Here's a short answer:
The pump is run and it's head is measured. Then, the suction pressure is slowly decreased (this can be accomplihed through the use of a valve or vacuum chamber).
The head is constantly measured as the suction pressure is decreased. When the head has dropped by 3% then that is recorded as the NPSHr for that flow.
The test is run again at a couple of different points and a curve is developed from that data.
Likewise for 0% head drop criteria - but 0% drop is usually only seen with critical service applications. One of the problems with this measurement is that normal fluctuations occur in measurement (you know, the +/- tolerances..) and it's difficult to discern between the head having dropped from the low NPSHa and the normal fluctuations in the pressure gauges. So, more expensive test loops are used - which raises costs but also accuracy...
So, whenever you see a performance curve in a catalog, the NPSH curve is actually the 3% curve. I know of no pump manufacturer that publish a 0% NPSHr curve as part of their standard catalog curves.
So, the moral of the story is if you are sufficiently close to the NPSH values on the catalog curve, but still below them, you are already cavitating.
Some manufacturers use a rule of thumb of adding half a meter to the NPSHr or 10%, whichever is greater. But the topic of NPSH goes even further which I have not seen adequately discussed in the literature and that is one of impeller material selection.
The NPSHr curves are generally generated using an impeller of the manufacturer's standard material; however, it is well-known (amoing pump designers but not pump salespersons) that impeller material selection can greatly affect the NPSH characteristics of your pump. For example, if you were to measure the NPSH using a ductile iron, bronze, and stianless steel impeller, you would get 3 different NPSH curves using the same test loop, the same instrumentation, and the same pump casing.
Of course now, if you were to slow the pump down or if the NPSHr was already pretty low, the measured values might very well be within the accuracy of the gauges so you will never see it - but get to values of around 25 feet and you will see some change.
But realistically speaking, smaller diameter impellers also are not as susceptible to this problem - it's when you get into splitcase pumps that you need to really pay attention to it.
So, one might do well to ask the pump manufacturer for an NPSH curve based on their specific impeller material; however, you would probably get a few blank stares and maybe even a chuckle from the ill-informed salesman.