Understanding Total QPS from named stats

[JINMEI Tatuya / 神明達哉]

At Tue, 25 May 2010 07:48:48 -0700 (PDT),
ivan jr sy <ivan_jr at yahoo.com> wrote:

> As a follow up question, the stats "queries resulted in successful
> answer" - does this counter only cover queries that were answered
> with DNS data?

See the BIND 9 ARM
(http://www.isc.org/files/arm96.html#statistics_counters).
To quote the relevant part:

  Queries resulted in a successful answer.  This means the query which
  returns a NOERROR response with at least one answer RR.

> how about DNS queries that where the responded with SERVFAIL,
> NXDOMAIN, timed-out due to delegation, dropped, or non-successful
> answers - should those be included to the Total QPS as well?

There's no simple answer to this question.  If you use a relatively
naive tool like queryperf, it measures qps based on any responses
regardless of the RCODE.  But if the server returns a response with
SERVFAIL because it's too busy to handle all incoming queries (rather
than, e.g. it was due to a bogus response from an external
authoritative server), you wouldn't like to include those responses to
measure the effective performance of the server.

On the other hand, if a query results in "dropped" or "duplicate", no
response is sent to the client, so a tool like queryperf would
consider this a "lost query" (although "duplicate" isn't likely to
happen with queryperf).  However, it may not make sense to exclude
"dropped" queries because it doesn't mean the server loses the query
with its performance problem.

So, again, there's no simple answer.  But, in general, I think you can
get a reasonable measurement of performance even if you include (or
incorrectly exclude) these questionable cases.

---
JINMEI, Tatuya
Internet Systems Consortium, Inc.