benchmark the effect of C/J metadata caching

something like, create a table with many columns, add some data, benchmark select * where pk=const. May be run few times with a different number of columns in the table.

what else?