* Library cache: This is used to locate objects in the shared pool and to ensure their proper relationship is maintained. While the library cache can be represented as a hashing structure (complete with buckets and chains), it is actually much more complicated than this. Efficient and unfettered library cache operations are central to a well-performing Oracle database system.

* Row cache: Commonly called the dictionary cache, tuning of this cache was automated starting in Oracle 7. The row cache caches Oracle's data dictionary rows (not blocks). Objects like sys.tab$, sys.col$, and sys.auth$ are cached within the shared pool's row cache, rather than in the buffer cache. While Oracle must retrieve an entire block to retrieve a single dictionary table row, the result is significant memory savings and a highly efficient cache. In fact, the row cache can be monitored through the v$rowcache view and typically has an extremely high cache hit ratio of around 99%. While DBAs were initially concerned about Oracle removing manual optimization control, Oracle has done a fantastic job at dictionary object cache management.

* In-memory undo (IMU): This was introduced in Oracle Database 10g Release 1. It increases DML and read-consistent-intensive operational performance by keeping as much undo as possible using in-memory objects, instead of traditional undo segments. My tests have shown a 21% CPU consumption reduction when using IMU with a read-consistent-intensive load.