This is apparently faster than doing things the other way around when
the scale factor is large.

Along the way, adjust -n to suppress vacuuming during initialization
as well as during test runs.

Jeff Janes, with some small changes by me.