[eatonphil]

> And generated two datasets: one with 20 columns and 1M rows, and one with 1K columns and 10K rows.

And if that's not big enough for you, the observed effects grow as I increased columns and rows.