BIND 4 / 8 / 9 performance

[Kevin Darcy]

Jim Reid wrote:

> >>>>> "John" == John Jetmore <jetmore at networkwcs.com> writes:
>
>     John> ... data on server load and hardware snipped ...
>
>     John> As these values continue to rise, the toll named is taking
>     John> on the machines begins to rise also.  We are very aware of
>     John> the need to upgrade to at least BIND 8, but we need some
>     John> data to convince management to release manpower to do this.
>
> The load on your servers is almost insignificant, so it's unlikely
> you'll see any noticeable performance improvements by upgrading. You
> might find that the logging and control hooks in BIND[89] are much
> more useful for efficiently managing and configuring the name
> server. Operating a BIND4 server is very clumsy and crude compared
> with BIND[89]. How can you live without features like incremental zone
> reload and refresh or (for BIND8) reconfiguring the server without
> having to restart it?
>
> I'd try telling management that they simply must upgrade from long-dead
> code. BIND4 has been dead for years and BIND8 is in the departure
> lounge. That - not quantifying the references to performance
> improvements in the release notes - should be justification enough for
> upgrading. I suspect those references are likely to be corner cases
> rather than generic ones that would make a substantial difference to
> throughout or CPU utilization. Those major optimizations should/would
> probably have been done years ago.
>
>     John> Specifically, does anyone have numbers on the performance
>     John> increase realized from a transition from 4 to 8, or from 8
>     John> to 9?  Numbers in terms of percentage of porocessor used
>     John> would be most useful, but any numbers could be leveraged to
>     John> give perspective.
>
> BIND9 is quite a bit slower than BIND8 because of the internal
> threading overheads. The performance hit should only matter to very
> busy servers - root server territory of 1000+ queries/second - or
> servers with tens of thousands of zones. How much slower BIND9 is
> depends on the OS. OTOH, because BIND9 is multi threaded it does mean
> that different threads can execute in parallel on a multi-processor.

Note that the pthreads implementation is broken on Solaris 2.6. So the
original poster would need to upgrade before being able to exploit that
feature of BIND 9.


- Kevin