Duane and all,
Perhaps we should call it a "benchmark site".
The content on the benchmark site should have characteristics that
closely resemble what is out there (in terms of file sizes, popularity,
temporal locality, update frequencies, mix of static vs dynamic, etc.)
There are many studies on characterizing these aspects. However, I
believe that it may be better to go for a vertical model, whereby the
benchmark site reflects an e-business model (news vs brokerage vs
multimedia, etc.)
I believe that the location of the benchmark site will have only a
secondary effect. To balance the effect of location, we could have the
benchmark site "distributed". Our measurements here at BU suggest that
"interesting" characteristics of Internet topology tend to be discovered
at a distribution level of ~ 10 or so sites (more on this on an upcoming
TR from our group).
My 2c,
--Azer
___________________________________________________________
Azer Bestavros, Associate Professor
Computer Science Dept, Boston University, Boston MA 02215
http://www.cs.bu.edu/~best Phone/Fax:(617)353-9726/6457
-----Original Message-----
From: cdd-owner@measurement-factory.com
[mailto:cdd-owner@measurement-factory.com]On Behalf Of Duane Wessels
Sent: Wednesday, May 31, 2000 9:59 PM
To: cdd@measurement-factory.com
Subject: some random questions
At the BOF in Lisbon, there seemed to be some agreement that a good way
to test CDNs is to create a "fake" content provider and have the CDNs
provide service for it. Measurements could be taken from a number of
dedicated clients scattered throughout the Internet. To further this
discussion, I'll ask a bunch of questions....
Does the fake content provider need to have real content? Does it need
to be HTML, JPG, GIF, and MP3s? Or can we use blobs of random bits?
How much content do we need? Whats the size of a typical content
provider? 100MB, 1GB, 1GB?
How dynamic is the content? Do pages/objects change during the test?
Do we want to measure freshness and consistency?
Is the location of the fake content provider significant? Does it need
at least T1 connectivity? Is a 10baseT interface sufficient?
How many measurement points do we need? 10? 100? more?
How long should the test run? one hour? one day?
One of the things we've found useful in cache benchmarking
is performance/price ratios. Are there "costs" that
should be included in CDN results, other than subscriptions?
Do some schemes require additional hardware?
Duane W.
This archive was generated by hypermail 2b29 : Mon Feb 06 2006 - 12:00:02 MST