[bind10-dev] a couple thoughts about performance testing of servers

[Michael Graff]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

For the last 10 years, off and on, I've had to test servers for
performance.  Usually this was related to tracking down a bug in BIND 9
or to determine if a feature worked and did not slow things down.

Queryperf or resperf are decent enough tools, but not really perfect.
They can quite easily hit the server harder than it could possibly
respond, which is good only to determine the melting point.

We have talked in the past of how exactly one measures query
performance, and query loss.  While some queries would eventually
respond, how useful is one that takes 20 seconds?

I'd love to hear the technical details of exactly how you measured, what
you found in those measurements, and what you think we should do.
However, I don't want to get in the way of your UI work, so perhaps we
can discuss this stuff after that is complete?

- --Michael

On 2010-08-26 3:22 PM, Jerry Scharf wrote:
> Hi,
> 
> About 4 years ago I did a round of nameserver performance testing. I
> just wanted to add a couple thoughts from the sidelines based on this
> experience. Use them as they help you.
> 
> Expecting every query to be answered gives really low numbers for a
> server. A server will start to drop queries at as little as 10% of full
> query throughput. So we need to be clear what we are testing to. Having
> some marketing background, I always use the test that gives the highest
> number. :)
> 
> On the server I was testing, there was a predictable point at about
> 1-1.5% query drop rate where the total throughput stopped going up and
> most additional queries were dropped. I called that the peak performance
> for the test. I also had to do some nasty things to decide exactly how
> fast to send queries at. I had a tool (I have no code now) that would
> send bursts of queries at intervals. This allowed for much higher query
> output rates than timing between each query. I had a pstat logging to a
> file for the whole time of the run on both machines, it helped to
> understand many the things as I tuned up the tests.
> 
> I assume we are testing out of cache for the primary numbers. It's the
> only fair way to compare the apples and oranges of BIND 9 and 10 (or any
> other server for that matter.)
> 
> hope that helps,
> jerry
> 
> _______________________________________________
> bind10-dev mailing list
> bind10-dev at lists.isc.org
> https://lists.isc.org/mailman/listinfo/bind10-dev

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.8 (Darwin)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAkx24NMACgkQ+NNi0s9NRJ0KAgCeNOp3EKmPY3N7l0Ld7ubnxYJb
/L0AnRei2R6SHCwerZsv8ObvFIgAmx11
=UmkD
-----END PGP SIGNATURE-----