for an exmple of somewhat more complete and better designed benchmarks of this type:
We make web measurements from over 800 points (POPs) on the Internet for roughly 20 ISP's using a standard basket of 10 popular web pages (porn pages excluded given there is the believe it will drive down the productivity of our operation staff :) We control the test by having every ISP runs through the same set of pages at roughly the same time. (By the way, to closely emulate consumer end user's experience, we specifically use a [painful] test harness to drive Netscape 3.0 on Windows95 over boring 33.6 modem line ... would have love to use a UNIX script with our own custom browser over our direct link into the net instead ... but then it is not what consumer users are doing ... oh well). Even with the extensive matrix of measurements, the only claim we dare to make is that it approximate a end user experience index a la Dow Jones's basket of companies. (By the way, with few interesting exceptions, mostly limited by the size of the final pipe - the modem link). We think there are way too much variables involved to allow us to defintively partition out the components such as the page structure (how many other embedded objects we have to retrieve, compressibility by modem of those objects etc.), the Web server load at the time the measurement was made, the load of the web server's LAN, the load and size of that LAN's pipe into the server's ISP, that ISP's network performance, its peering performance with the end user's ISP (and transit links if any), the user's ISP's network performance .... etc. etc. etc. Even with extensive analysis, it will be difficult for us to derive the backbone performance component from our end to end measurement. If it is the underlying network performance one wants I think a different measurment infrastructure such as those being developed by Paxon/Mathis/Jamshid (NIMI), Almes/Barber (Surveryor), and Graig Labovitz (Merit) ... using tools such as Jacobson's Pathchar will be much more appropriate. Regards, John Leong