I actually think that as long as you start splitting files across directories to keep from getting more than 1000 in a single dir, both file system and Berkeley would scale very far without a huge difference in performance. Remember, BerkeleyDB handles databases with terabytes of data.

There are definitely many advabtages to having things in normal files, especially for text content, and it's the only choice for NFS or other file servers.

I did use the DB_INIT_CDB flag, which initialized the concurrency methods. If I leave that off, BerkeleyDB gets faster, but you lose the ability to do concurrent access. I didn't think the test would be very interesting if it used options that didn't allow for concurrency.