Speed of BIND vs. W2k DNS

[Jim Reid]

>>>>> "Jozef" == Jozef Skvarcek <jozef at photonfield.net> writes:

    Jozef> I am testing BIND v9.1.0b1 (SPARC, 450MHz) and W2k DNS
    Jozef> (PIII 500MHz).  The W2k seems to be 50% faster currently. I
    Jozef> am doing the tests on isulated network, the objective is to
    Jozef> find out the right platfrom especially for the round-robin
    Jozef> answers.

I'm confused. How can the speed of the name server have any bearing
on round-robin answers?

    Jozef> Does someone know how to tune up BIND to catch up
    Jozef> with W2k DNS?

There's not much you can do to make BIND answer queries any faster,
except from maybe eliminating long access control lists in allow-query
clauses or view{} statements. If the server has the answer, all that
happens is the incoming query is decoded, a hash lookup is done and a
reply encoded and sent back. There's not much scope for tuning that or
optimising that already optimised code path. There are no options in
named.conf that can directly influence this. You could optimise the
OS: like tune its TCP/IP stack and not run memory-hungry or CPU-hungry
applications on the box.

BIND9 is a bit slower than BIND8 because of the extra internal
overheads for threading. But this could be compensated by having the
BIND9 server run two or more threads concurrently on different CPUs.

Even so, the speed of an answer probably doesn't matter much. Your
name server will maybe get a few queries per second: perhaps 5-10 at
best. Who cares if the server turns round those queries in 100us or
200us? Does it *really* matter? If you had a root server that was
getting battered with thousands of queries a second, then yes, it
would matter. But for the typical name server, I seriously doubt it.
Functionality and ease of administration are probably more important
considerations than performance when picking a name server platform.
Packet throughput should only matter for extreme cases, like Internet
root servers.

    Jozef> Also, I noticed that the round-robin answers from the BIND
    Jozef> are generated by random, i.e. the first values are not
    Jozef> consecutive.  Is there any way to make it that way?

The rrset-order clause controls this. If the ordering of round-robin
is your overriding concern, shouldn't you be assessing which name
server implementations have the best features that meet your
requirements in that area?