Here is one alternative metric. I have a monitor program that measures
single-packet ping times to name servers that are registered as in-addr.arpa
authorities. I figure that people should put their name servers in a "good"
position on their network and that they should always be up and running. My
goal is to monitor O(1000) hosts at regular intervals (currently 10 minutes),
but the list currently hast just over 400 sites. I started with the web
access logs for a very large site with a worldwide user base, looked up the
registered in-addr.arpa servers, then hand-pruned the list to weed out sites
that block ICMP, etc. ICMP echo obviously isn't the best metric in the world,
but it has low overhead and allows me to monitor a large number of sites
without being disruptive. (Most people shouldn't mind a single ICMP echo
packet every 10 minutes.)
It's been very interesting to watch the graphs as various backbones have
glitches. For awhile I was comparing connectivity from a single-homed site to
a 5-way multi-homed site. The difference in fault tolerance was dramatic.
The next step is to combine some policy routing with some /24 network
announcements that are only announced from one backbone to compare
connectivity via MCI, Sprint, BBN, and UUNET. (Ie, run parallel copies of the
monitor in an environment that simulates single-homed connectivity from each
provider.) The results should be interesting, but I wouldn't want to claim
that they represented anything more than a measure of connection quality from
the particular sites where the tests were run.
At the moment, the tool and the data are proprietary since I wrote it for a
particular client, but I'm hoping to get permission to release the data once
there are some results that are closer to "research quality." (We have been
collecting data continuously since early March.)
-dpm