Re: [wg-c] Initial Numbers

[Karl Auerbach <karl@CaveBear.com>]

> > Dr. Huitema's paper, which was based on
> > empirical research, reported that between 13%
> > and 16% of total web latency was due to DNS
> > latency, and that there is no reliable, empirical
> > evidence on the elasticity of DNS latency in
> > respect of enlarging the root.

If one looks at the pattern of web page access (as opposed to other forms
of net activity) one sees a strong pattern of re-resolution of the same
DNS names over and over again from embedded URL's pointing to the same
hosts.  Thus, if there is a delay on an initial fetch of a web page from a
newly visited site, a delay in name resolution will be multiplied by the
number of URL's (embedded gifs and the like) that the page includes from
that same site.  On subsequent fetches, the gethostbyname() lookup will
tend to be resolved from caches and not engender any long-distance net
traffic.  The net effect is that the numbers can be skewed.

As an aside, if one is really concerned with the latency of the web, then
I suggest that people go to their local web servers and turn on "expires"
headers.  That little change (amounting to a couple of lines in an Apache
configuration file) could reduce the redundant traffic flowing on the net
by a truely massive amount.  At the most recent IETF there was a very good
presentation on this at the plenary session before the crypto discussion.

		--karl--