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.
