[dns-operations] 86198 queries per second from BIND 9

John.Dickinson at nominet.org.uk John.Dickinson at nominet.org.uk
Tue May 30 09:57:29 UTC 2006


"Olaf M. Kolkman" <olaf at NLnetLabs.nl> wrote on 26/05/2006 19:51:10:

> >
> > I would be very interested to hear about other peoples experience 
> > of load
> > testing nameservers on this and other hardware. Any questions and
> > suggestions for other tests as also welcome.
> 
> 
> Hi John,
> 
> At NLnet Labs we  have a testlab called DISTEL that uses real-time 
> queries and fires them against an authoritative server[*]. The speed 
> at which these traces are replayed is varied to simulate higher query 
> rates. Since we capture the traces on the listener we are able to do 
> both performance as well as perform regression testing on the answers 
> (which was one of the main reasons to build the lab in the first place).

Thanks for the info. I am also planning on doing something similar using 
tcpreplay. Is that what you are using to replay the traffic?

> 
> The experiments using a root-zone and a root-trace towards a 1.6Ghz 
> Athlon single CPU on a 100Mb/s network is in the attached plot. In 
> the plot you can see the percentage of dropped packets against the 
> number of queries per second. The green line (echo) is the response 
> of a small server that just echos the incoming packet. Essentially 
> that is what the stack can do most efficiently.
> 
> One think I would like to know about your measurements is the amount 
> of packages dropped according to queryperf. I am not sure if 
> queryperf will try to figure out the sweet spot where the answer rate 
> is maximized while the drops are still zero.

According to queryperf no queries were lost. I think that queryperf trys 
to find the sweet spot. You can force it to run faster but then it seems 
to drop queries. I will try to produce a plot similar to the one you gave 
using both queryperf and tcpreplay. 

> 
> Obviously I would be more than interested to see a test with NSD on 
> your hardware with queryperf or netperf besides we are more than 
> happy to test other implementations in our setup as well.

I will be doing this as soon as I can.

Thanks
John



More information about the dns-operations mailing list