A pump cannot pull
Here is the sentence the whole lesson hangs on: a centrifugal pump does not suck. It lowers the pressure at its eye, and something else — atmosphere, a static level, a pressurised vessel — pushes the liquid in. If the pressure at that eye ever falls to the liquid's own vapour pressure, the liquid boils where it stands, and the bubbles collapse a few millimetres downstream with enough violence to pit stainless steel. That is cavitation, and the margin that prevents it is NPSH.
— read aloud N-P-S-H available equals h-atm plus h-s minus h-f minus h-v-p, and every term is in metres of the liquid. is the absolute pressure on the liquid's surface written as head — 10.3 m for an open vessel at sea level. is the static suction head, and it is the one signed term on the page: positive when the level is above the pump (a flooded suction) and negative when the pump is above the level (a suction lift). is the friction head lost in the suction piping, always subtracted. is the liquid's vapour pressure written as head, also always subtracted, because the margin is measured ABOVE boiling.
Two of those terms are where the trouble lives. A suction LIFT subtracts, which is why the atmosphere's 10.3 m is a hard ceiling on how far below a pump you may draw — and in practice the ceiling is far lower once friction and vapour pressure take their cut. And vapour pressure climbs fast with temperature: for water it is worth 0.2 m at 20 °C, 2.0 m at 60 °C and 4.8 m at 80 °C. A condensate pump that ran perfectly all winter can cavitate in August for no other reason.
Then the verdict. The pump's own curve prints , the required figure — measured on a test rig as the suction head at which that pump has already lost 3 % of its head to bubbles. The rule is , with margin. Positive alone is not the test.
The same margin has a dimensionless form worth meeting: the cavitation number , where is the reference absolute pressure in pascals, the vapour pressure in pascals, the density and the reference velocity in m/s. It counts the margin in velocity heads. There is no universal critical value — the threshold belongs to the SHAPE, roughly 0.2 for a good hydrofoil and several for a sharp-edged throttling valve — which is exactly why valve makers publish it per trim.