Read the paper here:

nms.csail.mit.edu/~stavros/pub...

There are many places where such overhead comes from, but all are necessary for a standard database to work.

Having a faster database is a matter of trade-offs:

no buffers -> keep the database entirely in memory
no locking or latching -> single-threaded system
no logs -> replication.

- Taking care of buffers (moving pages from a cache pool to the disk)
- Locking (row-level locks that manage concurrent multiple transactions)
- Latching (lightweight locks keep the internals of data structures safe)
- Logging (writing down operations before they happen so they can be recovered)