Вилд вилд вест
A long time ago, when computers remembered using little donuts made of rust, I worked on on a mainframe computer system (CP/V) that supported batch, timesharing, realtime, the works. It had performance monitoring tools, and a large basketload of parameters for sys admins to twiddle.
One of our favorite parameters was SL:BB, documented as batch bias, an input to the process scheduler. When someone called or wrote to us saying they were having problems with performance tuning, we usually suggested they redo their tests varying the setting of SL:BB and let us know what happened. Try different values, 0, 1, 5, 20, 50, 100, things like that. Try it and get back to us.
And lo, they would go off and redo performance runs, and report back.
And we would collect their results and go and muse over them, usually over beer.
SL:BB told us a lot about the user, because SL:BB was a knob that wasn't connected to anything. Oh, the value was range-checked by the parameter setting tool, and dutifully stored in memory, and displayed on performance displays, but it didn't change system performance in any way at all.
That's not what the documentation said, but who believes documentation? We had plans for SL:BB, we just hadn't gotten around to writing the code yet.
So if the user reported that setting SL:BB to 25, but not 24 or 26 gave them incredibly better (or worse) results, we definitely factored that into our analysis.
Those that reported back that the setting of SL:BB didn't make a damn bit of difference, and there were some, we honored as brothers, took into our confidences, and shared beer with at the soonest opportunity. Their bug reports and feature requests received far more attention, for they had passed an important test. Current Mood: amused