cathya at isc.org
Thu Feb 9 10:48:05 UTC 2017
On 09/02/2017 08:23, 이원영(IP망설계팀) wrote:
> We have DNS servers (Venders are Dell & HP & Fujitsu)
> Recently, we upgraded bind version from 9.9.8-P3 to 9.9.9-P4.
> We find out server query performance droped from our test.(Used DNS
> perf. program)
> Especially, dell server droped minus 61%.
> I wonder why server performance droped.
> And why dell server droped high percentage than HP servers.
> ㅇServer performance test result (Based Maximum Unique Query/Sec)
> - server name (CPU spec): Bind 9.9.8-P3 -> Bind 9.9.9-P4
> - HP DL360 G5 (3.0Ghz dual*2) : 10,000pps -> 6,896pps (-31%)
> - HP DL360 G7 (2.17Ghz 4core*1) : 14,000pps -> 13,938pps
> - HP DL380 G7 (2.4Ghz 4core*2): 14,000pps -> 14,262pps
> - HP DL360 G8 (2.3Ghz 6core*1): 24,000pps -> 21,318pps (-11%)
> - Dell R610 (2.27Ghz 4core*2): 17,000pps-> 6,557pps (-61%)
> Thank you,
> Wonyoung, Lee
This is surprising. Can you give us the details of your set-up for
Then some sanity-checking questions to make sure that the anomalies are
not due to other differences.
1. Not directly performance-related, but I would anyway recommend
9.9.9-P6 over 9.9.9-P4 - you'll be getting several security fixes, some
of them more important than others.
2. You didn't mention which OS and kernel you're running - please
provide those for each box.
3. Make sure that you have a threaded build. If this didn't happen
automatically, then when building BIND, use --enable-threads with
4. Check that all the versions that you're comparing are running with
the same values of -n (number of worker threads).
(We changed the default for -n between versions, so if you're defaulting
you may not be comparing like with like - although I admit I have not
minutely checked what changes we made when. The start up logs will tell
you how many worker threads named has started with.)
5. Check what value of -U (number of UDP listeners) named has defaulted
to using with the 9.9.9 builds. You may need to adjust your startup
Having multiple UDP listeners is good for avoiding an i/o throughput
bottle-neck in many situations, but if you already had the workaround of
configuring multiple virtual interfaces on the same NIC, could have the
opposite effect by creating more UDP listeners than you actually need.
Tuning both -n and -U to optimise performance for your specific set up
falls into a bit of a 'black arts and try it and see' category - it's
worth some experimentation because the defaults are not correct for
everyone (and what we set as the defaults are therefore a necessary
More information about the bind-workers