Just monitor a quantity such as, for example, the pressure at one position. Then run the simulation with different grid sizes. Once the monitored value does not change any more also by reducing the grid size, you have founded the appropriated grid that you can use.

Franco

The thing that has not been mentioned here is that you should focus your attention especially on regions with large gradients of physical quantities. Doubling the cell count in homogeneous regions won't bring you much benefit whereas doing the same in a large-gradient region may have significant impact on the final result. That is, you should not refine the mesh uniformly since you could end up with a large cell count in regions where it is not needed. That would increase your computational time to no avail.