I would say linear and predictable = good scalability, exponential and unpredictable = poor.
This is regardless of actual figures - what is important is whether the maximum expected users falls within your "good" scalability model. Every system will become non-linear at some point, but if that point is at double your max users then you're OK.
I'd suggest looking at the number of requests sent vs. responses received over some time interval. As long as the server can scale to keep up with the requested workload, the values will be on par. At the point that it can't, they'll begin to deviate and you'll know that your server has reached its scalability limit.
Well - I suppose the problem really is a matter of the difference between server capacity and application scalability. How does one extrapolate out the results from my example? Even if you can compute the approximate point that a server will start responding with a result of XX seconds under YY user load, it is nearly impossible to predict with any accuracy if any given hardware configuration will support that load.
[This message has been edited by zboss (edited 03-20-2002).]