I've found JDBM3 a pleasure to use; it's fast and stable, with an excellent API. It's really quite easy to start using (it exposes Java collection interfaces), but you can configure it to do some powerful things under the hood. I'm not using the library in a really high-performance setting (just need a persistent key-value store), so I can't quite comment on the extreme scaling qualities.
I'm excited to switch to MapDB as it matures.
I'm on an old XP machine with very little memory and a slow disk, and I get quite good performance out of it, despite not being able to effectively use mapped or off-heap memory. I expect that moving to MapDB on JDK7 (with its concurrent collections and better memory allocation) will improve performance quite a bit.