How They Are Created
Currently the predictions rest upon an accurate TCP model which will predict how long a file of a certain size will take to download given both latency and a loss rate. This TCP model was tested and found to be very accurate for latencies between 25ms and 200ms.
The rest of the predictions come from logic expressing how browsers currently download objects. The main difference is that IE 7 only makes 2 connections per host while IE 8 makes 6 per host. Also IE 7 will halt all downloading when pulling down scripts while IE 8 will download scripts concurrently.
The TCP model, browser logic, and observed server processing time make up the current response time prediction implementation.
What To Take Away
Right now the prediction’s accuracy depends on the makeup of your application. If you are looking at simple, static, sites, you should find the predictions to be quite accurate. If you work with giant/complex web applications you should probably expect to see some “funny” numbers.
In the near future, the predictions will be tuned much more accurate. The goal is to also have predictions for non-browser HTTP traffic such as WCF thick clients. For the next release with updated prediction we will try and provide as much test data as possible showing the improvements and accuracy against other prediction methods.
In the end, regional response time is most heavily affected by your site’s design and settings. By running the performance report and tuning your app based on its suggestions (and having no significant backend delays) your users should have a great experience, no matter where they are located.