Preliminary data indicates that SQLite does slightly better for tmpfs databases for writes and much better for reads. On normal disks, sqlite seems to do a whole lot better for writes. I'll post the test program in a couple of days. I noticed that the SQLite Btree take up only one third the size of Berkeley DB Btree on disk. That may explain the performance difference because the test program is I/O bound. Please note that these results are very preliminary, I am still learning the APIs.
Speaking about APIs, Berkeley DB is obviously well documented. For SQLite I had to go by the (well) commented source and a test program. Berkeley DB's transaction based API is a pain to use. Despite the lack of documentation, I found the SQLite API much more cleaner and simpler to use. In my next blog I'll present some numbers and the test program.