The graph was based on NLANR data, specifically the analysis of packet flows
at the FIX West. I have (indirectly) received feedback that mentioned that
the overhead in typical ISP network is much lower.
There is however a real problem with the DNS, one that is probably not paid
enough attention today. If you pick a valid DNS name at random, for example
by looking deep into the results of search engines such as Altavista, there
are currently more than 25% chances that "gethostbyname" will not return a
result in less than 2 seconds. I attribute this rather horrendous result to
two main causes. First, the root servers appear to be severely overloaded. I
refrain to run a measurement tool against these servers for fear of adding
to the overload, but when I did run a test I found that some servers failed
to complete as many as 40% of the transaction attempts thrown at them. The
second cause is the prevalent packet loss rate in the "public Internet"
(i.e. when you have to cross several inter-AS connections). This loss rate
hovers between 1% and 5%. The effect on UDP based protocols is obvious; when
you pile up several successive transactions, as the DNS does, the results
are cumulative...
Christian Huitema
-----Original Message-----
From: Craig Simon [mailto:cls(_at_)flywheel(_dot_)com]
Sent: Friday, May 26, 2000 1:25 PM
To: ietf(_at_)ietf(_dot_)org
Subject: Cite on DNS-related traffic.
I recall once seeing a graph shown by Christian Huitema
indicating that
DNS-related overhead accounted for about 5 to 10 percent of
Internet traffic.
Can anyone provide a link for this or equivalent documentation?
Thanks.
Craig Simon