We (PAE - Performance Engineering) recently completed a study to understand database performance with ZFS. Read on more details and recommendations. You can also read Roch's blog on the same study
Databases stress the filesystem in unique ways. Depending on the workload and configuration, you can have thousands of IO operations per second. The size of these IO is usually small (database block size). All the writes are synchronized writes. Reads can be random or sequential. Some writes are also more critical than others. Depending on the configuration, Reads are cached by the database program or the filesystem (if supported/requested). In many cases where filesystems are used, the IO is spread over a few files. This causes the single writer lock to be very hot under certain configurations like Buffered UFS.
