Falling Short

Initial targets were met, but it became clear as more sessions came on-line that there was a scaling issue.

Benchmarking with load testing software indicated that we should be able to handle five times the traffic we were choking on.

Why?

Our benchmarking code establishes a bus-load of sessions then pushes data through those sessions.  Hardware load testing used a similar pattern.

In pactice there turns to be more “dead” traffic than expected: users appear to connect and disconnect more frequently than our experience with cable would lead us to expect (used to dial-up?).