Then why invoke the pragma on $dbh and not $dbh2 (or is does the pragma have global effect)?

Because $dbh2 does a commit (and thus an fflush) only at the end of the entire set of inserts (all 30K of them).

Actually that's probably a bit of a benchmark skew right there, since I really should take the commit into account with the benchmark. Drat, now look what you've gone and done ;-)

Update: I moved the $dbh2->commit into the benchmark, removed the $dbh->{AutoCommit} = 0 (because you probably wouldn't do that in a real app), and also moved the $cdb->finish into the benchmark. Results are pretty much the same, with CDB slowing down a little for the inserts, and SQLite1 slowing down a little for selects (I have no idea why). The commit rolled into the benchmark for SQLite2 seemed to make no difference whatsoever.