in this area. The important question is the effect that
randomization and retry strategies have when faced with unreliable
networks and slow servers.

Compare dnscache with BIND 9 ( which has a more efficient retry and
selection mechanism ) in
http://dns.measurement-factory.com/writings/wessels-pam2004-paper.pdf

-P.